A Smart Whole Soy Milk Manufacturing System with Excellent Content Retention and Nutrient Preservation Features Through AI-Based Process Optimization

Abstract

A smart whole soy milk manufacturing system that produces whole soy milk with excellent solid content retention by processing data obtained from multiple sensors in real time based on AI comprises: a classification module that determines and classifies premium soybeans among multiple soybeans; an immersion module that immerses the determined premium soybeans in a liquid inside an immersion tank; a grinding module that produces a soybean mixture by grinding the immersed premium soybeans together with water inside a grinder; a temperature-changing module that introduces the soybean mixture into a reaction tank and heats it until it reaches a predetermined viscosity, and cools the heated soybean mixture until it reaches a predetermined temperature; a concentration module that concentrates the cooled soybean mixture under a vacuum until it reaches a predetermined solid content concentration; a homogenization module that operates a homogenizer so that soybean particles contained in the concentrated soybean mixture are finely dispersed; and a control unit connected to the classification module, the immersion module, the grinding module, the temperature-changing module, the concentration module, and the homogenization module, and transmitting or receiving data in real time with at least one module.

Description

Technology Field

[0001] The disclosed invention relates to a smart whole soy milk manufacturing system. Background Technology

[0002] Traditionally, whole soy milk is produced by washing and soaking soybeans, followed by grinding, heating, and cooling. However, if the quality of the soybeans and soaking conditions are inconsistent during this process, problems arise in maintaining uniform solid content, particle size, taste, and flavor in the final product. Furthermore, if temperature, viscosity, or pH is not properly controlled during the process, precipitation or protein denaturation may occur, potentially degrading the smooth texture and flavor preferred by consumers.

[0003] Although some automated equipment is introduced in the whole soy milk manufacturing process to control temperature, pressure, and time, it often remains at the level of single sensors (PID control) or limited mechanical control. Even if sensors are used to measure the weight, temperature, and viscosity of the soybeans, it is difficult to dynamically optimize process conditions by comprehensively analyzing real-time data. As a result, process deviations and quality variations between products increase, making it difficult to produce homogenized whole soy milk without sedimentation while maintaining a stable solid content.

[0004] Recently, smart manufacturing systems that combine machine learning models such as deep learning with sensor networks to monitor and control the entire process in real time, from raw material selection to final homogenization, are attracting attention.

[0006] The disclosed invention is intended to provide a smart manufacturing system for producing whole soy milk with excellent solid content retention and nutrient preservation functions using AI trained for process optimization.

[0007] A smart whole soy milk manufacturing system according to one embodiment of the present disclosure may include: a classification module that determines and classifies premium soybeans among a plurality of soybeans; an immersion module that immerses the determined premium soybeans in a liquid inside an immersion tank; a grinding module that prepares a soybean mixture by grinding the immersed premium soybeans together with water inside a grinder; a temperature-changing module that introduces the soybean mixture into a reaction tank and heats it until it reaches a predetermined viscosity, and cools the soybean mixture after heating until it reaches a predetermined temperature; a concentration module that concentrates the soybean mixture after cooling until it reaches a predetermined solid content concentration under vacuum; a homogenization module that operates a homogenizer so that soybean particles contained in the concentrated soybean mixture are finely dispersed; and a control unit connected to the classification module, the immersion module, the grinding module, the temperature-changing module, the concentration module, and the homogenization module, and transmitting or receiving data in real time with at least one module.

[0008] According to one aspect of the disclosed invention, since data measured by a plurality of sensors is integratedly analyzed and controlled by AI, the solid content is stably adjusted to a level of 9-10% without variation between batches, thereby ensuring homogeneous quality with consistent concentration and taste of the final whole soy milk.

[0009] According to one aspect of the disclosed invention, optimal conditions are dynamically reflected by utilizing real-time sensor data such as temperature, viscosity, pH, and pressure at each process step, including immersion, heating, and cooling, thereby significantly reducing protein denaturation or fat precipitation, which improves the soft texture preferred by consumers and hygienic safety.

[0010] According to one aspect of the disclosed invention, a machine learning model analyzes real-time sensor data to automatically adjust the time for each step, and by reducing unnecessary process delays or excessive energy consumption, the time and cost required for the entire process can be efficiently reduced.

[0011] According to one aspect of the disclosed invention, by combining an image sensor, a weight sensor, and a spectrum analyzer to evaluate the quality of soybeans from various angles, whole soy milk can be produced based on high-protein, high-quality soybeans, thereby directly contributing to enhancing the nutrition and flavor of the whole soy milk.

[0012] According to one aspect of the disclosed invention, since each part, such as the immersion module, concentration module, and homogenization module, is organically connected to and operated with an AI control unit, it is possible to flexibly respond to future process additions or production scale expansions, thereby enabling the development of a variety of high-value-added whole soy milk products.

[0013] The effects obtainable from the present disclosure are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art to which the present disclosure belongs from the description below. Brief explanation of the drawing

[0015] FIG. 1 illustrates an example of a smart whole soy milk manufacturing system according to one embodiment. FIG. 2 illustrates an example of the configuration of a control unit of a smart whole soy milk manufacturing system according to one embodiment. FIG. 3 illustrates an example of a block diagram showing the data flow between a classification module and a control unit according to one embodiment. FIG. 4 illustrates an example of a block diagram showing the data flow between an immersion module and a control unit according to one embodiment. FIG. 5 illustrates an example of a block diagram showing the data flow between a grinding module and a control unit according to one embodiment. FIG. 6 illustrates an example of a block diagram showing the data flow between a temperature change module and a control unit according to one embodiment. FIG. 7 illustrates an example of a block diagram showing the data flow between a concentration module and a control unit according to one embodiment. FIG. 8 illustrates an example of a block diagram showing the data flow between a homogenization module and a control unit according to one embodiment. FIG. 9 illustrates an example of a block diagram showing the data flow between a drying module and a control unit according to one embodiment. Specific details for implementing the invention

[0016] The various embodiments of this document and the terms used therein are not intended to limit the technical features described in this document to specific embodiments, and should be understood to include various modifications, equivalents, or substitutions of said embodiments.

[0017] In relation to the description of the drawings, similar reference numerals may be used for similar or related components.

[0018] The singular form of the noun corresponding to the item may include one or multiple items, unless the relevant context clearly indicates otherwise.

[0019] In this document, each of the phrases such as "A or B", "at least one of A and B", "at least one of A or B", "A, B or C", "at least one of A, B and C", and "at least one of A, B, or C" may include any one of the items listed together in the corresponding phrase, or all possible combinations thereof.

[0020] The term “and / or” includes a combination of multiple related described components or any of the multiple related described components.

[0021] For example, a phrase such as "A, B, and / or C" may include any one of the items listed together in the corresponding phrase, or all possible combinations thereof.

[0022] Terms such as "first," "second," or "first" or "second" may be used simply to distinguish a component from another component and do not limit the components in other aspects (e.g., importance or order).

[0023] Where any (e.g., 1st) component is referred to as "coupled" or "connected" to another (e.g., 2nd) component, with or without the terms "functionally" or "communicationly," it means that said any component may be connected to said other component directly (e.g., via a wire), wirelessly, or through a third component.

[0024] Terms such as “include” or “have” are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in this document, and do not preclude the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0025] When it is said that a component is "connected," "combined," "supported," or "in contact" with another component, this includes not only cases where the components are directly connected, combined, supported, or in contact, but also cases where they are indirectly connected, combined, supported, or in contact through a third component.

[0026] When it is said that a component is located "on" another component, this includes not only cases where one component is in contact with the other, but also cases where another component exists between the two components.

[0027] A smart whole soy milk manufacturing system according to various embodiments will be described in detail below with reference to the attached drawings.

[0028] FIG. 1 illustrates an example of a smart whole soy milk manufacturing system according to one embodiment.

[0029] FIG. 2 illustrates an example of the configuration of a control unit of a smart whole soy milk manufacturing system according to one embodiment.

[0030] Referring to FIG. 1, a smart whole soy milk manufacturing system according to one embodiment may include a classification module, an immersion module, a grinding module, a temperature change module, a concentration module, a homogenization module, and a control unit connected to each module to transmit and receive real-time data and control each module.

[0031] Accordingly, each module and control unit may include a communication unit to transmit and receive real-time data to and from each other. The communication unit may support wired or wireless methods for data transmission and reception.

[0032] Wired communication between each module and the control unit can be used for stable and rapid transmission of data. Wired communication methods may include Ethernet, RS-485, RS-232, CAN (Controller Area Network) bus, Modbus, etc.

[0033] Wireless communication between each module and the control unit is a method that enables data transmission and reception between each module and the control unit without physical cable connections, providing installation flexibility. Wireless communication methods may include Wi-Fi (IEEE 802.11), Bluetooth (IEEE 802.15.1), Zigbee (IEEE 802.15.4), Z-Wave, NFC (Near Field Communication), and cellular communication (LTE, 5G).

[0034] The communication unit of each module and the control unit may include a transmitting and receiving device and a protocol processing device. In addition, the smart whole soy milk manufacturing system according to the present invention can operate in an environment where wired and wireless communication methods are mixed.

[0035] For example, sensor data requiring real-time performance can be transmitted to the control unit via wired communication (Ethernet, RS-485, etc.), while the module's operating status can be transmitted to a central monitoring system via wireless communication (Wi-Fi, Zigbee, etc.). Additionally, the system can be configured to use an alternative communication method in the event of a communication failure. Furthermore, encryption protocols (WPA3, TLS, etc.) can be applied in wireless communication to maintain data integrity and security. In wired communication as well, technologies such as Cyclic Redundancy Check (CRC) can be applied to verify data authentication and integrity. This prevents data loss or external attacks that may occur during the process.

[0036] The classification module can automate the process of determining and classifying premium soybeans from among multiple soybeans. The classification module may include at least one image sensor and at least one weight sensor.

[0037] Image data of each of a plurality of soybeans can be obtained from at least one image sensor, and weight data of each of a plurality of soybeans can be obtained from at least one weight sensor.

[0038] As described above, image data and weight data acquired from each sensor can be transmitted to a control unit. The control unit may include at least one machine learning model. Based on the image data and weight data, the control unit can analyze the appearance, weight, internal components, etc., of the soybeans through the machine learning model to acquire related data. Through this process, high-quality soybeans, so-called premium soybeans, can be finally determined, thereby allowing for the classification of premium soybeans among multiple soybeans.

[0039] Among multiple types of soybeans, premium soybeans can be transferred from the sorting module to the immersion module. The immersion module can immerse the premium soybeans in the liquid inside the immersion tank. By dynamically adjusting the expansion rate and immersion temperature of the soybeans, the immersion module can perform an automated immersion process under optimal conditions. The immersion process is a procedure in which soybeans are soaked in water for a certain period to increase their moisture content and induce expansion and component activation.

[0040] The immersion module may include at least one ultrasonic sensor, at least one image sensor, etc., to measure the expansion rate of soybeans. Additionally, it may include at least one temperature sensor to measure the temperature inside the immersion tank or the temperature of the liquid inside the immersion tank.

[0041] Data related to the expansion rate of soybeans can be obtained from at least one ultrasonic sensor or at least one image sensor. Additionally, temperature data related to the inside of the immersion tank or the liquid inside can be obtained through at least one temperature sensor.

[0042] As described above, data related to the expansion rate and internal temperature of the soaking tank obtained from each sensor can be transmitted to the control unit. Based on the received data regarding the expansion rate and temperature, the control unit can compare and analyze the soaking state of the premium soybeans with the target state through a machine learning model, and then correct the soaking temperature and soaking time in real time, thereby enabling the premium soybeans to be soaked under optimal conditions.

[0043] The premium soybeans that have completed soaking can be transferred to a grinding module. The grinding module can produce a soybean mixture by grinding the soaked premium soybeans together with water. The grinding module may include a water supply unit that supplies water inside the grinder, at least one weight sensor that acquires weight data of the soaked premium soybeans, and at least one optical sensor or at least one ultrasonic sensor that acquires data related to the average particle size of the soybean mixture.

