An automated water suspension granulation method and system based on stirred flow field

Abstract

This invention discloses an automated water suspension granulation method and system based on stirred flow field, belonging to the field of chemical granulation technology. This method uses multi-source sensing technology to acquire the dynamic flow field morphology of materials in a stirred tank in real time and non-contactly, and combines it with process parameters such as online viscosity and stirring torque to construct a multi-dimensional feature dataset. Based on this dataset, a process control model is trained and continuously optimized online using machine learning algorithms. This model can accurately map the complex nonlinear relationship between the real-time flow field and material state and the optimal process parameters. In the actual granulation process, the system dynamically and collaboratively adjusts various process parameters according to the real-time collected multi-source data, performing multi-parameter coordinated adjustment, resulting in faster and more accurate response, and achieving closed-loop intelligent control driven by data and model. This invention significantly improves the uniformity, stability, and automation level of the granulation process.

Description

Technical Field

[0001] This invention relates to the fields of chemical process control and powder technology. Specifically, it relates to an automated control method and system for water suspension granulation process, and more particularly to a technical solution that achieves intelligent dynamic control of process parameters by analyzing the state of the stirring flow field in real time based on multi-source sensor data fusion and machine learning model. Background Technology

[0002] Aqueous suspension granulation is a wet granulation process that uses agitation to agitate fine powder into spherical particles by adding a liquid binder (and possibly water) dropwise to an aqueous system. It is widely used in the preparation of granular products in industries such as energetic materials, pharmaceuticals, food, catalysts, and ceramics. Key control parameters include stirring speed, binder addition rate, and water addition. These parameters directly affect the flow field within the reactor (e.g., shear force, mixing uniformity, vortex morphology) and the nucleation, aggregation, and breakup processes of the particles, ultimately determining key quality attributes such as particle size distribution, bulk density, and strength of the final product.

[0003] Currently, water suspension granulation processes in industrial production heavily rely on operator experience. They typically employ fixed or segmented process parameter curves, lacking the ability to perceive and respond to real-time, dynamically changing material states within the reactor. This open-loop or semi-open-loop control method has significant drawbacks: 1. Lagging and crude control: It is impossible to perceive changes in the flow field caused by fluctuations in material properties and uneven local concentration of binder in real time. When quality deviations are found through offline sampling, it is too late to make adjustments, resulting in poor uniformity within and between batches.

[0004] 2. Limited state perception: Although existing technologies use torque, power, or single viscosity sensors for monitoring, they can only reflect macroscopic mechanical load or average viscosity and cannot obtain spatial morphological information inside the flow field (such as vortex size and material level fluctuations). This morphological information is closely related to the collision frequency of particles and the growth environment.

[0005] 3. Process optimization relies on trial and error: For new materials or new formulas, a large amount of time-consuming trial production is required to explore process parameters, resulting in high trial and error costs and difficulties in process transfer and scale-up.

[0006] Although some studies have attempted to introduce visual monitoring (such as JP2020156782A which mentions observing the flow field), they usually remain at the level of qualitative observation or simple parameter extraction. They have failed to establish a precise and adaptive dynamic control model that integrates the quantitative characteristics of the flow field morphology with multiple process parameters. Furthermore, they lack the ability to integrate with multi-dimensional data such as online viscosity and torque, as well as the self-learning capability of the model based on data-driven principles.

[0007] Therefore, developing a water suspension granulation system and method that can deeply integrate multi-source real-time sensing data, analyze the intrinsic relationship between flow field state and process parameters through intelligent models, and achieve closed-loop adaptive optimization control is of urgent need and great significance for improving product quality consistency, production efficiency and process intelligence. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an automated water suspension granulation method and system based on a stirred flow field. This method aims to comprehensively quantify the dynamic state of the granulation process through non-contact image sensing combined with online property measurement; and to utilize machine learning to construct a data-driven intelligent control model to achieve real-time, collaborative, and adaptive optimization of process parameters, ultimately improving granulation uniformity, stability, and the ability for autonomous process optimization.

[0009] The technical solution of this patent is as follows: An automated water suspension granulation method based on a stirred flow field, comprising the following steps: Step 1. Real-time acquisition of multi-source data: During the granulation process, dynamic flow field image data, online viscosity data of materials, and torque data of the stirring shaft are acquired simultaneously in the mixing tank; Step 2. Feature extraction and data fusion: The dynamic flow field image is processed to extract morphological feature parameters, including at least the material level height H, the diameter D and depth d of the vortex at the stirring center; the extracted morphological feature parameters, online viscosity data and stirring torque data are timestamped and normalized, and fused to form a comprehensive feature vector representing the current material state; Step 3. Model-based intelligent decision-making: The comprehensive feature vector is input into a pre-trained process control model, which outputs an optimized adjustment of process parameters for the current state. The process parameters include at least the stirring speed, the binder dripping speed, and the amount of water added. Step 4. Closed-loop execution and feedback control: Adjust the corresponding actuators in real time according to the adjustment amount output by the process control model; and continue to repeat steps 1 to 4 to form closed-loop control until the granulation process ends; Step 5. Online iterative optimization of the model: After the granulation operation of a single batch is completed, the final product quality index data of this batch is correlated with the production process data, and the process control model is incrementally learned to update the model parameters.

