System and method for monitoring, regulating and controlling massecuite crystallization process in real time

Abstract

The invention discloses a massecuite crystallization process real-time monitoring and regulation system, comprising: a high-pressure and high-temperature resistant microscopic imaging module, which collects crystal growth images in a crystallization tank in real time through a high-temperature resistant camera at a crystallization tank sight glass to obtain basic image data, the high-temperature resistant camera being equipped with a coaxial illumination system and an antifouling cleaning device; the image processing module receives the basic image data, performs segmentation processing on a crystal image through an improved U-Net neural network model, extracts crystal particle size distribution, roundness and coalescence rate parameters, and obtains crystal morphological parameters, and the improved U-Net neural network model comprises an attention mechanism and a weighted Dice loss function; the closed-loop control module is used for dynamically adjusting the vacuum degree, the temperature and the seed crystal adding rate of the crystallizing tank through a fuzzy PID algorithm according to the crystal morphological parameters to form a closed-loop control loop; the high-pressure-resistant and high-temperature-resistant microscopic imaging module, the image processing module and the closed-loop control module realize data interaction through an industrial bus.

Description

Technical Field

[0001] This invention belongs to the field of automation technology in the sugar industry, and in particular relates to a real-time monitoring and control system and method for the crystallization process of sugar paste, which is especially suitable for solving technical problems such as lagging crystal morphology monitoring and inaccurate control in traditional crystallization processes. Background Technology

[0002] Sugar crystallization is a core process in the sugar industry, and the quality of crystal growth directly affects the purity, yield, and energy consumption of the sugar product. Traditional crystallization processes rely heavily on manual experience for control, which presents the following technical bottlenecks: Lagging monitoring methods: Existing technologies generally employ manual intermittent sampling combined with offline microscopic observation, resulting in process adjustments lagging significantly behind the actual crystallization state; Insufficient imaging reliability: Industrial cameras are prone to sealing failure or sensor thermal drift in high temperature and high pressure environments, with a mean time between failures (MTBF) of less than 500 hours. Low image analysis accuracy: Traditional threshold segmentation methods have an error rate of up to 35% in identifying small crystals and adhered crystals, and cannot effectively handle image blurring caused by sugar paste adhesion; Disconnected control strategy: Existing PID control algorithms rely on indirect parameters such as temperature and vacuum level, and are not directly linked to crystal morphology parameters such as particle size distribution and aggregation rate, resulting in particle size uniformity (CV value) being higher than 25% for a long time.

[0003] While some patents have attempted to address the aforementioned issues, none have achieved a complete technological breakthrough across the entire chain of high-pressure, high-temperature imaging, high-precision image analysis, and real-time closed-loop control. Therefore, a systematic solution integrating specialized imaging technologies, deep learning, and intelligent control algorithms is urgently needed. Summary of the Invention

[0004] The main objective of this invention is to provide a real-time monitoring and control system and method for the crystallization process of sugar paste, aiming to solve four levels of technical problems: first, the long cycle of manual sampling and detection leads to a lag in process adjustment; second, industrial cameras have poor reliability and short mean time between failures under high temperature and high pressure environments; third, traditional image processing methods are insufficient for the segmentation accuracy of small crystals; and fourth, the control strategy is disconnected from real-time monitoring data.

[0005] To achieve the above objectives, the present invention provides a real-time monitoring and control system and method for the crystallization process of sugar paste, the system comprising the following modules: The high-pressure and high-temperature resistant microscopic imaging module acquires real-time images of crystal growth inside the crystal tank through a high-temperature resistant camera at the sight glass of the crystallizer to obtain basic image data. The high-temperature resistant camera is equipped with a coaxial lighting system and a contamination prevention and cleaning device. The image processing module receives the basic image data acquired by the high-pressure and high-temperature microscopic imaging module, performs segmentation processing on the crystal image through an improved U-Net neural network model, extracts crystal grain size distribution, roundness and aggregation rate parameters, and obtains crystal morphology parameters. The improved U-Net model includes an attention mechanism and a weighted Dice loss function. The closed-loop control module dynamically adjusts the vacuum level, temperature, and seed addition rate of the crystallizer based on the crystal morphology parameters output by the image processing module using a fuzzy PID algorithm, thus forming a closed-loop control circuit. The high-pressure and high-temperature microscopic imaging module, image processing module, and closed-loop control module interact with each other via an industrial bus.

