Formate salt crystallization solution packaging intelligent identification and collaborative control method and system

Abstract

The application relates to the technical field of crystallization process control, and discloses a formate salt crystallization solution package intelligent recognition and collaborative control method and system, wherein the method comprises the following steps: acquiring multi-modal original data and performing source separation noise reduction and time synchronization; generating dynamic fusion features based on cross-modal attention weighting; generating an abnormal enhanced thickness feature map in a flow field crystallization coupling prediction mode; obtaining a package state recognition result based on an encoder-decoder segmentation mode; and generating a collaborative control instruction according to the package state recognition result. Compared with the prior art which relies on single sensing data or static control rules, the technical problem that accurate package state recognition and timely adjustment cannot be realized under the conditions of dynamic flow, bubble shielding and crystal adhesion is solved. Due to the multi-modal dynamic fusion, thickness abnormality enhancement and recognition result feedback control, real-time recognition and collaborative adjustment of the package abnormality are realized, and the crystallization process stability and product consistency are improved.

Description

Technical Field

[0001] This invention relates to the field of crystallization process control technology, and in particular to a method and system for intelligent identification and collaborative control of formate crystallization solution encapsulation. Background Technology

[0002] Formate crystallization solution encapsulation has important applications in high-end chemical, pharmaceutical, and fine materials preparation fields. The uniformity of the encapsulation layer thickness and the crystal morphology directly affect the purity, solubility, and stability of the product. This process typically involves the coupling of multiple physical fields, including solution flow, mass transfer, crystal growth, and mechanical stirring. The encapsulation state is influenced by various factors such as flow rate distribution, solute concentration, temperature, and bubble generation, exhibiting strong nonlinearity and time-varying characteristics, which poses a significant challenge to real-time accurate monitoring and effective control.

[0003] Currently, commonly used monitoring methods in industrial settings include high-speed imaging, ultrasonic detection, and conductivity measurement. High-speed visual images can acquire information on the morphology and size of encapsulated particles, but are easily affected by image noise caused by air bubbles and solution turbidity. Ultrasonic reflection signals can reflect the location and stratification of crystallization interfaces; however, their signal-to-noise ratio is low, and interface echo characteristics are easily submerged by flow field disturbances. Solution conductivity data can indicate the overall concentration change trend, but has poor spatial resolution and is not sensitive to local encapsulation anomalies. Traditional methods are mostly based on single sensors or simple feature thresholds, lacking effective integration of multimodal data in the spatiotemporal dimensions, making it difficult to comprehensively capture the dynamic characteristics of the encapsulation process. In addition, existing control strategies mostly use fixed-parameter PID or empirical rules, which cannot adjust the stirring rate and solution replenishment in real time according to changes in the flow field and abnormal encapsulation states, resulting in uneven encapsulation layer thickness, frequent occurrence of abnormal agglomerates, and difficulty in ensuring process stability and product qualification rate. Therefore, there is an urgent need for a method that can integrate multi-source sensor information and achieve collaborative intelligent control to meet the requirements for accurate identification and stable control of the encapsulation process of formate crystallization solutions under complex working conditions. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the present invention aims to propose an intelligent identification and collaborative control method for formate crystal solution encapsulation. This method addresses the technical problem that existing technologies rely on single sensor data or static control rules, particularly under dynamic flow, bubble obstruction, and crystal adhesion conditions, making it impossible to accurately identify and promptly adjust the encapsulation state.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a method for intelligent identification and collaborative control of formate crystal solution encapsulation.

[0006] The method for intelligent identification and collaborative control of formate crystal solution encapsulation includes:

[0007] Step S10: Obtain multimodal raw data of the formate crystallization solution encapsulation process, and perform data preprocessing tasks based on the multimodal raw data using source-specific noise reduction and time synchronization to output synchronized denoised multimodal data;

[0008] Step S20: Based on the synchronously denoised multimodal data, a cross-modal attention weighting method is used to perform a feature fusion task and output dynamic fused features;

[0009] Step S30: Based on the dynamic fusion features, the flow field crystallization coupling prediction method is used to perform the encapsulation anomaly enhancement task, and the anomaly enhancement thickness feature map is output;

[0010] Step S40: Based on the abnormally enhanced thickness feature map, perform the package status recognition task using an encoder-decoder segmentation method, and output the package status recognition result;

[0011] Step S50: Based on the package status recognition result, the crystallization process control task is executed using a fuzzy PID collaborative adjustment method, and a collaborative control command is output.

[0012] Preferably, step S10, which involves acquiring multimodal raw data of the formate crystallization solution encapsulation process, performing data preprocessing based on the multimodal raw data using a source-specific noise reduction and time synchronization method, and outputting synchronized denoised multimodal data, specifically includes:

[0013] Step S101: Simultaneously acquire high-speed visual images, ultrasonic reflection signals and solution conductivity data during the encapsulation process of formate crystallization solution, and record the acquisition timestamps corresponding to each sampled data to obtain time-stamped source sampling data;

[0014] Step S102: Perform source-specific noise reduction processing on the timestamped source-specific sampling data, wherein the high-speed visual image is subjected to adaptive median filtering to remove bubble occlusion noise, the ultrasonic reflection signal is subjected to wavelet threshold denoising to retain the echo characteristics of the crystal interface, and the solution conductivity data is subjected to Kalman filtering for smoothing to obtain source-specific noise reduction data;

[0015] Step S103: Based on the acquisition timestamp, perform time alignment and interpolation on the source-specific denoised data to obtain synchronous denoised multimodal data.

