An automatic control system and method for lithium carbonate crystalline particle size and purity regulation

Abstract

The application relates to the technical field of inorganic chemical crystallization process control, and discloses an automatic control system and method for lithium carbonate crystallization particle size and purity regulation, wherein the system comprises: a data acquisition module configured to acquire multi-source heterogeneous data; a data preprocessing module configured to perform a preprocessing operation on the multi-source heterogeneous data and output a unified time sequence feature vector; a deep learning and prediction module configured to process the unified time sequence feature vector based on a multi-modal fusion deep learning architecture and output a prediction result; and a DCS automatic control module configured to compare the prediction result with a preset target threshold, generate a regulation instruction according to a priority, adjust a crystallization process operation parameter of a reaction kettle, and form a closed-loop control. Through online real-time monitoring, deep learning modeling and DCS closed-loop control, the application can realize accurate regulation of particle size, peritectic inhibition, early prediction of purity and stable standard reaching.

Description

Technical Field

[0001] This invention relates to the field of inorganic chemical crystallization process control technology, and in particular to an automatic control system and method for regulating the particle size and purity of lithium carbonate crystals. Background Technology

[0002] With the rapid development of the global new energy industry, the demand for battery-grade lithium carbonate in the power battery and energy storage sectors continues to grow, while higher requirements are being placed on product quality stability, particle size uniformity, and purity consistency. As a core raw material in the lithium battery industry chain, the control level of the crystallization process of lithium carbonate directly determines the performance indicators of the final product and also affects the operating efficiency and processing costs of subsequent solid-liquid separation, drying, and other processes. Currently, industrial lithium carbonate preparation mainly covers two technical routes: lithium extraction from ore and lithium extraction from brine in salt lakes. Both rely on the lithium precipitation crystallization process as the core preparation step. During the crystallization process, the crystal growth rate, morphological regularity, particle size distribution, and the degree of peritectic inclusion are key factors determining the purity and application performance of the product.

[0003] Currently, there are still many shortcomings in the control of crystal state and product quality during the lithium deposition crystallization process. On the one hand, the particle size, morphology, and peritectic state of crystals mainly rely on offline testing after manual sampling. The testing cycle is long, and the feedback of results is delayed. It is impossible to reflect the crystallization dynamics in the reactor in real time, making it difficult to intervene in the process in a timely manner for crystallization anomalies. This can easily lead to problems such as uneven particle size and large differences in morphology between batches of products, increasing the processing difficulty of subsequent processes. On the other hand, the determination of product purity relies on laboratory testing methods. The test results have a long time lag and lack the ability to predict the final purity in advance. Often, the problem of substandard purity can only be discovered after the product is produced, resulting in waste of raw materials and energy consumption and increasing the production defect rate.

[0004] At the process control level, current production largely relies on manually setting parameters or using fixed process curves. Key parameters such as feed ratio, stirring intensity, reaction temperature, and residence time are difficult to dynamically optimize based on real-time crystallization conditions. The control effectiveness is highly dependent on the operator's experience, resulting in significant batch-to-batch quality fluctuations and difficulty in ensuring consistency. Furthermore, existing technologies lack an integrated system for online detection, intelligent analysis, and automatic control. Data from each stage is independent, failing to form a collaborative closed-loop management system. This hinders the stable production of high-purity, uniformly sized lithium carbonate products that meet the requirements of high-end lithium battery applications, thus restricting the improvement of production automation and intelligence. Summary of the Invention

[0005] To address the technical problems of lagging particle size control, lack of real-time monitoring of peritectic state, and inability to timely adjust purity detection in the intermittent and continuous lithium carbonate precipitation process, this invention proposes an automatic control system and method for controlling the particle size and purity of lithium carbonate crystals. Through online real-time monitoring, deep learning modeling, and DCS (distributed control system) closed-loop control, precise particle size control, peritectic suppression, and early prediction and stable achievement of purity standards can be achieved.

[0006] The technical solution adopted in this invention is as follows: An automatic control system for regulating the particle size and purity of lithium carbonate crystals includes: The data acquisition module is configured to acquire multi-source heterogeneous data, including morphological image data and particle size detection data of lithium carbonate crystals in the reactor, as well as process variables of the lithium precipitation reaction. The data preprocessing module is configured to perform preprocessing operations on the multi-source heterogeneous data and output a unified temporal feature vector; the preprocessing operations include data cleaning, feature extraction and temporal alignment; The deep learning and prediction module is configured to process the unified temporal feature vector based on a multimodal fusion deep learning architecture and output prediction results; the prediction results include purity prediction value, granularity compliance score and peritectic risk level; The DCS automatic control module is configured to compare the predicted results with preset target thresholds, generate control commands according to priority, and adjust the crystallization process operating parameters of the reactor to form a closed-loop control; the crystallization process operating parameters include feed parameters, stirring speed, reaction temperature and residence time.

[0007] Furthermore, the collection of multi-source heterogeneous data includes: The morphology and peritectic state data of lithium carbonate crystals are collected by an online microscopic observation unit installed on the inner wall of the lithium precipitation reactor. The online microscopic observation unit has automatic focus adjustment and self-cleaning functions. Particle size data of lithium carbonate crystals were collected using an online particle size analysis unit. The process variables of the lithium deposition reaction are accessed through a process data interface. These process variables include parameters such as feed, reaction temperature, stirring intensity, residence time, and slurry density.

