Intelligent control method and system for lithium carbonate preparation process

Abstract

The invention provides an intelligent control method and system for a lithium carbonate preparation process, and the method comprises the steps: collecting the temperature field, ion concentration, feeding flow velocity and turbulence intensity data of a reaction kettle through a sensor array, and generating a state feature vector through multivariate analysis; inputting the feature vector into a depth prediction model to output a particle size distribution probability cloud picture and an impurity trend; inputting the probability cloud picture and the impurity trend into a multi-target optimizer to generate a temperature compensation value, a feeding correction coefficient and a stirring adjustment instruction; inputting the instruction into an anti-interference execution module and converting the instruction into an equipment driving signal; actual crystal data are obtained through an online particle size analyzer, and the prediction model is updated; and outputting a closed-loop optimization instruction to the execution terminal through the digital twin platform. The product consistency and the production stability are synchronously improved through full-process closed-loop intelligent control, and reliable technical support is provided for continuous production of battery-grade lithium carbonate.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to an intelligent control method and system for a lithium carbonate preparation process. Background Technology

[0002] The rapid development of new energy vehicles and energy storage industries has driven a surge in demand for lithium battery materials. As a core cathode material precursor, lithium carbonate is undergoing a transformation in its preparation process towards large-scale continuous production. In the mainstream processes of lithium extraction from salt lakes and ore, the carbonation precipitation process generates battery-grade lithium carbonate by controlling the crystallization process of lithium ions and carbonate ions in the reactor. This process directly determines key performance indicators such as product crystal particle size distribution, impurity content, and tap density, thereby affecting the material's application efficiency in high-end power batteries.

[0003] Current industrial production generally adopts an operation mode based on fixed parameter settings or simple feedback control, which has three major drawbacks: First, key quality indicators such as crystal particle size distribution rely on manual sampling and offline testing, resulting in quality feedback delays of several hours and an inability to respond to dynamic changes in the reaction system in real time; Second, the strong coupling effect of multiple variables such as temperature, concentration, and feeding rate causes traditional PID control to frequently become unstable, resulting in significant particle size fluctuations in different batches of products and persistently high impurity exceedance rates; Third, manual experience-based adjustments lack quantitative basis and are prone to explosive nucleation or crystal agglomeration under fluctuations in raw material composition or environmental interference, leading to increased steam energy consumption and insufficient product qualification rate. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides an intelligent control method and system for lithium carbonate preparation processes. By embedding crystal growth kinetic constraints into a deep neural network to achieve real-time prediction of particle size distribution, using the Pareto front algorithm to dynamically optimize multi-objective control parameters, and designing a boil-out suppression collaborative control strategy, this invention effectively solves three major problems: quality feedback delay, control instability, and lag in human intervention. It improves product consistency to battery-grade standards while reducing production energy consumption, providing reliable technical support for continuous lithium carbonate production.

[0005] In a first aspect, the present invention provides an intelligent control method for a lithium carbonate preparation process, comprising: Data on temperature field, ion concentration, feed flow rate and turbulence intensity of the reactor are collected by a sensor array, and a state feature vector is generated through multivariate analysis. The feature vector is input into the depth prediction model to output a granularity distribution probability cloud map and impurity trend. The probability cloud map and impurity trend are input into a multi-objective optimizer to generate temperature compensation values, feeding correction coefficients, and stirring adjustment commands. The instruction input to the anti-interference execution module is converted into a device drive signal; The actual crystal data is obtained using an online particle size analyzer, and the prediction model is updated accordingly. The closed-loop optimization instructions are output to the execution terminal through the digital twin platform.

[0006] Furthermore, the feature vector is input into the deep prediction model, which outputs a granular distribution probability cloud map and impurity trends. The deep prediction model is implemented using a dual-channel deep neural network architecture, including: The first channel processes real-time feature vectors through a three-dimensional convolutional layer to capture the spatial distribution characteristics of local oversaturated regions within the reactor. The second channel analyzes the temporal evolution of historical crystal growth trajectories through attention-enhanced long short-term memory networks; By introducing a population equilibrium equation constraint describing crystal growth dynamics into the feature fusion layer, the output particle size distribution probability cloud map conforms to physical laws.

[0007] Furthermore, the population equilibrium equation is:

[0008] The equation describes the dynamic evolution of the crystal population size distribution, where This represents the number density of a crystal of size L at time t. This represents the linear growth rate of the crystal, driven by supersaturation. As a nucleation rate source term, it quantifies the rate of new crystal formation; The agglomeration and sedimentation loss term characterizes the loss caused by crystal agglomeration. By embedding the equation as a physical constraint into the neural network loss function, the prediction model maintains high accuracy even when the raw material composition fluctuates, and the crystal particle size prediction error remains stable.

[0009] Furthermore, the probability cloud map and impurity trend are input into a multi-objective optimizer to generate temperature compensation values, feeding correction coefficients, and stirring adjustment commands. The multi-objective optimizer operates based on a Pareto front dynamic optimization mechanism and includes: A dynamic target range function for the change of median crystal size over time was established, with the range remaining relatively wide in the early stage of the reaction to promote nucleation; We construct constraints to maintain a narrow, unimodal distribution across the granularity range, and suppress bimodal anomalies through a penalty function. With the goal of minimizing the probability of co-precipitation of alkali metal ions, an improved non-dominated sorting algorithm is used to generate coordinated control commands for temperature, feeding, and stirring.