[0006] Optionally, the high-pressure, high-temperature resistant microscopic imaging module includes: Ceramic substrate CMOS image sensor, operating temperature range -40℃ to 125℃; Sapphire crystal viewing lens assembly with a light transmittance of over 90% and a thickness of 8mm; It features a double-layer fluororubber sealing structure and is internally filled with nitrogen to maintain a positive pressure environment of 0.05MPa.

[0007] Optionally, the improved U-Net model in the image processing module includes: The encoder uses a pre-trained ResNet18 network; Add a channel attention mechanism at the jump connection; A joint loss function of weighted Dice loss and Focal Loss is adopted, where the weighting coefficient is 1+5·Y_true.

[0008] Optionally, the closed-loop control module includes: The crystal morphology parameter analysis unit calculates the average particle size d_avg and aggregation rate R in real time. The fuzzy PID controller has a built-in control rule base. When d_avg < 50μm, it outputs a vacuum adjustment signal of 0.01·(80-d_avg)MPa. The actuator interface connects to the vacuum pump, heater, and seed crystal addition device of the crystallizer.

[0009] Optionally, the image processing module further includes: The image preprocessing unit performs nonlocal mean denoising and CLAHE contrast enhancement processing on the basic image data acquired by the high-pressure and high-temperature microscopic imaging module. The post-processing unit uses a watershed algorithm to separate adhered crystals and filter out impurity regions with a roundness C < 0.3.

[0010] Optionally, the high-pressure, high-temperature resistant microscopic imaging module further includes: Rotary tungsten carbide scraper periodically cleans the surface of the sight glass; The compressed air jetting device has a pulse frequency of 0.1 seconds every 5 minutes.

[0011] Optionally, the system further includes: The data storage module records historical crystal morphology parameters and control logs; The remote monitoring interface supports real-time viewing of system status on both PC and mobile devices.

[0012] Furthermore, to achieve the above objectives, this invention also proposes a method for real-time monitoring of the sugar paste crystallization process, comprising the following steps: The high-pressure and high-temperature microscopic imaging module is used to acquire crystal images inside the crystallizer in real time to obtain basic image data; The basic image data is transmitted to the image processing module, where the improved U-Net model is used for crystal segmentation and parameter calculation to obtain crystal morphology parameters. Based on the crystal morphology parameters, the process parameters are dynamically adjusted through the closed-loop control module, and the control effect data is monitored in real time. The control effect data is fed back to the image acquisition module to form a closed-loop control cycle.

[0013] Optionally, the training process of the improved U-Net model includes: Collect 5000 labeled crystal images as the basic training set; For the basic training set, data augmentation is performed using rotation ±30° and Gaussian noise (σ=0.01) to obtain augmented data; For the augmented data, the AdamW optimizer was used with an initial learning rate of 0.0003, and training continued until the Dice coefficients on the validation set converged.

[0014] Optionally, the closed-loop control cycle process includes: The image acquisition-analysis-control cycle is completed once per second; When the coalescence rate R > 15%, a seed pulse is triggered, increasing the flow rate by 5%. It displays the crystal size distribution curve and the trend of process parameters in real time.

[0015] Compared with the existing technology, the technical solution provided by the embodiments of the present invention has the following advantages and features: First, the processing speed is fast, the complete process takes less than 80ms / frame, which meets the real-time requirement of 15fps, while the traditional OpenCV method takes 400ms / frame, which is 5 times faster than the traditional method; the accuracy is also significantly improved; it has strong engineering applicability, automatically adapts to changes in sugar paste concentration, and the focus blur tolerance reaches ±15μm. To more clearly illustrate the specific embodiments of the present invention, the accompanying drawings used in the description of the specific embodiments will be briefly introduced below. Attached Figure Description Figure 1 This is a flowchart of the real-time monitoring and control system for the crystallization process of sugar paste according to the present invention; Figure 2 This is a diagram of the control closed-loop architecture of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0017] Reference Figure 1 , Figure 1 This is a flowchart of the real-time monitoring and control system for the crystallization process of sugar paste according to the present invention.

