A series of konjac-based spleen-strengthening, stomach-nourishing, and tonifying dietary foods and their production process.

Abstract

This invention provides a konjac-based spleen-strengthening and stomach-nourishing dietary supplement and its production process, belonging to the field of food processing technology. The invention employs a stepwise extraction method, using supercritical extraction of fat-soluble components from ingredients A and B followed by hot water extraction of polysaccharide components. The polysaccharide extract is then combined with steamed yam and spray-dried to obtain a composite plant powder. Subsequently, the composite plant powder is mixed with konjac powder, and an alkaline gradient micro-injection is performed using a gelation timing control system, simultaneously collecting pH values ​​at various locations. The collected data is then input into a thermodynamic-kinetic dual-engine driven neural network model to obtain a gelation state assessment value. Based on the gelation state assessment value and the gelation uniformity evaluation index, it is determined whether additional homogenization treatment is needed. This invention solves the technical problem of uncontrolled gelation uniformity caused by the inability to perceive the dynamic coupling relationship of multiple processes during the gelation of konjac glucomannan.

Description

Technical Field

[0001] This invention belongs to the field of food processing technology, and specifically relates to a konjac-based spleen-strengthening, stomach-nourishing, and tonifying diet and its production process. Background Technology

[0002] The production of konjac-based spleen-strengthening and stomach-nourishing dietary supplements involves multiple continuous sub-processes, including the extraction of effective components from traditional Chinese medicinal materials, spray drying for powder production, and konjac glucomannan gelation. In existing processes, supercritical fluid extraction and hot water extraction typically rely on operator experience to determine the endpoint. The matching of inlet and outlet air temperatures and feed concentration during spray drying is also based on historical parameters. The subsequent gelation stage is controlled by manually sampling and testing pH values ​​at regular intervals and manually adjusting the amount of alkali added. These methods are applicable to a certain extent under steady-state conditions in individual sub-processes.

[0003] However, in existing technologies, the concentration deviation of the effective components in the supercritical extraction stage directly affects the physicochemical properties of the spray-dried powder, and the residual moisture content in the spray drying stage alters the swelling state and gelation rate of konjac glucomannan. These three sub-processes have a strongly coupled kinetic transmission relationship, and neither manual experience nor offline detection methods can capture this cross-process kinetic disturbance in real time. This leads to a failure to respond promptly to fluctuations in upstream operating conditions, resulting in frequent problems such as premature gelation or uneven gelation caused by excessively large local pH gradients. In other words, existing technologies suffer from a technical problem of uncontrolled gelation uniformity due to the inability to perceive the kinetic coupling relationship of multiple sub-processes during konjac glucomannan gelation in real time. Summary of the Invention

[0004] In view of this, the present invention provides a series of konjac-based spleen-strengthening and stomach-nourishing dietary supplements and their production process, which can solve the technical problem in the prior art where the gelation uniformity is out of control due to the inability to perceive the dynamic coupling relationship of multiple processes in the gelation process of konjac glucomannan in real time.

[0005] This invention is achieved as follows: This invention provides a konjac-based spleen-strengthening, stomach-nourishing, and tonifying dietary supplement and its production process, comprising the following steps:

[0006] Weigh each component according to the formula, and extract ingredient A and ingredient B in a stepwise sequence. First, extract the fat-soluble components, and then extract the polysaccharide components from the remaining material with hot water.

[0007] Ingredient C was steamed and then combined with polysaccharide extract, followed by spray drying to obtain compound plant powder.

[0008] The compound plant powder was mixed with ingredient D, and the alkaline solution was injected in a gradient micro-injection through a gelation timing control system, while the pH value at each location was collected simultaneously.

[0009] The pH values ​​and gelation kinetics data at each location were input into a thermodynamic-kinetic dual-engine driven neural network model to obtain the gelation state assessment value.

[0010] Based on the gelation state assessment value and gelation uniformity evaluation index, it is determined whether additional homogenization treatment is needed, followed by molding, sterilization and packaging.

[0011] Ingredient A consists of coix seed, astragalus, codonopsis, poria, euryale seed, licorice, and dried ginger; ingredient B consists of citron and amomum villosum; ingredient C is yam; and ingredient D is konjac flour.

[0012] The ingredients A are in the form of 10-30 parts by mass, the ingredients B are in the form of 10-30 parts by mass, the ingredients C are in the form of 30-60 parts by mass, the ingredients D are in the form of 10-30 parts by mass, and the sum of the mass parts of all components is 100 parts.

[0013] The product is prepared by any one of the production processes according to claims 1 to 3, comprising ingredients A, B, C and D, wherein the sum of the mass parts of each component is 100 parts, and each herb is distributed in proportion within its respective ingredient group according to the experimentally determined content of the active ingredient.

[0014] The specific steps of the stepwise sequential extraction are as follows: First, the mixture of ingredient A and ingredient B is subjected to supercritical fluid extraction. Extraction was performed, and the resulting fat-soluble extract was collected separately. The remaining material was subjected to hot water extraction to extract the polysaccharide components, and the polysaccharide extract was collected by filtration.

[0015] Among them, the supercritical The extraction pressure, temperature, and time parameters, as well as the temperature and time parameters for hot water extraction, were determined by analyzing multiple batches of control experiments using response surface methodology.

[0016] The gelation timing control system consists of an online distributed pH electrode array and an alkaline gradient micro-injection device. The online distributed pH electrode array is evenly distributed along the longitudinal direction of the mixing container wall and outputs the pH value at each location in real time. The alkaline gradient micro-injection device dynamically adjusts the alkaline injection rate according to the pH value at each location.

[0017] The thermodynamic-kinetic dual-engine driven neural ordinary differential equation network model uses neural ordinary differential equations as its core framework. It establishes a set of ordinary differential equation systems for each of the supercritical extraction subprocess, spray drying subprocess, and konjac glucomannan gelation subprocess, and the three sets of ordinary differential equation systems are coupled in series.

[0018] In this system, the right-hand function of each ordinary differential equation system is composed of a weighted fusion of the first engine and the second engine; the first engine is a multilayer perceptron; the second engine is a preset thermodynamic constraint term; the supercritical extraction subprocess introduces the Arrhenius rate equation; the spray drying subprocess introduces Fick's second diffusion law; and the konjac glucomannan gelation subprocess introduces the cross-linking reaction kinetic equation.

[0019] In the thermodynamic-kinetic dual-engine driven neural network model, a gelation state evaluation function is designed. The gelation state evaluation value is calculated based on the concentration of konjac glucomannan, the pH value at each location and the system temperature, and then the low-speed anchor stirring speed is dynamically adjusted.

[0020] Specifically, when the gelation state assessment value reaches the upper gelation threshold, the low-speed anchor stirring speed is reduced to the low-speed range and the alkali injection is suspended; when the gelation state assessment value is in the middle range, the low-speed anchor stirring speed is maintained in the medium-speed range; when the gelation state assessment value is below the lower gelation threshold, the low-speed anchor stirring speed is increased to the high-speed range and the alkali injection rate is moderately increased.

[0021] The gelation uniformity evaluation index is the coefficient of variation of the gelation state assessment value. When the gelation uniformity evaluation index does not exceed the uniformity threshold, the uniformity is determined to be up to standard and the molding process can proceed directly. When the gelation uniformity evaluation index exceeds the uniformity threshold, additional homogenization treatment is required.