[0010] Furthermore, the dynamic objective interval function is:

[0011] The exponential decay function simulates the time-varying characteristics of the control target during the reaction process. and respectively represent the lower and upper limits of the target particle size at time t. and are the initial boundary values, set according to the product grade. and are the shrinkage rate constants, controlling the boundary shrinkage speed; by setting the lower limit shrinks faster than the upper limit, providing sufficient nucleation space in the early stage of the reaction and strictly constraining the crystal growth quality in the later stage, reducing the particle size fluctuation range between batches.

[0012] Furthermore, the instruction is input into the anti-interference execution module and converted into a device drive signal. The anti-interference execution module implements a boiling suppression collaborative control strategy, including: When the supersaturation probability breaks through the critical threshold, immediately start the three-stage gradient cooling protocol. Synchronously reset the feeding correction coefficient to zero and maintain the minimum material flow, while increasing the stirring power to the turbulent enhancement zone. Ensure millisecond-level linkage response of temperature regulation, feeding truncation, and turbulent enhancement through hardware-level timestamp synchronization.

[0013] Furthermore, the determination of the turbulent enhancement zone adopts the Reynolds number criterion:

[0014] The formula quantifies the influence of fluid turbulent intensity on crystal dispersion. is the Reynolds number. is the solution density. represents the stirring speed. is the diameter of the stirring impeller. represents the solution viscosity. is the critical value of turbulent transition; when the fluid enters the fully turbulent state, providing sufficient shear force to disperse crystal aggregates and improving the product qualification rate.

[0015] Furthermore, actual crystal data is obtained through an online particle size analyzer and the prediction model is updated. The update of the prediction model adopts a joint algorithm of residual backpropagation and state estimation, including: First, obtain the actual D50 value and Span value of the crystals through the online particle size analyzer, and calculate the KL divergence residual between them and the prediction result. The convolutional neural network channel adjusts the weights of the convolutional kernels through the backpropagation algorithm based on the residual distribution, and focuses on optimizing the feature extraction layer sensitive to supersaturation. The long short-term memory network channel adopts the extended Kalman filter algorithm to calibrate the timing prediction deviation of the hidden state vector using the innovation covariance matrix. The entire model parameters are updated every 15 minutes, and an emergency correction mode is triggered immediately when a sudden change in raw material composition is detected.

[0016] Furthermore, closed-loop optimization instructions are output to the execution terminal through the digital twin platform, wherein the digital twin platform includes: The material balance engine tracks changes in lithium-ion concentration gradient in real time, calculates the deviation between the current conversion rate and the target value, and predicts the reaction endpoint. The energy optimization module integrates temperature field distribution and steam flow data to establish a thermodynamic efficiency model and dynamically adjust the steam supply. The virtual detection unit simulates the crystal collision and aggregation process through discrete element simulation, outputs an animation of crystal morphology evolution, and generates a quality warning. Every 5 minutes, the platform compares the predicted state of the digital image with the actual sensor data and outputs commands for axial temperature difference compensation, dynamic feeding ratio, and anti-agglomeration shear force control.

[0017] Secondly, the present invention also provides an intelligent control system for a lithium carbonate preparation process, applied to the intelligent control method for the lithium carbonate preparation process as described in the first aspect; the intelligent control system for the lithium carbonate preparation process includes: State perception module: Collects data on temperature field, ion concentration, feed flow rate and turbulence intensity of the reactor through sensor array, and generates state feature vector through multivariate analysis; Growth prediction module: Inputs the feature vector into the depth prediction model and outputs a particle size distribution probability cloud map and impurity trend; Optimization Decision Module: Inputs the probability cloud map and impurity trend into the multi-objective optimizer to generate temperature compensation value, feeding correction coefficient and stirring adjustment command; Execution conversion module: converts the instruction input to the anti-interference execution module into a device drive signal; Model update module: Acquires actual crystal data using an online particle size analyzer and updates the prediction model; Closed-loop control module: Outputs closed-loop optimization instructions to the execution terminal through the digital twin platform.

[0018] The intelligent control method for lithium carbonate preparation process provided in this invention integrates physical mechanisms with artificial intelligence prediction models to achieve real-time and precise control of the lithium carbonate crystallization process, effectively eliminating product fluctuations caused by quality feedback lag in traditional control. It employs a multi-objective dynamic optimization algorithm to coordinate temperature, feeding, and stirring parameters, overcoming the control instability problem caused by strong coupling of multiple variables. A millisecond-level boiling suppression strategy is designed to simultaneously trigger gradient cooling, feeding cutoff, and turbulence enhancement operations, completely avoiding production accidents caused by delayed human intervention. Combined with a digital twin platform, a closed-loop optimization mechanism is constructed for the entire process, ensuring precise control of crystal morphology to battery-grade standards while significantly reducing production energy consumption and raw material loss, providing a highly consistent and low-cost intelligent manufacturing solution for lithium carbonate materials in the new energy industry. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of an intelligent control method for lithium carbonate preparation process provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a smart control method for a lithium carbonate preparation process provided in an embodiment of the present 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] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0023] Reference Figure 1 , Figure 1 This is a flowchart illustrating an intelligent control method for a lithium carbonate preparation process provided by the present invention. The intelligent control method for a lithium carbonate preparation process includes: Step 10: Collect data on the temperature field, ion concentration, feed flow rate, and turbulence intensity of the reactor using a sensor array, and generate a state feature vector through multivariate analysis; Step 20: Input the feature vector into the depth prediction model to output the granularity distribution probability cloud map and impurity trend; Step 30: Input the probability cloud map and impurity trend into the multi-objective optimizer to generate temperature compensation value, feeding correction coefficient and stirring adjustment command; Step 40: The instruction input to the anti-interference execution module is converted into a device drive signal; Step 50: Obtain actual crystal data and update the prediction model using an online particle size analyzer; Step 60: Output closed-loop optimization instructions to the execution terminal through the digital twin platform.