[0010] Further, in step 1, the dynamic flow field image is acquired by a high-speed industrial camera set at a transparent window on the top or side wall of the stirred tank, with a sampling frequency of not less than 60 frames / second; the online viscosity data is measured by an online viscometer installed on the bottom or side wall of the stirred tank; and the stirring torque data is measured by a torque sensor installed on the stirring motor or stirring shaft.

[0011] Furthermore, in step 2, the feature extraction of the dynamic flow field image includes: using the Canny edge detection algorithm to identify the material liquid surface contour, and calculating its average height H and volatility; using a method based on morphological opening and closing operations and contour extraction to identify the central vortex region, and then calculating its equivalent diameter D, maximum depth d and contour symmetry index.

[0012] Furthermore, the process control model is trained using a deep neural network or a gradient boosting decision tree algorithm; the training objective function of the model is to minimize the difference between the particle size distribution dispersion of the particles obtained under the predicted process parameters and the target distribution; in step 5, the model is incrementally updated using online learning or mini-batch retraining.

[0013] Furthermore, in step 3, the adjustment logic output by the process control model includes: When the integrated feature vector indicates that the vortex diameter exceeds the first threshold and the stirring torque is lower than the second threshold, the model outputs a command to simultaneously reduce the stirring speed and increase the binder dripping speed. When the comprehensive feature vector indicates that the material level fluctuation rate exceeds the third threshold and the online viscosity changes, the model outputs instructions to adjust the water addition and fine-tune the stirring speed to stabilize the flow field. When the integrated eigenvector indicates a sharp increase in online viscosity and a simultaneous significant increase in torque, the model outputs a command to temporarily increase the stirring speed and reduce the binder dripping speed to prevent excessive agglomeration.

[0014] Furthermore, it also includes abnormal operating condition handling steps: when image acquisition failure or abnormal online viscosity data and stirring shaft torque data of the material are detected, the system automatically switches to a rule base based on historical similar operating condition data to perform emergency adjustment of process parameters and issues an alarm.

[0015] Furthermore, in step 3, the process control model is a hybrid model based on the fusion of physical empirical rules and a data-driven model, wherein the physical empirical rule part adopts the following control formula: 1. Formula for adjusting stirring speed based on vortex size: ; Where Δω is the stirring speed adjustment, Kp1 and Kd1 are the proportional and derivative coefficients, respectively, and D is the real-time vortex diameter. set To define the vortex diameter, dD / dt is the rate of change of the vortex diameter; 2. Adhesive flow rate adjustment formula based on viscosity and torque: ; Where, ΔQ bKp2 and Ki2 are the proportional and integral coefficients, respectively, and τ is the real-time stirring torque. set To set the torque, η is the real-time viscosity. set To set the viscosity; 3. Formula for adjusting water dosage based on material level fluctuations: ; Where, ΔQ w For the water addition adjustment, Kp3 and Kd3 are the proportional and differential coefficients, respectively, and σ is the coefficient of variation. H σ represents the standard deviation of material level fluctuation. Hset To set the standard deviation of the fluctuation, dσ H / dt represents the rate of change of the standard deviation of the fluctuation.

[0016] Furthermore, the fusion method of the hybrid model is as follows: when the amount of historical data is insufficient, the output of physical empirical rules is the primary method; as data accumulates, the weight of the data-driven model is gradually increased; finally, a weighted fusion method is adopted. ; Wherein, ΔP final The final process parameter adjustment amount, ΔP rule The output is the physical empirical rule, ΔP. model The output of the data-driven model is α, which is a dynamic weight coefficient that gradually decreases from 1 to 0.2 as the model confidence increases.

[0017] Furthermore, an automated aqueous suspension granulation system for realizing an automated aqueous suspension granulation method includes: The multi-source sensing module includes: a high-speed image acquisition unit for acquiring images of the flow field inside the stirred tank, an online viscosity sensing unit for measuring the viscosity of the material, and a torque sensing unit for measuring the stirring resistance. The data fusion and processing module, connected to the multi-source sensing module, is used to preprocess, extract and fuse the collected raw data to generate a comprehensive feature vector. The intelligent process decision-making module has an embedded process control model, which receives the comprehensive feature vector and outputs optimized process parameter adjustment instructions. The execution control module is connected to the process intelligent decision-making module and precisely controls the actions of the stirring motor driver, binder metering pump and water addition regulating valve according to the adjustment instructions. The model management and iteration module is used to store production data and product quality data, and provides the function of retraining and updating the process control model online.