[0022] In step S03, the collaborative decision between the gradient micro-injection of alkali solution and the uniformity of gelation is optimized using a game theory model. The upper-level model aims to minimize the total amount of alkali solution injected, while the lower-level model aims to maximize the uniformity of gelation. The two models achieve Nash equilibrium by coupling the standard deviation of pH value at each location in the system and solving the problem in a rolling manner.

[0023] The training of the thermodynamic-dynamic dual-engine driven neural ordinary differential equation network model uses the Adam optimizer, the loss function is the sum of the mean square error between the predicted and measured values ​​of each subprocess state, and a cosine annealing learning rate scheduling and early stopping mechanism are adopted.

[0024] Among them, supercritical Extraction pressure is 15–35. The temperature is 35–55℃; for hot water extraction, the water volume is 8–12 times the mass, the temperature is 80–95℃, and the time is 60–120 seconds. The spray drying inlet air temperature is 150–200℃, and the outlet air temperature is 70–90℃; the spacing between adjacent pH electrodes is 5–15 mm. The system pH was maintained at 7.5–8.5; the low-speed anchor stirring speed was 5–30 rpm. The upper gelation threshold was 0.8, the lower gelation threshold was 0.5, and the uniformity threshold was 0.1; the homogenization pressure was 10–30. The homogenization time is 3–10 minutes. The solution cycle for the game theory model is 300–500. The total number of training rounds is 200-500. The early stopping condition is that the validation set loss does not decrease for 20 consecutive rounds, and the initial learning rate is... .

[0025] This invention employs a thermodynamic-kinetic dual-engine driven ordinary differential equation network model, which connects the ordinary differential equation systems of the three sub-processes of supercritical extraction, spray drying, and konjac glucomannan gelation in series. This allows the state output of the upstream sub-process to serve as the initial condition for the downstream sub-process, thus enabling real-time transmission and prediction of dynamic disturbances across sub-processes.

[0026] This invention embeds the Arrhenius rate equation, Fick's second diffusion law, and crosslinking reaction kinetic equation as thermodynamic constraints within the framework of the neural network differential equations. This allows the model to still obey physical laws even under data-scarce conditions, effectively suppressing overfitting caused by insufficient training samples in purely data-driven models. Furthermore, by using an online distributed pH electrode array to collect pH values ​​at various locations in real time, continuously correcting the gelation state assessment value, and dynamically adjusting the alkali injection rate and low-speed anchor stirring speed based on the gelation state assessment value, the gelation process always operates within a controlled kinetic trajectory.

[0027] In summary, the present invention solves the technical problem mentioned in the background art of the loss of control over gelation uniformity caused by the inability to perceive the dynamic coupling relationship of multiple processes in real time during the gelation of konjac glucomannan. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention.

[0029] Figure 2 This is a graph showing the changes in outlet air temperature and residual moisture content of powder over time during the spray drying process.

[0030] Figure 3 This is a graph showing the change of pH value at each measuring point over time in step S03.

[0031] Figure 4 The graph shows the correlation between the gelation state assessment value and the rotation speed of the low-speed anchor stirring device.

[0032] Figure 5 The graph shows the change in the alkali injection rate command over time in the game theory model. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0034] like Figure 1 The diagram shown is a flowchart of a konjac-based spleen-strengthening and stomach-nourishing dietary supplement and its production process provided by the present invention. The method includes the following steps:

[0035] S01. Weigh each component according to the formula, and extract ingredient A and ingredient B in a stepwise sequence. First, extract the fat-soluble components, and then extract the polysaccharide components with hot water from the remaining material.

[0036] S02. The ingredient C is steamed and cooked, then combined with the polysaccharide extract obtained from S01, and spray-dried to obtain a compound plant powder.

[0037] S03. Mix the compound plant powder with ingredient D, and inject alkaline solution gradient micro-volume through the gelation timing control system, while simultaneously collecting pH values ​​at various locations.

[0038] S04. Input the pH values ​​and gelation kinetic data collected in S03 into the thermodynamic-kinetic dual-engine driven neural network model to obtain the gelation state evaluation value.

[0039] S05. Based on the gelation state assessment value and gelation uniformity evaluation index, determine whether to add homogenization treatment, and then proceed with molding, sterilization and packaging.

[0040] The specific components and mass ratios of the formula are as follows: Ingredient A consists of Coix seed, Astragalus membranaceus, Codonopsis pilosula, Poria cocos, Euryale ferox, Glycyrrhiza uralensis, and Zingiber officinale, with a total mass fraction of 10-30 parts; Ingredient B consists of Citrus medica and Amomum villosum, with a total mass fraction of 10-30 parts; Ingredient C is Dioscorea opposita, with a mass fraction of 30-60 parts; Ingredient D is konjac flour, with a mass fraction of 10-30 parts; The sum of the mass fractions of all components is 100 parts; Each herb is distributed proportionally within its respective ingredient group according to the experimentally determined content of its active ingredient.

[0041] The specific steps of the stepwise sequential extraction are as follows: First, the mixture of ingredient A and ingredient B is extracted at an extraction pressure of 15-35. At temperatures of 35–55℃, using supercritical water Lipid-soluble components were extracted and the resulting lipid-soluble extracts were collected separately. The remaining material was mixed with 8 to 12 times its weight of water and hot-extracted at 80 to 95°C for 60 to 120 minutes. The polysaccharide extract was then collected by filtration. The temperature, pressure, and time parameters for both extraction steps were determined by multiple batches of control experiments to measure the extraction rate of each component. The optimal range was determined by response surface methodology.

[0042] The gelation timing control system consists of an online distributed pH electrode array and an alkali gradient micro-injection device. The online distributed pH electrode array is evenly distributed longitudinally along the wall of the mixing container, with an adjacent electrode spacing of 5-15 cm, and outputs the pH value at each location in real time. The alkali gradient micro-injection device dynamically adjusts the alkali injection rate according to the pH value at each location to maintain the pH of the system within the range of 7.5-8.5. The low-speed anchor stirring speed is set to 5-30 rpm. The pH range and speed range are determined by collecting gelation time data at different pH and speeds through konjac glucomannan gelation kinetic experiments, with the goal of minimizing the probability of premature gelation.