[0024] Specifically, in step 10, a three-dimensional spatial grid monitoring network is constructed on the inner wall of the reactor using distributed optical fiber temperature sensors. The sensor nodes are deployed in an axially layered and radially angularly distributed manner to capture the temperature gradient changes at different heights and radial positions of the reactor in real time. An immersion lithium-ion selective electrode array is used, with electrode probes inserted vertically into the reaction solution at equal intervals to continuously monitor the dynamic migration process of lithium-ion concentration in the liquid phase. An ultrasonic Doppler velocimeter is installed on a specific shear plane to emit high-frequency sound waves into the fluid and receive the scattered signals, thereby quantifying the distribution of turbulent kinetic energy intensity in a micro-region. An optical turbidity sensor is integrated into the sight glass of the reactor to identify the critical state of nucleation phase transition by the attenuation rate of transmitted light intensity.

[0025] Multi-source sensor data is transmitted to edge computing nodes via an industrial bus. A spatiotemporal alignment algorithm is used to eliminate the acquisition delay of different sensors and to establish a four-dimensional dataset consisting of a temperature field distribution matrix, an ion concentration gradient vector, a flow velocity turbulence intensity tensor, and a turbidity change curve.

[0026] The multivariate coupling analysis engine first performs principal component analysis on the temperature matrix to extract axial and radial temperature uniformity indices; Next, the Pearson correlation coefficient between the concentration gradient and the temperature field is calculated to quantify the influence weight of the thermodynamic driving force on the mass transfer rate. Then, a transfer function model of turbulence intensity and turbidity variation is established to identify the nonlinear mapping relationship between fluid shear force and nucleation rate; Finally, all features are fused to generate a structured state feature vector that includes thermodynamic stability factor, mass transfer efficiency coefficient, and turbulence-nucleation coupling parameters.

[0027] This step, through spatially distributed sensor networks and high-frequency data fusion, completely eliminates the blind spots of traditional single-point detection, providing a complete physical field information foundation for crystal growth prediction.

[0028] In one specific embodiment, when the operator starts the carbonation precipitation process, he first activates the 24 distributed fiber optic sensors arranged on the inner wall of the reactor. Each sensor collects the temperature distribution in the axial range of 0.5 meters to 3.5 meters at a frequency of 10 Hz, generating a temperature field topology map containing 72 spatial nodes. Eight sets of immersion lithium-ion selective electrodes were started simultaneously. Each set of electrodes continuously recorded ion concentration values ​​at three depth levels (upper / middle / lower layers). The concentration gradient change rate was calculated using a sliding window algorithm. An ultrasonic Doppler velocity meter emits 2 MHz sound waves at a tangential plane 0.3 times the diameter of the impeller blade, and analyzes the fluid shear stress distribution based on the Doppler frequency shift. The turbidity sensor monitors the intensity of transmitted light at a sampling frequency of 5 times per second, and marks a nucleation event when the light intensity attenuation rate exceeds a set threshold.

[0029] All sensor data is aligned to millisecond-level precision via a timestamp synchronization module, and the edge computing node performs the following analysis process: Three-dimensional interpolation reconstruction of temperature field data was performed to calculate the extreme values ​​of axial temperature difference and the proportion of radial hot spot area; The ion concentration data is spatiotemporally correlated with the temperature field to output the concentration gradient slope and mass transfer driving force coefficient; the spectral characteristics of the flow velocity signal are analyzed to extract the dominant turbulence frequency and eddy scale parameters. By correlating turbidity abrupt change points with turbulent spectra, a nucleation tendency assessment index was established.

[0030] The final generated state feature vector includes: temperature uniformity index ( Range), concentration gradient slope ( ), turbulence dominant frequency ( Nucleation tendency coefficient ( This vector is transmitted to the prediction model in real time.

[0031] This embodiment demonstrates the complete transformation process from raw physical quantities to high-order eigenvectors, providing a quantifiable description of the system state for process control.

[0032] The following points should be noted during implementation: During the sensor deployment phase, it should be ensured that the distributed optical fiber is tightly attached to the inner wall of the reactor without air gaps to avoid temperature monitoring distortion; the lithium-ion selective electrode needs to be calibrated monthly with lithium chloride standard solution to prevent electrode drift from causing concentration monitoring deviation; the transmitting and receiving probes of the ultrasonic Doppler flow meter must be kept clean and regularly cleaned with dilute nitric acid to remove lithium carbonate crystals; the optical window of the turbidity sensor needs to be equipped with an automatic rinsing device to prevent crystal deposition from blocking the light path.

[0033] An outlier filtering mechanism needs to be set up during the data acquisition process: when the temperature sensor readings deviate for three consecutive cycles... The self-diagnostic procedure is triggered at 5%; if the ion concentration value shows a step change, redundant electrode cross-verification needs to be started; when high-frequency noise appears in the flow rate data, the wavelet noise reduction mode is automatically switched.

[0034] When performing multivariate coupling analysis, the following correlations should be monitored: when the spatial overlap between the temperature field hotspot distribution and the concentration gradient exceeds 75%, it indicates the risk of local oversaturation; when the turbulence dominance frequency is below 2Hz, it immediately warns of the tendency for crystal precipitation; when the nucleation tendency coefficient increases by more than 40% within 10 minutes, it predicts the possibility of explosive boiling.

[0035] During implementation, a sensor health status dashboard needs to be established to display equipment parameters such as fiber optic breakpoint location, electrode sensitivity attenuation rate, and ultrasonic signal-to-noise ratio in real time, ensuring the reliability of the data source.

[0036] This protection system ensures the accurate generation of state feature vectors, laying a solid data foundation for subsequent intelligent control.