[0018] Furthermore, the high-speed image acquisition unit includes a high-speed industrial camera, a ring LED fill light, and a dustproof and splashproof viewing window cover for protecting the high-speed industrial camera; the online viscosity sensing unit is a vibration or rotational online viscometer, whose sensor probe extends into the slurry inside the vessel; the torque sensing unit includes a torque sensor mounted on a stirring motor or stirring shaft.

[0019] Furthermore, the process intelligent decision-making module and the model management and iteration module are integrated into an industrial computer or an embedded edge computing device; the execution control module adopts a multi-loop PID controller or a more advanced model predictive controller, with a response time of less than 0.2 seconds.

[0020] Furthermore, it also includes a human-machine interaction and data interface module for parameter setting, status monitoring, report generation, and supports data exchange with the upper-level manufacturing execution system via OPC UA or Modbus protocol to achieve full lifecycle management of production batches.

[0021] Technical advantages: Compared with the prior art, the present invention has the following significant advantages: For the first time in the granulation process, the system achieves collaborative online perception of the microscopic granulation environment and macroscopic mechanical state of materials through the fusion of multi-source information such as images, viscosity, and torque. This provides rich information dimensions, and the image perception method does not interfere with the process. For new materials, the system can quickly train a basic model using limited experimental data in the early stages and continuously optimize itself in subsequent production, thereby greatly reducing the number of experiments and time required by traditional methods and accelerating the production of new products.

[0022] Based on data-driven machine learning models, this system can uncover and utilize the complex process patterns hidden within massive amounts of historical data to achieve optimized control that surpasses human experience and traditional PID control. The model can perform multi-parameter coordinated adjustments based on real-time conditions, resulting in faster and more precise responses. Closed-loop dynamic control ensures that the process always operates on an optimal or near-optimal trajectory, effectively suppressing various disturbances and significantly reducing the dispersion of key indicators such as particle size distribution and density within and between batches. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall architecture of the automated water suspension granulation system provided in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the overall process and closed-loop control of the method of the present invention.

[0025] Figure 3 A flowchart of the algorithm for extracting flow field morphology features for the image processing module.

[0026] In the diagram: 1-Agitator; 2-Agitator paddle; 3-Agitator motor; 4-High-speed industrial camera; 5-Ring LED fill light; 6-Online viscometer; 7-Binder storage tank; 8-Water addition pipeline and electric proportional control valve; 9-Core controller; 10-Human machine interface; 11-Data server. Detailed Implementation

[0027] Example 1 This embodiment discloses an automated water suspension granulation device and method. The device includes a 50L covered stainless steel stirring tank 1, and a stirring paddle 2 driven by a 7.5kW variable frequency stirring motor 3. A torque sensor is installed on the stirring motor 3 or the stirring shaft, which can output torque value and measure the stirring torque in real time (range 0-200 N·m).

[0028] An opening with a pressure-resistant glass window is made in the center of the lid of the stirred tank 1. A high-speed industrial camera 4 (resolution 2048×1536, frame rate 100 fps) is installed directly above it. A ring of LED supplementary lights 5 is integrated around the high-speed industrial camera 4 to ensure uniform lighting inside the tank. A dustproof and splashproof viewing window cover is installed on the high-speed industrial camera 4. A vibration-type online viscometer 6 (measuring range 10-10000 mPa·s) is installed on the lower half of the side wall of the stirred tank 1, with its sensor probe extending into the slurry inside the stirred tank 1.

[0029] The binder is pumped from the binder storage tank 7 by a high-precision plunger metering pump. The pump flow rate can be precisely controlled via pulse frequency (range 0-500 mL / min, control accuracy ±0.5%). Process water is added via a water addition line and an electric proportional control valve 8. All actuators (stirring motor 3, plunger metering pump, electric proportional control valve, etc.) are coordinated and controlled by the core controller 9.

[0030] An automated water suspension granulation method based on a stirred flow field includes the following steps: Step 1. Real-time acquisition of multi-source data: During the granulation process, the viscosity sensor data η is measured and read synchronously through the online viscometer 6 at a frequency of 100Hz, the torque τ is measured and read through the torque sensor, and the high-speed industrial camera 4 is triggered to take pictures, synchronously acquiring dynamic flow field image data in the stirring tank 1, with a sampling frequency of not less than 60 frames / second; and each frame image is accurately timestamped.