[0008] Furthermore, the data cleaning performed by the data preprocessing module includes: The morphological image data of lithium carbonate crystals in the reactor were screened for quality, and invalid image frames that were blurred or out of focus were removed. Uneven lighting background was corrected and local image contrast was enhanced. The particle size detection data of lithium carbonate crystals in the reactor are processed for noise smoothing, and abnormal measurement values ​​are detected and interpolated for correction. Missing values ​​were filled in for the process variables of the lithium precipitation reaction, and abnormal data in invalid time periods were marked and removed.

[0009] Furthermore, the feature extraction performed by the data preprocessing module includes: From the morphological image data of lithium carbonate crystals in the reactor, we extract the shape dimension features of single crystals, the statistical features of the proportion of crystals with different morphologies, and the quantitative features of the area proportion and thickness of peritectic regions. From the particle size detection data of lithium carbonate crystals in the reactor, the particle size distribution morphology characteristics and the dynamic change characteristics of particle size over time are extracted. The process variables of the lithium precipitation reaction are standardized in a dimensionless manner to extract time-series process features at a uniform scale.

[0010] Furthermore, the timing alignment performed by the data preprocessing module includes: Match the acquisition frequency of various data sources and set a unified reference time step; take the statistical mean of the time period for image features acquired at high frequency according to the reference time step; take the average of the time period for process variables according to the reference time step; match the effective measurement value of the corresponding time step for granular features acquired at low frequency; stitch various feature data according to the time dimension to generate a unified time-series feature vector corresponding to each time step, and form a continuous time-series feature sequence.

[0011] Furthermore, the multimodal fusion deep learning architecture includes: The image feature extraction subnetwork is configured to receive morphological image-related features and output an image semantic feature vector. The temporal process feature encoding subnetwork is configured to receive continuous temporal feature sequences, extract temporal dynamic features, and output temporal feature vectors; The multimodal feature fusion layer is configured to concatenate the image semantic feature vector and temporal feature vector, and to assign dynamic weights to different feature channels through an attention mechanism; The multi-task output layer is configured to output multi-dimensional prediction results in parallel based on the fused features.

[0012] Furthermore, the multi-task output layer employs a multi-task joint loss mechanism to complete model training, including: For the purity prediction task, a regression loss function robust to outliers is set. Set a binary classification cross-entropy loss function for the granularity achievement task; Set up a classification loss function that adapts to class imbalance for tasks with peritectic risk levels; The weight coefficients are set to balance the training contributions of the three types of tasks; the model training adopts a two-stage strategy of transfer learning pre-training combined with full fine-tuning, and expands the training samples with image and time series data augmentation methods.

[0013] Furthermore, the DCS automatic control module generates control commands according to priority, including: Set purity assurance as the first priority. When the predicted purity value is lower than the target threshold range, adjust the feed ratio and stirring speed first. Set peritectic suppression as the second priority. When the peritectic risk level reaches the high risk level, adjust the stirring intensity and reaction temperature. Particle size optimization is set as the third priority. When the particle size compliance rate is lower than the set standard, the feed flow rate gradient and stirring speed are adjusted. Set steady-state maintenance as the fourth priority. When all three types of prediction indicators meet the target threshold, fine-tune and optimize within the current parameter range.

[0014] Furthermore, the DCS automatic control module adjusts the crystallization process operating parameters of the reactor and forms a closed-loop control, including: Adjust the feed parameters to control the supply rate of reactants and the proportion of reactant components; Adjust the stirring speed to control the shear strength of the fluid inside the vessel and the suspension and dispersion state of the crystals; Adjusting the reaction temperature to control the crystallization reaction rate and the supersaturation of the system; Adjusting the material residence time to control the crystal growth cycle; The process operation data and crystal state data after parameter adjustment are transmitted back to the data acquisition module, forming a closed-loop feedback control for the entire process of monitoring, prediction and regulation.

[0015] An automatic control method for regulating the particle size and purity of lithium carbonate crystals, comprising: Collect multi-source heterogeneous data, including morphological image data and particle size detection data of lithium carbonate crystals in the reactor, as well as process variables of the lithium precipitation reaction; Preprocessing operations are performed on the multi-source heterogeneous data to output a unified temporal feature vector; the preprocessing operations include data cleaning, feature extraction and temporal alignment. The unified temporal feature vector is processed based on a multimodal fusion deep learning architecture, and the prediction results are output. The prediction results include the purity prediction value, the granularity compliance score, and the peritectic risk level. The predicted results are compared with preset target thresholds, and control instructions are generated according to priority to adjust the crystallization process operation parameters of the reactor, forming a closed-loop control. The crystallization process operation parameters include feed parameters, stirring speed, reaction temperature, and residence time.