[0037] Specifically, in step 20, the state feature vector generated in step 10 is input into a dual-channel deep neural network architecture for processing.

[0038] The first channel uses a three-dimensional convolutional neural network. Its convolutional kernel scans the temperature field distribution matrix and ion concentration gradient map in the spatial dimension, and captures the spatial distribution pattern of local oversaturated hot spots through a multi-layer feature pyramid structure. The second channel employs an attention-enhanced long short-term memory network to analyze historical crystal growth trajectory data and dynamically focus on key growth stages through temporal attention weights.

[0039] After tensor splicing of the dual-channel outputs in the fusion layer, a population equilibrium equation is introduced as a physical constraint:

[0040] in This represents the number density of a crystal of size L at time t. This represents the linear growth rate of the crystal, driven by supersaturation. As a nucleation rate source term, it quantifies the rate of new crystal formation; The agglomeration and sedimentation loss term characterizes the loss caused by crystal agglomeration. The equation is embedded in the model in the following ways: Equation constraint loss terms are superimposed on the output tensor of the feature fusion layer to make the prediction results satisfy the crystal population evolution law.

[0041] A grain size distribution probability cloud map is generated in the output layer. The cloud map contains the probability density function surface of the crystal size and the impurity concentration gradient distribution.

[0042] The entire prediction process is completed at the edge computing node, and the prediction results are updated every 30 seconds, enabling advanced perception of the crystal growth path.

[0043] This step breaks through the limitations of traditional empirical models in predicting complex working conditions by deeply integrating physical mechanisms with data-driven models.

[0044] In one specific embodiment, when the state feature vector includes a temperature uniformity index of 0.88 and a concentration gradient slope... When parameters are equal, the 3D convolutional neural network first performs a convolution operation on the temperature-concentration composite matrix to extract features of local oversaturated regions; the long short-term memory network processes historical crystal growth data to predict growth trends. The feature fusion layer performs tensor product operations on spatial and temporal features and superimposes population equilibrium equation constraints:

[0045] Crystal growth rate Based on the current supersaturation, The frequency of nucleation events monitored by the correlation turbidity sensor, Bind turbulence intensity parameters.

[0046] After physical constraint correction, the probability cloud map generated by the output layer shows the crystal grain size distribution characteristics in the future time period. For example, the probability that the D50 value will enter the target range after 20 minutes is 85%, and the distribution of magnesium ions at hot spots enriched on the crystal surface. This prediction result directly guides the optimization module to adjust operating parameters, avoiding the risk of impurities exceeding the standard from the source.

[0047] In the specific implementation process, the following should also be noted: The input data for 3D convolutional neural networks needs to be standardized and preprocessed to ensure uniform feature scale.

[0048] The historical data window length of a long short-term memory network should cover the entire reaction cycle.

[0049] Parameter calibration of the physical constraint equations is crucial: Crystal growth rate coefficient It is necessary to establish its functional relationship with supersaturation through offline experiments every quarter. ,in , For supersaturation, This represents the real-time concentration of lithium ions in the solution. The concentration is at thermodynamic equilibrium. Nucleation source term The correlation factors should be dynamically adjusted according to the raw material grade; Agglomeration Settlement Item A mapping model between the current impeller configuration and the Reynolds number needs to be established.

[0050] The model inference process is monitored in real time: data verification is initiated when the entropy value of the feature map is abnormal; model diagnosis is triggered if the variance of the hidden state increases; and a safe mode is switched when the confidence of the output layer is insufficient. The effectiveness of physical constraints is periodically verified: the equation residuals are calculated by comparing the predicted cloud map with the actual detection results.

[0051] When the residual continues to increase, the model is retrained.

[0052] This quality assurance system ensures that the prediction model maintains reliable accuracy under complex conditions such as raw material fluctuations.

[0053] Specifically, in step 30, the granularity distribution probability cloud map and impurity trend output in step 20 are input into the multi-objective optimization decision-maker, and control commands are dynamically generated based on the Pareto front algorithm.

[0054] First, an optimization objective function is established to make the median crystal grain size (D50) approach a dynamic target range. This target range is adjusted in real time using a dynamic target range function.

[0055] in and These represent the lower and upper limits of the target granularity at time t, respectively. and These are the initial boundary values, set according to the product grade. and is the contraction rate constant.

[0056] Simultaneously, a hard constraint condition is constructed to maintain a narrow unimodal distribution of particle size distribution span (Span value), and a threshold for the standard deviation of Span value is set as the trigger boundary of the penalty function; with minimizing the probability of co-precipitation of alkali metal ions as the second optimization objective, a calculation model for the impurity adsorption energy barrier is established.

[0057] An improved non-dominated sorting genetic algorithm is adopted. Searching for Pareto optimal solutions in the three-dimensional target space: Initializing the population includes a random combination of temperature compensation values, feed correction coefficients, and stirring power adjustment commands; generating offspring solutions through crossover and mutation operations; calculating the fitness of each solution on the D50 objective function, Span constraints, and impurity objectives; and using reference point-guided non-dominated sorting to select the optimal solution set.

[0058] The final output layer selects the most balanced solution vector from the Pareto front through fuzzy decision-making, and generates temperature field equalization commands, feeding pulse modulation strategies and turbulence intensity adjustment schemes in real time.

[0059] This step achieves a coordinated balance among multiple objectives, completely resolving the batch-to-batch fluctuation problem caused by traditional fixed-parameter control.

[0060] In one specific embodiment, the optimizer initiates the decision-making process when the probability cloud map displays the predicted D50 values ​​and their distribution characteristics for future time periods.