[0018] This invention provides a real-time monitoring and control system for the crystallization process of sugar paste, comprising the following modules: The high-pressure and high-temperature resistant microscopic imaging module acquires real-time images of crystal growth inside the crystal tank through a high-temperature resistant camera at the sight glass of the crystallizer to obtain basic image data. The high-temperature resistant camera is equipped with a coaxial lighting system and a contamination prevention and cleaning device. The image processing module receives the basic image data acquired by the high-pressure and high-temperature microscopic imaging module, performs segmentation processing on the crystal image through an improved U-Net neural network model, extracts crystal grain size distribution, roundness and aggregation rate parameters, and obtains crystal morphology parameters. The improved U-Net model includes an attention mechanism and a weighted Dice loss function. The closed-loop control module dynamically adjusts the vacuum level, temperature, and seed addition rate of the crystallizer based on the crystal morphology parameters output by the image processing module using a fuzzy PID algorithm, thus forming a closed-loop control circuit. The high-pressure and high-temperature microscopic imaging module, image processing module, and closed-loop control module interact with each other via an industrial bus, forming a complete real-time monitoring-analysis-control system.

[0019] Furthermore, the high-pressure and high-temperature resistant microscopic imaging module includes an image acquisition unit, an optical system, and an environmental protection system.

[0020] To withstand the high-temperature environment inside the crystallization tank, the image acquisition unit utilizes a ceramic substrate CMOS image sensor. Specifically, the image sensor is encapsulated on a ceramic substrate and electrically connected via gold wire bonding. The interior is filled with high-purity nitrogen to isolate moisture and corrosive gases. For example, the ON Semiconductor PYTHON500 CMOS sensor can be used, featuring a resolution of 2448×2048 and a frame rate of 15fps, clearly capturing the dynamic process of sugar crystallization. The ceramic substrate has high thermal conductivity, rapidly conducting heat from the chip to the heat dissipation structure, while the ceramic material itself possesses high-temperature resistance. Utilizing the physical property of gold wire's melting point of 1064℃, gold wire bonding prevents open circuits caused by metal migration at high temperatures. Maintaining a slightly positive pressure nitrogen filling method prevents sugar vapor from seeping into and corroding the circuitry; a optimal solution is to maintain a slightly positive pressure of 0.05MPa. Furthermore, the coefficient of thermal expansion (CTE) of the ceramic package (CTE=7ppm / ℃) matches that of the silicon chip (CTE=4ppm / ℃), effectively preventing solder joint cracking caused by temperature cycling.

[0021] The optical system consists of sapphire lens components. A preferred design uses a 5-element sapphire lens group with a 400-700nm broadband anti-reflective coating. The lens uses an 8mm thick sapphire window with >90% light transmittance, paired with a fixed-focus design with a numerical aperture (NA) of 0.65. Illumination uses a 450nm coaxial LED, whose wavelength matches the refractive index of sugar crystals (n=1.54), enhancing edge contrast. Utilizing sapphire's Mohs hardness of approximately 9, it resists scratches from sugar particles, and its higher thermal conductivity than ordinary glass prevents thermal stress cracking. Coaxial illumination eliminates interference from vapor scattering within the can, and the 450nm blue light can penetrate suspended sugar particles.

[0022] The environmental protection system is achieved through a double-layer fluororubber sealing structure. The double-layer fluororubber O-rings can withstand temperatures up to 200℃, and the interior is filled with 0.05MPa nitrogen to form a positive pressure barrier. The heat dissipation system consists of copper fins and a water-cooling circulation system. For example, a preferred solution is to use copper fins with a surface area of ​​500cm² and a 25℃ water-cooling circulation system, which can ensure that the sensor temperature is ≤80℃.

[0023] A preferred failure prevention mechanism is to employ a double-sealing layer. When the first O-ring fails, nitrogen leakage will trigger a pressure alarm, while the second layer can maintain a short-term seal. Furthermore, positive pressure nitrogen can prevent sugar paste from adhering to the sight glass; at 0.05 MPa, the amount of paste adhering to the surface can be reduced by approximately 60%.

[0024] Furthermore, the image processing module includes an image preprocessing unit and a post-processing unit: The image preprocessing unit performs nonlocal mean denoising and adaptive CLAHE enhancement processing on the basic image data acquired by the high-pressure and high-temperature microscopic imaging module. The post-processing unit uses a watershed algorithm to separate adhered crystals and filter out impurity regions with a roundness C < 0.3.

[0025] For nonlocal means denoising, this invention employs an improved NLM algorithm to process sugar syrup crystallization images, the core formula of which is:

[0026] in, The denoised pixel value at position x; Normalization factor; The neighborhood of pixel x is a 5×5 pixel region; The neighborhood block of pixel y is a 15×15 search window; Original image pixel values; ||N(x)-N(y)||² calculates the Euclidean distance between two neighboring blocks, 0.1 is a smoothing parameter to control weight decay.