[0016] Preferably, step S20, which involves performing a feature fusion task based on the synchronously denoised multimodal data using a cross-modal attention weighting method and outputting dynamically fused features, specifically includes:

[0017] Step S201: Input the high-speed visual image, ultrasonic reflection signal and solution conductivity data in the synchronous denoised multimodal data into the visual branch convolutional network, the ultrasonic branch one-dimensional convolutional network and the conductivity branch fully connected network respectively to extract visual features, ultrasonic features and conductivity features;

[0018] Step S202: Based on the visual features, the ultrasonic features, and the electrical conductivity features, a cross-modal attention module is used to calculate the first modal weight, the second modal weight, and the third modal weight, wherein the first modal weight corresponds to the visual features, the second modal weight corresponds to the ultrasonic features, and the third modal weight corresponds to the electrical conductivity features;

[0019] Step S203: The visual features, ultrasonic features, and electrical conductivity features are weighted and fused according to the first modal weight, the second modal weight, and the third modal weight to obtain dynamic fused features.

[0020] Preferably, in step S202, the first modal weight, the second modal weight, and the third modal weight are determined according to the following formula:

[0021] ;

[0022] in, Modal number, This represents the first mode corresponding to a high-speed visual image. This represents the second mode corresponding to the ultrasonic reflection signal. This indicates the third mode corresponding to the solution conductivity data; For the first Modal weights; For the first Modal attention baseline score; For the first Signal confidence level of the modality; For the first The amount of interference penalty for the modality; The flow field disturbance is obtained by normalizing the rate of change of flow velocity, the rate of change of shear stress, and the bubble obstruction rate. For the first Confidence adjustment coefficient for the modality; For the first The interference modulation coefficient; For the first Modal flow field disturbance adjustment coefficient.

[0023] Preferably, step S30, which involves performing the encapsulation anomaly enhancement task based on the dynamic fusion features using a flow field crystallization coupling prediction method and outputting anomaly enhancement thickness feature maps, specifically includes:

[0024] Step S301: Extract particle boundary response, echo phase response and conductance gradient response based on the dynamic fusion features, and concatenate the particle boundary response, echo phase response and conductance gradient response to obtain coupled input features;

[0025] Step S302: Based on the coupled input features, the Noyes Whitney mass transfer growth equation and shear stripping correction term are used to predict the coating thickness to obtain the predicted thickness field;

[0026] Step S303: Based on the residual between the predicted thickness field and the measured thickness field obtained by inversion from the dynamic fusion feature, perform anomaly enhancement on the dynamic fusion feature to obtain an anomaly enhanced thickness feature map; wherein, the anomaly enhanced thickness feature map is determined according to the following formula:

[0027] ;

[0028] in, pixel position The abnormally enhanced thickness characteristic value at the location; pixel position Dynamic fusion feature values ​​at the location; pixel position The predicted thickness value at that location; pixel position The measured thickness value at the location; The thickness residual threshold; This is the abnormal enhancement coefficient; To prevent the stability constant from having a denominator of zero; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0029] Preferably, step S40, which involves performing package status recognition based on the abnormally enhanced thickness feature map using an encoder-decoder segmentation method and outputting the package status recognition result, specifically includes:

[0030] Step S401: Input the abnormally enhanced thickness feature map into the U-Net encoder-decoder network, and use the encoder to extract the wrapping layer boundary features, crystal contour features and aggregate texture features to obtain a multi-scale wrapping characterization;

[0031] Step S402: The multi-scale encapsulation representation is input to the decoder and upsampled step by step. Based on the preset thickness consistency constraint, cross-entropy loss and Dice coefficient loss, the encapsulation layer region, crystal region and abnormal agglomeration region are jointly segmented to obtain the joint segmentation result.

[0032] Step S403: Generate package status identification results based on the joint segmentation results. The package status identification results include a package layer thickness segmentation map, a crystal distribution map, and a binary mask of abnormal aggregates.

[0033] Preferably, step S50, which involves executing the crystallization process control task using a fuzzy PID collaborative adjustment method based on the package status identification result and outputting collaborative control commands, specifically includes:

[0034] Step S501: Calculate the thickness deviation between the current average thickness of the package layer and the set thickness value based on the package layer thickness segmentation map in the package status recognition result, and calculate the area ratio of abnormal aggregates based on the binary mask of abnormal aggregates in the package status recognition result to obtain the control input quantity;

[0035] Step S502: Input the control input quantity into the fuzzy PID controller, and adjust the proportional parameter, integral parameter and derivative parameter in real time according to the thickness deviation and the area ratio of the abnormal agglomerates to obtain the candidate control quantity;

[0036] Step S503: Based on the preset allowable speed range of the stirring equipment, the allowable range of solution concentration, and the maximum adjustment range per unit time, the candidate control quantity is subjected to amplitude limiting processing to obtain a coordinated control command. The coordinated control command includes a stirring rate adjustment command and a solution concentration fine-tuning command.

[0037] This invention also provides a smart identification and collaborative control system for formate crystallization solution encapsulation, comprising:

[0038] The data preprocessing module is used to acquire multimodal raw data of the formate crystallization solution encapsulation process, and to perform data preprocessing tasks based on the multimodal raw data using source-specific noise reduction and time synchronization methods, outputting synchronized denoised multimodal data;

[0039] The feature fusion module is used to perform a feature fusion task based on the synchronously denoised multimodal data using a cross-modal attention weighting method, and output dynamic fused features;

[0040] The coupling enhancement module is used to perform the encapsulation anomaly enhancement task based on the dynamic fusion features using the flow field crystallization coupling prediction method, and outputs the anomaly enhancement thickness feature map.