[0061] First, calculate the current reaction time point. Target range:

[0062] The D50 optimization objective is set to minimize the absolute deviation between the predicted value and the midpoint of the target interval; the Span constraint is set to the standard deviation threshold of the distribution span; and the impurity objective is set to the maximum enrichment probability threshold.

[0063] The initial population contains multiple solutions and vector combinations. For example, solution A: positive temperature compensation, negative adjustment of the feeding coefficient, and positive adjustment of the stirring power; solution B: negative adjustment of temperature compensation, positive adjustment of the feeding coefficient, and a significant increase in stirring power.

[0064] After multiple generations of evolution, the Pareto front retains several non-dominated solutions, among which the optimal compromise solution should simultaneously satisfy: the D50 prediction value is within the dynamic target range, the Span standard deviation is lower than the set threshold, and the impurity enrichment probability is less than the target value.

[0065] Control commands are sent to the execution system via industrial communication protocols to achieve precise control of the reaction system.

[0066] This optimized decision-making process is completed efficiently at edge computing nodes, achieving an order-of-magnitude speed improvement compared to traditional manual decision-making.

[0067] The following points should be noted during implementation: The parameter settings of the dynamic target interval function need to match the process characteristics: initial boundary values and The shrinkage rate constant should be set according to the product grade and classification. and It needs to be fitted using historical data and maintained. Ensure the lower limit contracts faster.

[0068] The Pareto optimization algorithm requires proper parameter configuration: It is recommended that the population size be an appropriate multiple of the number of control variables; Set the crossover probability to a reasonable range; The probability of mutation is set as the reciprocal of the number of variables; The number of reference points grows exponentially with respect to the target dimension.

[0069] The constraint handling adopts the feasible solution priority criterion: when the proportion of feasible solutions in the population is insufficient, the constraint threshold is automatically relaxed.

[0070] Real-time monitoring of Pareto solution set quality: Calculating the solution set distribution entropy.

[0071] Increase the mutation intensity when the entropy value is lower than the baseline level; evaluate the convergence of the solution set, and terminate the iteration early if the rate of change of the optimal solution is less than the threshold for multiple consecutive generations.

[0072] Regularly calibrate algorithm parameters: By comparing actual production data with optimization results, adjust the weight coefficients of the objective function in reverse to ensure continuous optimization of control commands.

[0073] This refined management system ensures the stability and adaptability of optimized decision-making.

[0074] Specifically, in step 40, the control commands (temperature compensation values) generated by the multi-objective optimizer are... Feeding correction coefficient Adjusting the stirring power percentage The input anti-interference execution module converts the signal into a device drive signal. The temperature compensation command is converted into a control signal for the heating unit via a pulse width modulation controller.

[0075] in This is the temperature-duty cycle conversion factor.

[0076] Feeding correction factor Encoded as the switching pulse sequence of a pneumatic diaphragm valve:

[0077] in Minimum feed baseline, This represents the maximum pulse width.

[0078] Stirring power adjustment command The speed control curve is mapped to that of a variable frequency drive. First, the required speed in the turbulence enhancement region is calculated based on the Reynolds number criterion.

[0079] when When the indication needs to enter the turbulence enhancement region, the solution satisfies minimum speed Generate S-shaped acceleration / deceleration curves:

[0080] in The current rotational speed, This is the acceleration factor.

[0081] When implementing the boiling suppression strategy, a hardware synchronization bus ensures millisecond-level coordination of three-stage gradient cooling, feed cutoff, and turbulence enhancement: when supersaturation... Synchronous triggering negative step jump , Three operations.

[0082] This step enables a lossless conversion between control commands and device actions.

[0083] In one specific embodiment, when the optimizer outputs instructions: , , At that time, the execution module initiates the signal conversion process.

[0084] The temperature compensation unit calculates the PWM duty cycle adjustment:

[0085] Feed control unit processes correction coefficient :

[0086] Stirring power unit detection Calculate the rotational speed required for turbulence enhancement:

[0087] Generate acceleration curves:

[0088] If the supersaturation suddenly increases at this time Immediately triggers cooperative inhibition: Switch to negative signal. Reset to zero Increase to the maximum value.

[0089] This embodiment achieves rapid state switching by synchronizing the actions of multiple devices using timestamps.

[0090] The following points should be noted during implementation: Temperature signal conversion requires periodic calibration. coefficient:

[0091] Feeding pulse control needs to prevent frequent valve operation: set a minimum pulse interval time, when Enable moving average filtering when there are continuous fluctuations:

[0092] The stirring speed control must ensure a smooth transition of turbulence: acceleration coefficient Based on fluid inertia settings; Monitoring actual Value feedback adjusts the speed:

[0093] Set the speed change rate limit:

[0094] Cooperative control requires strict timing alignment: configure a global clock source on the hardware synchronization bus; set up an execution state interlock mechanism.

[0095] Monthly boiling suppression tests are conducted to verify response performance.

[0096] Specifically, in step 50, the actual crystal morphology data in the reaction system is acquired in real time by an online particle size analyzer. The analyzer uses Raman spectroscopy scanning technology to non-invasively collect the spatial distribution information of the crystal group at the reaction vessel window.

[0097] The acquired spectral data, after preprocessing, is input into the feature extraction engine: first, baseline correction is performed to eliminate fluorescence background interference; then, Savitzky-Golay filtering is used to smooth noise; finally, peak position identification is performed to extract the characteristic peak intensity ratio. (Corresponding to crystal structure order) and full width at half maximum (FWHM) (Related crystal size distribution).

[0098] The residuals are calculated by comparing the processed actual particle size parameters (D50 value, Span value, and roundness coefficient) with the predicted particle size distribution probability cloud map.