[0043] The specific structure of the thermodynamic-kinetic dual-engine driven neural ordinary differential equation network model is as follows: Using neural ordinary differential equations as the core framework, a set of ordinary differential equation systems is established for each of the supercritical extraction sub-process, spray drying sub-process, and konjac glucomannan gelation sub-process; the right-hand function of each set of ordinary differential equation systems is formed by a weighted fusion of the first engine and the second engine; the first engine is a multilayer perceptron with 3-5 layers, each layer containing 32-128 neurons, with residual connections between adjacent layers, and the weight matrix between neurons uses a sparse attention mask to achieve dynamic distribution of neurons. Block allocation involves dividing the weight matrix into multiple sub-blocks based on the current batch input dimension and distributing them to different CUDA streams to achieve CUDA stream-level parallel computing. The number of CUDA thread blocks is adaptively allocated based on the current subprocess state vector dimension. Memory allocation uses a memory pool reuse mechanism to pre-allocate fixed memory blocks hierarchically according to subprocesses. The second engine uses preset thermodynamic constraints: the supercritical extraction subprocess introduces the Arrhenius rate equation, the spray drying subprocess introduces Fick's second diffusion law, and the konjac glucomannan gelation subprocess introduces the cross-linking reaction kinetic equation. The parameters of each prior equation are... The input is fixed after pre-calibration through independent experiments; the outputs of the first engine and the second engine are weighted and summed with learnable fusion weights, which are adaptively adjusted based on the current state and the prediction residual of the second engine; each set of ordinary differential equations uses the adaptive step-size Dormand-Prince method for iterative integration, and the integration step size is automatically encrypted when the rate of change of the state exceeds a set threshold; the effective component concentration output from the supercritical extraction subprocess is used as the initial input condition for the spray drying subprocess, and the residual moisture content output from the spray drying subprocess is used as the initial condition for the konjac glucomannan gelation subprocess. The three sets of ordinary differential equations for the subprocesses are coupled in series to achieve end-to-end dynamic prediction of the entire process; the network parameters are updated through the gradient link of the ordinary differential equation solver via backpropagation during the training phase; during the inference phase, the gelation state evaluation value is rolled and corrected based on the real-time collected pH value at each location and the residual moisture content of the spray drying subprocess, and the gelation state evaluation value is output to the low-speed anchor stirring speed control interface; CUDA memory is statically divided into layers according to the three subprocesses, and the memory block size of each subprocess is pre-allocated based on the maximum state vector dimension of the historical batch to avoid memory fragmentation caused by dynamic allocation.

[0044] The specific steps for establishing the training dataset for the thermodynamic-kinetic dual-engine driven neural network model include: collecting concentration-time curve data of effective component concentration changing with time under different pressure and temperature conditions during supercritical extraction; collecting drying curve data of moisture changing with time under different inlet air temperature and material concentration during spray drying; collecting kinetic data of konjac glucomannan gelation and temperature field distribution data under different pH values, temperatures and konjac powder concentrations; classifying and organizing the above data according to sub-processes, and independently dividing each sub-process into training and validation sets with a ratio of 8:2; and inputting all data into the model for training after normalization.

[0045] The specific steps for training the thermodynamic-dynamic dual-engine driven neural network model for ordinary differential equations include: iteratively updating the network parameters using the Adam optimizer, with the initial learning rate set to... The loss function is the sum of the mean square errors between the predicted and measured values ​​of each subprocess state; cosine annealing learning rate scheduling is used, with a total training round of 200-500 rounds; an early stopping mechanism is triggered when the validation set loss does not decrease for 20 consecutive rounds; after training, the model weights are fixed and deployed to the control system, and real-time data is collected and input during inference.

[0046] The technical advantages of the thermodynamic-kinetic dual-engine driven neural network model are as follows: Traditional production relies on manual experience to judge the extraction endpoint and gelation timing, making it impossible to perceive the dynamic coupling relationship between multiple sub-processes in real time; the thermodynamic-kinetic dual-engine driven neural network model embeds thermodynamic prior equations into the neural network as structural constraints, enabling the model to still obey physical laws under data-scarce conditions and reducing the risk of overfitting in purely data-driven models; the neural network framework replaces discrete-time step prediction with continuous-time dynamics, enabling interpolation and prediction of the state at any time, adapting to variable-step production conditions; the three-sub-process serial coupling structure allows deviations in the upstream extraction conditions to be propagated downwards in the prediction link and provide early warnings, achieving coordinated control of the entire process; CUDA streaming parallelism and hierarchical memory allocation ensure that real-time inference latency meets online control requirements.

[0047] The thermodynamic-kinetic dual-engine driven neural network model incorporates a gelation state evaluation function. This function calculates the gelation state evaluation value based on the konjac glucomannan concentration, pH value at various locations, and system temperature, thereby dynamically adjusting the low-speed anchor stirring speed. The gelation state evaluation function is described below: ;in This is a dimensionless assessment value for the gelation state. This represents the ratio of the current konjac glucomannan concentration to the initial konjac glucomannan concentration. This is the normalized value of the current pH. This is the normalized value of the current system temperature. and For learnable fusion weights and satisfying , , , , The pH lower limit, pH upper limit, system temperature lower limit, and system temperature upper limit are the values ​​determined by the experiment; when the gelation state assessment value is... At this time, reduce the speed of the low-speed anchor stirrer to 5-10 rpm and stop the alkali injection; when the gelation state assessment value At that time, the low-speed anchor stirring speed was maintained at 10-20 rpm, and the alkali injection rate was executed normally according to the predicted value of the thermodynamic-kinetic dual-engine driven neural network model; when the gelation state assessment value At that time, the low-speed anchor stirring speed was increased to 20-30 rpm, and the alkali injection rate was moderately increased; the threshold values ​​of 0.5 and 0.8 were determined by minimizing the occurrence rate of gelation lumps under different gelation state evaluation values ​​through multiple batches of konjac glucomannan gelation experiments; the input of the gelation state evaluation function is the current konjac glucomannan concentration. pH value at various locations and system temperature The output is a dimensionless gelation state assessment value. .

[0048] The game theory model is used for collaborative decision optimization between the incremental injection of alkali solution gradient and the uniformity of gelation in step S03; the objective function of the upper-level model, which aims to minimize the total amount of alkali solution injected, is expressed as follows: The objective function of the lower-level model, which aims to maximize gelation uniformity, is expressed as follows: ;in This is the ratio of the current alkali injection volume to the baseline alkali injection volume. This is the ratio of the standard deviation of pH at each location within the system to the standard deviation of the reference pH. The penalty coefficient is... The coefficient of variation for the gelation state assessment value. This represents the mean value of the gelation state assessment. The standard deviation of the gelation state assessment value is used; the upper-level model constraint is the gelation state assessment value. The lower-level model has the following constraints: The coupling terms of the two objective functions are: The upper-level model affects the effect by adjusting the alkali injection rate. The lower-level model is based on Update the assessment values ​​of gelation status at each location. This feedback is then sent to the upper-level constraints, forming a two-layer iterative coupling. After determining the product molding quality under different combinations of alkali injection volume and gelation uniformity through multiple batch experiments, the dual objective of minimizing alkali injection volume and maximizing gelation uniformity was determined. A game theory model was used to solve the problem on a rolling basis with a period of 300–500 ms, outputting the optimal alkali injection rate command to the alkali gradient micro-injection device. The function input is the current alkali injection volume. Standard deviation of pH value at various locations within the system The output is the target value for upper-level optimization; The function input is the standard deviation of the gelation state assessment value. Mean value of gelation state assessment The output is the gelation uniformity index.

[0049] The technical effect of the game theory model is as follows: there is an inherent contradiction between the amount of alkali injected and the uniformity of gelation. Injecting too quickly leads to excessively large local pH gradients, causing premature gelation; injecting too slowly results in uneven overall gelation. The two-layer game framework decomposes the two mutually constraining objectives into independent optimization problems at upper and lower layers, through coupling terms... This approach enables information transmission and collaborative decision-making, avoiding the problem of ignoring constraints and interactions in single-objective optimization. The upper-level model penalizes the amount of alkali injected, making the system tend to conserve alkali, while the lower-level model drives the maximization of gelation uniformity. The two reach Nash equilibrium in the iteration through the pH standard deviation coupling term, thus finding the dynamic optimal balance between alkali injection efficiency and gelation uniformity.