[0027] By setting a 5×5 pixel neighborhood window and a 15×15 pixel search window, large-scale similar block matching is achieved while ensuring computational efficiency. A specially designed attenuation coefficient h=0.1 effectively optimizes the edge preservation effect of the sugar crystals.

[0028] The adaptive CLAHE enhancement module employs an 8×8 grid, a size perfectly matching typical crystal sizes of 20-100μm, corresponding to 10-50 pixels. A contrast limit of 2.0 effectively enhances the image while avoiding local over-enhancement. A configuration of 256 histogram bins ensures smooth grayscale transitions. Innovative improvements include a dynamic adjustment function that automatically switches to a 4×4 grid when a crystal density greater than 50 crystals / grid is detected, and a dedicated grayscale mapping algorithm specifically designed to enhance the 400-550nm wavelength band, which corresponds to the characteristic absorption peaks of sugar crystals.

[0029] The improvements to the watershed algorithm of this invention include two key points: First, L2 distance metric is used for distance transformation optimization, and a neighborhood radius of 5 pixels is set to accurately separate crystals with a spacing greater than 5μm; Second, U-Net prediction probability map is used as the initial label, with a threshold >0.8, and over-segmented regions with an area less than 10 pixels are eliminated.

[0030] The morphological filtering employs a three-stage filtering mechanism: area filtering is set to a minimum of 10 pixels, corresponding to 5μm crystals, effectively eliminating artifacts caused by noise; roundness screening is set with a threshold of C=0.3 to filter amorphous impurities; and the aspect ratio is limited to a maximum of 2.5:1 to exclude fibrous contaminants. Actual measurements show that this post-processing workflow improves crystal counting accuracy from 78% to 94%, with an average particle size measurement error of less than 0.7μm.

[0031] Furthermore, the improved U-Net model in the image processing module includes: The encoder uses a pre-trained ResNet18 network; Add a channel attention mechanism (SE module) at the jump connection; A joint loss function of weighted Dice loss and Focal Loss is adopted, where the weighting coefficient is 1+5·Y_true.

[0032] In terms of network architecture, the encoder uses a pre-trained ResNet18 network, and the kernel size of the first layer is adjusted from 3 to 5 to better accommodate the microscopic features of sugar crystals. A channel attention mechanism module is embedded at each skip connection, with a compression ratio of r=16, improving feature selection capability while maintaining computational efficiency, especially for small crystals below 20μm, where channel weights can be improved by 3-5 times. The decoder uses transposed convolutional upsampling instead of traditional bilinear interpolation, and the batch normalization momentum coefficient is set to 0.1 to better adapt to mini-batch training.

[0033] The hybrid loss function adopts a combination of L=0.7L_Dice + 0.3L_Focal. The weighted Dice coefficient W=1+5·Y_true increases the weight of small crystals by 6 times. The Focal Loss parameters are set to α=0.8 and γ=2, which effectively adjusts the balance between positive and negative samples and suppresses easily separable samples. Actual training results show that this design improves the Dice coefficient for segmenting tiny crystals smaller than 20μm from 0.62 to 0.83, and increases the IoU at the boundary of adhered crystals by 28%.

[0034] In terms of training strategies, data augmentation methods included ±30° rotation and the addition of Gaussian noise with σ=0.01. The optimizer used was AdamW, with a weight decay of 0.01 to prevent overfitting. An initial learning rate of 0.0003 was used, and training was conducted for 100 epochs. The final validation set loss converged to 0.153.

[0035] Furthermore, the closed-loop control module includes: The crystal morphology parameter analysis unit calculates the average particle size d_avg and aggregation rate R in real time. The fuzzy PID controller has a built-in control rule base. When d_avg < 50μm, it outputs a vacuum adjustment signal of 0.01·(80-d_avg)MPa. The actuator interface connects to the vacuum pump, heater, and seed crystal addition device of the crystallizer.

[0036] like Figure 2 This is a diagram of the closed-loop control architecture of the present invention. The closed-loop control module of the present invention dynamically adjusts crystallization process parameters, such as vacuum level, temperature, and seed crystal flow rate, by analyzing crystal morphology parameters (particle size, aggregation rate, and roundness) in real time, thus forming a closed-loop control module. The specific data flow is as follows: The high-pressure, high-temperature microscopic imaging module acquires raw 2448×2048 images at a refresh rate of 15fps. After preprocessing, U-Net segmentation, and post-processing, three key parameters are extracted: average particle size (d_avg), aggregation rate (R), and average roundness (C_avg). These parameters are input to the control core at an update frequency of 15 times per second, forming the first stage of data input for the control closed loop. These parameters are transmitted in real time to the fuzzy PID controller via an industrial bus.