[0044] As described above, weight data of the immersed premium soybeans and average particle size data of the soybean mixture being ground, obtained from each sensor, can be transmitted to the control unit. Based on the received weight data and average particle size data, the control unit can compare and analyze the grinding state of the soybean mixture with the target state through a machine learning model, and then correct the grinding speed or stop the operation of the grinder.

[0045] The soybean mixture produced in the grinding module can be transferred to the temperature-changing module. The temperature-changing module can introduce the soybean mixture into a reaction vessel and heat it until it reaches a predetermined viscosity. Additionally, the temperature-changing module can cool the heated soybean mixture until it reaches a predetermined temperature.

[0046] The temperature-changing module may include a reaction vessel for holding a soybean mixture when heating or cooling, at least one temperature sensor for measuring the heating and / or cooling state, at least one viscosity sensor, at least one pH sensor, etc.

[0047] At least one temperature sensor can acquire data regarding the temperature of the soybean mixture. At least one viscosity sensor can acquire data regarding the viscosity of the soybean mixture. At least one pH sensor can acquire data regarding the pH of the soybean mixture.

[0048] As described above, data related to the temperature, viscosity, and pH of the soybean mixture obtained from each sensor can be transmitted to the control unit. Based on the received data, the control unit can operate a heater or cooler through a machine learning model until the soybean mixture reaches a target viscosity. Additionally, once the soybean mixture reaches the target viscosity by removing moisture through heating in the variable temperature module, the control unit can operate the cooler to lower the temperature of the soybean mixture to a target temperature for concentration to be performed in the concentration module.

[0049] The soybean mixture, having completed the temperature-changing process, can be transferred to a concentration module and concentrated under vacuum. The concentration module can increase the concentration of solids by removing moisture from the soybean mixture under vacuum. The concentration module may include a temperature sensor for acquiring data regarding the internal temperature of the vacuum concentrator, a viscosity sensor for acquiring viscosity data of the soybean mixture, a pressure sensor for acquiring data regarding the internal pressure of the vacuum concentrator, and a refractometer for acquiring data regarding the solid concentration of the soybean mixture.

[0050] As described above, data acquired from each sensor or refractometer can be transmitted to the control unit. Based on the received data, the control unit can dynamically correct the internal pressure and concentration time of the vacuum concentrator through a machine learning model.

[0051] The homogenization module can perform a process of producing whole soy milk of uniform quality by finely dispersing soybean particles contained in a concentrated soybean mixture. The homogenization module may include a high-pressure homogenizer, at least one temperature sensor for acquiring temperature data of the soybean mixture, at least one viscosity sensor for acquiring viscosity data of the soybean mixture, and at least one optical sensor for acquiring data on soybean particle size and dispersion state.

[0052] A high-pressure homogenizer can apply high pressure to a soybean mixture to grind the particles, thereby reducing the average size of the soybean particles and making the particle distribution uniform. At least one temperature sensor can acquire temperature data of the soybean mixture inside the high-pressure homogenizer. At least one viscosity sensor can acquire viscosity data of the soybean mixture inside the high-pressure homogenizer. At least one optical sensor can measure the average particle size, particle distribution, and dispersion state data of the soybean mixture in real time.

[0053] As described above, temperature data, viscosity data, and average particle size data obtained from each sensor can be transmitted to the control unit. Based on the received data, the control unit can determine the homogenization pressure, flow rate, and number of repetitions after comparing and analyzing the homogenization state and target state of the soybean mixture through a machine learning model.

[0054] For example, if the control unit determines that the average particle size data received from the optical sensor approaches or reaches a predetermined particle size reference value, it can transmit a signal to the high-pressure homogenizer to stop the homogenization process. In addition, by analyzing viscosity data or temperature data, the homogenization pressure or flow rate can be adjusted in real time to prevent viscosity from rising due to excessive pressure or quality degradation due to heat accumulation.

[0055] Referring to FIG. 2, the control unit processes data collected from a plurality of sensors, transmits control commands to each module, and can operate as a key element for dynamically controlling the entire process. The control unit may include at least one processor to process data in real time and execute control commands.

[0056] The processor can be composed of a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), or an artificial intelligence accelerator (AI Accelerator). For example, high-performance GPUs or dedicated AI semiconductor chips (TPU, NPU, etc.) can be utilized to process deep learning models. The processor can handle core computational tasks of the control unit, such as data analysis, machine learning model execution, sensor signal processing, and control command generation.

[0057] The control unit may include at least one memory for storing and processing data. The memory may include volatile memory (e.g., DRAM, SRAM) for storing temporary data and non-volatile memory (e.g., flash memory, SSD) for storing long-term data. For storing temporary data, the memory may store real-time data collected from sensors or intermediate calculation results of machine learning models. For storing long-term data, the memory may store pre-trained machine learning models, process history data, sensor setting values, etc.

[0058] The control unit may include a user interface for various inputs and outputs for data transmission and reception with sensors and module control. The user interface may acquire user input. The user interface may provide various information regarding user actions. The user interface may include an input interface and an output interface.

[0059] The input interface can convert sensory information received from a user into an electrical signal. The electrical signal can correspond to user input. User input may include various commands. The input interface can transmit the electrical signal (voltage or current) corresponding to the user input to at least one processor.

[0060] The input interface may include various input devices capable of converting tactile information into electrical signals. For example, the input interface may be provided as a physical button or a touch screen. The input interface may include a microphone capable of converting auditory information into electrical signals.

[0061] The output interface can output information related to user actions. The output interface can display information entered by the user or information provided to the user on various screens. The output interface can display information related to user actions as at least one of an image or text. For example, the output interface can output the interface of a sleep management application. In addition, the output interface can display a Graphic User Interface (GUI) that enables user control. That is, the output interface can display UI elements such as icons.

[0062] For example, the output interface may include a Liquid Crystal Display Panel (LCD Panel), a Light Emitting Diode Panel (LED Panel), an Organic Light Emitting Diode Panel (OLED Panel), or a Micro LED Panel. The output interface may include a touch display that also functions as an input device.

[0063] The output interface and the input interface may be provided as separate devices or as a single device (e.g., a touch display).

[0064] The control unit may include at least one machine learning model. Accordingly, the control unit may include at least one of a first machine learning model, a second machine learning model, a third machine learning model, a fourth machine learning model, a fifth machine learning model, a sixth machine learning model, a seventh machine learning model, an eighth machine learning model, a ninth machine learning model, and a tenth machine learning model.

[0066] FIG. 3 illustrates an example of a block diagram showing the data flow between a classification module and a control unit according to one embodiment.

[0067] Referring to FIG. 3, the classification module can perform a process of determining and classifying premium soybeans among a plurality of soybeans. The classification module includes at least one image sensor and at least one weight sensor, and can acquire image data and weight data of each soybean through these sensors.

[0068] The image sensor can provide image data for identifying the external characteristics of the soybean, and the first weight sensor can provide data containing information closely related to the weight of the soybean. Thus, the control unit can collect basic data necessary for quality classification of the soybean.

[0069] As described above, data acquired from each sensor can be transmitted to the control unit in real time. The control unit may include at least one machine learning model to process this data.

[0070] According to one embodiment, the control unit inputs acquired image data into a first machine learning model to calculate at least one of volume data, cross-sectional area data, and surface reflectance data for each soybean. The volume data indirectly indicates the size and density of the soybean, and the cross-sectional area data and surface reflectance data can be used to determine the appearance quality and condition of the soybean.

[0071] The first machine learning model can be trained to extract external data, such as volume, cross-sectional area, and surface reflectance, from image data (external data) of soybeans. The image data may consist of photographs of soybeans of various varieties and sizes taken from multiple angles. Each image may be labeled with external characteristics specific to the variety (e.g., surface smoothness, reflectance, color, shape, etc.). During training, actual measured values ​​of the soybean's external appearance (volume, cross-sectional area) are used as labels, allowing the model to learn the correlation between the image data and the actual values.

[0072] By taking image data as input and utilizing a deep learning model such as a Convolutional Neural Network (CNN), it can be trained to extract feature data such as the volume, cross-sectional area, and surface reflectance of soybeans. For example, the trained model can be designed to calculate the diameter, volume (assuming a spherical shape), and surface reflectance of a soybean when an image of the soybean taken from a specific angle is input.

[0073] The trained first machine learning model can receive image data collected in real-time from the actual process and produce data related to the appearance of soybeans.

[0074] The control unit can additionally input image data, weight data, volume data, cross-sectional area data, and surface reflectance data into a second machine learning model to calculate component data scores for each soybean. The component data score is an indicator that reflects the quality of internal components of the soybean, such as protein content, moisture content, and fat content.

[0075] These component data scores are calculated based on a dataset (comparison of component data between high-quality and low-quality soybeans) that the second machine learning model was trained on in advance.

[0076] The second machine learning model may have learned the correlation between external data of soybeans (output values ​​of the first model) and internal component data (weight, protein content, etc.). The training dataset may include internal component data, such as protein content, fat content, moisture content, and carbohydrate content, obtained through laboratory analysis for soybeans of various varieties and sizes. In addition, it may also include external data, such as the weight, volume, cross-sectional area, and surface reflectance of soybeans, which are the output data of the first machine learning model.

[0077] To train a second machine learning model, internal component data (protein content, moisture content, etc.) can be set as target labels, and external data (volume, cross-sectional area, surface reflectance, weight, etc.) can be set as inputs to train the machine learning model.

[0078] For example, it may have learned patterns between appearance data and component data using Random Forest or Gradient Boosting algorithms.

[0079] Accordingly, the second machine learning model learns that the input appearance data correlates with specific component content, enabling it to predict this even in new data. The trained second machine learning model receives appearance data (volume, weight, etc.) from the actual process and can predict the internal components of each soybean. By scoring these prediction results, soybeans exceeding a specific threshold can be classified as premium soybeans. For example, soybeans can be classified as premium based on the component data score corresponding to protein content exceeding a certain threshold.

[0080] In this embodiment, the machine learning model applied to the classification module may include a first machine learning model that analyzes soybean images to extract external characteristics, and a second machine learning model that combines the extracted external characteristics and internal component data obtained from laboratory analysis to calculate a final component data score.

[0081] First, the first machine learning model is designed to receive image data of soybeans and infer volume, cross-sectional area, surface reflectance, etc. This model may be a Convolutional Neural Network (CNN) trained with approximately 20,000 or more images of soybeans (under various angle and illumination conditions). For example, samples with measured soybean diameters of 7mm, 8mm, and 9.5mm are photographed at various angles from 0° to 180° and in an illumination range of 300 to 500 lux, and actual labels (diameter, reflectance, etc.) are assigned to each image. By training the CNN with this image dataset, the model recognizes the correlation between pixel information and physical size.

[0082] As a result of learning, the first machine learning model is able to infer the diameter, volume, and surface reflectance of soybeans within an error range of approximately ±0.3 mm when receiving an image taken during the actual process. For example, since the surface reflectance of soybeans is also learned based on labeling data, the first machine learning model detects differences in reflection intensity at an illuminance of 300 to 500 lux and calculates basic information regarding the appearance quality of soybeans.

[0083] Next, the second machine learning model receives appearance data (volume, reflectance, etc.) inferred by the first machine learning model and weight data obtained from a weight sensor (precision ±0.1g) as input, predicts internal components of the soybean such as protein content, fat content, and moisture content, and calculates a component data score.

[0084] To this end, the second machine learning model is trained based on approximately 10,000 or more pieces of laboratory analysis data and can utilize Random Forest or other regression-based algorithms. For example, the second machine learning model learns the correlation between appearance, weight, and composition on its own, such as predicting that the internal protein content is relatively high when "the soybean diameter is within a specific range, the reflectance is above a certain level, and the weight exceeds 15g." The resulting composition data score can be converted into a range of 0 to 100, and, for instance, it can be set to classify as "premium soybeans" if the score is 70 or higher.

[0085] In one embodiment, when more than 5,000 soybeans were tested in an actual process, it was confirmed that the classification accuracy of the second machine learning model reached over 92%. This is an improvement of about 20%p compared to the conventional method of considering only soybeans weighing 15g or more as "high quality," and the effect of the present invention is prominent in that it comprehensively evaluates the external appearance and internal components of the soybeans.