[0041] The real-time recognition module is used to perform package status recognition tasks based on the abnormal enhanced thickness feature map using an encoder-decoder segmentation method, and output the package status recognition results.

[0042] The collaborative control module is used to execute the crystallization process control task using a fuzzy PID collaborative adjustment method based on the package status identification result, and output collaborative control commands.

[0043] The present invention also provides a formate crystal solution encapsulation intelligent identification and collaborative control device, the formate crystal solution encapsulation intelligent identification and collaborative control device comprising: a memory, a processor, and a formate crystal solution encapsulation intelligent identification and collaborative control program stored in the memory and executable on the processor, the formate crystal solution encapsulation intelligent identification and collaborative control program implementing the above method when executed by the processor.

[0044] The present invention also provides a computer program product, the computer program product including a formate crystal solution encapsulation intelligent identification and collaborative control program, the formate crystal solution encapsulation intelligent identification and collaborative control program implementing the above method when executed by a processor.

[0045] The beneficial effects of this invention are as follows:

[0046] 1. This invention employs source-specific denoising and time-synchronized preprocessing to address the different noise characteristics of high-speed visual images, ultrasonic reflection signals, and solution conductivity. It utilizes adaptive median filtering, wavelet threshold denoising, and Kalman filtering respectively, and performs precise alignment and interpolation based on the acquisition timestamp. This effectively eliminates heterogeneous noise and time deviation in the original multimodal data. The output synchronously denoised multimodal data exhibits high fidelity and spatiotemporal consistency, providing a reliable data foundation for subsequent feature fusion.

[0047] 2. This invention, through cross-modal attention-weighted fusion, not only extracts deep features from vision, ultrasound, and conductivity, but also adaptively calculates the modal weights based on the attention baseline score, signal confidence, interference penalty, and flow field disturbance. Under interference such as bubble occlusion and solution turbidity, the fusion strategy is dynamically adjusted. The generated dynamic fusion features fully complement the advantages of each modality, suppress misleading information from contaminated modalities, and significantly enhance the robustness of feature representation and the ability to describe the crystallization encapsulation state.

[0048] 3. This invention employs a flow field crystallization coupling prediction method, combining dynamic fusion features with the Noyes-Whitney mass transfer growth equation and shear stripping correction terms to predict the coating thickness. The predicted-measured thickness residual is used to enhance the feature map, amplifying local abnormal signals that do not conform to the normal growth pattern. The generated abnormally enhanced thickness feature map effectively highlights the uneven coating and abnormal agglomeration areas, overcoming the shortcomings of traditional features in distinguishing subtle anomalies.

[0049] 4. This invention utilizes an encoder-decoder segmentation network and introduces thickness consistency constraints and a joint loss function to perform multi-scale segmentation of the encapsulation layer, crystals, and agglomerates in the abnormally enhanced thickness feature map. The resulting encapsulation layer thickness segmentation map, crystal distribution map, and abnormal agglomerate binary mask have clear boundaries and continuous thickness, providing a precise and comprehensive quantitative basis for control decisions. Based on this, fuzzy PID collaboratively adjusts the stirring rate and solution concentration, and adjusts control parameters in real time according to the encapsulation state, followed by amplitude limiting processing. This achieves multi-variable coordinated optimization, significantly improving the uniformity and process safety of the encapsulation process, and reducing human intervention and batch-to-batch variations. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the first embodiment of a method for intelligent identification and collaborative control of formate crystallization solution encapsulation according to the present invention.

[0051] Figure 2 This is a flowchart illustrating a second embodiment of the intelligent identification and collaborative control system for formate crystallization solution encapsulation according to the present invention;

[0052] Figure 3 This is a flowchart of cross-modal attention dynamic weighted fusion, representing the fifth embodiment of a method for intelligent recognition and collaborative control of formate crystal solution encapsulation according to the present invention. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0054] 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.

[0055] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the intelligent identification and collaborative control method for formate crystal solution encapsulation according to the present invention. The first embodiment of the intelligent identification and collaborative control method for formate crystal solution encapsulation according to the present invention is presented.

[0056] In the first embodiment, the method for intelligent identification and collaborative control of formate crystallization solution encapsulation includes:

[0057] Step S10: Obtain multimodal raw data of the formate crystallization solution encapsulation process, and perform data preprocessing tasks based on the multimodal raw data using source-specific noise reduction and time synchronization to output synchronized denoised multimodal data;

[0058] The multimodal raw data in this step refers to high-speed visual images, ultrasonic reflection signals, and solution conductivity data collected synchronously by a high-speed camera, ultrasonic transducer, and conductivity sensor during the encapsulation process of formate crystallization solution. All data are accompanied by high-precision acquisition timestamps. Source-specific denoising and time-synchronized preprocessing include: applying adaptive median filtering to the high-speed visual images to remove salt-and-pepper noise and random noise caused by bubble occlusion; applying wavelet thresholding to the ultrasonic reflection signals to preserve the abrupt changes in the crystallization interface echoes and suppress background noise; and applying Kalman filtering to smooth the solution conductivity sequence and eliminate high-frequency fluctuations in the measurements. Subsequently, using the timestamps of each sensor as a reference, interpolation and time alignment algorithms are used to map the source-specific denoised data onto a unified time grid, generating spatiotemporally strictly synchronized synchronous denoised multimodal data. This data will be directly used as the input for cross-modal feature fusion in step S20.