[0031] Step 2. Feature Extraction and Data Fusion: Process the dynamic flow field image and run the algorithm on the acquired image in real time (see flowchart). Figure 3 First, Gaussian filtering is performed to remove noise. Then, the Canny operator is used to detect the boundary between the material and the air (i.e., the material surface profile). The liquid surface line is fitted by Hough transform, and the average height H and the standard deviation of fluctuation σ are calculated.H The extracted morphological feature parameters, online viscosity data, and stirring torque data are timestamped and normalized, and then fused to form a comprehensive feature vector representing the current state of the material. Subsequently, the lower region of the image is binarized and subjected to morphological opening and closing operations and contour extraction. The central concave vortex region is extracted, its minimum circumscribed circle diameter is calculated as the equivalent diameter D, and the gray-level gradient of the profile is analyzed to calculate the maximum depth d.

[0032] After identifying the contour of the vortex region, in order to further quantify the depth information d of the vortex, this invention adopts a method based on gray-scale gradient analysis or structured light three-dimensional reconstruction.

[0033] Method 1 (Gray-level Gradient Method): This method utilizes the specific light reflection characteristics of the ring-shaped LED supplementary light 5 on the liquid surface. The bottom of the vortex, due to its greatest curvature change, typically forms a highlight area; the sidewalls of the vortex form gray-level gradient areas. By analyzing the gray-level gradient changes of pixels within the vortex region and the morphology of the highlight areas, a mapping model between gray-level gradient and physical depth is established. Specifically, the gray-level profile line from the center to the edge within the vortex region is calculated, and using a pre-calibrated "gray-level-depth" curve (obtained through calibration at different known depths), the gray-level gradient change rate is converted into the maximum depth d of the vortex.

[0034] Method 2 (Stereo Vision Method): If a binocular high-speed camera or structured light sensor is installed on the top of the stirred tank 1, active stereo vision technology can be used. By projecting coded structured light onto the liquid surface, the local three-dimensional point cloud of the vortex region is directly reconstructed based on the deformation of the light stripe using the principle of triangulation, thereby accurately calculating the maximum depth d and volume of the vortex. This embodiment preferentially uses the gray-scale gradient method to balance real-time performance and cost.

[0035] Step 3. Model-based intelligent decision-making: The comprehensive feature vector is input into a pre-trained process control model, which outputs an optimized adjustment of process parameters for the current state. The process parameters include at least the stirring speed, the binder dripping speed, and the water addition amount.

[0036] The adjustment logic output by the process control model includes: When the integrated feature vector indicates that the vortex diameter exceeds the first threshold and the stirring torque is lower than the second threshold, the model outputs a command to simultaneously reduce the stirring speed and increase the binder dripping speed. When the comprehensive feature vector indicates that the material level fluctuation rate exceeds the third threshold and the online viscosity changes, the model outputs instructions to adjust the water addition and fine-tune the stirring speed to stabilize the flow field. When the integrated eigenvector indicates a sharp increase in online viscosity and a simultaneous significant increase in torque, the model outputs a command to temporarily increase the stirring speed and reduce the binder dripping speed to prevent excessive agglomeration.

[0037] The current-time features H, σH, D, and d are Z-score normalized with the synchronized viscosity η and torque τ to form a six-dimensional feature vector. This vector is then input into a loaded deep neural network (DNN) model. This DNN model has two hidden layers (32 and 16 nodes respectively) and a three-node output layer, corresponding to the stirring speed adjustment Δω (rpm) and the binder flow rate adjustment ΔQ, respectively. b (mL / min), water flow rate adjustment ΔQ w (mL / min).

[0038] Step 4. Closed-loop execution and feedback control: The adjustment value output by the model is superimposed with the current setpoint to generate a new target setpoint. The control signal is calculated using the built-in PID control algorithm and sent to the drivers of the frequency converter (stirring motor 3), plunger metering pump, and electric proportional control valve, respectively. The corresponding actuators are adjusted in real time according to the adjustment value output by the process control model; and steps 1 to 4 are repeated continuously to form a closed-loop control until the granulation process ends.

[0039] Step 5. Online iterative optimization of the model: After the granulation operation of a single batch is completed, the final product quality index data of this batch is correlated with the production process data, and the process control model is incrementally learned to update the model parameters.

[0040] The process control model is trained using a deep neural network or gradient boosting decision tree algorithm; the training objective function of the model is to minimize the difference between the particle size distribution dispersion of the particles obtained under the predicted process parameters and the target distribution; in step 5, the model is incrementally updated using online learning or mini-batch retraining.

[0041] All raw data, feature data, control commands, and operational events are stored in a local database in time-series format. The core controller 9 is connected to the host data server 11 via Ethernet and periodically uploads production data. The data server 11 is responsible for periodically retraining and versioning the DNN model using the accumulated data.