[0037] At the control algorithm layer, the system employs an improved fuzzy PID control strategy. The innovation of this strategy lies in combining the linear control of traditional PID with the nonlinear processing of fuzzy logic: firstly, the precise crystal parameters are transformed into fuzzy variables through the membership function, with the particle size deviation E_d mapped to the domain of discourse [-1,1] and the aggregation rate deviation E_R mapped to the domain of discourse [0,2].

[0038] Where E_d = (d_target - d_avg) / d_target, E_R = (R - 15%) / 15% Where E_d is the particle size deviation, d_avg is the average particle size, d_target is the target average particle size, E_R is the agglomeration rate deviation, and R is the agglomeration rate.

[0039] These fuzzy variables are then processed in a rule base containing 27 control rules, among which the core rules include: When d_avg < 50μm, the output vacuum adjustment ΔV = 0.01·(80-d_avg)MPa; When R > 15%, a pulse command is triggered to increase the seed flow rate by 5%. When C_avg < 0.8, the temperature setpoint is increased by 1℃.

[0040] Control commands enable real-time adjustment of process parameters through three types of actuators: Vacuum pump control system: Driven by Siemens S7-1200 PLC, with an adjustment range of -0.1~0MPa and an accuracy of ±0.005MPa. Its control algorithm integrates feedforward compensation and feedback correction. The feedforward term predicts vacuum requirements based on a crystallization kinetic model, while the feedback term is fine-tuned based on real-time image analysis results. Each 0.01MPa increase in vacuum level can increase the crystallization rate by 12%, but excessively high vacuum can lead to crystal breakage; therefore, the control algorithm sets a safety upper limit of -0.085MPa.

[0041] Heater temperature control system: A closed-loop system is formed using a PT100 temperature sensor (accuracy ±0.1℃) and an SSR solid-state relay, with a temperature control range of 60~100℃. Temperature regulation adopts an adaptive PID algorithm, whose parameters dynamically change with the crystallization stage: during the nucleation stage (t<2h), more conservative parameters (Kp=0.8, Ki=0.05) are used to avoid excessive temperature fluctuations; during the growth stage, a more aggressive regulation is adopted (Kp=1.2, Ki=0.03) to accelerate crystal growth.

[0042] Seed addition device: It adopts a high-precision peristaltic pump with a flow rate range of 0.1-5L / min and a repeatability of ±0.5%. Its control strategy has two innovative features: First, pulse addition based on coalescence rate, which triggers a 5% flow rate step when R>15%; Second, preventive addition based on historical data learning, which uses LSTM network prediction to increase the seed flow rate by 1-2% in advance before coalescence is predicted to occur.

[0043] Furthermore, the high-pressure and high-temperature resistant microscopic imaging module also includes: a rotating tungsten carbide scraper for periodically cleaning the surface of the sight glass; and a compressed air jetting device with a pulse frequency of 0.1 seconds every 5 minutes.

[0044] Furthermore, the system also includes: a data storage module for recording historical crystal morphology parameters and control logs; and a remote monitoring interface for real-time viewing of system status on PCs and mobile devices.

[0045] Accordingly, this invention also provides a method for real-time monitoring of the crystallization process of sugar paste, comprising the following steps: The high-pressure and high-temperature microscopic imaging module is used to acquire crystal images inside the crystallizer in real time to obtain basic image data; The basic image data is transmitted to the image processing module, where the improved U-Net model is used for crystal segmentation and parameter calculation to obtain crystal morphology parameters. Based on the crystal morphology parameters, the process parameters are dynamically adjusted through the closed-loop control module, and the control effect data is monitored in real time. The control effect data is fed back to the image acquisition module to form a closed-loop control cycle.

[0046] Furthermore, the training process of the improved U-Net model includes: collecting 5000 labeled crystal images as the basic training set; performing data augmentation on the basic training set by rotating ±30° and adding Gaussian noise (σ=0.01) to obtain augmented data; and using the AdamW optimizer with an initial learning rate of 0.0003 on the augmented data until the Dice coefficients of the validation set converge.

[0047] Furthermore, the closed-loop control cycle includes: completing an image acquisition-analysis-regulation cycle once per second; when the coalescence rate R>15%, triggering the addition of seed crystal pulses, increasing the flow rate by 5%; and displaying the crystal particle size distribution curve and the trend of process parameter changes in real time.