[0099] in For the predicted probability vector, This is the measured distribution vector.

[0100] The residual signal updates the prediction model in two ways: the convolutional neural network channels adjust the network weights through backpropagation, and the loss function is defined as:

[0101] in This is the actual value. For predicted values, The regularization coefficient is used. The Long Short-Term Memory (LSTM) network channel uses the Extended Kalman Filter (EPF) algorithm, through the state update equation:

[0102] and observation update equations:

[0103] Dynamically calibrate the hidden state vector.

[0104] Cross-validation is performed immediately after the model is updated: the updated prediction accuracy is validated using historical data from the most recent 24 hours, and a rollback mechanism is triggered when the mean absolute percentage error (MAPE) exceeds a threshold.

[0105] This closed-loop correction system enables the prediction model to continuously adapt to fluctuations in raw materials and changes in operating conditions.

[0106] In one specific embodiment, when the online particle size analyzer detects an actual D50 value of 46.2 μm and a Span value of 0.62, while the prediction contour map shows a predicted D50 value of 42.8 μm (probability 85%) and a predicted Span value of 0.51, the residual calculation module outputs: ,

[0107] Convolutional Neural Network Channels Initiate Backpropagation: Calculate the Gradient of the Loss Function

[0108] The key adjustment is to the weights of the third convolutional layer, which is sensitive to temperature gradients; Kalman filtering is performed on the long short-term memory channel to update the predicted state covariance.

[0109] Kalman gain calculation

[0110] Status update

[0111] The updated model re-predicts: the D50 value is adjusted to 45.5. (Error <2%), Span value 0.60 (error) 3%.

[0112] The model version management record shows this update, which will be used for the next 3 MAPE predictions. If the update is confirmed to be effective at 5%, otherwise revert to the previous stable version.

[0113] The following points should be noted during implementation: Online particle size analyzers require regular calibration: monthly calibration using standard polystyrene microspheres (diameter...) Calibrate and verify the D50 measurement error. 1.5%; Replace the anti-fouling film on the spectrometer's optical window quarterly to ensure light transmittance. 95%.

[0114] Residual calculations require a validity check: when Updates are paused temporarily (due to anomaly detection); when Manual review is triggered at that time.

[0115] Multiple layers of protection are implemented during the model update process: Backpropagation constraint: Gradient clipping threshold set to To prevent sudden weight changes; the learning rate is adaptively adjusted. ; Kalman filter configuration: process noise covariance Positively correlated with the intensity of raw material fluctuations; observation noise covariance Dynamically adjusts according to sensor accuracy; Update frequency control: Incremental updates are performed every 15 minutes in normal mode; a full model update is triggered when raw material batches are switched; the number of updates per day shall not exceed 8. Version rollback mechanism: Establishing a model performance evaluation matrix

[0116] Automatic regression occurs when the score falls below the baseline twice consecutively.

[0117] Stress tests are conducted quarterly to simulate step changes in raw material composition (such as lithium concentration ±20%) and verify the model's adaptive capabilities.

[0118] This rigorous implementation system ensures the safety and reliability of the calibration process.

[0119] Specifically, in step 60, the entire process data flow from steps 10 to 50 is integrated through a digital twin platform to construct a high-fidelity digital mirror of process control.

[0120] The material balance engine processes ion concentration gradient data and cumulative feed amount in real time, calculates lithium ion conversion rate deviation, and predicts reaction endpoint; the energy optimization module integrates temperature field distribution, steam flow rate, and stirring power data to establish a thermodynamic efficiency model and dynamically adjust heat energy supply; the virtual detection unit simulates crystal collision and aggregation process based on discrete element simulation technology, generates crystal morphology evolution animation, and outputs quality warning signals.

[0121] The platform adopts a layered data fusion architecture: the bottom layer receives raw sensor data and prediction model output; the middle layer performs spatiotemporal alignment and feature correlation analysis; and the application layer displays the thermal map of the internal state of the reactor through a 3D visualization interface.

[0122] The closed-loop optimization instruction generation mechanism includes: comparing the deviation matrix between the digital mirror prediction state and the actual sensor data; identifying abnormal patterns such as excessive axial temperature difference, inaccurate feeding ratio, and high risk of agglomeration; and generating a set of compensation instructions based on a rule engine and reinforcement learning algorithm.

[0123] The instruction distribution system pushes optimized instructions to the execution terminal in real time through a distributed control bus, and at the same time establishes an instruction execution feedback loop to verify the compliance of the terminal status update with the instruction requirements.

[0124] This digital twin system enables visualized monitoring and advanced optimization of process status, providing a decision-making brain for continuous production.

[0125] In one specific embodiment, when the material balance engine detects that the lithium-ion conversion rate is lower than the expected value of 5% for three consecutive cycles, the digital twin platform initiates the diagnostic process.

[0126] The virtual detection unit retrieved the current crystal morphology simulation data and found that there was a tendency for crystal agglomeration in the southeast quadrant; the energy optimization module analyzed the temperature field distribution and identified hot spots with excessive axial temperature difference; the platform judged that the local supercooling was caused by inaccurate feeding ratio.

[0127] The rule engine generates three levels of compensation instructions: the first level sends an axial temperature difference equalization instruction to the temperature control system, requiring an increase in bottom heating power while reducing top power; the second level sends a dynamic ratio adjustment instruction to the feeding unit, reducing the supply of carbonate solution; and the third level sends an anti-agglomeration control instruction to the stirring system, enhancing the fluid shear force in the southeast quadrant.