[0059] Employing a source-specific noise reduction strategy maximizes the preservation of key information related to the encapsulation state in each modality, such as particle contours in visual images, reflection peaks at crystal interfaces in ultrasonic signals, and trends in conductivity changes, while eliminating noise interference from different sources. Precise time synchronization avoids feature misalignment caused by sampling time deviations from multiple sensors, ensuring that different modalities describe the encapsulation state at the same instant during subsequent fusion. This guarantees the instantaneous consistency of the dynamic process representation, laying a solid foundation of data quality for high-precision intelligent recognition.

[0060] Traditional preprocessing methods typically employ uniform filtering for multimodal data, ignoring the differences in noise characteristics among different sensors. This can easily lead to over-smoothing of useful signals or residual noise. Simply relying on system clock alignment introduces significant time registration errors during the rapidly changing crystal wrapping process, due to varying sensor response delays. This invention matches a dedicated noise reduction algorithm to the physical characteristics of each modality and performs interpolation synchronization based on timestamps. Compared to existing technologies, this significantly improves the signal-to-noise ratio and timing synchronization accuracy of multimodal data, effectively avoiding subsequent feature distortion and recognition errors caused by image blurring, signal aliasing, and time misalignment.

[0061] For example, in a stirred crystallizer equipped with a high-speed camera, an ultrasonic probe, and a conductivity electrode, the vortex bubbles generated by the rotating impeller frequently adhere to the viewing window surface, resulting in significant random occlusion of the visual image. Simultaneously, the ultrasonic signal is affected by turbulent flow, producing spikes, and the conductivity curve oscillates due to local concentration fluctuations. After preprocessing in step S10, the bubbles in the visual image are cleanly etched away, and the edges of the crystal particles are clearly visible. The reflection waveform of the deposition layer in the ultrasonic A-scan signal becomes smooth, and the echo peak stably reflects the interface of the encapsulation layer. The conductivity data forms a smooth, gradual curve, and all three are precisely aligned to the same microsecond-level timeline. This synchronously denoised data stream provides ideal input for the next step of multidimensional feature extraction.

[0062] Step S20: Based on the synchronously denoised multimodal data, a cross-modal attention weighting method is used to perform a feature fusion task and output dynamic fused features;

[0063] This step first inputs the high-speed visual image from the synchronously denoised multimodal data into a lightweight convolutional neural network branch to extract visual features; inputs the ultrasonic reflection signal into a one-dimensional convolutional network branch to extract ultrasonic features; and inputs the solution conductivity data into a fully connected network branch to extract conductivity features. Then, the cross-modal attention module uses these three features as input, combined with the flow field disturbance *f* obtained by normalizing the flow velocity change rate, shear stress change rate, and bubble occlusion rate, to calculate the first modal weight corresponding to the visual feature, the second modal weight corresponding to the ultrasonic feature, and the third modal weight corresponding to the conductivity feature, according to the formula. This formula comprehensively considers the attention baseline score, signal confidence, interference penalty, and flow field disturbance adjustment coefficient, enabling adaptive weight allocation under different operating conditions. Finally, based on these weights, the three features are summed element-wise to output a dynamically fused feature, which serves as the core input for the flow field crystallization coupling prediction in step S30.

[0064] The cross-modal attention weighting mechanism empowers the network to autonomously select high-value information under various disturbances. When the quality of a mode deteriorates due to occlusion, noise, or failure, its corresponding weight automatically decreases; conversely, the weight of reliable modes is increased. The fused features always focus on the combination of modes that best reflects the true state of the package at that moment. This dynamic adjustment makes the fused features inherently robust to disturbances such as changes in illumination, bubble transients, and conductivity drift, significantly improving the generalization performance of subsequent prediction models under different batches and process parameters.

[0065] Most existing multimodal fusion methods employ equal-weighted splicing or fixed-weight learning, which cannot adapt to changes in online operating conditions once determined offline. When the bubble occlusion rate suddenly increases, the fixed high weight of the visual branch injects a large amount of invalid information into the fused features; the same applies when the ultrasonic or conductivity branches are disturbed. This invention introduces an adaptive weight formula that is instantaneously correlated with the amount of flow field disturbance, adjusting the fusion strategy based on real-time signal quality. This overcomes the limitations of traditional static fusion, ensuring high-fidelity fusion features are consistently obtained in industrial environments with drastically fluctuating flow fields. This is the key reason why this invention improves recognition reliability compared to existing technologies.

[0066] For example, in a typical crystallization scenario, during the initial stage of accelerated stirring, a large number of bubbles rise, severely obscuring the visual image. The signal confidence (ci) of the visual features drops sharply, while the interference penalty (di) increases. Simultaneously, the flow field disturbance (f) increases. According to the weighting formula, the weight of the visual modality is rapidly reduced, while the weights of ultrasound and conductivity are correspondingly increased. The fusion feature mainly relies on the interface echo from ultrasonic thickness measurement and changes in conductivity concentration to infer the encapsulation layer. When the bubbles dissipate and the visual image becomes clear again, the visual features regain high confidence, and their weights automatically recover, allowing the fusion feature to regain rich morphological details. In this way, the fusion feature can always accurately reflect the encapsulation state without being obscured by transient interference.