[0086] As a result, by having the first and second machine learning models cooperate through a step-by-step process of "soybean appearance data -> internal component prediction -> component data score," the limitations of existing weight-based classification methods can be overcome, and premium soybeans can be selected more accurately.

[0087] Furthermore, if new soybean varieties or exceptional cases occur during the actual process, the data can be additionally labeled to perform online learning or periodic retraining; thus, the classification module of the present invention has the potential to continuously improve classification accuracy over time. This enables a significant increase in the efficiency of the core process for producing whole soy milk with a high solid content.

[0088] The control unit can determine whether each soybean corresponds to a premium soybean based on the calculated component data score. To this end, the control unit compares the output data (component data score) of the machine learning model with a specific threshold, classifying soybeans above the threshold as premium soybeans and soybeans below the threshold as ordinary soybeans.

[0089] The data related to this classification result is transmitted back to the classification module, where the task of physically separating premium soybeans and regular soybeans can be performed. For example, the classification module can operate by moving the soybeans via a device such as a conveyor belt, discharging premium soybeans to one side and regular soybeans to the other.

[0090] In addition, if specific patterns or abnormal data are detected during the classification process, the control unit can store such data to be used as additional training data for the machine learning model or dynamically correct the classification criteria. Through this, the classification module and the control unit can continuously improve performance and increase classification accuracy.

[0091] A classification module configured in this manner analyzes comprehensive data on the external and internal quality of soybeans and, in cooperation with the control unit's machine learning model, can efficiently select high-quality premium soybeans. This process contributes to stably securing the quality of raw materials essential for manufacturing whole soy milk with excellent solid content.

[0092] In addition, actual process data obtained through the first and second machine learning models (appearance data and component analysis results acquired during mass processing) can be continuously added to the dataset to improve the accuracy of the models. If new soybean varieties or characteristics are discovered during the process, they can be added to the datasets of the first and second machine learning models for retraining; this allows for the continuous improvement of classification accuracy and adaptation to various environments and varieties.

[0093] The first machine learning model generates data based on the external characteristics of soybeans, while the second machine learning model calculates component scores by utilizing the correlation between external data and internal components. These two models precisely evaluate the quality of soybeans through pre-trained datasets and real-time collected data, serving as key factors for the production of high-quality whole soy milk.

[0095] FIG. 4 illustrates an example of a block diagram showing the data flow between an immersion module and a control unit according to one embodiment.

[0096] Referring to FIG. 4, the immersion module can perform a process of immersing premium soybeans classified from a plurality of soybeans into a liquid inside an immersion tank to induce an optimal expansion rate and activation of internal components. The immersion module may include at least one ultrasonic sensor and at least one temperature sensor to monitor the expansion rate of the soybeans and the temperature inside the immersion tank during the immersion process.

[0097] The ultrasonic sensor projects an ultrasonic signal onto the surface of the soybean and collects the signal reflected from the surface to provide the data necessary to calculate the expansion rate of the soybean in real time. The temperature sensor measures the liquid temperature inside the immersion tank and can provide related data.

[0098] For example, the ultrasonic sensor projects a signal onto the surface of the bean at a frequency of about 40 kHz, allowing it to track the increase in volume (expansion rate) of the bean with an error range of ±2% or less. The first temperature sensor measures the internal temperature of the soaking tank at a level of ±0.3°C, allowing it to detect sudden temperature changes in advance and transmit them to the control unit.

[0099] Data acquired from the ultrasonic sensor and the temperature sensor can be transmitted to the control unit in real time. The control unit includes at least one processor and memory, and may include a third machine learning model and a fourth machine learning model to optimize the immersion process.

[0100] The third machine learning model takes the component data scores of premium soybeans (e.g., protein content, moisture content, etc.) as input and can calculate the initial soaking conditions (soaking temperature and soaking time) and target expansion rate suitable for each soybean. This third machine learning model operates based on a pre-trained dataset (correlation between component data by soybean variety and optimal soaking conditions / target expansion rates) and can dynamically set soaking conditions according to the components and quality characteristics of the soybeans.

[0101] Once the initial soaking conditions are set, the control unit transmits the soaking temperature and soaking time to the soaking module to start the soaking process. During the soaking process, the expansion rate of the soybeans is continuously calculated based on the reflected signals collected by the ultrasonic sensor.

[0102] The control unit analyzes changes in ultrasonic signals to determine the real-time expansion status of soybeans and can input the related expansion rate data into the fourth machine learning model. Additionally, the control unit can input the initial immersion conditions (immersion temperature and immersion time) and the target expansion rate, which are output data of the third machine learning model, into the fourth machine learning model. The fourth machine learning model analyzes the expansion rate data and evaluates the difference between the target expansion rate and the current expansion rate, and can correct the immersion temperature and immersion time in real time if necessary. For example, if the expansion rate falls short of the target value, the correction process can be performed by extending the immersion time or slightly increasing the temperature.

[0103] The third machine learning model can receive soybean component data scores as input before starting the soaking process and calculate the appropriate initial soaking temperature and soaking time for each soybean. To achieve this, the results of soaking experiments conducted in advance on various soybean varieties can be incorporated into the pre-training.

[0104] For example, it may have been trained on a dataset of thousands of cases containing correlations such as "soybeans with a protein content of 40% or more expand most uniformly when soaked at 15°C for 8 hours."

[0105] During the learning process, soybean protein and moisture content, variety characteristics, and past optimal soaking conditions are provided in a labeled state, enabling the third machine learning model to recognize patterns between these component data scores and optimal temperature and time. Consequently, when component score data for premium soybeans selected by the classification module in the actual process is received, the third model immediately determines the initial soaking temperature and time and transmits them to the soaking module.

[0106] The fourth machine learning model can dynamically correct immersion conditions by identifying the deviation between the expansion rate measured in real-time and the target expansion rate after immersion has begun in earnest. The control unit can calculate the current expansion rate within an error range of ±2% or less by projecting ultrasound onto the surface of the soybeans using an ultrasonic sensor and analyzing the reflected signal.

[0107] For example, the fourth machine learning model can combine the trend of change in this expansion rate and the first temperature data for the immersion solution obtained from the first temperature sensor, and suggest to the control unit to perform corrections such as raising the temperature by about 1°C or shortening the immersion time by 10 minutes if the target value is not reached or is excessively exceeded.

[0108] The fourth machine learning model can utilize the results learned from multiple prior immersion experiments to determine how much temperature or time must be varied in the current situation to efficiently reach the target expansion rate. To this end, "expansion rate(t+ A method of learning a function such as t) = f(current expansion rate, temperature, elapsed time, soybean components, etc.) can be used, and prediction accuracy is improved by utilizing actual process data and simulation results together.

[0109] In other words, the entire soaking process can be operated flexibly by the third machine learning model setting initial soaking conditions based on internal component data of soybeans and the fourth machine learning model tracking and correcting deviations in real-time expansion rate data during soaking.

[0110] For example, when the goal is to increase the volume of soybeans by more than 30%, if the expansion rate is 5% lower than expected during the process, the sluggish expansion can be compensated for by slightly increasing the temperature or extending the soaking time. Conversely, if the expansion is already at a rate that seems likely to exceed the target, it is recommended to lower the soaking temperature or reduce the time to prevent excessive deterioration and loss of quality.

[0111] This entire process can be automated and integratedly controlled by a control unit, enabling the realization of a uniform and optimized expansion state that was difficult to achieve with conventional fixed soaking times and temperatures alone, and consequently, stably contributing to the production of whole soy milk with a high solid content.

[0112] When the soaking process is completed, the control unit can analyze the final data received from the ultrasonic sensor and temperature sensor to confirm that the soybeans have reached the target expansion state. Once the target state is confirmed, the soaking module can terminate the soaking process.

[0113] In addition, the control unit can control the soaking module to automatically transfer the soaked soybeans to the grinding module. Furthermore, the control unit can store data such as the expansion rate, soaking time, and soaking temperature recorded during the soaking process so that it can be utilized in subsequent processes (e.g., grinding, concentration).

[0114] The cooperation between the soaking module and the control unit can contribute to minimizing quality variations in soybeans and ensuring uniform performance. In particular, by analyzing the real-time status of the soybeans during the soaking process and dynamically adjusting soaking conditions, the soybeans can be fed into the next process in a state where they are properly expanded and their component activation is optimized. This provides an important foundation for ultimately producing high-quality whole soy milk.

[0116] FIG. 5 illustrates an example of a block diagram showing the data flow between a grinding module and a control unit according to one embodiment.

[0117] Referring to FIG. 5, the grinding module can perform a process of grinding premium soybeans that have finished the soaking process together with water to form a soybean mixture. The grinding module may include at least one water feeder, at least one weight sensor, and at least one optical sensor.

[0118] The water feeder is a device for supplying water to the inside of the grinder. The water feeder can adjust the amount of water to be mixed with the soaked premium soybeans based on a signal from the control unit. In other words, the control unit can determine the amount of water supplied from the water feeder based on the weight of the soaked premium soybeans. Here, the amount of water may refer to the volume or weight of the water.

[0119] The weight sensor can acquire weight data of the immersed premium soybeans and transmit it to the control unit. This can serve as the basis for determining the amount of water supplied by calculating the water supply ratio relative to the premium soybeans.

[0120] The first optical sensor can acquire first average particle size data of the premium soybean mixture in real time during the grinding process and transmit it to the control unit. Through this, the control unit can determine how much the grinding process is progressing.

[0121] Data acquired from each sensor can be transmitted to the control unit in real time. The control unit includes at least one processor and may include a fifth machine learning model to comprehensively analyze weight data and average particle size data.

[0122] The fifth machine learning model may be pre-trained to predict optimal grinding conditions based on various sensor data and target particle size information when grinding premium soybeans soaked in water during the grinding process. Based on the pre-trained dataset, the fifth machine learning model can be designed to calculate the optimal water supply amount when premium soybeans are ground at a specific volume-to-mass ratio. To this end, the control unit can receive weight data of the soaked soybeans and target particle size information as input to dynamically determine the volume of water to be supplied to the water dispenser.

[0123] The fifth machine learning model learns the water supply ratio of premium soybeans and the pattern of particle size change of the soybean mixture based on a pre-acquired dataset, and can provide optimal control values ​​to the control unit in the grinding process.

[0124] For example, a pre-collected dataset may include changes in average particle size, grinding time, and final texture characteristics when soybeans of various varieties and weights are ground at different water ratios.

[0125] By training this dataset with various machine learning algorithms (e.g., Random Forest, Deep Learning, Support Vector Machine, etc.), the fifth machine learning model recognizes correlations such as "optimal water supply amount and grinding time for the grinding particle size to reach a target range when the soybean weight is in a specific range."

[0126] When the weight of premium soybeans soaked in the actual grinding process is transmitted from the sensor and a reference value for the target particle size is set, the machine learning model can predict "how much water supply is required for soybeans within that weight range to reach a certain particle size" based on pre-trained results.

[0127] Furthermore, by analyzing changes in particle size of the soybean mixture collected in real time from the optical sensor, the control unit can determine to adjust the water supply amount or dynamically correct the grinding time if grinding proceeds faster or slower than expected.

[0128] For example, if grinding occurs faster than the target, the machine learning model may decide to lower the grinding speed or vary the water ratio within a certain range to prevent loss of texture.

[0129] As such, since the fifth machine learning model operates based on a pre-trained dataset, it can calculate the most suitable grinding conditions by comprehensively considering the physical characteristics of the soybean mixture (weight, particle size, grinding speed, etc.), even in situations where complex correlations between process variables are difficult to control easily using human experience or simple rules.

[0130] By feeding back data accumulated in the actual process over time into the model, such prediction and control can be expected to continuously improve the accuracy and efficiency of the grinding process.

[0131] In addition, the fifth machine learning model can pre-learn data related to grinding intensity (such as the rotational speed of the grinder blades or the stirring speed) in addition to the amount of water supplied under grinding conditions, and provide it as output data.