[0016] The beneficial effects of this invention are as follows: 1. This invention constructs a complete closed-loop control system for the lithium deposition crystallization process, from state perception, data analysis, quality prediction to parameter control. It breaks through the limitations of relying on manual detection and manual parameter adjustment in traditional production. It can obtain the crystal state and process operation information in real time during the crystallization process, predict product quality indicators in advance, and automatically complete the optimization of process parameters, realizing the automated and intelligent control of the crystallization process, and effectively improving the response speed and control accuracy of the production process.

[0017] 2. This invention, through online monitoring and standardized preprocessing of multi-source heterogeneous data, can continuously acquire comprehensive data on crystal morphology, peritectic state, particle size distribution, and process operation without interfering with the crystallization reaction, eliminating process interference and detection lag caused by manual sampling. Targeted data cleaning, multi-dimensional feature extraction, and time-series alignment effectively filter out noise interference and abnormal data in on-site detection, unifying the time dimension and data scale of different data sources. This provides high-quality, highly reliable input data for subsequent intelligent prediction, ensuring the accuracy and stability of the prediction results.

[0018] 3. This invention employs a multimodal fusion deep learning architecture, which can simultaneously mine the spatial features of crystal morphology images and the temporal dynamic features of the process. Through feature fusion and attention weight allocation, it achieves multi-task synchronous prediction of crystal grain size compliance, peritectic risk level, and product purity. This prediction mode can predict quality indicators in advance before product output, replacing the traditional lagging offline testing method and reserving sufficient response time for process control. At the same time, the multi-task joint training model architecture can fully explore the coupling correlation between various parameters in the crystallization process, adapt to application scenarios with limited data in industrial sites, and has good generalization ability and prediction accuracy.

[0019] 4. This invention utilizes a graded priority DCS automatic control logic, which prioritizes product purity as the core indicator when multiple quality metrics deviate, while simultaneously addressing peritectic suppression and particle size optimization needs. This avoids system fluctuations caused by simultaneous adjustments of multiple parameters, ensuring stable operation of the crystallization process. Through coordinated adjustment of core process parameters such as feed, stirring, temperature, and residence time, the crystallization process is precisely controlled from multiple dimensions, including reaction components, flow field environment, reaction kinetics, and crystal growth cycle. Combined with a closed-loop feedback mechanism, the control effect is continuously verified and parameters are dynamically corrected. This effectively improves batch consistency, reduces peritectic entrainment issues, stably controls product particle size distribution, and enhances the operational efficiency of subsequent solid-liquid separation processes. Attached Figure Description

[0020] Figure 1 This is a flowchart of an automatic control method for regulating the particle size and purity of lithium carbonate crystals according to Embodiment 2 of the present invention. Detailed Implementation

[0021] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] Example 1 This embodiment provides an automatic control system for regulating the particle size and purity of lithium carbonate crystals, including a data acquisition module, a data preprocessing module, a deep learning and prediction module, and a DCS automatic control module, wherein: The data acquisition module is configured to acquire multi-source heterogeneous data, including morphological image data and particle size detection data of lithium carbonate crystals in the reactor, as well as process variables of the lithium precipitation reaction. The data preprocessing module is configured to perform preprocessing operations on the multi-source heterogeneous data and output a unified temporal feature vector; the preprocessing operations include data cleaning, feature extraction and temporal alignment; The deep learning and prediction module is configured to process the unified temporal feature vector based on a multimodal fusion deep learning architecture and output prediction results; the prediction results include purity prediction value, granularity compliance score and peritectic risk level; The DCS automatic control module is configured to compare the predicted results with preset target thresholds, generate control commands according to priority, and adjust the crystallization process operating parameters of the reactor to form a closed-loop control; the crystallization process operating parameters include feed parameters, stirring speed, reaction temperature and residence time.

[0023] Preferably, the acquisition of multi-source heterogeneous data includes: acquiring lithium carbonate crystal morphology and peritectic state data through an online microscopic observation unit located on the inner wall of the lithium precipitation reactor; acquiring lithium carbonate crystal particle size detection data through an online particle size analysis unit; and accessing process variables of the lithium precipitation reaction through a process data interface.

[0024] Specifically, the probe of the online microscopic observation unit is embedded in the inner wall of the reactor, directly contacting the material inside the reactor without interfering with the flow field inside the reactor. The probe has a built-in automatic focusing component and a self-cleaning component, which can automatically adjust the imaging focal length and periodically clean the crystals adhering to the probe surface; the online particle size analysis unit is connected to the discharge pipeline or the internal circulation pipeline of the reactor to detect crystal particle size-related data in real time; the process data interface communicates with the field control system to read the entire process parameters such as feed, reaction temperature, stirring intensity, residence time, and slurry density.

[0025] It should be noted that by adopting the above-mentioned online detection method, the morphology, peritectic and grain size of the crystals can be obtained in real time without manual sampling, avoiding the interference of the sampling process on the crystallization system, and solving the problem of strong lag in offline detection, thus providing a data foundation for subsequent real-time control.

[0026] Preferably, the data cleaning performed by the data preprocessing module includes: quality screening and image enhancement processing of morphological image data, noise smoothing and outlier correction of granularity detection data, and missing value filling and invalid data removal of process variables.