[0067] Step S30: Based on the dynamic fusion features, the flow field crystallization coupling prediction method is used to perform the encapsulation anomaly enhancement task, and the anomaly enhancement thickness feature map is output;

[0068] Based on dynamic fusion features, local response extraction is first performed: particle boundary response, ultrasonic echo phase response, and conductivity gradient response are obtained through gradient operators and phase analysis, and these three are concatenated into a coupled input feature. Next, the coupled input feature is fed into a coupled prediction network, which embeds the Noyes-Whitney mass transfer growth equation to describe the growth rate of the crystalline coating under a concentration gradient, and adds a shear stripping correction term to simulate the effect of flow field shear force on surface crystal erosion, thereby predicting the coating thickness at each pixel location, forming a predicted thickness field. Simultaneously, the measured thickness field is obtained through an inversion branch using the dynamic fusion features. Finally, the pixel-wise residual between the predicted and measured thickness fields is calculated to obtain an abnormally enhanced thickness feature map, which is directly fed into the segmentation network in step S40.

[0069] The flow-field crystallization coupling prediction combines inherent mass transfer dynamics knowledge with data-driven learning, incorporating the physical expectation of normal crystallization trends into the predicted thickness field. The anomaly enhancement mechanism, based on the deviation between prediction and reality, precisely amplifies anomalous regions that do not conform to the mass transfer-stripping equilibrium, significantly enhancing localized over-thickness, under-thinness, or agglomeration precursors caused by flow dead zones in the feature map. This not only improves anomaly visibility but also provides more discriminative features for the segmentation network, reducing the dependence on the number of training samples.

[0070] In existing technologies, abnormal package areas are typically identified using only conventional features, which easily leads to missed detection when the anomaly is in its early stages or has low contrast with normal areas. This invention utilizes a growth mechanism model to generate prior predictions, reflecting abnormal behavior that violates physical laws through residuals, and adaptively enhancing feature responses. This is equivalent to providing the segmentation network with an "anomaly heatmap"-like attention guidance. Compared to schemes that solely rely on data-driven features, this invention can generate sufficiently strong feature signals when the anomaly is not yet obvious, significantly improving the early detection rate and localization accuracy of package anomalies—something difficult to achieve with traditional methods.

[0071] For example, suppose that insufficient stirring leads to localized concentration enrichment in a certain area at the bottom of a reactor, causing the coating layer to grow at a rate far exceeding the thickness predicted by the Noyes-Whitney equation based on the average concentration. The measured inversion thickness has already increased significantly, with the residual exceeding a preset threshold τ. This area, originally weak in brightness and difficult to distinguish from neighboring normal areas in the dynamic fusion feature, is enhanced by multiplying its feature value by a (1+η)·polynomial factor, resulting in a several-fold increase in intensity and the formation of a bright, high-brightness patch. When the corresponding anomalously enhanced thickness feature map is fed into the segmentation network, even very slight non-uniform coatings are accurately segmented and marked as anomalous.

[0072] Step S40: Based on the abnormally enhanced thickness feature map, perform the package status recognition task using an encoder-decoder segmentation method, and output the package status recognition result;

[0073] This step inputs the anomalous thickness feature map into the classic U-Net encoder-decoder network. The encoder progressively extracts features at different scales through convolution and downsampling operations, simultaneously capturing low-level wrapper boundaries, crystal contour details, and high-level agglomerate texture distribution information to form a multi-scale wrapper representation. The decoder receives the multi-scale feature map from the encoder and performs progressive upsampling through transposed convolution and skip connections to restore the original resolution. During training and inference, the decoder introduces a preset thickness consistency constraint, requiring smooth and continuous changes in wrapper thickness between adjacent pixels. It also combines cross-entropy loss and Dice coefficient loss to jointly optimize the segmentation of the wrapper region, crystal region, and anomalous agglomerate region. The final output is a wrapper thickness segmentation map (containing the thickness value of each pixel), a crystal distribution map (showing the location and morphology of crystal particles), and a binary mask of anomalous agglomerates (marking agglomerate regions). These results will serve as quantitative inputs for the fuzzy PID control in step S50.

[0074] The design of the multi-scale encoding and decoding structure and the joint constraint loss enables the network to simultaneously consider detailed contours and region consistency. The thickness consistency constraint effectively suppresses thickness abrupt artifacts in the segmentation image, resulting in natural transitions and smooth boundaries within the wrapping layer region. The combination of cross-entropy and Dice loss enhances the network's ability to segment small targets (such as tiny clusters) and imbalanced categories. The resulting wrapping state recognition results are spatially accurate and metrically reliable, directly providing fine-grained and unambiguous quantitative indicators for process control.

[0075] Traditional image segmentation methods often suffer from discontinuous boundaries or isolated noise when processing images encased in crystalline solutions due to low contrast and irregular target shapes, and they struggle to infuse continuous variation information in the thickness dimension. This invention introduces a thickness consistency constraint, unifying pixel-by-pixel thickness regression and the segmentation task within the same network. This ensures that the segmentation result not only includes region labels but also implies the physical thickness distribution. This improvement makes the conversion from the segmentation image to control variables more accurate, avoiding the error accumulation caused by the need to build a separate thickness regression model after traditional segmentation. Consequently, the end-to-end performance of the overall system is significantly improved.