[0132] For example, preliminary experiments accumulate results such as the average particle size being 500 μm when grinding 10 kg of soybeans with 8 liters of water at a speed of 1,000 rpm per minute, and reaching the target value of 300 μm after a certain period of time; as such data accumulates, the model becomes more sophisticated in identifying the correlation between grinding time, water ratio, rotation speed, and final particle size. The model can be designed to derive an output (amount of additional water supply or grinding intensity adjustment amount) for external inputs (weight of soaked soybeans, already achieved particle size, grinding intensity) by utilizing one of machine learning algorithms such as Random Forest, Deep Learning, or Support Vector Machine. At this time, the model also learns experimental causal relationships such as "if the weight of soaked soybeans is excessive and the water supply is insufficient, the grinding time becomes longer" or "if the grinding intensity is too high, the product temperature rises and the texture deteriorates," and dynamically derives control values ​​by observing the trend of average particle size change coming from the sensor in real time during the grinding process, lowering the grinding intensity or adjusting the water supply if the grinding is faster than the target, and increasing the grinding intensity if it is slower than the target.

[0133] Subsequently, when the average particle size data transmitted from the first optical sensor reaches a predetermined particle size reference value, the control unit can send a command to the grinding module to stop the operation of the grinder or reduce the speed.

[0134] In this process, the fifth machine learning model further analyzes trends in particle size changes and can dynamically adjust the grinding time or speed if grinding occurs faster than expected. When the grinder stops, the soybean mixture produced within the grinding module can be transferred to a subsequent process or stored.

[0135] In addition, the control unit records weight data, water supply ratios, and changes in average particle size collected during the grinding stage, and can transmit these records to higher-level modules (e.g., classification module, immersion module) or subsequent modules (e.g., temperature-changing module, concentration module). For instance, the classification module can refer to particle distribution information after grinding to identify soybean characteristics, while the concentration module can improve processing efficiency by subdividing concentration conditions based on the ground particle size. Through this, real-time data is shared throughout the process, and mutual complementarity between machine learning models can be achieved, thereby improving the quality of the final whole soy milk.

[0137] FIG. 6 illustrates an example of a block diagram showing the data flow between a temperature change module and a control unit according to one embodiment.

[0138] Referring to FIG. 6, the temperature change module may include a heater for raising the internal temperature of the reaction vessel, a cooler for lowering the internal temperature of the reaction vessel, at least one temperature sensor for acquiring temperature data of the soybean mixture, at least one viscosity sensor for acquiring viscosity data of the soybean mixture, and at least one pH sensor for acquiring pH data of the soybean mixture.

[0139] The temperature-changing module can adjust the viscosity and temperature of the soybean mixture to a target range by introducing the soybean mixture transferred from the grinding process into a reaction vessel and performing heating and cooling processes. During the heating step, heat can be supplied through a heater installed inside or outside the reaction vessel. Additionally, a second temperature sensor installed inside the reaction vessel can acquire real-time temperature data of the soybean mixture and transmit it to the control unit.

[0140] The control unit can analyze this second temperature data to dynamically control the heating temperature and heating time. In addition, it can monitor the first viscosity data transmitted from the viscosity sensor to prevent excessive thermal denaturation caused by an excessively long heating time.

[0141] The control unit can determine the heating temperature and heating time to operate the heater based on the second temperature data and the first viscosity data. The control unit can acquire these data from each sensor in real time and correct the heating temperature and heating time in real time.

[0142] The control unit can control the heater to stop based on the fact that the first viscosity data has reached a predetermined viscosity (target viscosity threshold).

[0143] The 6th machine learning model may be a pre-trained model designed to predict and control changes in temperature, viscosity, and pH of the soybean mixture during the heating phase. To this end, data is collected through various experiments regarding how viscosity increases and at what point protein denaturation occurs when the soybean mixture is heated to a specific temperature range.

[0144] For example, the soybean mixture is heated for a set period of time in several intervals between 60°C and 90°C, and the viscosity, pH fluctuations, and the point at which fat components separate are accurately recorded. Then, each scenario is labeled to allow the model to learn. The 6th machine learning model receives hundreds to thousands of such accumulated samples as input and is trained to predict what level of viscosity will be reached when a specific soybean mixture is heated at 80°C for 10 minutes, and how many additional degrees or minutes are needed to reach the target viscosity.

[0145] When the 6th machine learning model recognizes this relationship, it can immediately calculate the necessary correction amount (temperature rise / fall range, heating time reduction / extension) and transmit it to the control unit even if the temperature in the actual process is higher or lower than predicted.

[0146] Thus, the 6th machine learning model, based on a pre-trained dataset, recognizes the viscosity increase pattern of the soybean mixture within a specific temperature range and can predict the heating temperature and time required to reach the target viscosity. As these prediction results are calculated and reflected within the control unit, the heating process can proceed with minimal protein denaturation or separation of fat components.

[0147] During the heating process, the control unit inputs the first viscosity data received in real-time from the first viscosity sensor into the sixth machine learning model to determine the extent to which the soybean mixture approaches the target viscosity range. For example, if the viscosity rises faster than expected, the control unit can prevent excessive denaturation by immediately lowering the heating temperature or shortening the heating time. On the other hand, if it is determined that there is insufficient heat supply to raise the viscosity, the control unit can transmit a signal to the temperature control module instructing it to increase the heating temperature or extend the heating time.

[0148] In this way, by the control unit continuously comparing and analyzing the model's predictions with actual sensor data, deviations and errors in the heating stage can be reduced, and homogeneous product quality can be secured.

[0149] After heating is complete, the variable temperature module can perform a cooling process. The cooler gradually lowers the internal temperature of the reaction vessel to stabilize the soybean mixture, and the control unit can set an appropriate cooling temperature and cooling time by combining data obtained from the first viscosity sensor and the pH sensor.

[0150] The control unit can determine the cooling temperature and cooling time using the seventh machine learning model based on the first viscosity data and pH data. The control unit can acquire these data from each sensor in real time and correct the cooling temperature and cooling time in real time.

[0151] The control unit can control the cooler to stop based on the second temperature data reaching a predetermined temperature (target temperature threshold).

[0152] The seventh machine learning model can propose an optimal cooling curve to the control unit by utilizing the pre-learned results on how the viscosity and pH of the soybean mixture change as the temperature decreases.

[0153] The seventh machine learning model can learn in advance the changes in viscosity and pH according to the temperature decrease during the cooling phase. The seventh machine learning model may have been created by collecting data through multiple experiments and simulations regarding at what point protein precipitation or microbial risk increases, and the curve following which pH decreases, when soybean mixtures of each variety are cooled from high temperatures at a specific rate, and then labeling this data to enable the model to learn.

[0154] For example, the 7th machine learning model recognizes how viscosity changes when cooled from 90°C to 50°C at a constant rate, and how pH drops below a critical value when the cooling rate is excessively fast or slow. As a result, when real-time data is received from the temperature, viscosity, and pH sensors of the soybean mixture in the actual process, the cooling method (cooler operation time, cooling rate) required to reach the target temperature, viscosity, and pH can be immediately calculated and provided to the control unit.

[0155] In other words, the calculations of the seventh machine learning model can help systematically adjust cooling times to prevent the soybean mixture from coagulating rapidly due to thermal shock or remaining in the microbial risk zone for a long time.

[0156] During the cooling process, the control unit compares the prediction result received from the 7th machine learning model with the actual measurement data received from the temperature, viscosity, and pH sensors, and can correct the operation of the cooler immediately if an error occurs in the set cooling temperature.

[0157] For example, if the pH of the soybean mixture drops below a certain range, the risk of protein precipitation may increase, so the cooling speed may be slowed down or the cooling time shortened. Since the interaction between the control unit, the 6th and 7th machine learning models, and each sensor takes place in real time, the soybean mixture with appropriately controlled viscosity and stabilized is finally transferred to the subsequent process, the concentration stage.

[0158] The results of the variable temperature module's operation directly affect the final quality of the whole soy milk. In particular, if protein and fat components are properly denatured during the heating stage, the homogenization process becomes easier, and if viscosity and pH are maintained uniformly during the cooling stage, sedimentation or off-flavors can be prevented during the concentration and homogenization processes. The control unit stores historical data on temperature, viscosity, and pH collected from the variable temperature module and, when necessary, feeds this data back to preceding stages such as the classification module or grinding module, thereby supporting the organic optimization of the entire process.

[0160] FIG. 7 illustrates an example of a block diagram showing the data flow between a concentration module and a control unit according to one embodiment.

[0161] Referring to FIG. 7, the concentration module may include a vacuum concentrator, at least one temperature sensor installed inside the vacuum concentrator to acquire data on the temperature of the soybean mixture, at least one viscosity sensor installed inside the vacuum concentrator to acquire data on the viscosity of the soybean mixture, at least one pressure sensor installed inside the vacuum concentrator to acquire data on the internal pressure of the vacuum concentrator, a refractometer installed inside the vacuum concentrator to acquire data on the solid content concentration of the soybean mixture, and a stirrer that mixes the liquefied soybean mixture inside the vacuum concentrator.

[0162] The concentration module can play a key role in increasing the solid content concentration of whole soy milk by concentrating the soybean mixture, upon completion of the temperature-variable process, under vacuum conditions. A third temperature sensor, a second viscosity sensor, a first pressure sensor, and a refractometer may be installed inside the vacuum concentrator. The third temperature sensor can acquire third temperature data of the soybean mixture, the second viscosity sensor can acquire second temperature data of the soybean mixture, the first pressure sensor can acquire first pressure data indicating the internal pressure of the vacuum concentrator, and the refractometer can acquire solid content concentration data of the soybean mixture in real time. This data can be transmitted to the control unit.

[0163] The control unit inputs this data into the eighth machine learning model to dynamically correct the internal pressure and concentration time of the vacuum concentrator. For example, the eighth machine learning model comprehensively analyzes the viscosity increase pattern and changes in solid content concentration of the soybean mixture, and can instruct the pressure to be relieved if the concentration rises faster than expected, or conversely, to maintain a stronger vacuum if the concentration does not rise well.

[0164] The 8th machine learning model may be a pre-trained model designed to predict changes in temperature, viscosity, internal pressure, and solid content concentration of the soybean mixture during the concentration process, and to dynamically control the pressure and concentration time of the vacuum concentrator based on these predictions. Prior to the operation of the 8th machine learning model, viscosity rise curves, the time at which solid content concentration is reached, and heat and mass transfer characteristics of the mixture when stirring intensity is changed can be systematically collected through actual experiments and simulations when the soybean mixture is concentrated under various temperature and pressure conditions.

[0165] For example, when the temperature is varied in the range of 50°C to 70°C while maintaining the soybean mixture at 0.1 bar or less, specific data can be collected, such as the time it takes to reach specific standards for solid content concentrations of 8%, 10%, 12%, 15%, etc., how the viscosity increases during the process, and how much the protein aggregates.

[0166] These preliminary datasets can be accumulated for each scenario in a labeled form, including 'what the initial pressure was,' 'how much the stirring speed or intensity was adjusted in which range,' and 'at what point the solid content concentration and viscosity consequently entered the target range.'

[0167] The 8th machine learning model can use this label to learn, via regression or time series analysis algorithms, how quickly the solid content concentration rises under specific ranges of pressure, temperature, and viscosity conditions, and at what level the mixture exhibits bubbling (excessive foaming) or aggregation.

[0168] Tree-based models such as Random Forest or XGBoost can be used as learning algorithms, and time series models such as Multilayer Neural Networks (MLP) or LSTM / RNN are applied to simultaneously process multiple variables such as temperature, pressure, viscosity, stirring speed, and solid content concentration.

[0169] The learning process can be designed to specifically consider both 'how much the time to reach the target solid content concentration can be reduced' and 'the impact of rapid pressure release or increased stirring intensity on product quality (color, aroma, viscosity).' For example, in one experiment, increasing stirring intensity raised the concentration rate but resulted in an excessively high final viscosity, causing micro-protein aggregation; conversely, rapidly lowering the pressure led to excessive bubble generation, resulting in a deterioration of the mixture's quality.

[0170] The eighth machine learning model comprehensively analyzes these conflicting factors to identify rules such as, "Reducing internal pressure below a certain speed enables sufficiently rapid concentration while minimizing bubble generation," and can store these as pre-learned parameters. Through this learning process, it is also possible to output predictive guidelines to the control unit to gradually adjust stirring speed or pressure when the solid content concentration is about to reach a target value.