[0128] After the instructions were distributed via the OPC-UA protocol, the platform monitored the execution feedback in real time: the temperature sensor showed that the axial temperature difference decreased from 5 degrees to 1 degree; the ion electrode showed that the lithium-ion conversion rate had returned to the expected range; and the online particle size analyzer confirmed that the crystal agglomeration trend had disappeared. The entire process was completed within 2 minutes, avoiding batch quality incidents.

[0129] The following points should be noted during implementation: The construction of a digital twin platform should follow the principle of layered decoupling: The data acquisition layer is physically isolated from the equipment control system and protected against network attacks by an industrial firewall. The model calculation layer is deployed in a containerized manner to ensure that the material balance, energy optimization, and virtual detection modules operate independently. The application layer provides cross-platform visualization support.

[0130] Data timeliness management mechanisms include: The validity period of sensor data is marked as 3 seconds after acquisition; The prediction model output is valid for 2 minutes. Optimize instruction lifecycle by 1 minute.

[0131] Virtual simulation accuracy assurance measures: The coefficient of restitution of the crystal collision model is calibrated quarterly. The viscosity-temperature correlation parameters of the fluid dynamics model are updated annually. The heat transfer model is dynamically corrected according to the degree of fouling in the vessel.

[0132] Command distribution security policy: Important operational instructions require double confirmation (such as adjustments to the feed rate). 10%) The execution terminal feedback timeout will automatically trigger command withdrawal; Establish an instruction history traceability database to save the generation logic and execution effects of all optimized instructions.

[0133] Monthly full-system stress tests are conducted to simulate extreme scenarios such as large-scale sensor failures, network outages, and model inaccuracies, verifying the platform's ability to operate under degraded conditions.

[0134] This rigorous implementation framework ensures the reliability of digital twin systems in complex industrial environments.

[0135] This invention addresses three major industry pain points in lithium carbonate production processes—delayed quality feedback, control instability, and the risk of boiling over—by constructing a comprehensive intelligent control system encompassing "sensing—prediction—optimization—execution—correction—closed-loop." Its core innovative value lies in: First, the combination of multi-source sensor networks and physical mechanism-embedded predictive models enables minute-level precise sensing of crystal growth states, completely replacing the traditional lag-driven model relying on manual offline detection, and upgrading process control from experience-driven to data-driven. Second, the synergy between time-varying target intervals and Pareto dynamic optimization algorithms overcomes the control instability problem caused by strong coupling of multiple variables, maintaining high stability in crystal particle size and impurity content even under fluctuating raw material composition. Third, the millisecond-level collaborative control mechanism for boiling over suppression, through multi-actuator linkage involving temperature gradient adjustment, precise feed cutoff, and enhanced turbulence intensity, transforms traditional passive accident handling into proactive defense, significantly improving production safety boundaries. More importantly, the digital twin-driven closed-loop optimization architecture uses virtual mirrors to map physical reaction processes in real time, achieving comprehensive synergistic optimization of material conversion, energy transfer, and crystal morphology. This drives a leap in production models from "post-event correction" to an intelligent paradigm of "pre-event prediction—in-event control—post-event evolution." This technology system not only significantly enhances the application value of products in the high-end lithium battery field but also, through continuous enhancement of self-learning and adaptive capabilities, builds an evolvable intelligent foundation for new energy material manufacturing, marking a milestone in the green and high-end transformation of the industry.

[0136] Furthermore, the intelligent control system for the lithium carbonate preparation process provided by this invention will be described below. The intelligent control system for the lithium carbonate preparation process described below can be referred to in correspondence with the intelligent control method for the lithium carbonate preparation process described above. Optionally, referencing... Figure 2 , Figure 2 This is a schematic diagram of the intelligent control system for the lithium carbonate preparation process provided by the present invention. The intelligent control system for the lithium carbonate preparation process includes...

[0137] State perception module 210: Collects data on temperature field, ion concentration, feed flow rate and turbulence intensity of the reactor through sensor array, and generates state feature vector through multivariate analysis; Growth prediction module 220: Inputs the feature vector into the depth prediction model and outputs a particle size distribution probability cloud map and impurity trend; Optimization decision module 230: Inputs the probability cloud map and impurity trend into the multi-objective optimizer to generate temperature compensation value, feeding correction coefficient and stirring adjustment command; Execution conversion module 240: converts the instruction input anti-interference execution module into a device drive signal; Model update module 250: acquires actual crystal data through an online particle size analyzer and updates the prediction model; Closed-loop control module 260: Outputs closed-loop optimization instructions to the execution terminal through the digital twin platform.

[0138] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart control method for a lithium carbonate preparation process, characterized in that, It includes: Collect data on the temperature field, ion concentration, feeding flow rate, and turbulence intensity of the reactor through a sensor array, and generate a state feature vector through multivariate analysis; Input the feature vector into a deep prediction model to output a probability cloud map of particle size distribution and impurity trends; Input the probability cloud map and impurity trends into a multi-objective optimizer to generate a temperature compensation value, a feeding correction coefficient, and a stirring adjustment instruction; Input the instruction into an anti-interference execution module to convert it into a device drive signal; Obtain actual crystal data through an online particle size analyzer and update the prediction model; Output a closed-loop optimization instruction to the execution terminal through a digital twin platform.

2. The intelligent control method for lithium carbonate preparation process according to claim 1, characterized in that, Input the feature vector into a deep prediction model to output a probability cloud map of particle size distribution and impurity trends. The deep prediction model is implemented using a dual-channel deep neural network architecture, including: The first channel processes the real-time feature vector through a three-dimensional convolutional layer to capture the spatial distribution characteristics of local supersaturated regions in the reactor; The second channel analyzes the temporal evolution law of the historical crystal growth trajectory through an attention-enhanced long short-term memory network; Introduce a population balance equation constraint that describes crystal growth kinetics in the feature fusion layer to make the output probability cloud map of particle size distribution conform to physical laws.