[0050] The evaluation index for gelation uniformity is: The standard deviation of the system pH value was calculated using pH values ​​collected at various locations by an online distributed pH electrode array. Simultaneously, the thermodynamic-kinetic dual-engine driven neural network model outputs the gelation state evaluation values ​​at each location. Calculate the standard deviation of the gelation state assessment value Mean value of gelation state assessment Then, the gelation uniformity evaluation index was calculated. When the gelation uniformity evaluation index If the uniformity is deemed satisfactory, the process proceeds directly to the molding stage; if the gelation uniformity evaluation index is not met... Homogenization was added at the time, with the homogenization pressure set at 10-30 MPa and the homogenization time at 3-10 min; the threshold of 0.1 was determined by maximizing the score pass rate after determining the internal structure uniformity score of the molded product under different gelation uniformity evaluation indicators through multiple batch experiments.

[0051] The konjac glucomannan is a water-soluble polysaccharide component extracted from konjac flour, with a molecular weight of approximately [missing information]. The product, at the Da scale, forms a thermally irreversible gel after deacetylation under alkaline conditions. The gelation process is highly sensitive to pH and system temperature. The spray drying method involves atomizing the liquid extract and then instantly evaporating moisture in a high-temperature airflow to produce powder. The inlet air temperature ranges from 150 to 200°C, and the outlet air temperature ranges from 70 to 90°C. The Arrhenius rate equation describes the exponential relationship between the reaction rate constant and temperature, characterizing the kinetics of the dissolution rate of the active ingredient during supercritical extraction. Fick's second diffusion law describes the diffusion equation of the concentration of a substance in a medium as a function of time and space, characterizing the migration kinetics of moisture within the material during spray drying. The Dormand-Prince method is an adaptive step-size Runge-Kutta numerical integration method that automatically adjusts the integration step size by comparing the differences between fourth-order and fifth-order approximate solutions. The response surface methodology is an experimental design and statistical analysis method that uses multi-factor, multi-level experiments and fits the polynomial surface relationship between response values ​​and factors to find the optimal combination of factors. The Nash equilibrium is a stable state in game theory where none of the participants have an incentive to unilaterally change their own strategies given that the other's strategies are fixed. In the two-layer game model, this is represented by a stable solution where the upper-layer alkali injection rate command and the lower-layer gelation uniformity feedback reach a mutually consistent agreement. The memory pool reuse mechanism is a memory management strategy that pre-allocates fixed-size memory blocks and reuses them across multiple computations, rather than dynamically allocating and releasing them each time. This reduces memory fragmentation and lowers the time overhead of dynamic memory allocation. The residual connection is a connection method that directly adds the input of a layer to the output of that layer, used to alleviate the gradient vanishing problem in deep network training. The sparse attention mask is an operation that resets weights below a threshold to zero in the attention weight matrix, used to reduce redundant connections between neurons and achieve dynamic block allocation of GPU memory.

[0052] The specific implementation method of step S01 is as follows: First, accurately weigh each component according to the formula. Ingredient A (Coix seed, Astragalus, Codonopsis, Poria, Euryale seed, Licorice, Dried ginger) and Ingredient B (Citrus medica, Amomum villosum) are mixed and allocated in equal proportions according to the experimentally determined active ingredient content of each individual herb. The step-by-step extraction is carried out in two stages. The first stage uses supercritical fluid extraction. Extraction technology, with extraction pressures of 15–35 ppm. Extraction of the fat-soluble components of the mixture was performed at temperatures of 35–55°C, and the resulting fat-soluble extracts were collected and stored separately; supercritical fluid extraction was also performed. Extraction and utilization Exhibiting both gas diffusion and liquid solubility under supercritical conditions, it can selectively dissolve heat-sensitive components such as volatile oils and fat-soluble active ingredients at relatively low temperatures, avoiding high-temperature decomposition. In the second stage, 8–12 times the mass of water is added to the residue after extraction, and the polysaccharide components are extracted using hot water at 80–95°C for 60–120 seconds. The polysaccharide extract was collected after filtration. Hot water extraction utilized the principle that the dissolution rate of polysaccharide components is significantly increased in high-temperature aqueous solutions, achieving targeted separation from lipid-soluble components. The temperature, pressure, and time parameters for both extraction stages were determined by collecting extraction rate data through multiple batches of control experiments. The optimal range was then determined by fitting a polynomial surface relationship using the response surface methodology, ensuring the scientific validity and repeatability of the extraction process parameters.

[0053] The specific implementation of step S02 is as follows: Yam (ingredient C) is steamed to gelatinize the starch and dissolve the polysaccharides, which are then combined with the polysaccharide extract obtained in S01 to form a mixed liquid rich in polysaccharides. Subsequently, the combined liquid is dried using spray drying technology. The inlet air temperature is controlled at 150–200℃, and the outlet air temperature is controlled at 70–90℃. After atomization, the liquid evaporates moisture instantly in the high-temperature airflow, yielding a composite plant powder. During spray drying, the migration of moisture within the material follows Fick's second diffusion law. The uniformity of the atomized droplet size determines the consistency of the drying rate. A narrower droplet size distribution helps obtain powder with uniform residual moisture content. The residual moisture content of the spray-dried composite plant powder is a key parameter for the initial conditions of the subsequent konjac glucomannan gelation process. This parameter directly affects the degree of swelling of konjac glucomannan and the subsequent gelation rate; therefore, the stability of the drying process parameters must be strictly controlled.

[0054] The specific implementation of step S03 is as follows: After mixing the compound plant powder with ingredient D (konjac powder), place the mixture in a mixing container, and implement gradient micro-injection of alkali solution using a gelation timing control system. The gelation timing control system consists of an online distributed pH electrode array and an alkali gradient micro-injection device. The online distributed pH electrode array is evenly distributed longitudinally along the wall of the mixing container, with an adjacent electrode spacing of 5-15 mm. The system outputs the pH value at each location in real time; the alkali gradient micro-injection device dynamically adjusts the alkali injection rate according to the pH value at each location to maintain the system pH within the range of 7.5 to 8.5, and the low-speed anchor stirring speed is set between 5 and 30. The pH value is dynamically adjusted based on the gelation state assessment value within the specified range. The vertically uniform design of the distributed pH electrode array allows for real-time sensing of local pH values ​​at different heights within the container, avoiding the limitations of single-point detection in reflecting the spatial distribution of pH values ​​within the container. This provides a data foundation for subsequent assessment of the spatial uniformity of the gelation state.