[0171] Finally, the 8th machine learning model can calculate and transmit to the control unit, based on prior acquired correlations and patterns, "how much more pressure needs to be lowered or raised," "whether stirring intensity needs to be adjusted now," and "how many degrees the temperature can be raised to shorten the time to reach the target concentration," based on the inputs of temperature, pressure, viscosity, and solid content concentration transmitted in real time by each sensor during the concentration process.

[0172] For example, if the eighth machine learning model determines that the concentration is rising faster than expected, it may instruct the stirring intensity to be reduced or the internal pressure to be gradually returned to atmospheric pressure to prevent excessive evaporation or quality degradation. On the other hand, if the concentration does not increase, the control unit may operate by commanding a stronger vacuum state to be maintained to increase concentration efficiency.

[0173] The eighth machine learning model can periodically retrain itself by accumulating this data during the process and additionally labeling exceptional situations (special varieties, unexpected temperature changes, etc.), so the prediction accuracy and stability of the concentration process can be improved over time.

[0174] In other words, during the concentration process, the control unit can initiate a procedure to gradually return the internal pressure to atmospheric pressure when the solid content concentration transmitted from the refractometer approaches a preset target value. At this time, the control unit can analyze the viscosity sensor data together to finely adjust the stirring intensity or temperature to prevent the internal mixture from boiling or generating excessive bubbles due to sudden pressure changes. In particular, to prevent the phenomenon of proteins or dietary fibers clumping when the solid content reaches a high concentration, the eighth machine learning model can predict changes in the viscosity of the liquefied soybean mixture and control the pressure and temperature in detail accordingly.

[0175] If a stirrer is included inside the vacuum concentrator, the control unit can analyze viscosity data and adjust the stirring intensity in steps. For example, if the viscosity exceeds a predetermined threshold, the stirring intensity can be increased to ensure the uniformity of the mixture and facilitate heat and mass transfer.

[0176] On the other hand, by relaxing stirring just before the solid content concentration reaches the target value, losses or product discoloration caused by excessive vortexing can be prevented. The various data collected from the concentration module (temperature, viscosity, pressure, and solid content concentration) are also used as important data for future process analysis and quality improvement.

[0177] When the concentration process is completed and the solid content of the whole soy milk becomes sufficiently high, the control unit can stop the operation of the vacuum concentrator and gradually restore the interior to atmospheric pressure. At this time, the stirrer can be operated appropriately to stabilize the viscosity, and the concentrated soybean mixture can be finally transferred to the homogenization module. The high-concentration soybean mixture produced in the concentration module enables more efficient energy use and fine dispersion in the homogenization process, and as a result, the richness and flavor of the whole soy milk can be maximized.

[0179] FIG. 8 illustrates an example of a block diagram showing the data flow between a homogenization module and a control unit according to one embodiment.

[0180] Referring to FIG. 8, the homogenization module may include a high-pressure homogenizer, at least one temperature sensor installed inside the high-pressure homogenizer to acquire temperature data of the soybean mixture, at least one viscosity sensor installed inside the high-pressure homogenizer to acquire viscosity data of the soybean mixture, and at least one optical sensor installed inside the high-pressure homogenizer to acquire average particle size data of the soybean mixture.

[0181] The homogenization module can perform the final process that determines the texture and stability of whole soy milk by applying high pressure to the soybean mixture that has undergone the concentration step and finely dispersing the particles.

[0182] The high-pressure homogenizer may include a fourth temperature sensor, a third viscosity sensor, and a second optical sensor. The fourth temperature sensor can acquire fourth temperature data of the soybean mixture. The third viscosity sensor can acquire third viscosity data of the soybean mixture. The second optical sensor can acquire second average particle size data of the soybean mixture. Accordingly, the homogenization module can measure the temperature, viscosity, and average particle size of the soybean mixture in real time during the homogenization process.

[0183] The control unit can input this data into the ninth machine learning model to calculate or correct the homogenization pressure, homogenization flow rate, and number of iterations. These calculations may be performed based on a dataset that has been comprehensively learned from various varieties of beans, concentration levels, crushed particle sizes, and heating and cooling variables accumulated in advance.

[0184] The ninth machine learning model is a pre-trained model that predicts changes in temperature, viscosity, and average particle size of a soybean mixture during the homogenization process and determines the homogenization pressure, homogenization flow rate, and number of iterations based on this information. The ninth machine learning model may have learned prior data regarding how quickly particles are micronized under various pressure and flow rate conditions, and how the temperature and viscosity fluctuate accordingly, for a soybean mixture that has already been concentrated prior to homogenization.

[0185] For example, when a soybean mixture is subjected to high-pressure processing with varying pressures ranging from 100 to 200 MPa, the time it takes for the average particle size to drop below a specific threshold and the amount of heat accumulated during this process, which causes viscosity to rise, are systematically recorded and labeled. At this time, the ninth machine learning model is also trained in advance to recognize trade-offs, such as the fact that while a mixture with a high protein content easily becomes more fluid as pressure increases, the temperature also rises, increasing the risk of denaturation.

[0186] The input variables of the ninth machine learning model include the temperature, viscosity, and average particle size detected by an optical sensor inside the high-pressure homogenizer, and can also make a judgment such as "if the ground particles are already sufficiently small, too many repeated homogenizations are unnecessary" by referring to data accumulated from previous processes (grinding, concentration). The ninth machine learning model uses one of a random forest, a regression-based neural network, or a time series analysis algorithm to predict the most efficient pressure, flow rate, and number of iterations to reach the target value based on the trends of temperature, viscosity, and particle size changes collected at each high-pressure homogenization point.

[0187] For example, if the ninth machine learning model detects that "the current average particle size is 10% larger than the target but the temperature is rising excessively," it transmits a calculated value to the control unit to suppress the rise in temperature while approaching the target particle size by maintaining the pressure as is but slightly reducing the flow rate or adding the number of iterations.

[0188] In this pre-training process, the ninth machine learning model learns all kinds of scenarios, such as soybean mixtures with different mixing ratios, various varieties with different protein contents, and conditions with high concentration levels. For example, the model will be able to recognize patterns such as "for soybean mixtures with a concentration of 12% or more, viscosity increases rapidly when the pressure exceeds 150 MPa, so the homogenization time must be shortened" or "when the temperature exceeds 60°C, particles are rapidly fined, but there is a risk that the taste may change."

[0189] In addition, when the ninth machine learning model receives real-time information from the fourth temperature sensor, the third viscosity sensor, and the second optical sensor during the actual process, it can compare this with pre-learned correlations to suggest an optimal solution such as, "If the pressure is increased slightly, the particle size will reach the target level faster, but since there is a concern about denaturation due to the increase in viscosity, increase the number of iterations while maintaining the pressure constant."

[0190] In other words, the ninth machine learning model comprehensively considers factors such as temperature, viscosity, and additional effects caused by stirring friction during the high-pressure homogenization process to calculate flexible control values ​​that achieve target particle size and viscosity without excessive energy input. If the ground particle size is larger than expected, it recommends slightly higher pressure and an appropriate flow rate initially; if the particle size is decreasing rapidly, it gradually lowers the pressure to suppress product degradation and temperature rise. As a result, the final whole soy milk maintains a smooth texture while achieving a stable dispersion state.

[0191] As a result, the control unit can determine the homogenization pressure, homogenization flow rate, and number of repetitions by inputting the fourth temperature data, the third viscosity data, and the second average particle size data into the ninth machine learning model. In addition, whenever homogenization is repeated, the data obtained from the previous operation is input again into the ninth machine learning model to correct the homogenization pressure, homogenization flow rate, and number of repetitions in real time.

[0192] In addition, the control unit can control the homogenization module to stop the high-pressure homogenizer based on the second average particle size data reaching a predetermined particle size. In addition, the control unit can control the homogenization module to stop the high-pressure homogenizer based on the third viscosity data reaching a predetermined viscosity.

[0193] In this way, during the homogenization process, the control unit can use the ninth machine learning model to simultaneously check the heat accumulation state transmitted by the temperature sensor and the fluidity data transmitted by the viscosity sensor, thereby preventing excessive temperature rise or particle rearrangement.

[0194] When the second average particle size data transmitted from the second optical sensor falls below a specific threshold, the control unit can send a command to the homogenization module to stop the homogenizer operation or lower the pressure. As a result, the whole soy milk obtains the target soft texture and uniform quality without unnecessary energy consumption. If the viscosity unexpectedly increases significantly during the homogenization process, the control unit minimizes product degradation by controlling the temperature while adjusting the flow rate or increasing the number of homogenization cycles.

[0195] The final viscosity and particle distribution data for whole soy milk produced by the homogenization module are stored within the control unit and can be used to strengthen linkages with preceding stages, such as the classification module and the immersion module. For example, if the difference in texture or taste based on particle size detected by the optical sensor shows a certain pattern, the immersion time or concentration time can be readjusted to improve the overall process more efficiently. As the control unit continuously updates the machine learning model based on this data, product quality gradually improves with each batch repetition.

[0196] Other products may be produced by applying the smart whole soy milk manufacturing system according to the present invention or by adding additional modules thereto. For example, the manufacturing system can be designed so that after concentration is performed in the concentration module to a solid content concentration higher than that required for manufacturing whole soy milk, a homogenization process is performed in the homogenization module, and then a drying process is carried out to produce soy milk powder.

[0197] Therefore, to design a system that incorporates an optimized process for manufacturing soy milk powder, additional pre-training may be required for the 8th and 9th machine learning models. This is because further moisture must be removed from the liquid soybean mixture to produce soy milk powder. Since removing more moisture from the soybean mixture for whole soy milk production generally leads to an increase in viscosity and solid content, pre-training based on these factors is necessary to fully achieve the optimized process according to the machine learning models.

[0199] FIG. 9 illustrates an example of a block diagram showing the data flow between a drying module and a control unit according to one embodiment.

[0200] The drying module can perform drying processes. Drying processes include spray drying, freeze drying, and drum drying.

[0201] Spray drying is a method in which whole soy milk is sprayed in the form of a fine mist and exposed to hot air to instantaneously evaporate the moisture. It is suitable for mass production, offers a fast drying speed, and allows for easy control of powder particle size uniformly. Since it is carried out at high temperatures, some heat-sensitive components (such as vitamins) may be lost, but it is the most commonly used process in industrial settings.

[0202] Freeze-drying is a method in which whole soybeans are first frozen to sub-zero temperatures, and the frozen water is removed by sublimation under a vacuum. While it minimizes the loss of nutrients and flavor, it entails high equipment costs and long processing times. It is suitable for premium powdered products or when maximizing the preservation of heat-sensitive ingredients.

[0203] The drum drying method involves applying whole soy milk in the form of a thin film to the surface of a rotating hot drum and drying it by rapidly evaporating the moisture. While this method can be achieved with relatively simple equipment, it is difficult to form powder particles evenly, and there is a risk of significant nutrient destruction due to heat.

[0204] Since an important effect among the goals to be achieved by the present invention is to minimize the destruction of nutrients, an example of a system for manufacturing soy milk powder by applying a freeze-drying method will be described below.

[0206] Referring to FIG. 9, the freeze-drying module may include a chamber for holding a liquid soybean mixture, a rapid freezer for lowering the temperature inside the chamber to change the soybean mixture into a solid state, a vacuum pump for lowering the pressure inside the chamber to near a vacuum, at least one temperature sensor for measuring the temperature inside the chamber, at least one pressure sensor for measuring the pressure inside the chamber, and at least one moisture sensor (e.g., infrared humidity sensor, NIR sensor, refractometer, etc.) capable of measuring the moisture content of the soybean mixture.

[0207] The fifth temperature sensor can acquire fifth temperature data inside the chamber and transmit it to the control unit. The sixth temperature sensor can acquire sixth temperature data representing the surface temperature of the solidified soybean mixture and transmit it to the control unit. The second pressure sensor can measure the internal pressure (vacuum degree) inside the chamber and transmit second pressure data to the control unit. The moisture sensor is used to indirectly determine how much moisture remains in the soybean mixture while it is frozen; for example, the moisture sensor can calculate changes in moisture content from reflection or transmission when infrared light is projected onto the sample surface, or measure changes in conductivity and transmit moisture data to the control unit.