[0076] For example, in a test image covering a large area of ​​encapsulated particles, the encapsulation layer thickness on some particles gradually thins from the center to the edge, while an anomalous agglomeration formed by the adhesion of fine crystals exists in the upper left corner. During encoder extraction, shallow features capture the particle outline and the texture of the thin-layer gradient, while deep features capture the chaotic arrangement pattern within the agglomeration. The decoder, combined with thickness consistency constraints, reconstructs an encapsulation layer thickness segmentation map that exhibits a smooth thickness gradient, an accurate crystal distribution map that delineates each individual particle, and a binary mask of the anomalous agglomeration that completely covers the adhered area. The control module can thus obtain precise current average encapsulation layer thickness of 1.2 μm and agglomeration area percentage of 8.5%, providing solid data for further adjustments.

[0077] Step S50: Based on the package status recognition result, the crystallization process control task is executed using a fuzzy PID collaborative adjustment method, and a collaborative control command is output.

[0078] Based on the encapsulation status recognition results, the deviation between the average encapsulation thickness and the process-set thickness value is first calculated from the encapsulation layer thickness segmentation map. The percentage of abnormal agglomerate area in the effective image region is then calculated from the binary mask of abnormal agglomerates. These two factors constitute the control input. This input is fed into a fuzzy PID controller. Based on a preset fuzzy rule table, the controller infers the adjustment amounts of the proportional parameter Kp, integral parameter Ki, and derivative parameter Kd according to the magnitude and rate of change of the sum. It then calculates candidate control quantities in real time—suggested stirring rate and solution concentration supplement values. Subsequently, considering the minimum / maximum permissible speed of the reactor agitator, the permissible upper and lower limits of the solution concentration, and the maximum safe adjustment range of the actuator per unit time, the candidate control quantities are limited. Finally, stirring rate adjustment commands and solution concentration fine-tuning commands are generated, directly driving the frequency converter and metering pump to achieve coordinated closed-loop control of the crystallization process.

[0079] Fuzzy PID coordinated control essentially integrates expert knowledge with adaptive control, utilizing key deviation variables of the encapsulation state to optimize controller parameters online, enabling coordinated action between the stirring and concentration loops. Adjusting the stirring rate affects flow field shear and mass transfer boundary layer thickness, while fine-tuning the solution concentration directly alters the crystallization driving force. The combination of these two methods effectively suppresses uneven encapsulation and abnormal aggregation. A limiting mechanism ensures that control actions do not exceed the physical limits of the equipment or cause secondary fluctuations, guaranteeing process safety and actuator lifespan.

[0080] Traditional PID control parameters remain fixed once tuned, making it difficult to adapt to the slow drift of encapsulation dynamics or sudden changes in operating conditions during crystallization. When abnormal agglomeration occurs suddenly, the controller often responds with a delay, causing batch product quality fluctuations. This invention feeds back the real-time encapsulation status identification results to a fuzzy inference engine, dynamically coordinating and adjusting the proportional, integral, and derivative actions to jointly drive a dual-loop system. It possesses both the steady-state accuracy of PID and the adaptability of fuzzy control to nonlinearity and uncertainty. Amplitude limiting further prevents overshoot, and the synergistic effect allows the encapsulation layer thickness to quickly return to the target and the area of ​​abnormal agglomerates to gradually converge. Compared with conventional single-loop fixed-parameter control, this invention has significant advantages in convergence speed, steady-state error, and robustness.

[0081] For example, during continuous production, the identification results showed that the average thickness of the coating layer was 15% lower than the set value, and the area of ​​agglomerates had reached the warning level of 12%. The fuzzy PID controller determined that it was necessary to simultaneously increase stirring to thin the diffusion layer and promote growth, and appropriately reduce the supersaturation of the solution to suppress abnormal nucleation. The resulting candidate control variables required increasing the stirring speed from 200 rpm to 260 rpm and reducing the concentration replenishment rate from 0.5 mL / min to 0.2 mL / min. The 60 rpm increase in stirring speed was close to the maximum allowable step size of the equipment. After processing by the limiting module, the actual output command was to increase the stirring speed to 250 rpm (within the safety limit) and fine-tune the concentration to 0.3 mL / min, and continue to adjust in the next control cycle. Finally, the coating thickness approached the set value within 5 minutes, and the agglomerate percentage also dropped below 3%.

[0082] Example 2: Figure 2 The diagram shown is a flowchart illustrating a second embodiment of the intelligent identification and collaborative control system for formate crystallization solution encapsulation according to the present invention. The intelligent identification and collaborative control system for formate crystallization solution encapsulation provided by the present invention employs a method for intelligent identification and collaborative control of formate crystallization solution encapsulation in the above embodiments, and can solve the technical problem of intelligent identification and collaborative control of formate crystallization solution encapsulation. The beneficial effects of the intelligent identification and collaborative control system for formate crystallization solution encapsulation provided by the present invention are the same as those of the method for intelligent identification and collaborative control of formate crystallization solution encapsulation provided in the above embodiments, and other technical features in the intelligent identification and collaborative control system for formate crystallization solution encapsulation are the same as those disclosed in the method of the above embodiments, and will not be repeated here.