[0027] Specifically, for morphological image data, out-of-focus and blurry image frames are filtered out and removed through sharpness calculation. A background correction algorithm is used to eliminate imaging deviations caused by uneven lighting inside the reactor. Then, an image enhancement algorithm is used to improve the contrast between crystal edges and peritectic regions. For particle size detection data, a filtering algorithm is used to remove random noise from the detection signal. An outlier identification algorithm is used to locate measurement data that deviates from the normal range and correct it using interpolation. For process variables, short-term missing data is filled in using interpolation, and abnormal data segments that have been invalid for a long time are marked and excluded from the valid data.

[0028] It should be noted that targeted data cleaning can eliminate noise, interference, and missing data in on-site detection, improve the reliability and accuracy of input data, and avoid abnormal data causing deviations in subsequent feature extraction and model prediction results.

[0029] Preferably, the feature extraction performed by the data preprocessing module includes: extracting crystal shape, morphology proportion and peritectic quantization features from morphological image data; extracting distribution morphology and dynamic change features from particle size detection data; and standardizing process variables and extracting time-series process features.

[0030] Specifically, after crystal segmentation of the morphological image, shape dimension features such as major axis, minor axis, aspect ratio, and roundness of the single crystal are extracted, and the proportion of crystals with different morphologies is statistically analyzed. At the same time, peritectic regions in the image are identified, and the quantitative values ​​of peritectic area proportion and peritectic thickness are calculated. For particle size detection data, morphological features such as width, skewness, and kurtosis of particle size distribution are extracted, and dynamic features such as the rate of change of particle size over time are calculated. For process variables, a standardization algorithm is used to perform dimensionless processing to eliminate dimensional differences between different parameters and form process features of a unified scale.

[0031] It should be noted that transforming the original images, granularity, and process data into quantifiable feature indicators not only preserves the core information of crystal state and process operation but also adapts to the data input requirements of deep learning models, providing support for multimodal feature fusion.

[0032] Preferably, the time alignment performed by the data preprocessing module includes: setting a unified reference time step, matching data from different acquisition frequencies in terms of time dimension, and generating a continuous time-series feature sequence.

[0033] Specifically, a unified reference time step is set according to the dynamic characteristics of the lithium deposition crystallization process. For image-type features with high acquisition frequency, the statistical mean of the features within the time period is calculated according to the reference time step. For process variables, the average value of the parameters within the time period is calculated according to the reference time step. For particle-type features with low acquisition frequency, the effective measurement value within the corresponding time step is matched. Finally, the various features under the same time step are spliced ​​together to generate a unified time-series feature vector corresponding to each time step, and then arranged into a continuous time-series feature sequence in chronological order.

[0034] It should be noted that by using time-series alignment processing, the problem of inconsistent acquisition frequencies of different detection devices is solved, and multi-source heterogeneous data is integrated into a feature sequence with a unified time dimension. This ensures that the input data format of time-series deep learning models is compliant, while also ensuring the accuracy of the time correspondence between features.

[0035] Preferably, the multimodal fusion deep learning architecture includes an image feature extraction subnetwork, a temporal process feature encoding subnetwork, a multimodal feature fusion layer, and a multi-task output layer.

[0036] Specifically, the image feature extraction subnetwork receives preprocessed image-related features, extracts deep semantic information from the image through multi-layer convolution operations, and outputs a fixed-dimensional image semantic feature vector; the temporal process feature encoding subnetwork receives continuous temporal feature sequences, extracts dynamic change features of process and crystal state through a temporal coding network, and outputs the corresponding temporal feature vector; the multimodal feature fusion layer concatenates the image semantic feature vector and the temporal feature vector, and automatically adjusts the weights of different feature channels through an attention mechanism; the multi-task output layer performs parallel computation based on the fused features and outputs multi-dimensional prediction results.

[0037] It should be noted that the dual-branch multimodal fusion architecture can simultaneously mine the spatial features of crystal morphology images and the temporal features of the process. By adaptively allocating feature weights through an attention mechanism, the efficiency of the model in utilizing information from different dimensions is improved, ensuring the accuracy of multi-task prediction.

[0038] Preferably, the multi-task output layer uses a multi-task joint loss mechanism to complete model training, sets loss functions and weight coefficients corresponding to three types of tasks, and adopts a two-stage training strategy combined with data augmentation techniques.

[0039] Specifically, for the purity prediction regression task, a regression loss function with strong robustness to outliers is adopted; for the granularity attainment binary classification task, a binary cross-entropy loss function is adopted; and for the pericrystalline risk level classification task, a classification loss function that can alleviate class imbalance is adopted. Corresponding weight coefficients are set for the three loss functions to balance the training priorities of different tasks. Model training is divided into two stages. The first stage involves pre-training and fine-tuning the image feature extraction sub-network based on a public dataset to reduce the risk of overfitting with small samples. The second stage involves end-to-end full fine-tuning of the entire network after accumulating sufficient production data. During training, corresponding augmentation methods are used to expand the training samples for both image data and time-series data.

[0040] It should be noted that the multi-task joint loss mechanism can enable a single model to output multiple prediction indicators simultaneously. The two-stage training and data augmentation strategy can be adapted to scenarios where the amount of data is insufficient in the early stages of industrial sites, improving the model's generalization ability and prediction accuracy, and ensuring the stable operation of the model under different production conditions.