[0208] The control unit periodically receives the 5th temperature data, 6th temperature data, 2nd pressure data, and moisture data received from each sensor, and then inputs them into a pre-trained 10th machine learning model to analyze the current freeze-drying situation and predict the sublimation rate.

[0209] The 10th machine learning model may be trained on a large dataset accumulated in advance regarding freeze-drying. For example, it may include experimental results in which different samples (or different batches of the same sample) were frozen in various temperature ranges, such as -40°C and -50°C, and then sublimation was carried out while varying the vacuum to 0.001 bar, 0.01 bar, etc. In each experiment, various labels may be attached, such as how much ice was removed from the sample when a specific temperature and pressure combination was maintained for a few minutes or hours, how cracks or tissue damage occurred on the sample surface, or what the microbial stability was.

[0210] The 10th machine learning model can learn by self-deriving correlations through prior learning of such data, such as “if the temperature is raised rapidly at excessively low pressure, only the surface of the sample dries first and ice may still remain inside” or “if the freezing temperature is not sufficiently lowered below -40 degrees, the ice crystals are not homogeneous even though it appears frozen, so the sublimation process takes a long time.”

[0211] In the freeze-drying process, when the fifth temperature data representing the internal temperature of the chamber, the second pressure data, the sixth temperature data representing the surface temperature of the soybean mixture, and the moisture data are input into the tenth machine learning model, the control unit can determine whether the current sample is sublimating slower than expected or if there is a possibility of overheating or tissue damage occurring due to it proceeding too fast by comparing it with rules acquired in advance.

[0212] In this way, the 10th machine learning model can output correction values ​​for the chamber internal pressure, internal temperature, and drying time. Since data measured from each sensor is input into the 10th machine learning model periodically, correction values ​​are output periodically, thereby enabling the control unit to control the drying module to continuously optimize the drying process.

[0213] For example, if the internal pressure is lowered excessively, the drying speed may be temporarily accelerated, but side effects may occur, such as the rapid sublimation of only the sample surface leading to microcracks or insufficient removal of internal ice, resulting in reduced quality.

[0214] In response, the 10th machine learning model combines the freeze-drying experiment data obtained so far with target values ​​(e.g., final moisture content of 2% or less, processing time within 5 hours, etc.) to determine what level of vacuum and temperature conditions can achieve uniform sublimation. If it receives a signal that the sample surface temperature (6th temperature data) has risen excessively above the target level, the 10th machine learning model may instruct the vacuum pump to operate with a slightly reduced intensity and the refrigerator to lower its operating temperature further in order to prevent a situation where only the surface dries quickly and internal ice remains. By doing so, sublimation does not occur too rapidly, which improves the balance between surface and internal temperature and moisture movement, and reduces the risk of bubbles or cracks.

[0215] On the other hand, if the rate of change in moisture data (drying progress rate) is already over 80% and the internal ice has almost disappeared, the 10th machine learning model may decide to secure a little more time at the end to dry the entire product uniformly. Ultimately, the 10th machine learning model receives temperature, pressure, and moisture-related data transmitted from each sensor in real time, predicts the sublimation rate, and dynamically calculates pressure and temperature scenarios, and the control unit adjusts the vacuum pump or refrigerator according to the scenario to ultimately maintain a uniform and high-quality freeze-drying state.

[0216] That is, the control unit can input at least one of the fifth temperature data, sixth temperature data, second pressure data, and moisture data acquired in real time into the tenth machine learning model to correct the internal pressure, internal temperature, and drying time (freeze-drying scenario) of the vacuum pump in real time. The control unit can control the drying module to stop the operation of the vacuum pump based on the moisture data reaching a predetermined value.

[0218] [Example]

[0219] 1) Raw Soybean and Classification Process

[0220] In one embodiment, 500 kg of soybeans with a protein content of about 40% is used as a raw material.

[0221] The appearance and weight of each soybean are measured using an image sensor (resolution 2,048X1,536) and a weight sensor (precision ±0.1g) in the classification module.

[0222] The first machine learning model (CNN-based) infers the diameter, surface reflectance, and cross-sectional area of ​​the soybean, and the second machine learning model utilizes pre-trained appearance-component correlations (more than 10,000 experimental data points) to calculate the component data score of the soybean (predicted protein value, etc.) in the range of 0 to 100.

[0223] If the score is 70 points or higher, it is determined to be "premium soybeans". Under the conditions of the example, about 400 kg out of 500 kg is classified as premium soybeans (classification accuracy about 92%).

[0224] 2) Immersion process

[0225] 400 kg of premium soybeans selected from the sorting module are added to the liquid inside the immersion tank (maintained at a temperature of 15 ± 0.5°C).

[0226] The immersion module is equipped with an ultrasonic sensor (40 kHz frequency, expansion rate ±2% error) and a temperature sensor (precision ±0.3 degrees), and transmits expansion rate and temperature data at 1-minute intervals.

[0227] The third machine learning model sets the initial soaking conditions (temperature 15°C, time 10 hours) based on each soybean variety and component score, and the fourth machine learning model determines whether the target expansion rate (30% volume increase) is reached relative to the real-time expansion rate and adjusts the soaking time in 5 to 15 minute increments.

[0228] As a result of the actual test, the ultrasonic sensor reported that it took 9 hours and 30 minutes to reach the target expansion rate from the initial setting of 10 hours, and the soaking was terminated. The soaked soybeans expanded (absorbed water) to about 420 kg.

[0229] 3) Grinding process

[0230] The soaked premium soybeans (420 kg) are transferred to a grinder. At this time, a second weight sensor measures the soybean-to-water ratio in real time (±0.1 g error).

[0231] The fifth machine learning model receives weight data based on the pre-trained correlation of "soybean weight-water supply amount-average particle size" and determines the water supply amount (approximately 3,150L at a 1:7.5 ratio).

[0232] An optical sensor (laser scattering, ±5μm error) continuously measures the average particle size and transmits it to the control unit. Grinding is stopped when the target particle size (e.g., 50μm) is reached (about 12 minutes), and the soybean mixture (target size ±5μm) is completed.

[0233] 4) Variable temperature (heating / cooling) process

[0234] The soybean mixture is introduced into a reaction vessel, and viscosity is controlled while inducing partial protein denaturation using a heater (target temperature: 80°C).

[0235] Data from the second temperature sensor and viscosity sensor are input into the sixth machine learning model to predict the time to reach the target viscosity (e.g., 300 cP). When the actual measurement result shows that the viscosity approaches 300 cP after heating for about 15 minutes, heating is stopped and the device is switched to a cooler.

[0236] When cooling, the 7th machine learning model simultaneously analyzes the pH sensor (±0.05 pH), viscosity sensor, and temperature sensor to automatically adjust the cooling speed (cooling temperature ~40 degrees). Since protein precipitation may occur with excessive cooling, the model sets the cooling time to about 10 minutes to achieve uniform cooling.

[0237] 5) Concentration process

[0238] The soybean mixture that has completed the temperature change process is transferred to a vacuum concentrator (internal pressure 0.05~0.07 bar).

[0239] Data collected through the third temperature sensor, the second viscosity sensor, the pressure sensor, and the refractometer (solid content measurement) is evaluated by the eighth machine learning model. Concentration is carried out for about 30 minutes by referring to the pre-trained "pressure-viscosity-solid content concentration" relationship, and a signal to stop concentration is issued when the solid content concentration reaches 9.5±0.1%.

[0240] At this time, the internal stirring intensity is also appropriately adjusted when viscosity rises rapidly to prevent bubble formation and fat aggregation. In fact, results showed that the concentration time was reduced by approximately 20% compared to the conventional method.

[0241] 6) Homogenization process

[0242] The concentrated soybean mixture (solid content about 9.5%) is transferred to a high-pressure homogenizer to further disperse the particles (pressure e.g., 150 MPa).

[0243] When the fourth temperature sensor, the third viscosity sensor, and the second optical sensor measure "temperature, viscosity, and average particle size" and input them into the ninth machine learning model, the model corrects the homogenization conditions in real time based on the previously accumulated "pressure-number of iterations-flow rate-final particle size" data.

[0244] As a result of the test, at the point of 2 repetitions, the average particle size (e.g., 1–2 μm) and viscosity (target 200 cP) are achieved, and homogenization is terminated without degradation of product quality by maintaining the temperature at 60°C or below. The final whole soy milk (approx. 4,800 L) has a smooth texture and uniform consistency.

[0245] According to the present embodiment, by immediately analyzing and processing sensor data generated during the entire process (classification, soaking, grinding, temperature control, concentration, and homogenization), the time required to reach target values ​​(e.g., expansion rate, particle size, viscosity, and solid content concentration) can be significantly shortened, and product quality deviations can be reduced. Furthermore, the machine learning model (CNN-based) of the classification module learns the correlation between soybean appearance, weight, and composition, thereby improving the accuracy of selecting high-quality soybeans to approximately 92%. This serves as a key foundation for reducing quality deviations at each stage from soaking to homogenization, along with reducing raw material loss. Additionally, as the machine learning model dynamically calculates the appropriate temperature, pressure, and time during soaking, concentration, and homogenization, an average 15–20% reduction in manufacturing time and energy consumption compared to conventional methods can be expected (concentration time was reduced by 20% in the embodiment). By appropriately controlling the expansion rate, viscosity, and solid content concentration, the whole soy milk possesses a uniform concentration (approx. 9.5±0.1%) and a smooth texture. Thermal denaturation and fat aggregation or precipitation are also reduced, enhancing the preservation of taste, flavor, and nutrients in the final product. Sensor data accumulated at each stage is fed back to the model for every batch, enabling rapid adaptation to exceptional situations or new soybean varieties. Over time, model prediction accuracy improves, and process stability is further strengthened.

[0247] According to one embodiment of the present disclosure, by using a combination of wired and wireless communication methods between each module according to the situation, sensor data and module status information for which real-time performance is important can be stably transmitted and received.

[0248] According to one embodiment of the present disclosure, by utilizing a machine learning model in a classification module to comprehensively analyze appearance and weight data, high-quality soybeans (premium soybeans) can be selected more accurately.

[0249] According to one embodiment of the present disclosure, an immersion module monitors the expansion rate and temperature in real time through an ultrasonic and temperature sensor, and based on this, a machine learning model dynamically corrects the immersion conditions, thereby minimizing quality variation of soybeans and achieving a uniform expansion state.

[0250] According to one embodiment of the present disclosure, sensor data is input into a machine learning model at each process step (grinding, temperature change, concentration, homogenization) to automatically optimize the temperature, pressure, time, etc., for reaching target values ​​(viscosity, particle size, solid content concentration, etc.), thereby improving the efficiency of the entire manufacturing process.

[0251] According to one embodiment of the present disclosure, real-time sensor data accumulated in a process is continuously fed back to a machine learning model, thereby improving the model's prediction accuracy and classification and control performance over time, and enabling adaptation to various environments and changes in varieties.

[0253] A smart whole soy milk manufacturing system according to one embodiment of the present disclosure may include: a classification module that determines and classifies premium soybeans among a plurality of soybeans; an immersion module that immerses the determined premium soybeans in a liquid inside an immersion tank; a grinding module that prepares a soybean mixture by grinding the immersed premium soybeans together with water inside a grinder; a temperature-changing module that introduces the soybean mixture into a reaction tank and heats it until it reaches a predetermined viscosity, and cools the soybean mixture after heating until it reaches a predetermined temperature; a concentration module that concentrates the soybean mixture after cooling until it reaches a predetermined solid content concentration under vacuum; a homogenization module that operates a homogenizer so that soybean particles contained in the concentrated soybean mixture are finely dispersed; and a control unit connected to the classification module, the immersion module, the grinding module, the temperature-changing module, the concentration module, and the homogenization module, and transmitting or receiving data in real time with at least one module.