[0083] Example 3: This invention provides a formate crystal solution encapsulation intelligent identification and collaborative control device. The device includes: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which, when executed, enable the at least one processor to perform the formate crystal solution encapsulation intelligent identification and collaborative control method described in Example 1. The formate crystal solution encapsulation intelligent identification and collaborative control device in this invention can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This formate crystal solution encapsulation intelligent identification and collaborative control device is merely an example and should not limit the functionality or scope of the invention. A formate crystal solution-encapsulated intelligent identification and collaborative control device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory or a program loaded from a storage device into a random access memory. The random access memory also stores various programs and data required for the operation of the formate crystal solution-encapsulated intelligent identification and collaborative control device. The processing unit, the read-only memory, and the random access memory are interconnected via a bus. An I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including, for example, magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the formate crystal solution-encapsulated intelligent identification and collaborative control device to communicate wirelessly or wiredly with other devices to exchange data. Although a formate crystal solution-encapsulated intelligent identification and collaborative control device with various systems has been described, it should be understood that implementation or possession of all described systems is not required. It can be implemented alternatively or with more or fewer systems.

[0084] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described intelligent identification and collaborative control method for formate crystal solution encapsulation. The computer program product provided by this invention can solve the technical problem of intelligent identification and collaborative control of formate crystal solution encapsulation. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the intelligent identification and collaborative control method for formate crystal solution encapsulation provided in the above embodiments, and will not be repeated here.

[0085] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0086] Example 5: Figure 3 The diagram shown is a flowchart of the cross-modal attention dynamic weighted fusion process of the fifth embodiment of the intelligent identification and collaborative control method for formate crystallization solution encapsulation of the present invention. It illustrates the process of dynamically calculating and weighting the fusion based on multi-dimensional evaluation indicators such as base score, confidence level, interference amount, and flow field disturbance after inputting three types of features (visual, ultrasonic, and conductivity) into the cross-modal attention module, and finally generating dynamic fusion features. This process achieves robust integration of multi-source heterogeneous information and provides high-quality input for subsequent crystallization encapsulation state identification.

[0087] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0088] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A method for intelligent identification and collaborative control of formate crystal solution encapsulation, characterized in that, The methods include: Step S10: Obtain multimodal raw data of the formate crystallization solution encapsulation process, and perform data preprocessing tasks based on the multimodal raw data using source-specific noise reduction and time synchronization to output synchronized denoised multimodal data; Step S20: Based on the synchronously denoised multimodal data, a cross-modal attention weighting method is used to perform a feature fusion task and output dynamic fused features; Step S30: Based on the dynamic fusion features, the flow field crystallization coupling prediction method is used to perform the encapsulation anomaly enhancement task, and the anomaly enhancement thickness feature map is output; Step S40: Based on the abnormally enhanced thickness feature map, perform the package status recognition task using an encoder-decoder segmentation method, and output the package status recognition result; Step S50: Based on the package status recognition result, the crystallization process control task is executed using a fuzzy PID collaborative adjustment method, and a collaborative control command is output.

2. The method for intelligent identification and collaborative control of formate crystallization solution encapsulation as described in claim 1, characterized in that, In step S10, the process of acquiring multimodal raw data of the formate crystallization solution encapsulation process, and performing data preprocessing tasks based on the multimodal raw data using source-specific noise reduction and time synchronization to output synchronized and denoised multimodal data, specifically includes: Step S101: Simultaneously acquire high-speed visual images, ultrasonic reflection signals and solution conductivity data during the encapsulation process of formate crystallization solution, and record the acquisition timestamps corresponding to each sampled data to obtain time-stamped source sampling data; Step S102: Perform source-specific noise reduction processing on the timestamped source-specific sampling data, wherein the high-speed visual image is subjected to adaptive median filtering to remove bubble occlusion noise, the ultrasonic reflection signal is subjected to wavelet threshold denoising to retain the echo characteristics of the crystal interface, and the solution conductivity data is subjected to Kalman filtering for smoothing to obtain source-specific noise reduction data; Step S103: Based on the acquisition timestamp, perform time alignment and interpolation on the source-specific denoised data to obtain synchronous denoised multimodal data.

3. The method for intelligent identification and collaborative control of formate crystallization solution encapsulation as described in claim 1, characterized in that, Step S20, which involves performing a feature fusion task based on the synchronously denoised multimodal data using a cross-modal attention weighting method and outputting dynamically fused features, specifically includes: Step S201: Input the high-speed visual image, ultrasonic reflection signal and solution conductivity data in the synchronous denoised multimodal data into the visual branch convolutional network, the ultrasonic branch one-dimensional convolutional network and the conductivity branch fully connected network respectively to extract visual features, ultrasonic features and conductivity features; Step S202: Based on the visual features, the ultrasonic features, and the electrical conductivity features, a cross-modal attention module is used to calculate the first modal weight, the second modal weight, and the third modal weight, wherein the first modal weight corresponds to the visual features, the second modal weight corresponds to the ultrasonic features, and the third modal weight corresponds to the electrical conductivity features; Step S203: The visual features, ultrasonic features, and electrical conductivity features are weighted and fused according to the first modal weight, the second modal weight, and the third modal weight to obtain dynamic fused features.

4. The method for intelligent identification and collaborative control of formate crystallization solution encapsulation as described in claim 3, characterized in that, In step S202, the first modal weight, the second modal weight, and the third modal weight are determined according to the following formula: ; in, Modal number, This represents the first mode corresponding to a high-speed visual image. This represents the second mode corresponding to the ultrasonic reflection signal. This indicates the third mode corresponding to the solution conductivity data; For the first Modal weights; For the first Modal attention baseline score; For the first Signal confidence level of the modality; For the first The amount of interference penalty for the modality; The flow field disturbance is obtained by normalizing the rate of change of flow velocity, the rate of change of shear stress, and the bubble obstruction rate. For the first Confidence adjustment coefficient for the modality; For the first The interference modulation coefficient; For the first Modal flow field disturbance adjustment coefficient.