[0055] The specific implementation of step S04 is as follows: The pH values ​​and gelation kinetic data collected in S03 are input into a thermodynamic-kinetic dual-engine driven neural ordinary differential equation (NDE) network model to obtain the gelation state assessment value. The model uses the neural NDE as its core framework, establishing NDE systems for three sub-processes: supercritical extraction, spray drying, and konjac glucomannan gelation. These three systems are coupled in series, with the output state of the upstream sub-process serving as the initial condition for the downstream sub-process. The right-hand side function of each NDE system is composed of a weighted sum of the first engine (multilayer perceptron) and the second engine (thermodynamic constraint term) with learnable fusion weights. These learnable fusion weights are adaptively adjusted based on the current state and the predicted residuals of the second engine. The expression for the gelation state assessment function is: Input is the current concentration of konjac glucomannan. pH value at various locations and system temperature The output is a dimensionless gelation state assessment value. , and For learnable fusion weights and satisfying The inference phase uses real-time collected pH values ​​and residual moisture content at various locations to continuously correct the gelation state assessment value, and transmits the output to the low-speed anchor stirring speed control interface to form a closed-loop control.

[0056] The specific implementation of step S05 is as follows: based on the gelation state assessment value Evaluation index of gelation uniformity Determine the subsequent processing method. When If the uniformity is deemed satisfactory, the process proceeds directly to the molding stage; if... At this time, additional homogenization treatment is performed, with the homogenization pressure set to 10–30. Homogenization time 3–10 Homogenization involves high-pressure shearing to break down and redisperse areas of high gelation in the gel system, resulting in a more uniform gel network structure. After molding, sterilization and packaging are performed sequentially. The sterilization process uses appropriate thermal sterilization parameters based on the product form to ensure that the product's microbiological indicators meet food safety requirements. The game theory model uses a range of 300-500 between SO3 and SO5. For periodic rolling solutions, the objective function of the upper-level model The objective function of the lower-level model is to minimize the total amount of alkali injected. With the goal of maximizing gelation uniformity, both methods are used to measure the standard deviation of pH values ​​at various locations within the system. The coupling process reaches Nash equilibrium during iteration, outputting the optimal alkali injection rate command to the alkali gradient micro-injection device.

[0057] It should be noted that the key technologies of this invention include: a thermodynamic-kinetic dual-engine driven neural network model that embeds thermodynamic prior equations into a neural network as structural constraints, enabling the model to still obey physical laws under data-scarce conditions, overcoming the inherent defect of overfitting in small-sample scenarios of pure data-driven models; the neural network model describes system evolution with continuous-time dynamics, enabling interpolation and prediction of states at any time, naturally adapting to variable-step production conditions, while discrete-time step models require redesigning the time embedding when the sampling interval changes, resulting in structural mismatch; a three-subprocess cascade coupling structure that allows the effective component concentration deviation in the supercritical extraction stage to be transmitted to the spray drying and gelation stages through the prediction link, achieving full-process collaborative early warning; and a two-layer game model that incorporates the two mutually constraining objectives of alkali injection efficiency and gelation uniformity into a unified framework for iterative solution through a pH standard deviation coupling term, avoiding the problem of local optima caused by ignoring constraint interactions in single-objective optimization. The synergistic effect of the aforementioned key technologies enables the present invention to achieve real-time closed-loop control of gelation uniformity in multi-subprocess dynamic coupling scenarios, fundamentally making up for the structural defect that human experience judgment cannot perceive the dynamic coupling relationship across subprocesses.

[0058] It should be noted that when there are batch-to-batch fluctuations in the inlet air temperature or feed concentration during the spray drying stage, the residual moisture content of the composite plant powder will shift, thereby altering the swelling degree and gelation initiation rate of konjac glucomannan. If the gelation stage continues with a fixed alkali injection rate and stirring speed, it will lead to localized premature gelation or overall delayed gelation, ultimately resulting in an uneven internal structure of the molded product. The reason for the above technical problem is that moisture migration during spray drying is affected by the inlet air temperature, feed concentration, and atomized particle size. Batch-to-batch fluctuations in any of these parameters will change the residual moisture content of the powder. Furthermore, the gelation rate of konjac glucomannan under alkaline conditions is highly sensitive to the system's water activity; even a small shift in residual moisture content can cause a significant drift in the gelation kinetics curve, which traditional fixed-parameter processes cannot respond to in real time. The usual solution to this technical problem is to perform offline moisture testing on the powder after spray drying and manually revise the alkali addition plan for the subsequent gelation stage based on the test results. However, offline detection inherently has a time delay, typically taking several minutes to tens of minutes from sampling to obtaining results. The kinetic timescale of the konjac glucomannan gelation process is comparable to this, making manual correction impossible before deviations occur in the gelation process. Furthermore, manual judgment relies on the operator's accumulated experience in the coupling relationships between multiple parameters, making it difficult to eliminate systematic errors. This invention effectively solves this technical problem. A thermodynamic-kinetic dual-engine driven neural network model uses the residual moisture content output from the spray drying sub-process as the initial condition for the konjac glucomannan gelation sub-process. Simultaneously with the end of the spray drying stage, the prediction link of the gelation sub-process is re-integrated with the updated initial conditions, without waiting for offline detection results. An online distributed pH electrode array collects pH values ​​at various locations in real time, at a concentration of 300–500. The periodic rolling correction of the gelation state evaluation value and the game model synchronously output the optimal alkali injection rate command enable the alkali addition strategy to be adaptively adjusted before the gelation process deviates from the expected trajectory, fundamentally eliminating the adverse effects of offline detection time delay and human experience judgment error on gelation uniformity.

[0059] Specifically, the principle of this invention is:

[0060] The fundamental reason why this invention can solve the above-mentioned technical problems is that the ordinary differential equation framework of the divine process replaces discrete time step prediction with continuous time dynamics, which can interpolate and estimate the system state at any time. This makes the model naturally adaptable to production conditions with variable step size, rather than relying on discrete state machines with fixed sampling intervals. After the three sets of sub-process ordinary differential equation systems are coupled in series, the effective component concentration output by the supercritical extraction sub-process is directly used as the initial input condition of the spray drying sub-process, and the residual moisture content output by the spray drying sub-process is directly used as the initial condition of the konjac glucomannan gelation sub-process. This series coupling structure reproduces the mass transfer relationship between the three sub-processes in real production at the prediction link level, so that upstream disturbances can be transmitted downward in the prediction channel and trigger early warning in advance, rather than being detected only after macroscopic anomalies occur in the gelation stage.

[0061] The dual-engine fusion mechanism further ensures the physical rationality of the model's predictions. The second engine embeds the Arrhenius rate equation, Fick's second diffusion law, and cross-linking reaction kinetics equation as structural constraints into the right-hand side functions of the ordinary differential equations of each sub-process. The parameters of these prior equations are fixed inputs after being pre-calibrated through independent experiments, which is equivalent to imposing physical boundaries on the model, enabling it to output predictive values ​​that conform to thermodynamic laws even in operating conditions with insufficient training sample coverage. The first engine, the multilayer perceptron, captures the nonlinear residual dynamics that the prior equations fail to describe through residual connections and sparse attention masks. The two are weighted and summed with learnable fusion weights, which are adaptively adjusted according to the current state and the predicted residuals of the second engine. In areas with sufficient data, it favors data-driven approaches, while in areas with scarce data, it favors physical constraints, forming a dynamic balance.