[0254] The classification module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: an image sensor for acquiring image data of each of the plurality of soybeans; and a first weight sensor for acquiring first weight data representing the weight of each of the plurality of soybeans; and transmits the image data and the first weight data to the control unit, wherein the control unit inputs the image data to a first machine learning model to acquire at least one of volume data, cross-sectional area data, and surface reflectance data of each of the plurality of soybeans, inputs the image data, the first weight data, the volume data, the cross-sectional area data, and the surface reflectance data to a second machine learning model to acquire a component data score of each of the plurality of soybeans, and determines whether each of the plurality of soybeans corresponds to the premium soybean based on the acquired component data score.

[0255] The immersion module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: an ultrasonic sensor installed inside the immersion tank that projects an ultrasonic signal onto the surface of the premium soybean and acquires the reflected signal and transmits it to the control unit in real time; and a first temperature sensor installed inside the immersion tank that acquires first temperature data indicating the temperature of the immersion liquid and transmits it to the control unit in real time. The control unit may input the determined component data score of the premium soybean and the received first temperature data into a third machine learning model to acquire immersion temperature data and immersion time data, transmit the immersion temperature data and the immersion time data to the immersion module, determine the expansion rate of the premium soybean based on the reflected signal received in real time, input the expansion rate into a fourth machine learning model to correct the immersion temperature data and the immersion time data in real time, and transmit them to the immersion module.

[0256] The grinding module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: a water supply unit that supplies water inside the grinder; a second weight sensor installed inside the grinder that acquires second weight data of the immersed premium soybeans and transmits it to the control unit; and a first optical sensor installed inside the grinder that acquires first average particle size data of the premium soybeans being ground in real time and transmits it to the control unit; wherein the control unit inputs the second weight data of the immersed premium soybeans into a fifth machine learning model to determine the volume of water supplied by the water supply unit, and may stop the operation of the grinder based on the fact that the average particle size data received in real time reaches a predetermined size.

[0257] The temperature-changing module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: a heater that raises the internal temperature of the reaction vessel; a second temperature sensor installed in the reaction vessel that acquires second temperature data of the soybean mixture and transmits it to a control unit in real time; and a first viscosity sensor installed in the reaction vessel that acquires first viscosity data of the soybean mixture and transmits it to a control unit in real time; wherein the control unit determines a heating temperature and a heating time based on the second temperature data and the first viscosity data of the soybean mixture, inputs the first viscosity data received in real time into a sixth machine learning model to correct the heating temperature and heating time, and may stop the heater based on the first viscosity data received in real time reaching the predetermined viscosity.

[0258] The temperature-changing module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: a cooler that lowers the internal temperature of the reaction vessel; and a pH sensor installed in the reaction vessel that acquires pH data of the soybean mixture in real time and transmits it to the control unit; wherein the control unit inputs the first viscosity data and the pH data of the soybean mixture received in real time into a seventh machine learning model to correct the cooling temperature and cooling time, and may stop the cooler based on the second temperature data of the soybean mixture received in real time reaching a predetermined temperature.

[0259] The concentration module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: a vacuum concentrator that increases the solid content concentration of the soybean mixture in a vacuum state; a third temperature sensor installed inside the vacuum concentrator and acquiring third temperature data of the soybean mixture; a second viscosity sensor installed inside the vacuum concentrator and acquiring second viscosity data of the soybean mixture and transmitting it to a control unit in real time; a pressure sensor installed inside the vacuum concentrator and acquiring first pressure data indicating the internal pressure of the vacuum concentrator; and a refractometer installed inside the vacuum concentrator and acquiring solid content concentration data of the soybean mixture in real time and transmitting it to the control unit; wherein the control unit inputs the third temperature data of the soybean mixture, the second viscosity data, the first pressure data, and the solid content concentration data into an eighth machine learning model to dynamically correct the internal pressure and concentration time of the vacuum concentrator, and may stop the operation of the vacuum concentrator based on the solid content concentration data reaching a predetermined solid content concentration.

[0260] The control unit of the smart whole soy milk manufacturing system according to one embodiment of the present disclosure can adjust the internal pressure to gradually change and reach atmospheric pressure based on the solid content concentration data approaching a preset solid content concentration.

[0261] The vacuum concentrator of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure includes a stirrer that stirs the soybean mixture introduced into the vacuum concentrator; and the control unit can increase the stirring intensity of the stirrer when the second viscosity data of the soybean mixture obtained in real time exceeds a predetermined viscosity.

[0262] The homogenization module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: a high-pressure homogenizer that evenly disperses particles within the soybean mixture; a fourth temperature sensor installed inside the high-pressure homogenizer and acquiring fourth temperature data of the soybean mixture; and a third viscosity sensor installed inside the high-pressure homogenizer and acquiring third viscosity data of the soybean mixture. It includes a second optical sensor installed inside the high-pressure homogenizer and acquiring second average particle size data of the soybean mixture; and the control unit inputs the fourth temperature data, the third viscosity data, and the second average particle size data into a ninth machine learning model to determine the homogenization pressure, homogenization flow rate, and number of repetitions, and whenever homogenization is repeated, inputs the fourth temperature data, the third viscosity data, and the second average particle size data into the ninth machine learning model again to correct the homogenization pressure, the homogenization flow rate, and the number of repetitions, and can stop the high-pressure homogenizer based on the second average particle size data reaching a predetermined particle size and the third viscosity data reaching a predetermined viscosity.

[0263] Meanwhile, the disclosed embodiments may be implemented in the form of a recording medium that stores instructions executable by a computer. The instructions may be stored in the form of program code and, when executed by a processor, may generate a program module to perform the operation of the disclosed embodiments. The recording medium may be implemented as a computer-readable recording medium.

[0264] Computer-readable recording media include all types of recording media that store instructions that can be decoded by a computer. Examples include ROM (read-only memory), RAM (random access memory), magnetic tape, magnetic disk, flash memory, optical data storage devices, etc.

[0265] Additionally, computer-readable recording media may be provided in the form of non-transitory storage media. Here, 'non-transitory storage media' simply means that it is a tangible device and does not contain a signal (e.g., electromagnetic waves), and this term does not distinguish between cases where data is stored semi-permanently and cases where it is stored temporarily. For example, 'non-transitory storage media' may include a buffer in which data is stored temporarily.

[0266] According to one embodiment, the method according to the various embodiments disclosed herein may be provided as included in a computer program product. The computer program product may be traded between a seller and a buyer as a product. The computer program product may be distributed in the form of a device-readable recording medium (e.g., compact disc read-only memory (CD-ROM)), or distributed online (e.g., download or upload) through an application store (e.g., Play Store™) or directly between two user devices (e.g., smartphones). In the case of online distribution, at least a portion of the computer program product (e.g., downloadable app) may be temporarily stored or temporarily created on a device-readable recording medium, such as the memory of a manufacturer's server, an application store's server, or a relay server.

[0267] Specific embodiments have been illustrated and described above. However, the invention is not limited to the embodiments described above, and those skilled in the art may make various modifications without departing from the essence of the technical concept of the invention as described in the following claims.

[0270] According to one embodiment of the present disclosure, by using a combination of wired and wireless communication methods between each module according to the situation, sensor data and module status information for which real-time performance is important can be stably transmitted and received.

[0271] According to one embodiment of the present disclosure, by utilizing a machine learning model in a classification module to comprehensively analyze appearance and weight data, high-quality soybeans (premium soybeans) can be selected more accurately.

[0272] According to one embodiment of the present disclosure, an immersion module monitors the expansion rate and temperature in real time through an ultrasonic and temperature sensor, and based on this, a machine learning model dynamically corrects the immersion conditions, thereby minimizing quality variation of soybeans and achieving a uniform expansion state.

[0273] According to one embodiment of the present disclosure, sensor data is input into a machine learning model at each process step (grinding, temperature change, concentration, homogenization) to automatically optimize the temperature, pressure, time, etc., for reaching target values ​​(viscosity, particle size, solid content concentration, etc.), thereby improving the efficiency of the entire manufacturing process.

[0274] According to one embodiment of the present disclosure, real-time sensor data accumulated in a process is continuously fed back to a machine learning model, thereby improving the model's prediction accuracy and classification and control performance over time, and enabling adaptation to various environments and changes in varieties.

[0276] A smart whole soy milk manufacturing system according to one embodiment of the present disclosure may include: a classification module that determines and classifies premium soybeans among a plurality of soybeans; an immersion module that immerses the determined premium soybeans in a liquid inside an immersion tank; a grinding module that prepares a soybean mixture by grinding the immersed premium soybeans together with water inside a grinder; a temperature-changing module that introduces the soybean mixture into a reaction tank and heats it until it reaches a predetermined viscosity, and cools the soybean mixture after heating until it reaches a predetermined temperature; a concentration module that concentrates the soybean mixture after cooling until it reaches a predetermined solid content concentration under vacuum; a homogenization module that operates a homogenizer so that soybean particles contained in the concentrated soybean mixture are finely dispersed; and a control unit connected to the classification module, the immersion module, the grinding module, the temperature-changing module, the concentration module, and the homogenization module, and transmitting or receiving data in real time with at least one module.

[0277] The classification module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: an image sensor for acquiring image data of each of the plurality of soybeans; and a first weight sensor for acquiring first weight data representing the weight of each of the plurality of soybeans; and transmits the image data and the first weight data to the control unit, wherein the control unit inputs the image data to a first machine learning model to acquire at least one of volume data, cross-sectional area data, and surface reflectance data of each of the plurality of soybeans, inputs the image data, the first weight data, the volume data, the cross-sectional area data, and the surface reflectance data to a second machine learning model to acquire a component data score of each of the plurality of soybeans, and determines whether each of the plurality of soybeans corresponds to the premium soybean based on the acquired component data score.

[0278] The immersion module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: an ultrasonic sensor installed inside the immersion tank that projects an ultrasonic signal onto the surface of the premium soybean and acquires the reflected signal and transmits it to the control unit in real time; and a first temperature sensor installed inside the immersion tank that acquires first temperature data indicating the temperature of the immersion liquid and transmits it to the control unit in real time. The control unit may input the determined component data score of the premium soybean and the received first temperature data into a third machine learning model to acquire immersion temperature data and immersion time data, transmit the immersion temperature data and the immersion time data to the immersion module, determine the expansion rate of the premium soybean based on the reflected signal received in real time, input the expansion rate into a fourth machine learning model to correct the immersion temperature data and the immersion time data in real time, and transmit them to the immersion module.

[0279] The grinding module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: a water supply unit that supplies water inside the grinder; a second weight sensor installed inside the grinder that acquires second weight data of the immersed premium soybeans and transmits it to the control unit; and a first optical sensor installed inside the grinder that acquires first average particle size data of the premium soybeans being ground in real time and transmits it to the control unit; wherein the control unit inputs the second weight data of the immersed premium soybeans into a fifth machine learning model to determine the volume of water supplied by the water supply unit, and may stop the operation of the grinder based on the fact that the average particle size data received in real time reaches a predetermined size.

[0280] The temperature-changing module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: a heater that raises the internal temperature of the reaction vessel; a second temperature sensor installed in the reaction vessel that acquires second temperature data of the soybean mixture and transmits it to a control unit in real time; and a first viscosity sensor installed in the reaction vessel that acquires first viscosity data of the soybean mixture and transmits it to a control unit in real time; wherein the control unit determines a heating temperature and a heating time based on the second temperature data and the first viscosity data of the soybean mixture, inputs the first viscosity data received in real time into a sixth machine learning model to correct the heating temperature and heating time, and may stop the heater based on the first viscosity data received in real time reaching the predetermined viscosity.

[0281] The temperature-changing module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: a cooler that lowers the internal temperature of the reaction vessel; and a pH sensor installed in the reaction vessel that acquires pH data of the soybean mixture in real time and transmits it to the control unit; wherein the control unit inputs the first viscosity data and the pH data of the soybean mixture received in real time into a seventh machine learning model to correct the cooling temperature and cooling time, and may stop the cooler based on the second temperature data of the soybean mixture received in real time reaching a predetermined temperature.