[0041] Preferably, the DCS automatic control module generates control commands according to priority, and sets four control priorities in sequence: purity assurance, peritectic suppression, particle size optimization, and steady-state maintenance.

[0042] Specifically, purity assurance is set as the first priority. When the predicted purity value is lower than the target threshold range, the feed ratio and stirring speed are adjusted first to optimize the proportion of reaction components and the mixing effect in the reactor. Peritecosis suppression is set as the second priority. When the peritectic risk reaches a high risk level, the stirring intensity and reaction temperature are adjusted to destroy the environmental conditions for peritectic formation. Particle size optimization is set as the third priority. When the particle size compliance is lower than the set standard, the feed gradient and stirring speed are adjusted to guide uniform crystal growth. Steady-state maintenance is set as the fourth priority. When all three predicted indicators meet the target requirements, small adjustments are made within the current parameter range to maintain the stable operation of the crystallization process.

[0043] It should be noted that by adopting a graded priority control logic, the core indicator of product purity can be prioritized when multiple indicators deviate simultaneously, while also taking into account the needs of peritectic suppression and particle size control. This avoids system fluctuations caused by simultaneous adjustment of multiple parameters and ensures a stable and controllable crystallization process.

[0044] Preferably, the DCS automatic control module adjusts the crystallization process operating parameters of the reactor and forms a closed-loop control. It achieves crystallization regulation by adjusting four types of parameters: feed, stirring, temperature, and residence time, and forms a closed loop through data feedback.

[0045] Specifically, when adjusting the feed parameters, the feed flow rate and ratio of lithium liquid and precipitant are adjusted by controlling the opening of the feed pump, thereby controlling the supply rate of reactants and the proportion of reaction components. When adjusting the stirring speed, the shear intensity of the fluid inside the reactor is adjusted by controlling the speed of the stirring motor, ensuring uniform suspension and dispersion of crystals. When adjusting the reaction temperature, the reaction temperature inside the reactor is adjusted by controlling the flow rate of the heating / cooling medium, thereby controlling the supersaturation and reaction rate of the crystallization system. When adjusting the material residence time, the average residence time of the material in the reactor is controlled by adjusting the discharge rate, thereby controlling the crystal growth cycle. After the parameters are adjusted, the crystal state and process operation data inside the reactor are re-acquired by the data acquisition module, entering the next round of prediction and control cycle.

[0046] It should be noted that by precisely adjusting the four core crystallization process parameters, the crystal growth process can be synergistically controlled from multiple dimensions such as reaction components, fluid environment, reaction kinetics and growth cycle. At the same time, the closed-loop feedback mechanism can continuously verify the control effect and dynamically correct the parameters, so as to achieve continuous optimization and stable control of the crystallization process.

[0047] Example 2 This embodiment provides an automatic control method for regulating the particle size and purity of lithium carbonate crystals, including: Collect multi-source heterogeneous data, including morphological image data and particle size detection data of lithium carbonate crystals in the reactor, as well as process variables of the lithium precipitation reaction; Preprocessing operations are performed on the multi-source heterogeneous data to output a unified temporal feature vector; the preprocessing operations include data cleaning, feature extraction and temporal alignment. The unified temporal feature vector is processed based on a multimodal fusion deep learning architecture, and the prediction results are output. The prediction results include the purity prediction value, the granularity compliance score, and the peritectic risk level. The predicted results are compared with preset target thresholds, and control instructions are generated according to priority to adjust the crystallization process operation parameters of the reactor, forming a closed-loop control. The crystallization process operation parameters include feed parameters, stirring speed, reaction temperature, and residence time.

[0048] Specifically, after the lithium precipitation reaction is started, morphological image data, particle size detection data, and process variables of lithium carbonate crystals in the reactor are collected simultaneously. The collected multi-source heterogeneous data are sequentially cleaned, feature extracted, and time-aligned to generate a unified time-series feature vector. The unified time-series feature vector is input into a pre-trained multimodal fusion deep learning architecture to calculate the purity prediction value, particle size compliance score, and peritectic risk level. The prediction results are compared with the preset production target threshold, and process control instructions are generated according to the set priority logic to adjust the reactor feed parameters, stirring speed, reaction temperature, and residence time. The adjusted operating data is re-entered into the acquisition stage to form a closed-loop control cycle.

[0049] It should be noted that this method combines online detection, intelligent prediction and automatic control to achieve fully automated management of the lithium precipitation crystallization process. It eliminates the need for frequent manual sampling and parameter adjustment, and can stably control crystal size, reduce peritectic inclusions and ensure product purity, thereby improving batch consistency and production efficiency.

[0050] like Figure 1 As shown, the automatic control method of this embodiment can be implemented by the following steps: 1) The lithium precipitation reaction is started by continuously / intermittently feeding lithium source brine / solution and precipitant into the reactor; 2) Online microscopic observation acquires crystal morphology and peritectic state in real time; online particle size analyzer outputs median particle size and particle size distribution in real time; 3) Two data streams are simultaneously fed into the particle size-peritectic-purity coupled prediction model, which uses deep learning based on historical process and purity data to output real-time purity prediction values; 4) Compare the real-time crystal state and predicted purity with the target threshold to determine whether the standard is met; 5) If the standards are not met, the DCS will automatically adjust the following: feed flow rate, feed ratio, stirring speed, reaction temperature, and residence time to inhibit peritecticization and guide the directional growth of crystals; 6) Stable output after meeting the standards, achieving continuous / intermittent automated production with uniform particle size, light peritectic particles, and stable purity.