[0062] The gelation state evaluation function linearly integrates konjac glucomannan concentration, normalized pH value, and normalized system temperature with learnable weights into a dimensionless evaluation value. This evaluation value is directly mapped to the low-speed anchor stirring speed control interface, forming a closed-loop feedback. A two-layer game model uses minimizing the total amount of alkali injected and maximizing gelation uniformity as the upper and lower layer objectives. By using the pH standard deviation as a coupling term, it reaches Nash equilibrium during iteration, automatically solving for the dynamic optimal balance between alkali injection efficiency and gelation uniformity, rather than relying on manual experience to set a fixed injection rate. The combined effect of these mechanisms enables real-time control of gelation uniformity in a multi-process dynamic coupling scenario.

[0063] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0064] The specific implementation method of step S01 is as follows.

[0065] After weighing each component according to the formula, the mixture of ingredient A and ingredient B is extracted stepwise. The first step uses supercritical fluid extraction. Extraction of fat-soluble components at an extraction pressure range of 15–35. The temperature range is 35–55℃, and the resulting fat-soluble extract is collected separately. In the second step, 8–12 times the weight of water is added to the remaining material, and the polysaccharide component is extracted using hot water at 80–95℃ for 60–120 minutes. The polysaccharide extract was collected by filtration. The extraction rates of each component were determined through multiple batches of control experiments, and the optimal range was determined by fitting a polynomial surface using the response surface methodology. The response surface methodology involves designing multi-factor, multi-level experiments to fit the polynomial surface relationship between the response values ​​and the factors, in order to find the optimal combination of factors.

[0066] The specific implementation method of step S02 is as follows.

[0067] The yam was steamed and boiled, then combined with the polysaccharide extract obtained from SO1, and finally spray-dried to obtain a compound plant powder. The inlet air temperature range of the spray drying was 150-200℃, and the outlet air temperature range was 70-90℃. The liquid extract was atomized and the water was instantly evaporated in the high-temperature airflow to form powder.

[0068] The specific implementation method of step S03 is as follows.

[0069] After mixing the compound plant powder and konjac powder, an alkaline gradient micro-injection system was used to control the gelation timing. The gelation timing control system consisted of an online distributed pH electrode array and an alkaline gradient micro-injection device. The online distributed pH electrode array was evenly distributed longitudinally along the wall of the mixing container, with an adjacent electrode spacing of 5–15 mm. Output the pH value at each location in real time, denoted as ,in Electrode numbering index. The alkali gradient micro-injection device is based on each position... The alkali injection rate is dynamically adjusted to maintain the system pH within the range of 7.5–8.5. The low-speed anchor stirring speed is set to 5–30 rpm. The pH and rotation speed ranges mentioned above were determined through konjac glucomannan gelation kinetic experiments, collecting gelation time data at different pH and rotation speeds, with the goal of minimizing the probability of premature gelation.

[0070] The specific implementation method of step S04 is as follows.

[0071] The locations collected by S03 By inputting gelation kinetic data into a thermodynamic-kinetic dual-engine driven neural network model, gelation state assessment values ​​are obtained. The model uses ordinary differential equations (ODEs) as its core framework, establishing three sets of ODE systems for the supercritical extraction sub-process, spray drying sub-process, and konjac glucomannan gelation sub-process, respectively. These three systems are coupled in series. The right-hand side function of each ODE system is formed by a weighted fusion of the first and second engines, as specifically described below:

[0072] ;

[0073] In the formula, For the first The dimensionless state vector of the time-series subprocess is used for extracting subprocesses in supercritical conditions. For the spray drying sub-process, For the gelation process, All components have been dimensionless. The output of the first engine, the multilayer perceptron, is the unit and Consistency, that is Quantity based on reference time After normalization, it becomes dimensionless. The set of learnable parameters for a multilayer perceptron. For the output of the thermodynamic constraint terms of the second engine, the units are... Consistent, For the pre-calibrated set of prior equation parameters, For learnable fusion weights, satisfying The data is adaptively adjusted based on the current state and the residual predicted by the second engine, as described below:

[0074] ;

[0075] In the formula, for Activation function, mapping real numbers to interval, This is the dimensionless state vector predicted separately by the second engine. For learnable scalar weights, For learnable scalar bias, Let be the Euclidean distance between the two vectors, which is dimensionless.

[0076] The second engine of the supercritical extraction subprocess introduces the Arrhenius rate equation, which has the following functional form: In the formula For temperature The reaction rate constant at the given time, Pre-exponential factor, unit and Consistent, Activation energy, unit: , is the molar gas constant, with a value of 8.314. , Extraction temperature, unit: Substituting the Arrhenius rate equation into the extraction kinetics, the complete equation for the supercritical extraction subprocess is expressed as follows:

[0077] ;

[0078] In the formula, for Concentration of active ingredients in the extract at any given time, in units of , This represents the initial concentration of the active ingredient, in units of... , This represents the saturated solubility concentration of the active ingredient, in units of... , For reference time, the unit is... Used for time axis normalization. This is the error term in the supercritical extraction subprocess model, dimensionless, left-hand side. The derivative of dimensionless concentration with respect to dimensionless time is given; it is dimensionless and has the right-hand side. middle Multiplying the rate constant by the reference time makes the entire term dimensionless. and All are dimensionless.

[0079] The second engine of the spray drying sub-process introduces Fick's second diffusion law, whose functional form is: In the formula For position ,time The concentration of the substance at that location, in units of , The effective diffusion coefficient is expressed in units of 1000 ppm. , Spatial coordinates, in units of After normalization, the complete equation for the spray drying sub-process is expressed as follows:

[0080] ;

[0081] In the formula, For position ,time Moisture content of materials, in units of , This represents the initial moisture content, in units of... , For reference spatial scale, the unit is . The empirical value is taken as the droplet radius. The denoted scalar is a dimensionless diffusion number. The left side represents the dimensionless partial derivative of moisture content with respect to dimensionless time, which is also dimensionless. All terms on the right side are dimensionless. This is the error term in the spray drying subprocess model, and it is dimensionless. The data was obtained through experimental calibration. The calibration steps included: Step 1, setting different inlet air temperatures and feed concentrations, and collecting drying curve data; Step 2, fitting Fick's second diffusion law using the least squares method to obtain the data. .

[0082] The second engine of the konjac glucomannan gelation process introduces a cross-linking reaction kinetic equation, which is specifically expressed as follows:

[0083] ;

[0084] In the formula, for Concentration of konjac glucomannan at any given time, in units of , The initial concentration of konjac glucomannan is given in units of... , This is the gelation rate constant, in units of... , This is the reaction order, dimensionless, with an empirical value of 1 to 2. The pH value is the real-time value of the system, dimensionless. and These are the experimentally calibrated lower and upper limits of pH, dimensionless. The error term in the gelation subprocess model is dimensionless. The left side represents the dimensionless derivative of concentration with respect to dimensionless time, which is also dimensionless. The right side... Multiplying the rate constant by the reference time makes the entire term dimensionless. and All are dimensionless. and The results were obtained through multiple batches of gelation kinetic experiments.

[0085] The forward propagation of a multilayer perceptron is described in detail below:

[0086] ;

[0087] ;

[0088] In the formula, For the first The output vector of the layer linear transformation, For the first layer The dimensional weight matrix, after being processed by sparse attention masking, has elements below a threshold set to zero. For the first Number of neurons in a layer For the first Number of neurons per layer, number of layers The value ranges from 1 to 5, and the number of neurons per layer ranges from 32 to 128. For the first Layer bias vector, dimension is , For the first Layer activation output vector, residual connection will Add directly to , It is a linear rectified activation function, which takes elements of the input one by one. .