[0282] The concentration module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: a vacuum concentrator that increases the solid content concentration of the soybean mixture in a vacuum state; a third temperature sensor installed inside the vacuum concentrator and acquiring third temperature data of the soybean mixture; a second viscosity sensor installed inside the vacuum concentrator and acquiring second viscosity data of the soybean mixture and transmitting it to a control unit in real time; a pressure sensor installed inside the vacuum concentrator and acquiring first pressure data indicating the internal pressure of the vacuum concentrator; and a refractometer installed inside the vacuum concentrator and acquiring solid content concentration data of the soybean mixture in real time and transmitting it to the control unit; wherein the control unit inputs the third temperature data of the soybean mixture, the second viscosity data, the first pressure data, and the solid content concentration data into an eighth machine learning model to dynamically correct the internal pressure and concentration time of the vacuum concentrator, and may stop the operation of the vacuum concentrator based on the solid content concentration data reaching a predetermined solid content concentration.

[0283] The control unit of the smart whole soy milk manufacturing system according to one embodiment of the present disclosure can adjust the internal pressure to gradually change and reach atmospheric pressure based on the solid content concentration data approaching a preset solid content concentration.

[0284] The vacuum concentrator of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure includes a stirrer that stirs the soybean mixture introduced into the vacuum concentrator; and the control unit can increase the stirring intensity of the stirrer when the second viscosity data of the soybean mixture obtained in real time exceeds a predetermined viscosity.

[0285] The homogenization module of a smart whole soy milk manufacturing system according to one embodiment of the present disclosure comprises: a high-pressure homogenizer that evenly disperses particles within the soybean mixture; a fourth temperature sensor installed inside the high-pressure homogenizer and acquiring fourth temperature data of the soybean mixture; and a third viscosity sensor installed inside the high-pressure homogenizer and acquiring third viscosity data of the soybean mixture. It includes a second optical sensor installed inside the high-pressure homogenizer and acquiring second average particle size data of the soybean mixture; and the control unit inputs the fourth temperature data, the third viscosity data, and the second average particle size data into a ninth machine learning model to determine the homogenization pressure, homogenization flow rate, and number of repetitions, and whenever homogenization is repeated, inputs the fourth temperature data, the third viscosity data, and the second average particle size data into the ninth machine learning model again to correct the homogenization pressure, the homogenization flow rate, and the number of repetitions, and can stop the high-pressure homogenizer based on the second average particle size data reaching a predetermined particle size and the third viscosity data reaching a predetermined viscosity.

[0286] Meanwhile, the disclosed embodiments may be implemented in the form of a recording medium that stores instructions executable by a computer. The instructions may be stored in the form of program code and, when executed by a processor, may generate a program module to perform the operation of the disclosed embodiments. The recording medium may be implemented as a computer-readable recording medium.

[0287] Computer-readable recording media include all types of recording media that store instructions that can be decoded by a computer. Examples include ROM (read-only memory), RAM (random access memory), magnetic tape, magnetic disk, flash memory, optical data storage devices, etc.

[0288] Additionally, computer-readable recording media may be provided in the form of non-transitory storage media. Here, 'non-transitory storage media' simply means that it is a tangible device and does not contain a signal (e.g., electromagnetic waves), and this term does not distinguish between cases where data is stored semi-permanently and cases where it is stored temporarily. For example, 'non-transitory storage media' may include a buffer in which data is stored temporarily.

[0289] According to one embodiment, the method according to the various embodiments disclosed herein may be provided as included in a computer program product. The computer program product may be traded between a seller and a buyer as a product. The computer program product may be distributed in the form of a device-readable recording medium (e.g., compact disc read-only memory (CD-ROM)), or distributed online (e.g., download or upload) through an application store (e.g., Play Store™) or directly between two user devices (e.g., smartphones). In the case of online distribution, at least a portion of the computer program product (e.g., downloadable app) may be temporarily stored or temporarily created on a device-readable recording medium, such as the memory of a manufacturer's server, an application store's server, or a relay server.

[0290] Specific embodiments have been illustrated and described above. However, the invention is not limited to the embodiments described above, and those skilled in the art may make various modifications without departing from the essence of the technical concept of the invention as described in the following claims. Explanation of the symbols

[0292] 1: Smart Whole Soy Milk Manufacturing System 100: Control unit 101: Processor 102: Memory 103: Communications Department 104: User Interface 110: The first machine learning model 120: The second machine learning model 130: The Third Machine Learning Model 140: The 4th Machine Learning Model 150: The 5th Machine Learning Model 160: The 6th Machine Learning Model 170: The 7th Machine Learning Model 180: The 8th Machine Learning Model 190: The 9th Machine Learning Model 191: The 10th Machine Learning Model 200: Classification Module 210: Classifier 220: Image sensor 230: First weight sensor 300: Immersion Module 310: Immersion tank 320: Ultrasonic sensor 330: First temperature sensor 400: Crushing Module 410: Grinder 420: First optical sensor 430: Second weight sensor 500: Variable temperature module 510: Heater 520: Cooler 530: pH sensor 540: First viscosity sensor 550: Second temperature sensor 600: Concentration Module 610: Vacuum concentrator 620: Refractometer 630: First pressure sensor 640: Second viscosity sensor 650: Third temperature sensor 700: Homogenization Module 710: High-pressure homogenizer 720: Second optical sensor 730: Third viscosity sensor 740: 4th temperature sensor 800: Drying Module 810: Vacuum pump 820: Moisture sensor 830: Second pressure sensor 840: 5th viscosity sensor 850: 6th temperature sensor

Claims

Claim 1 A smart whole soy milk manufacturing system for producing whole soy milk with excellent solid content retention by processing data obtained from multiple sensors in real time based on AI, comprising: a classification module for determining and classifying premium soybeans among multiple soybeans; an immersion module for immersing the determined premium soybeans in a liquid inside an immersion tank; a grinding module for producing a soybean mixture by grinding the immersed premium soybeans together with water inside a grinder; a temperature-changing module for introducing the soybean mixture into a reaction tank and heating it until it reaches a predetermined viscosity, and cooling the heated soybean mixture until it reaches a predetermined temperature; a concentration module for concentrating the cooled soybean mixture under a vacuum until it reaches a predetermined solid content concentration; a homogenization module for operating a homogenizer so that soybean particles contained in the concentrated soybean mixture are finely dispersed; and a control unit connected to the classification module, the immersion module, the grinding module, the temperature-changing module, the concentration module, and the homogenization module, and transmitting or receiving data in real time with at least one module. Claim 2 In paragraph 1, the classification module is, An image sensor for acquiring image data of each of the above plurality of soybeans; and It includes a first weight sensor that acquires first weight data representing the weight of each of the plurality of soybeans; A smart whole soy milk manufacturing system characterized by transmitting the image data and the first weight data to the control unit, wherein the control unit inputs the image data into a first machine learning model to obtain at least one of the volume data, cross-sectional area data, and surface reflectance data of each of the plurality of soybeans, inputs the image data, the first weight data, the volume data, the cross-sectional area data, and the surface reflectance data into a second machine learning model to obtain a component data score of each of the plurality of soybeans, and determines whether each of the plurality of soybeans corresponds to the premium soybean based on the obtained component data score. Claim 3 In paragraph 2, the above-mentioned immersion module is, An ultrasonic sensor installed inside the above-mentioned immersion tank, which projects an ultrasonic signal onto the surface of the premium soybeans, acquires the reflected signal, and transmits it to the control unit in real time; A smart whole soy milk manufacturing system comprising: a first temperature sensor installed inside the immersion tank and acquiring first temperature data indicating the temperature of the immersion liquid and transmitting it to the control unit in real time; wherein the control unit inputs the determined premium soybean component data score and the received first temperature data into a third machine learning model to acquire immersion temperature data and immersion time data, transmits the immersion temperature data and the immersion time data to the immersion module, determines the expansion rate of the premium soybean based on the reflected signal received in real time, inputs the expansion rate into a fourth machine learning model to correct the immersion temperature data and the immersion time data in real time, and transmits them to the immersion module. Claim 4 In paragraph 3, the crushing module is, A water supply device that supplies water to the inside of the above-mentioned grinder; A second weight sensor installed inside the grinder and acquiring second weight data of the immersed premium soybeans and transmitting it to the control unit; and A smart whole soy milk manufacturing system comprising: a first optical sensor installed inside the grinder and acquiring first average particle size data of the premium soybeans being ground in real time and transmitting it to the control unit; wherein the control unit inputs second weight data of the immersed premium soybeans into a fifth machine learning model to determine the volume of water supplied by the water dispenser, and stops the operation of the grinder based on the average particle size data received in real time reaching a predetermined size. Claim 5 In paragraph 4, the above-mentioned temperature change module is, A heater that raises the internal temperature of the above reaction vessel; A second temperature sensor installed in the above reaction vessel, which acquires second temperature data of the soybean mixture and transmits it to a control unit in real time; and A smart whole soy milk manufacturing system comprising: a first viscosity sensor installed in the reaction vessel and acquiring first viscosity data of the soybean mixture and transmitting it to a control unit in real time; wherein the control unit determines a heating temperature and a heating time based on second temperature data and first viscosity data of the soybean mixture, inputs the first viscosity data received in real time into a sixth machine learning model to correct the heating temperature and heating time, and stops the heater based on the first viscosity data received in real time reaching a predetermined viscosity. Claim 6 In paragraph 5, the above-mentioned temperature change module is, A cooler that lowers the internal temperature of the above reaction vessel; A smart whole soy milk manufacturing system comprising: a pH sensor installed in the reaction vessel and acquiring pH data of the soybean mixture in real time and transmitting it to the control unit; wherein the control unit inputs the first viscosity data and the pH data of the soybean mixture received in real time into a seventh machine learning model to correct the cooling temperature and cooling time, and stops the cooler based on the second temperature data of the soybean mixture received in real time reaching a predetermined temperature. Claim 7 In paragraph 6, the above concentration module is, A vacuum concentrator that increases the solid content concentration of the above soybean mixture in a vacuum state; A third temperature sensor installed inside the vacuum concentrator and acquiring third temperature data of the soybean mixture; A second viscosity sensor installed inside the vacuum concentrator, which acquires second viscosity data of the soybean mixture and transmits it to the control unit in real time; A pressure sensor installed inside the vacuum concentrator and acquiring first pressure data indicating the internal pressure of the vacuum concentrator; and A smart whole soy milk manufacturing system comprising: a refractometer installed inside the vacuum concentrator and acquiring solid content concentration data of the soybean mixture in real time and transmitting it to the control unit; wherein the control unit inputs the third temperature data, the second viscosity data, the first pressure data, and the solid content concentration data of the soybean mixture into an eighth machine learning model to dynamically correct the internal pressure and concentration time of the vacuum concentrator, and stops the operation of the vacuum concentrator based on the solid content concentration data reaching a preset solid content concentration. Claim 8 In claim 7, the control unit is a smart whole soy milk manufacturing system that adjusts the internal pressure to gradually change and reach atmospheric pressure based on the solid content concentration data approaching a preset solid content concentration. Claim 9 In paragraph 7, the vacuum concentrator is, A smart whole soy milk manufacturing system comprising: a stirrer for stirring the soybean mixture introduced into a vacuum concentrator; wherein the control unit increases the stirring intensity of the stirrer when the second viscosity data of the soybean mixture obtained in real time exceeds a predetermined viscosity. Claim 10 In claim 9, the homogenization module is, A high-pressure homogenizer for evenly dispersing particles within the above-mentioned soybean mixture; A fourth temperature sensor installed inside the high-pressure homogenizer and acquiring fourth temperature data of the soybean mixture; A third viscosity sensor installed inside the high-pressure homogenizer and acquiring third viscosity data of the soybean mixture; A smart whole soy milk manufacturing system comprising: a second optical sensor installed inside the high-pressure homogenizer and acquiring second average particle size data of the soybean mixture; wherein the control unit inputs the fourth temperature data, the third viscosity data, and the second average particle size data into a ninth machine learning model to determine the homogenization pressure, homogenization flow rate, and number of repetitions; inputs the fourth temperature data, the third viscosity data, and the second average particle size data into the ninth machine learning model again each time homogenization is repeated to correct the homogenization pressure, the homogenization flow rate, and the number of repetitions; and stops the high-pressure homogenizer based on the second average particle size data reaching a predetermined particle size and the third viscosity data reaching a predetermined viscosity.

Images