[0042] The human-machine interface 10 enables two-way information exchange, enhances system security and reliability, and facilitates remote access and intelligent management. Specifically, users send commands to the system (such as setting the vortex diameter or torque) through input devices (e.g., keyboard, touchscreen, voice), and the system provides feedback on status or results through visual and auditory means, building an efficient communication mechanism. It can monitor equipment status in real time, issue alarms, assist in fault diagnosis, and reduce human error. It supports remote control and data acquisition via network, and combined with AI and big data, enables predictive maintenance, personalized services, and automated decision-making.

[0043] Regarding the labeling mechanism for model training data: In this invention, the labels for the training data of the process control model are not preset by humans, but are generated automatically or semi-automatically through the following two modes: Mode 1 (Reverse Labeling Based on Target Quality): During the offline training phase, multi-source sensor data (feature vector sequences) collected throughout the historical production process are correlated with the final product quality indicators (such as particle size distribution D50, D90, and coefficient of variation) for that batch. Using reinforcement learning or inverse reinforcement learning, the process parameter sequences corresponding to batches that "produce high-quality particles (e.g., particle size distribution dispersion below a set threshold)" are labeled as positive samples of the "optimal strategy"; batches that produce substandard particles are labeled as negative samples. Through comparative learning, the model establishes a mapping relationship between feature states and optimal control actions.

[0044] Mode 2 (Online Correction Based on Expert Experience): An "Expert Labeling Mode" is provided in the human-computer interaction interface 10. When the system is in semi-automatic operation or model fine-tuning stage, on-site experts can observe the system's recommended adjustment amount in real time through the interface. If the expert believes the recommendation is unreasonable and intervenes manually (such as manually adjusting the stirring speed), the system will record the state feature vector at this time and the "corrected parameters" finally executed by the expert, and store this data pair as high-confidence label data in the cache for subsequent incremental learning and optimization of the model.

[0045] Example 2 Application of formulas for granulation process optimization and control of energetic complex particles 1. Initial Model Construction and Parameter Setting: Energetic particles were used as the starting powder, and fluoropolymer was used as the binder. First, experienced engineers manually completed 10 successful batches of production, during which the system recorded all sensor data and manually adjusted parameters throughout the process. Using the data from these 10 batches (approximately 500,000 time-series data points) as the training set, an initial DNN model was trained on data server 11. The training objective was to minimize the difference between the process parameter sequence predicted by the model and the actual successful batch parameter sequence, while incorporating the variance of the finished particle size D90 as a regularization term. The trained model was then loaded into the core controller 9.

[0046] At the same time, based on historical data and expert experience, control parameters for physical experience rules are set: Vortex diameter setting value D set = 6.0 cm; Stirring torque setting value τ set = 45 N·m; Material viscosity setpoint η set = 1200 mPa·s; Standard deviation setting value σ for material level fluctuationHset = 1.5 mm; Control coefficients: Kp1 = 5.0 rpm / cm, Kd1 = 2.0 rpm·s / cm; Kp2 = 0.1 mL / (min·N·m), Ki2 = 0.02 mL / (min·mPa·s) 2 ); Kp3 = 10 mL / (min·mm), Kd3 = 3 mL·s / (min·mm).

[0047] 2. Implementation of Hybrid Control Strategy: Start a new batch of production. In the initial stage, due to the low confidence level of the data-driven model in the early stage of the new batch, the dynamic weight coefficient α=0.8 is set, that is, physical empirical rules are the main factor (80% weight), and the data-driven model is the auxiliary factor (20% weight).

[0048] Specific control process: 0-5 minutes (wetting stage): Initial stirring speed set to 300 rpm. The system monitored no vortices (D≈0) and low torque τ (approximately 15 N·m). Physical law calculation: Δω rule = -5.0×(0-6.0) = +30 rpm (increased speed recommended); Model output Δω model = +25 rpm. Δω after weighted fusion final = 0.8×30+0.2×25 = 29 rpm, actually adjust the stirring speed to 329 rpm. At the same time, adjust the amount of binder added according to viscosity and torque.

[0049] 6-20 minutes (particle growth period): Vortex forms and stabilizes at D≈5.8 cm, τ rises to 42 N·m, η≈1150 mPa·s, σ H =1.8 mm. At this point, the model has accumulated a certain amount of real-time data, the confidence level has increased, and α has been adjusted to 0.5.

[0050] Physical rule calculation: Δω rule = -5.0×(5.8-6.0)-2.0×(0.1) = +1.0-0.2=+0.8 rpm; ΔQb rule = 0.1×(42-45)-0.02×∫(1150-1200)dt ≈ -0.3+1.0=+0.7 mL / min; ΔQw rule = 10×(1.8-1.5)+3×(0.05) = +3.0+0.15=+3.15 mL / min.

[0051] Data model output: Δωmodel = +1.2 rpm, ΔQb model = +0.9 mL / min, ΔQw model = +2.8 mL / min.

[0052] The weighted merging process will be adjusted accordingly.