5. The method for intelligent identification and collaborative control of formate crystallization solution encapsulation as described in claim 1, characterized in that, Step S30, which involves performing the encapsulation anomaly enhancement task based on the dynamic fusion features using a flow field crystallization coupling prediction method and outputting anomaly enhancement thickness feature maps, specifically includes: Step S301: Extract particle boundary response, echo phase response and conductance gradient response based on the dynamic fusion features, and concatenate the particle boundary response, echo phase response and conductance gradient response to obtain coupled input features; Step S302: Based on the coupled input features, the Noyes Whitney mass transfer growth equation and shear stripping correction term are used to predict the coating thickness to obtain the predicted thickness field; Step S303: Based on the residual between the predicted thickness field and the measured thickness field obtained by inversion from the dynamic fusion feature, perform anomaly enhancement on the dynamic fusion feature to obtain an anomaly enhanced thickness feature map; wherein, the anomaly enhanced thickness feature map is determined according to the following formula: ; in, pixel position The abnormally enhanced thickness characteristic value at the location; pixel position Dynamic fusion feature values ​​at the location; pixel position The predicted thickness value at that location; pixel position The measured thickness value at the location; The thickness residual threshold; This is the abnormal enhancement coefficient; To prevent the stability constant from having a denominator of zero; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

6. The method for intelligent identification and collaborative control of formate crystallization solution encapsulation as described in claim 1, characterized in that, Step S40, which involves performing package status recognition based on the abnormally enhanced thickness feature map using an encoder-decoder segmentation method and outputting the package status recognition result, specifically includes: Step S401: Input the abnormally enhanced thickness feature map into the U-Net encoder-decoder network, and use the encoder to extract the wrapping layer boundary features, crystal contour features and aggregate texture features to obtain a multi-scale wrapping characterization; Step S402: The multi-scale encapsulation representation is input to the decoder and upsampled step by step. Based on the preset thickness consistency constraint, cross-entropy loss and Dice coefficient loss, the encapsulation layer region, crystal region and abnormal agglomeration region are jointly segmented to obtain the joint segmentation result. Step S403: Generate package status identification results based on the joint segmentation results. The package status identification results include a package layer thickness segmentation map, a crystal distribution map, and a binary mask of abnormal aggregates.

7. The method for intelligent identification and collaborative control of formate crystallization solution encapsulation as described in claim 1, characterized in that, Step S50, which involves executing the crystallization process control task using a fuzzy PID collaborative adjustment method based on the package status identification result and outputting collaborative control commands, specifically includes: Step S501: Calculate the thickness deviation between the current average thickness of the package layer and the set thickness value based on the package layer thickness segmentation map in the package status recognition result, and calculate the area ratio of abnormal aggregates based on the binary mask of abnormal aggregates in the package status recognition result to obtain the control input quantity; Step S502: Input the control input quantity into the fuzzy PID controller, and adjust the proportional parameter, integral parameter and derivative parameter in real time according to the thickness deviation and the area ratio of the abnormal agglomerates to obtain the candidate control quantity; Step S503: Based on the preset allowable speed range of the stirring equipment, the allowable range of solution concentration, and the maximum adjustment range per unit time, the candidate control quantity is subjected to amplitude limiting processing to obtain a coordinated control command. The coordinated control command includes a stirring rate adjustment command and a solution concentration fine-tuning command.

8. A formate crystallization solution encapsulation intelligent recognition and collaborative control system, applied to the formate crystallization solution encapsulation intelligent recognition and collaborative control method according to any one of claims 1 to 7, characterized in that, The intelligent identification and collaborative control system for formate crystallization solution encapsulation includes: The data preprocessing module is used to acquire multimodal raw data of the formate crystallization solution encapsulation process, and to perform data preprocessing tasks based on the multimodal raw data using source-specific noise reduction and time synchronization methods, outputting synchronized denoised multimodal data; The feature fusion module is used to perform a feature fusion task based on the synchronously denoised multimodal data using a cross-modal attention weighting method, and output dynamic fused features; The coupling enhancement module is used to perform the encapsulation anomaly enhancement task based on the dynamic fusion features using the flow field crystallization coupling prediction method, and outputs the anomaly enhancement thickness feature map. The real-time recognition module is used to perform package status recognition tasks based on the abnormal enhanced thickness feature map using an encoder-decoder segmentation method, and output the package status recognition results. The collaborative control module is used to execute the crystallization process control task using a fuzzy PID collaborative adjustment method based on the package status identification result, and output collaborative control commands.

9. A device for intelligent identification and collaborative control of formate crystallization solution encapsulation, characterized in that, The formate crystal solution encapsulation intelligent identification and collaborative control device includes: a memory, a processor, and a formate crystal solution encapsulation intelligent identification and collaborative control program stored in the memory and executable on the processor. When the formate crystal solution encapsulation intelligent identification and collaborative control program is executed by the processor, it implements the formate crystal solution encapsulation intelligent identification and collaborative control method according to any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a formate crystal solution encapsulation intelligent identification and collaborative control program, which, when executed by a processor, implements a formate crystal solution encapsulation intelligent identification and collaborative control method according to any one of claims 1 to 7.

Images