[0048] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A real-time monitoring and control system for the crystallization process of sugar paste, characterized in that, include: The high-pressure and high-temperature resistant microscopic imaging module acquires real-time images of crystal growth inside the crystal tank through a high-temperature resistant camera at the sight glass of the crystallizer to obtain basic image data. The high-temperature resistant camera is equipped with a coaxial lighting system and a contamination prevention and cleaning device. The image processing module receives the basic image data acquired by the high-pressure and high-temperature microscopic imaging module, performs crystal image segmentation processing through an improved U-Net neural network model, extracts crystal grain size distribution, roundness and aggregation rate parameters, and obtains crystal morphology parameters. The improved U-Net model includes an attention mechanism and a weighted Dice loss function. The closed-loop control module dynamically adjusts the vacuum level, temperature, and seed addition rate of the crystallizer based on the crystal morphology parameters output by the image processing module using a fuzzy PID algorithm, thus forming a closed-loop control circuit. The high-pressure and high-temperature microscopic imaging module, image processing module, and closed-loop control module interact with each other via an industrial bus.

2. The system according to claim 1, characterized in that, The high-pressure, high-temperature resistant microscopic imaging module includes: Ceramic substrate CMOS image sensor, operating temperature range -40℃ to 125℃; Sapphire crystal viewing lens assembly with a light transmittance of over 90% and a thickness of 8mm; It features a double-layer fluororubber sealing structure and is internally filled with nitrogen to maintain a positive pressure environment of 0.05MPa.

3. The system according to claim 1, characterized in that, The improved U-Net model in the image processing module includes: The encoder uses a pre-trained ResNet18 network; Add a channel attention mechanism at the jump connection; A joint loss function of weighted Dice loss and Focal Loss is adopted, where the weighting coefficient is 1+5·Y_true.

4. The system according to claim 1, characterized in that, The closed-loop control module includes: The crystal morphology parameter analysis unit calculates the average particle size d_avg and aggregation rate R in real time. The fuzzy PID controller has a built-in control rule base. When d_avg < 50μm, it outputs a vacuum adjustment signal of 0.01·(80-d_avg)MPa. The actuator interface connects to the vacuum pump, heater, and seed crystal addition device of the crystallizer.

5. The system according to claim 1, characterized in that, The image processing module further includes: The image preprocessing unit performs nonlocal mean denoising and adaptive CLAHE enhancement processing on the basic image data acquired by the high-pressure and high-temperature microscopic imaging module. The post-processing unit uses a watershed algorithm to separate adhered crystals and filter out impurity regions with a roundness C < 0.

3.

6. The system according to claim 1, characterized in that, The high-pressure, high-temperature resistant microscopic imaging module also includes: Rotary tungsten carbide scraper periodically cleans the surface of the sight glass; The compressed air jetting device has a pulse frequency of 0.1 seconds every 5 minutes.

7. The system according to claim 1, characterized in that, The system also includes: The data storage module records historical crystal morphology parameters and control logs; The remote monitoring interface supports real-time viewing of system status on both PC and mobile devices.

8. A method for real-time monitoring of the crystallization process of sugar paste, characterized in that, Includes the following steps: The high-pressure and high-temperature microscopic imaging module is used to acquire crystal images inside the crystallizer in real time to obtain basic image data; The basic image data is transmitted to the image processing module, where the improved U-Net model is used for crystal segmentation and parameter calculation to obtain crystal morphology parameters. Based on the crystal morphology parameters, the process parameters are dynamically adjusted through the closed-loop control module, and the control effect data is monitored in real time. The control effect data is fed back to the image acquisition module to form a closed-loop control cycle.

9. The method according to claim 8, characterized in that, The training process of the improved U-Net model includes: Collect 5000 labeled crystal images as the basic training set; For the basic training set, data augmentation is performed using rotation ±30° and Gaussian noise (σ=0.01) to obtain augmented data; For the augmented data, the AdamW optimizer was used with an initial learning rate of 0.0003, and training continued until the Dice coefficients on the validation set converged.

10. The method according to claim 8, characterized in that, The closed-loop control cycle process includes: The image acquisition-analysis-control cycle is completed once per second; When the coalescence rate R > 15%, a seed pulse is triggered, increasing the flow rate by 5%. It displays the crystal size distribution curve and the trend of process parameters in real time.

Images