[0089] The ordinary differential equation system of each subprocess is subjected to adaptive step-size iterative integration using the Doppler-Prince method, with fourth-order estimates. With fifth-order estimate The difference is used to estimate the local truncation error, as described below:

[0090] ;

[0091] In the formula, This is the adjusted integration step size, in units of... , This is the current integration step size, in units of , This is the error tolerance, dimensionless, and the default value is... , The dimensionless state vector is the fourth-order Runge-Kutta estimate. The dimensionless state vector is the fifth-order Runge-Kutta estimate. Let be the Euclidean distance between the two, which is dimensionless.

[0092] The three sets of subprocesses are coupled in series as follows: the supercritical extraction subprocess terminates at time... Output dimensionless active ingredient concentration As the initial input condition for the spray drying sub-process, the residual moisture content at the output of the spray drying sub-process is... (Unit is) () as the initial conditions for the gelation process of konjac glucomannan.

[0093] The network loss function is specifically described as follows:

[0094] ;

[0095] In the formula, For the first Subprocess number The model predictions for each sample are dimensionless. To correspond to the measured value, dimensionless. For the first The historical maximum value of the subprocess state variable, dimensionless, used for normalization. For the first Subprocess sample size Pick , , These correspond to the three sub-processes: supercritical extraction, spray drying, and gelation. The total loss value is dimensionless. Training uses an adaptive moment estimation optimizer to update parameters, with an initial learning rate of... The cosine annealing learning rate scheduling is adopted, with a total training round of 200-500 rounds. An early stopping mechanism is triggered when the validation set loss does not decrease for 20 consecutive rounds.

[0096] The specific description of the gelation state evaluation function is as follows:

[0097] ;

[0098] In the formula, This is a dimensionless assessment value for the gelation state. This represents the current concentration of konjac glucomannan, in units of... , The current pH value of the system is dimensionless. and For learnable fusion weights, satisfying Dimensionless The current system temperature, in units of , and These are the lower and upper limits of the system temperature as determined in the experiment, in units of... The two terms on the right are the dimensionless concentration-pH composite term and the dimensionless temperature term, respectively. The sum of the two terms constitutes the dimensionless evaluation value. .when At this time, the speed of the low-speed anchor mixer is reduced to 5-10. And suspend the injection of alkali solution; when At that time, the speed is maintained at 10-20. The alkali injection rate is operating normally; when At that time, the speed increases to 20-30. The alkali injection rate was moderately increased.

[0099] The objective function of the upper-level game theory model is specifically expressed as follows:

[0100] ;

[0101] The specific description of the lower-level objective function is as follows:

[0102] ;

[0103] In the formula, This represents the current alkali solution injection volume, in units of... , This is the baseline alkali solution injection volume, in units of , For dimensionless injection ratio, The standard deviation of pH value at various locations within the system is dimensionless. The standard deviation of the baseline pH value is dimensionless. This is a dimensionless pH standard deviation ratio. The penalty coefficient, dimensionless, was determined through multiple batch experiments with the dual objectives of minimizing the amount of alkali injected and maximizing the uniformity of gelation. The standard deviation of the gelation state assessment value is dimensionless. The mean value of the gelation state assessment is dimensionless. The coefficient of variation is dimensionless. and All are dimensionless optimization objective values. The upper-level constraints are: The lower-level constraint is The two objective functions are coupled through a term. Iteration, upper layer adjusts the alkali injection rate and affects According to the lower level Update all locations And feedback is sent to the upper level, in the range of 300-500. The algorithm iterates periodically until it reaches Nash equilibrium, at which point it outputs the optimal alkali injection rate command. Nash equilibrium represents a stable state where none of the participants have an incentive to unilaterally change their own strategies given that the strategies of the other participants are fixed. In this case, it is a stable solution where the upper alkali injection rate command and the lower gelation uniformity feedback are mutually consistent.

[0104] The specific implementation method of step S05 is as follows.

[0105] The evaluation index for gelation uniformity is Data collected at various locations by an online distributed pH electrode array calculate The positions output by the model calculate and .when If the uniformity is deemed satisfactory, proceed directly to the molding process; when Additional homogenization treatment is then performed, with the homogenization pressure set to 10–30 kcal / kg. Homogenization time 3–10 Then it is shaped, sterilized and packaged.

[0106] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effect of the invention, technicians built a pilot-scale testing environment. Based on a total mass of 100 portions per batch, the following components were weighed according to the formula: Ingredient A (8 portions of Coix seed, 6 portions of Astragalus membranaceus, 5 portions of Codonopsis pilosula, 4 portions of Poria cocos, 4 portions of Euryale ferox, 2 portions of Glycyrrhiza uralensis, 1 portion of Zingiber officinale, totaling 30 portions), Ingredient B (16 portions of Citrus medica, 4 portions of Amomum villosum, totaling 20 portions), Ingredient C (35 portions of Dioscorea opposita), and Ingredient D (15 portions of Konjac flour). The sum of the mass portions of each component was 100 portions. Each herb was proportionally allocated within its respective ingredient group according to the experimentally determined content of active ingredients.

[0107] The specific execution of step S01 is as follows: After mixing ingredient A and ingredient B, place them in a supercritical extraction device and set the extraction pressure to 25. Temperature 45℃ The flow rate is configured according to the material mass ratio, with an extraction rate of 60. The fat-soluble extract was then collected, mainly enriched with volatile oils and fat-soluble flavonoids. The remaining extract was mixed with 10 times its weight of water and hot-extracted at 88°C for 90 minutes. The polysaccharide extract was collected by filtration, and the concentration of solids in the extract was measured and recorded using a refractometer. This concentration was used as one of the input parameters for the spray drying sub-process, as shown in Table 1.

[0108] Table 1 Summary of extraction process parameters and extraction rate for step S01

[0109]

[0110] In step S02, the yam is cooked and gelatinized, then combined with the polysaccharide extract. The solid concentration of the combined liquid is approximately 12%. The mixture is then spray-dried to obtain a composite plant powder. The inlet air temperature is set at 175℃, and the outlet air temperature is controlled at 80℃. The residual moisture content of the resulting powder, determined by the Karl Fischer method, is 4.2%, and the particle size... 68 The residual moisture content of the composite plant powder is used as the initial condition input model for the ordinary differential equation system of the konjac glucomannan gelation process, such as... Figure 2 The figure shows the curves of the change of outlet air temperature and residual moisture content of powder over time during the spray drying process. It can be seen that the fluctuation range of outlet air temperature is controlled within ±2℃ in the steady state stage, and the batch stability of residual moisture content of powder is good.