3. The intelligent control method for the lithium carbonate preparation process according to claim 2, characterized in that, The population balance equation is: ； The equation describes the dynamic evolution of the crystal population size distribution, where This represents the number density of a crystal of size L at time t. Represents the linear growth rate of the crystal, driven by supersaturation. As a nucleation rate source term, it quantifies the rate of new crystal formation. The agglomeration and sedimentation loss term characterizes the loss caused by crystal agglomeration. By embedding the equation as a physical constraint into the neural network loss function, the prediction model can be ensured to maintain high accuracy when the raw material composition fluctuates, and the crystal particle size prediction error can be kept stable.

4. The intelligent control method for lithium carbonate preparation process according to claim 1, characterized in that, Input the probability cloud map and impurity trends into a multi-objective optimizer to generate a temperature compensation value, a feeding correction coefficient, and a stirring adjustment instruction. The multi-objective optimizer operates based on a Pareto front dynamic optimization mechanism, including: Establish a dynamic target interval function for the median crystal size over time. The interval remains relatively wide in the early stage of the reaction to promote nucleation; Construct a constraint condition for maintaining a single-peak narrow distribution of the particle size distribution span, and suppress bimodal abnormal distribution through a penalty function; Take the minimization of the co-precipitation probability of alkali metal ions as the optimization goal, and use an improved non-dominated sorting algorithm to generate coordinated control instructions for temperature, feeding, and stirring.

5. The intelligent control method for the lithium carbonate preparation process according to claim 4, characterized in that, The dynamic target interval function is: ； The exponential decay function simulates the time-varying characteristics of the control target during the reaction process. and These represent the lower and upper limits of the target granularity at time t, respectively. and These are the initial boundary values, set according to the product grade. and The contraction rate constant controls the boundary contraction speed; By setting This allows the lower limit to shrink faster than the upper limit, providing ample nucleation space in the early stages of the reaction and strictly controlling crystal growth quality in the later stages, thereby reducing the range of particle size fluctuations between batches.

6. The intelligent control method for the lithium carbonate preparation process according to claim 1, characterized in that, Input the instruction into an anti-interference execution module to convert it into a device drive signal. The anti-interference execution module implements a boiling suppression coordinated control strategy, including: When the supersaturation probability breaks through the critical threshold, immediately initiate a three-stage gradient cooling protocol; Simultaneously reset the feeding correction coefficient to zero and maintain the minimum material flow, while increasing the stirring power to the turbulent enhancement zone; Ensure millisecond-level linkage response of temperature regulation, feeding truncation, and turbulent enhancement through hardware-level timestamp synchronization.

7. The intelligent control method for the lithium carbonate preparation process according to claim 6, characterized in that, The determination of the turbulent enhancement zone uses the Reynolds number criterion: ； The formula quantifies the effect of fluid turbulence intensity on crystal dispersion. Let Reynolds number be 1. It is the solution density. This indicates the stirring speed. The diameter of the agitator blades. Represents solution viscosity. For the critical value of turbulent transition; when When the fluid enters a fully turbulent state, it provides sufficient shear force to disperse the crystal aggregates, thereby improving the product qualification rate.

8. The intelligent control method for lithium carbonate preparation process according to claim 1, characterized in that, Obtain actual crystal data through an online particle size analyzer and update the prediction model. The prediction model update uses a joint algorithm of residual backpropagation and state estimation, including: First, obtain the actual crystal D50 value and Span value through an online particle size analyzer, and calculate the KL divergence residual between them and the prediction result; The convolutional neural network channel adjusts the weights of the convolutional kernels through the backpropagation algorithm based on the residual distribution, and focuses on optimizing the feature extraction layer sensitive to supersaturation; The long short-term memory network channel uses the extended Kalman filter algorithm to calibrate the temporal prediction deviation of the hidden state vector using the innovation covariance matrix; Perform full model parameter update every 15 minutes, and immediately trigger an emergency correction mode when a raw material composition mutation is detected.

9. The intelligent control method for the lithium carbonate preparation process according to claim 1, characterized in that, The digital twin platform outputs closed-loop optimization instructions to the execution terminal, and the digital twin platform includes: The material balance engine tracks changes in lithium-ion concentration gradient in real time, calculates the deviation between the current conversion rate and the target value, and predicts the reaction endpoint. The energy optimization module integrates temperature field distribution and steam flow data to establish a thermodynamic efficiency model and dynamically adjust the steam supply. The virtual detection unit simulates the crystal collision and aggregation process through discrete element simulation, outputs an animation of crystal morphology evolution, and generates a quality warning. Every 5 minutes, the platform compares the predicted state of the digital image with the actual sensor data and outputs commands for axial temperature difference compensation, dynamic feeding ratio, and anti-agglomeration shear force control.

10. A system for implementing the intelligent control method of the lithium carbonate preparation process according to claims 1-9, characterized in that, include: State perception module: Collects data on temperature field, ion concentration, feed flow rate and turbulence intensity of the reactor through sensor array, and generates state feature vector through multivariate analysis; Growth prediction module: Inputs the feature vector into the depth prediction model and outputs a particle size distribution probability cloud map and impurity trend; Optimization Decision Module: Inputs the probability cloud map and impurity trend into the multi-objective optimizer to generate temperature compensation value, feeding correction coefficient and stirring adjustment command; Execution conversion module: converts the instruction input to the anti-interference execution module into a device drive signal; Model update module: Acquires actual crystal data using an online particle size analyzer and updates the prediction model; Closed-loop control module: Outputs closed-loop optimization instructions to the execution terminal through the digital twin platform.

Images