[0053] 25-35 minutes (stabilization period): The system is running stably, the model confidence is very high, and α is adjusted to 0.2. At this stage, fine-tuning of the model mainly relies on data-driven approaches, with physical rules serving as verification and safeguards.

[0054] 3. Effect Comparison: A comparison was made between the traditional fixed parameter method (constant speed of 300 rpm, binder added according to a fixed curve) and 10 batches of production.

[0055] Particle size distribution: Particles produced by conventional methods have an average D50 of 920 μm and an RSD (relative standard deviation) of 12.5%. Particles produced by the method of this invention have an average D50 of 950 μm, an RSD reduced to 4.8%, and a significantly more concentrated particle size distribution.

[0056] Batch consistency: The D50 range of 10 batches using the traditional method is 200 μm, while the D50 range of 10 batches using the method of this invention is 80 μm.

[0057] Example 3 The automated water suspension granulation method also includes online model iteration and abnormal operating condition handling steps: when image acquisition failure or abnormal online viscosity data and stirring shaft torque data of the material are detected, the system automatically switches to the rule base based on historical similar operating condition data to perform emergency adjustment of process parameters and issue an alarm.

[0058] Model Iteration Process: After 20 consecutive production batches, data server 11 used the 20 newly added batches of data to perform incremental learning (fine-tuning) on ​​the initial DNN model. The updated model showed faster adjustment speed and stability when dealing with minor differences in the flowability of raw materials. At the same time, the setpoints and control coefficients in the physical empirical rules were optimized based on the new data, making the rules closer to the actual optimal working conditions.

[0059] Anomaly Handling Demonstration: Simulating a situation where a high-speed industrial camera 4 is temporarily obscured by water mist. The system detects that the image quality score remains below the threshold for more than 10 seconds, triggering the anomaly handling procedure: The first alarm message is "weak visual signal".

[0060] The core controller 9 automatically switches to "degraded control mode". In this mode, the system only uses the changing trends of viscosity η and torque τ, combined with the current production stage (inferred from the running time), to call data patterns of similar materials and similar stages in the historical database to perform rule-based reasoning and PID control on the process parameters.

[0061] Specifically, a simplified control rule based on viscosity and torque is adopted: ; ; Where Kp1', Kd1', Kp2', and Ki2' are the control coefficients under the degradation mode.

[0062] Experiments showed that running in degraded mode for 5 minutes did not significantly worsen the production process, and torque and viscosity remained stable. After the camera was cleaned, the system automatically resumed full-function closed-loop control.

[0063] The hybrid model employs the following fusion approach: when historical data is insufficient, it primarily relies on physical empirical rules; as data accumulates, the weights of the data-driven model are gradually increased; finally, a weighted fusion method is used. Wherein, ΔP final The final process parameter adjustment amount, ΔP rule The output is the physical empirical rule, ΔP. model The output of the data-driven model is α, which is a dynamic weight coefficient that gradually decreases from 1 to 0.2 as the model confidence increases.

[0064] Regarding the quantification mechanism of the dynamic weight coefficient α: In order to achieve a smooth transition and intelligent integration between physical rules and data models, this invention defines the weight coefficient α as a function of the model confidence Cmodel: Specifically, Cmodel∈[0,1], α decreases linearly from 1 to 0.2.

[0065] The model confidence score (Cmodel) is comprehensively evaluated using the following three dimensions: Data Coverage (Cdata): Based on the comprehensive feature vector of the current input, nearest neighbor search is performed in the historical database. If the local sample density of the current condition in the feature space is high, the model's inference confidence in that region is considered high. Cdata is calculated by normalizing the average distance between the current feature point and the k nearest neighbor samples in the training set.

[0066] Forecast consistency (Cconsistency): The consistency of the adjustment amount output by the data-driven model over multiple consecutive control periods (e.g., the past 10 periods). Adjustment amount of physical rule output If the direction is consistent and the fluctuations are small, the model output is considered stable and reliable, and the confidence level is increased. Conversely, if the output oscillates violently or contradicts physical rules, the confidence level is decreased.

[0067] Model Freshness: Records the number of samples processed by the model since its last online update or retraining. The more samples processed, the more adapted the model has been to recent data, and the higher the confidence level.

[0068] Overall confidence level Where w is the weighting coefficient, satisfying The system calculates the Cmodel in real time in the core controller 9 and dynamically adjusts α to achieve an adaptive fusion mechanism where "the more complete and reliable the data, the greater the model weight".

[0069] Detailed explanation of the control formula: The control formula proposed in this invention is based on an understanding of the physical mechanism of the granulation process: 1. Relationship between vortex diameter and stirring speed: A vortex that is too large indicates insufficient shear force, requiring an increase in stirring speed; a vortex that is too small may indicate excessive shear, requiring a decrease in stirring speed. The differential term dD / dt should be considered in relation to its changing trend, allowing for an earlier response.