[0051] Specifically, the automatic control method of this embodiment can also be implemented using the following steps: 1. Online monitoring deployment 1) An online microprobe is installed inside the inner wall of the lithium deposition reactor to achieve undisturbed, real-time, in-situ acquisition of crystal morphology (proportion of needle-like / blocky / spherical crystals, crystal boundary integrity) and peritectic state (peritectic area ratio, peritectic thickness estimation) within the reactor. The microprobe is equipped with automatic focus adjustment and self-cleaning functions to prevent crystal adhesion from affecting imaging quality.

[0052] 2) The online particle size analyzer detects D50, D90 and distribution width in real time.

[0053] 3) Process variable acquisition The following key process variables should be collected through a DCS (Distributed Control System) at a frequency of no less than once per second:

[0054] 2. Data preprocessing and feature library establishment 1) Image data preprocessing The raw microscopic images are processed through the following steps before being fed into the feature extraction network: Image quality screening: The Laplacian operator is used to calculate the image sharpness score, and blurry frames below the threshold (e.g., variance <100) are automatically filtered out; Background homogenization: Top-Hat Transform is used to remove uneven lighting background and improve crystal edge contrast; Image enhancement: Adaptive histogram equalization (CLAHE) enhances local contrast; Morphological feature extraction: Extract the major axis, minor axis, aspect ratio (AR), circularity, solidity, and Fourier shape descriptor for each segmented crystal; Peripheral feature quantization: The periphery area ratio ρ is calculated through multi-scale texture analysis (GLCM gray-level co-occurrence matrix) and semantic segmentation of the periphery region. inclusion and the estimated peritectic thickness δ; Statistical Summary: Statistical vector of crystal morphology distribution for each frame: needle proportion f needle Block proportion f block Average AR, Average Circularity C mean ρ inclusion δ mean .

[0055] 2) Granularity data preprocessing Median filtering: Perform 5-point median filtering on the original granular time series to remove abrupt noise; Outlier detection: The 3σ criterion is used to detect and mark outlier points in granularity measurement, and the outlier points are interpolated with the mean before and after the outlier. Feature engineering: Calculate the rate of change of D50 ΔD50 / Δt, the ratio of D90 / D10, the Span value, and the particle size distribution skewness and kurtosis to reflect the dynamic evolution trend of particle size distribution.

[0056] 3) Process variable preprocessing and time-series alignment missing value handling: Linear interpolation fills in transient missing data in process variables (e.g., missing duration ≤ 30s). Data exceeding the threshold is marked as invalid time periods and removed during training. Normalization: Z-Score standardization is used to normalize all numerical process variables to eliminate the influence of dimensions; Time alignment: Since the acquisition frequency of microscopic images (e.g., 5 frames / s), the sampling frequency of particle size analyzer (e.g., 1 time / min) and the acquisition frequency of DCS (e.g., 1 time / s) are inconsistent, a unified time step of 1 minute is used as the base time step. The average value of image features within 1 minute is taken, the average value of process variables within minutes is taken, and the latest valid value of particle size data is taken to form a unified time feature vector X(t). Dataset construction: Based on historical test purity P lab (ICP-MS measurement, accuracy 0.001%) is used as the supervision label to construct training sample pairs (historical feature sequence X, future purity P). lab The recommended backtracking time window T is 30-60 min, and the purity test lag time Δ is usually 30-90 min, to reflect the relationship between the current crystal state and the prediction of future purity.

[0057] 4) Dataset Construction and Quality Assessment The training dataset is recommended to meet the following requirements:

[0058] 3. Deep learning model training 1) Overall Model Architecture This embodiment employs a multimodal fusion deep learning architecture, fusing image feature streams and temporal process variable streams at the feature level to jointly predict three output targets: purity, granularity attainment, and peritectic risk level. The overall model architecture is as follows: Image feature extraction subnetwork: A lightweight convolutional neural network (such as MobileNetV3 or EfficientNet-B0) serves as the backbone network. The input is a preprocessed microscopic image or image statistical feature vector, and the output is a 128-dimensional image semantic feature vector f. img .

[0059] Temporal process feature encoding sub-network (LSTM / Transformer branch): The input is a temporal feature matrix X(tT:t) of length T. It uses bidirectional LSTM (BiLSTM) or Transformer Encoder to extract temporal dynamic features and outputs a temporal feature vector f. seq .

[0060] Multimodal feature fusion layer: f img with f seq After being spliced, the data is fed into a fully connected layer, where dynamic weights are assigned to different feature channels through an attention mechanism.

[0061] Multi-task output layer: Parallel output of purity prediction value P predParticle size compliance score S size Peritec risk level R inclusion .