[0070] 2. Relationship between torque, viscosity, and binder addition: Torque reflects the degree of agglomeration, and viscosity reflects the consistency of the slurry. When the torque is lower than the set value, it indicates insufficient particle growth, and the binder can be added; when the viscosity remains lower than the set value, the binder addition should be gradually increased through the integral term.

[0071] 3. Relationship between material level fluctuation and water addition: Excessive material level fluctuation reflects flow field instability. The flow field can be stabilized by adding water to change the rheological properties. The differential term takes into account the trend of fluctuation.

[0072] The integration of these formulas with the data-driven model ensures basic control capabilities in the initial stage when no data is available, while also enabling continuously improved optimized control performance as data accumulates.

[0073] Example 4 An automated aqueous suspension granulation system for realizing an automated aqueous suspension granulation method includes: The multi-source sensing module includes: a high-speed image acquisition unit for acquiring flow field images inside the stirred tank 1, an online viscosity sensing unit for measuring material viscosity, and a torque sensing unit for measuring stirring resistance. The high-speed image acquisition unit includes a high-speed industrial camera 4, a ring LED fill light 5, and a dustproof and splashproof viewing window cover for protecting the camera; the online viscosity sensing unit is a vibration or rotational online viscometer 6, whose sensor probe extends into the slurry inside the vessel; the torque sensing unit includes a torque sensor mounted on the stirring motor 3 or the stirring shaft.

[0074] The data fusion and processing module, connected to the multi-source sensing module, is used to preprocess, extract and fuse the collected raw data to generate a comprehensive feature vector. The intelligent process decision-making module has an embedded process control model, which receives the comprehensive feature vector and outputs optimized process parameter adjustment instructions. The execution control module is connected to the process intelligent decision-making module and precisely controls the actions of the stirring motor driver, binder metering pump and water addition regulating valve according to the adjustment instructions. The model management and iteration module is used to store production data and product quality data, and provides the function of retraining and updating the process control model online.

[0075] The process intelligent decision-making module and the model management and iteration module are integrated into an industrial computer or an embedded edge computing device; the execution control module adopts a multi-loop PID controller or a more advanced model predictive controller, with a response time of less than 0.2 seconds.

[0076] The automated water suspension granulation system also includes a human-machine interaction and data interface module for parameter setting, status monitoring, report generation, and supports data exchange with the upper-level manufacturing execution system via OPC UA or Modbus protocol to achieve full lifecycle management of production batches.

[0077] It is evident that modifications and / or additions can be made to the above-described automated water suspension granulation device, system, and corresponding method based on stirred flow field without departing from the scope and domain of the present invention.

[0078] It is equally clear that, although the present invention has described in detail the automated water suspension granulation apparatus, system, and corresponding method based on stirred flow field, those skilled in the art will certainly be able to obtain many other equivalent forms of the automated water suspension granulation apparatus, system, and corresponding method based on stirred flow field, which have the features described in the claims and are therefore within the scope of protection defined herein.

Claims

1. An automated water suspension granulation method based on a stirred flow field, characterized in that, Includes the following steps: Step 1. Real-time acquisition of multi-source data: During the granulation process, dynamic flow field image data, online viscosity data of materials, and torque data of the stirring shaft in the mixing tank are acquired simultaneously; Step 2. Feature extraction and data fusion: The dynamic flow field image is processed to extract morphological feature parameters, including at least the material level height H, the diameter D and depth d of the vortex at the stirring center; the extracted morphological feature parameters, online viscosity data and stirring torque data are timestamped and normalized, and fused to form a comprehensive feature vector representing the current material state; Step 3. Model-based intelligent decision-making: The comprehensive feature vector is input into a pre-trained process control model, which outputs an optimized adjustment of process parameters for the current state. The process parameters include at least the stirring speed, the binder dripping speed, and the water addition amount. Step 4. Closed-loop execution and feedback control: Adjust the corresponding actuators in real time according to the adjustment amount output by the process control model; and continue to repeat steps 1 to 4 to form closed-loop control until the granulation process ends. Step 5. Online iterative optimization of the model: After the granulation operation of a single batch is completed, the final product quality index data of this batch is correlated with the production process data, and the process control model is incrementally learned to update the model parameters.

2. The method according to claim 1, characterized in that, In step 1, the dynamic flow field image is acquired by a high-speed industrial camera set at a transparent window on the top or side wall of the stirred tank, with a sampling frequency of not less than 60 frames / second; the online viscosity data is measured by an online viscometer installed on the bottom or side wall of the stirred tank; and the stirring torque data is measured by a torque sensor installed on the stirring motor or stirring shaft.

3. The method according to claim 1, characterized in that, In step 2, the feature extraction of the dynamic flow field image includes: using the Canny edge detection algorithm to identify the material liquid surface contour, and calculating its average height H and volatility; using a method based on morphological opening and closing operations and contour extraction to identify the central vortex region, and then calculating its equivalent diameter D, maximum depth d and contour symmetry index.

4. The method according to claim 1, characterized in that, The process control model is trained using a deep neural network or gradient boosting decision tree algorithm; the training objective function of the model is to minimize the difference between the particle size distribution dispersion of the particles obtained under the predicted process parameters and the target distribution; in step 5, the model is incrementally updated using online learning or mini-batch retraining.

5. The method according to claim 1, characterized in that, In step 3, the adjustment logic output by the process control model includes: When the integrated feature vector indicates that the vortex diameter exceeds the first threshold and the stirring torque is lower than the second threshold, the model outputs a command to simultaneously reduce the stirring speed and increase the binder dripping speed. When the comprehensive feature vector indicates that the material level fluctuation rate exceeds the third threshold and the online viscosity changes, the model outputs instructions to adjust the water addition and fine-tune the stirring speed to stabilize the flow field. When the integrated eigenvector indicates a sharp increase in online viscosity and a simultaneous significant increase in torque, the model outputs a command to temporarily increase the stirring speed and reduce the binder dripping speed to prevent excessive agglomeration.

6. The method according to claim 1, characterized in that, It also includes abnormal operating condition handling steps: when image acquisition failure or abnormal online viscosity data and stirring shaft torque data of materials are detected, the system automatically switches to a rule base based on historical similar operating condition data to perform emergency adjustment of process parameters and issue an alarm.

7. The method according to claim 1, characterized in that, In step 3, the process control model is a hybrid model based on the fusion of physical empirical rules and a data-driven model, wherein the physical empirical rule part adopts the following control formula: Formula for adjusting stirring speed based on vortex size: ； Where Δω is the stirring speed adjustment, Kp1 and Kd1 are the proportional and derivative coefficients, respectively, and D is the real-time vortex diameter. set To define the vortex diameter, dD / dt is the rate of change of the vortex diameter; Binder flow rate adjustment formula based on viscosity and torque: ； Where, ΔQ b Kp2 and Ki2 are the proportional and integral coefficients, respectively, and τ is the real-time stirring torque. set To set the torque, η is the real-time viscosity. set To set the viscosity; Water addition adjustment formula based on material level fluctuations: ； Where, ΔQ w For the water addition adjustment, Kp3 and Kd3 are the proportional and differential coefficients, respectively, and σ is the coefficient of variation. H σ represents the standard deviation of material level fluctuation. Hset To set the standard deviation of the fluctuation, dσ H / dt represents the rate of change of the standard deviation of the fluctuation.

8. The method according to claim 7, characterized in that, The hybrid model employs the following fusion method: when historical data is insufficient, it primarily relies on physical empirical rules; as data accumulates, the weights of the data-driven model are gradually increased; finally, a weighted fusion method is used. ； Where, ΔP final The final process parameter adjustment amount, ΔP rule The output is the physical empirical rule, ΔP. model The output of the data-driven model is α, which is a dynamic weight coefficient that gradually decreases from 1 to 0.2 as the model confidence increases.

9. An automated aqueous suspension granulation system for implementing the method according to any one of claims 1-8, characterized in that, include: The multi-source sensing module includes: a high-speed image acquisition unit for acquiring images of the flow field inside the stirred tank, an online viscosity sensing unit for measuring the viscosity of the material, and a torque sensing unit for measuring the stirring resistance. The data fusion and processing module, connected to the multi-source sensing module, is used to preprocess, extract and fuse the collected raw data to generate a comprehensive feature vector. The intelligent process decision-making module has an embedded process control model, which receives the comprehensive feature vector and outputs optimized process parameter adjustment instructions. The execution control module is connected to the process intelligent decision-making module and precisely controls the actions of the stirring motor driver, binder metering pump and water addition regulating valve according to the adjustment instructions. The model management and iteration module is used to store production data and product quality data, and provides the function of retraining and updating the process control model online.

10. The process intelligent decision-making module and the model management and iteration module are integrated into an industrial computer or an embedded edge computing device; the execution control module adopts a multi-loop PID controller or a more advanced model predictive controller, with a response time of less than 0.2 seconds.

11. The system according to claim 9, characterized in that, The high-speed image acquisition unit includes a high-speed industrial camera, a ring LED fill light, and a dustproof and splashproof viewing window cover for protecting the high-speed industrial camera; the online viscosity sensing unit is a vibration or rotational online viscometer, whose sensor probe extends into the slurry inside the vessel; the torque sensing unit includes a torque sensor mounted on a stirring motor or stirring shaft.

Images