[0062] 2) Loss Function Design The multi-task learning loss function is designed as follows: L total = λ1·L purity + λ2·L size + λ3·L inclusion in: L purity Purity regression loss, using Huber Loss (δ=0.1) instead of MSE, is more robust to outlier labels; L purity = Huber(P pred , P lab ); L size : Granularity meets the binary classification cross-entropy loss; L size = BCE(S pred , S label ); L inclusion Peripheral risk three-class classification cross-entropy loss, using Focal Loss to alleviate class imbalance; L inclusion = FocalLoss(R pred , R label , γ=2); The weight coefficients (λ1, λ2, λ3) are initially set to (1.0, 0.5, 0.5) and dynamically adjusted through multi-objective optimization on the validation set. It is recommended to use Pareto optimal search to determine the final weight combination.

[0063] 3) Model training process To address the issue of insufficient early-stage data in industrial settings, a two-stage training strategy is adopted: Phase 1 (Transfer Learning Pre-training): The CNN backbone network is fine-tuned based on the weights pre-trained on ImageNet, the first 3 convolutional blocks are frozen, and only the subsequent layers and fusion layers are trained to reduce the risk of overfitting to small sample data. Phase Two (Full Fine-Tuning): Once the factory has collected more than 500 valid test samples, all layers are thawed for end-to-end fine-tuning.

[0064] To expand the training data and improve the model's generalization ability, the following data augmentation techniques were used: Image enhancement: random horizontal / vertical flipping, random rotation, random cropping, random brightness / contrast perturbation, Gaussian noise injection, simulating different lighting and focal length conditions of microscopic images; Time series data augmentation: time series jitter, time step random masking, time axis random stretching / compression.

[0065] 4) Model evaluation index system The model performance evaluation uses the following indicator system:

[0066] 4. DCS Closed-Loop Control Logic 1) Communicates with DCS to achieve real-time prediction and output of control commands within seconds. If the purity is predicted to be too low or the peritectic content is too high: appropriately reduce the feed rate, optimize stirring, fine-tune the temperature, and extend the stable residence time.

[0067] The DCS control module receives the output (P) of the deep learning model every minute. pred , S size , R inclusion The triplet, and the set target value (P) target , S target , R target The comparison is performed, and control instructions are generated according to the following priority logic: Priority 1 (Purity Guarantee): When P pred <P target - At 0.05%, the purity restoration control strategy is triggered, prioritizing the adjustment of the feed ratio and stirring speed; Priority 2 (peritectic suppression): When R inclusion = In high-risk situations, a peritectic suppression strategy is triggered, adjusting the stirring intensity and reaction temperature; Priority 3 (Granularity Adjustment): When S size When the particle size is less than 0.85, a particle size adjustment strategy is triggered to adjust the feed flow rate gradient and stirring speed. Priority 4 (Steady-state maintenance): When all three indicators meet the target, maintain the current parameter range and perform fine-tuning and optimization.

[0068] 2) Particle size too fine / too wide: Adjust the stirring intensity and feeding gradient to promote uniform crystal growth.

[0069] 3) Specific regulatory rules

[0070] In summary, this invention, through online real-time monitoring, deep learning modeling, and DCS closed-loop control, can achieve precise granularity control, peritectic suppression, early purity prediction, and stable achievement of standards.

[0071] Example 3 This embodiment is based on embodiment 2: This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the automatic control method for regulating the particle size and purity of lithium carbonate crystals as described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form.

[0072] Example 4 This embodiment is based on embodiment 2: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the automatic control method for regulating the particle size and purity of lithium carbonate crystals as described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0073] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

[0074] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. An automatic control system for regulating the particle size and purity of lithium carbonate crystals, characterized in that, include: The data acquisition module is configured to acquire multi-source heterogeneous data, including morphological image data and particle size detection data of lithium carbonate crystals in the reactor, as well as process variables of the lithium precipitation reaction. The data preprocessing module is configured to perform preprocessing operations on the multi-source heterogeneous data and output a unified time-series feature vector. The preprocessing operations include data cleaning, feature extraction, and temporal alignment. The deep learning and prediction module is configured to process the unified temporal feature vector based on a multimodal fusion deep learning architecture and output prediction results; the prediction results include purity prediction value, granularity compliance score and peritectic risk level; The DCS automatic control module is configured to compare the predicted results with preset target thresholds, generate control commands according to priority, and adjust the crystallization process operating parameters of the reactor to form a closed-loop control; the crystallization process operating parameters include feed parameters, stirring speed, reaction temperature and residence time.

2. The automatic control system for adjusting the particle size and purity of lithium carbonate crystals according to claim 1, characterized in that, The collection of multi-source heterogeneous data includes: The morphology and peritectic state data of lithium carbonate crystals are collected by an online microscopic observation unit installed on the inner wall of the lithium precipitation reactor. The online microscopic observation unit has automatic focus adjustment and self-cleaning functions. Particle size data of lithium carbonate crystals were collected using an online particle size analysis unit. The process variables of the lithium deposition reaction are accessed through a process data interface. These process variables include parameters such as feed, reaction temperature, stirring intensity, residence time, and slurry density.

3. The automatic control system for adjusting the particle size and purity of lithium carbonate crystals according to claim 1, characterized in that, The data cleaning performed by the data preprocessing module includes: The morphological image data of lithium carbonate crystals in the reactor were screened for quality, and invalid image frames that were blurred or out of focus were removed. Uneven lighting background was corrected and local image contrast was enhanced. The particle size detection data of lithium carbonate crystals in the reactor are processed for noise smoothing, and abnormal measurement values ​​are detected and interpolated for correction. Missing values ​​were filled in for the process variables of the lithium precipitation reaction, and abnormal data in invalid time periods were marked and removed.

4. The automatic control system for regulating the particle size and purity of lithium carbonate crystals according to claim 1, characterized in that, The feature extraction performed by the data preprocessing module includes: From the morphological image data of lithium carbonate crystals in the reactor, we extract the shape dimension features of single crystals, the statistical features of the proportion of crystals with different morphologies, and the quantitative features of the area proportion and thickness of peritectic regions. From the particle size detection data of lithium carbonate crystals in the reactor, the particle size distribution morphology characteristics and the dynamic change characteristics of particle size over time are extracted. The process variables of the lithium precipitation reaction are standardized in a dimensionless manner to extract time-series process features at a uniform scale.

5. The automatic control system for adjusting the particle size and purity of lithium carbonate crystals according to claim 1, characterized in that, The timing alignment performed by the data preprocessing module includes: Match the acquisition frequency of various data sources and set a unified reference time step; take the statistical mean of the time period for image features acquired at high frequency according to the reference time step; take the average of the time period for process variables according to the reference time step; match the effective measurement value of the corresponding time step for granular features acquired at low frequency; stitch various feature data according to the time dimension to generate a unified time-series feature vector corresponding to each time step, and form a continuous time-series feature sequence.

6. The automatic control system for adjusting the particle size and purity of lithium carbonate crystals according to claim 1, characterized in that, The multimodal fusion deep learning architecture includes: The image feature extraction subnetwork is configured to receive morphological image-related features and output an image semantic feature vector. The temporal process feature encoding subnetwork is configured to receive continuous temporal feature sequences, extract temporal dynamic features, and output temporal feature vectors; The multimodal feature fusion layer is configured to concatenate the image semantic feature vector and temporal feature vector, and to assign dynamic weights to different feature channels through an attention mechanism; The multi-task output layer is configured to output multi-dimensional prediction results in parallel based on the fused features.

7. The automatic control system for adjusting the particle size and purity of lithium carbonate crystals according to claim 6, characterized in that, The multi-task output layer employs a multi-task joint loss mechanism to complete model training, including: For the purity prediction task, a regression loss function robust to outliers is set. Set a binary classification cross-entropy loss function for the granularity achievement task; Set up a classification loss function that adapts to class imbalance for tasks with peritectic risk levels; The weight coefficients are set to balance the training contributions of the three types of tasks; the model training adopts a two-stage strategy of transfer learning pre-training combined with full fine-tuning, and expands the training samples with image and time series data augmentation methods.

8. The automatic control system for adjusting the particle size and purity of lithium carbonate crystals according to claim 1, characterized in that, The DCS automatic control module generates control commands according to priority, including: Set purity assurance as the first priority. When the predicted purity value is lower than the target threshold range, adjust the feed ratio and stirring speed first. Set peritectic suppression as the second priority. When the peritectic risk level reaches the high risk level, adjust the stirring intensity and reaction temperature. Particle size optimization is set as the third priority. When the particle size compliance rate is lower than the set standard, the feed flow rate gradient and stirring speed are adjusted. Set steady-state maintenance as the fourth priority. When all three types of prediction indicators meet the target threshold, fine-tune and optimize within the current parameter range.

9. The automatic control system for adjusting the particle size and purity of lithium carbonate crystals according to claim 1, characterized in that, The DCS automatic control module adjusts the crystallization process operating parameters of the reactor and forms a closed-loop control, including: Adjust the feed parameters to control the supply rate of reactants and the proportion of reactant components; Adjust the stirring speed to control the shear strength of the fluid inside the vessel and the suspension and dispersion state of the crystals; Adjusting the reaction temperature to control the crystallization reaction rate and the supersaturation of the system; Adjusting the material residence time to control the crystal growth cycle; The process operation data and crystal state data after parameter adjustment are transmitted back to the data acquisition module, forming a closed-loop feedback control for the entire process of monitoring, prediction and regulation.

10. An automatic control method for regulating the particle size and purity of lithium carbonate crystals, characterized in that, include: Collect multi-source heterogeneous data, including morphological image data and particle size detection data of lithium carbonate crystals in the reactor, as well as process variables of the lithium precipitation reaction; Preprocessing operations are performed on the multi-source heterogeneous data to output a unified time-series feature vector; The preprocessing operations include data cleaning, feature extraction, and temporal alignment. The unified temporal feature vector is processed based on a multimodal fusion deep learning architecture, and the prediction results are output. The prediction results include the purity prediction value, the granularity compliance score, and the peritectic risk level. The predicted results are compared with preset target thresholds, and control instructions are generated according to priority to adjust the crystallization process operation parameters of the reactor, forming a closed-loop control. The crystallization process operation parameters include feed parameters, stirring speed, reaction temperature, and residence time.

Images