[0111] In step S03, after mixing the compound plant powder and konjac powder, an appropriate amount of water is added. The gelation timing control system is activated under low-speed anchor stirring. An online distributed pH electrode array is arranged with a total of 8 measuring points, with an adjacent electrode spacing of 10 mm. The pH value at each measuring point is collected in real time. The alkaline solution gradient micro-injection device dynamically adjusts the alkaline solution injection rate according to the pH value at each location to maintain the system pH within the range of 7.8 to 8.2. The initial low-speed anchor stirring speed is set to 20. ,like Figure 3 The figure shows the pH value at each measuring point over time in step S03. It can be seen that the standard deviation of the pH value at each measuring point is controlled by the game theory model. The pH level was consistently maintained at a low level, effectively suppressing excessively large local pH gradients. The pH data collected at each measurement point are shown in Table 2.

[0112] Table 2. pH value record at each measuring point at typical moments during step S03.

[0113]

[0114] In step S04, the pH values ​​and gelation kinetic data at each measuring point are input into a thermodynamic-kinetic dual-engine driven ordinary differential equation network model. The model iteratively integrates the three sets of cascaded coupled ordinary differential equations using the adaptive step-size Dormand-Prince method, with a step size of 300. To periodically output the gelation state assessment value ,like Figure 4 The values ​​shown are the gelation state assessment values. The curve showing the correlation between the low-speed anchor stirring speed and the time-varying curves indicates that when... When the speed approaches 0.8, the low-speed anchor mixer speed automatically decreases to 8. The injection of alkali solution was then paused, and the gelation process smoothly converged. During the gelation process, The adaptive adjustment result stabilized around 0.68. The value stabilized around 0.32, indicating that under the current operating conditions, the combined contribution of konjac glucomannan concentration and pH value to the characterization of the gelation state has a significantly higher weight than that of the temperature term.

[0115] In step S05, the gelation uniformity evaluation index is calculated after gelation is completed. The result was 0.063, which is lower than the uniformity threshold of 0.1, indicating that the uniformity meets the standard and the product can proceed directly to the molding process without additional homogenization treatment. After molding, sterilization and packaging are completed according to the predetermined sterilization parameters. The uniformity score of the internal gel network structure of the final product was rated as excellent by microscopic observation of cross-sections. The evaluation indexes for gelation uniformity at each stage are summarized in Table 3.

[0116] Table 3 Summary of uniformity evaluation indicators at each stage of the gelation process

[0117]

[0118] like Figure 5The figure shows the curve of the alkali injection rate command changing with time in step S03 of the game model. It can be seen that the penalty mechanism imposed by the upper model on the amount of alkali injected causes the alkali injection rate to decrease significantly in the middle and late stages of gelation. Meanwhile, the lower model's drive to maximize the uniformity of gelation keeps the standard deviation of pH value at each measuring point at a low level. The two reach Nash equilibrium in the iteration through the pH standard deviation coupling term. The total amount of alkali injected is saved compared with the fixed rate addition scheme, and the uniformity of gelation does not deteriorate.

[0119] Compared to traditional methods that rely on manual experience to determine the extraction endpoint and gelation timing, this invention represents a fundamental advancement in its technical principles. Manual judgment depends on the operator's qualitative understanding of the coupling relationships between multiple sub-processes, failing to quantify the transmission path of effective component concentration deviations during supercritical extraction on gelation kinetics. This invention, however, incorporates these transmission relationships into a continuous-time kinetic framework using a series-coupled ordinary differential equation system of three sub-processes, enabling numerical prediction of the gelation state at any given time under physical constraints. The embedding of thermodynamic prior equations ensures the model adheres to physical laws even in pilot-scale trials with limited batch sizes, avoiding the prediction distortion caused by insufficient training samples in purely data-driven models. The online distributed pH electrode array and 300... The game theory model with periodic rolling solutions together constitutes a closed-loop control link with high time resolution, which fundamentally eliminates the adverse effects of offline detection time delay on gelation uniformity.

[0120] It should be noted that the variables involved in this invention are explained in detail in Tables 4 and 5.

[0121] Table 4. Variable Explanation Table (Part 1)

[0122]

[0123] Table 5. Variable Explanation Table (Part Two)

[0124]

[0125] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A konjac-based dietary supplement for strengthening the spleen and stomach, characterized in that, It includes ingredients A, B, C and D, wherein ingredient A is composed of coix seed, astragalus, codonopsis, poria, euryale seed, licorice and dried ginger, ingredient B is composed of citron and amomum, ingredient C is yam and ingredient D is konjac flour.

2. The konjac-based spleen-strengthening and stomach-nourishing dietary supplement according to claim 1, characterized in that, The mass fraction of ingredient A is 10-30 parts, the mass fraction of ingredient B is 10-30 parts, the mass fraction of ingredient C is 30-60 parts, the mass fraction of ingredient D is 10-30 parts, and the sum of the mass fractions of all components is 100 parts.

3. A konjac-based dietary supplement for strengthening the spleen and stomach, characterized in that... In ingredients A and B, each herb is distributed in equal proportions within its respective ingredient group according to the experimentally determined content of its active ingredient.

4. A production process for a series of konjac-based spleen-strengthening, stomach-nourishing, and dietary supplements, characterized in that... Includes the following steps: Weigh each component according to the formula, and extract ingredient A and ingredient B in a stepwise sequence. First, extract the fat-soluble components, and then extract the polysaccharide components from the remaining material with hot water. Ingredient C was steamed and then combined with the polysaccharide extract, followed by spray drying to obtain a compound plant powder. The compound plant powder was mixed with ingredient D, and the alkaline solution was injected in a gradient micro-injection through a gelation timing control system, while the pH value at each location was collected simultaneously. The pH value and gelation kinetic data at each location were input into the thermodynamic-kinetic dual-engine driven neural network model to obtain the gelation state evaluation value. Based on the gelation state assessment value and gelation uniformity evaluation index, it is determined whether additional homogenization treatment is needed, followed by molding, sterilization and packaging.

5. The production process method according to claim 4, characterized in that, The specific steps of the stepwise sequential extraction are as follows: First, the mixture of ingredient A and ingredient B is subjected to supercritical fluid extraction. Extraction was performed, and the resulting fat-soluble extract was collected separately. The remaining material was subjected to hot water extraction to extract polysaccharide components, and the polysaccharide extract was collected by filtration.

6. The production process method according to claim 5, characterized in that, The supercritical The extraction pressure, temperature, and time parameters, as well as the temperature and time parameters for hot water extraction, were determined by analyzing multiple batches of control experiments using response surface methodology.

7. The production process method according to claim 6, characterized in that, The gelation timing control system consists of an online distributed pH electrode array and an alkaline gradient micro-injection device. The online distributed pH electrode array is evenly distributed along the longitudinal direction of the mixing container wall and outputs the pH value at each location in real time. The alkaline gradient micro-injection device dynamically adjusts the alkaline injection rate according to the pH value at each location.

8. The production process method according to claim 7, characterized in that, The thermodynamic-kinetic dual-engine driven neural ordinary differential equation network model uses neural ordinary differential equations as its core framework. It establishes a set of ordinary differential equation systems for each of the supercritical extraction subprocess, spray drying subprocess, and konjac glucomannan gelation subprocess, and the three sets of ordinary differential equation systems are coupled in series.

9. The production process method according to claim 8, characterized in that, The right-hand side function of each ordinary differential equation system is composed of a weighted fusion of the first engine and the second engine; the first engine is a multilayer perceptron; the second engine is a preset thermodynamic constraint term. The supercritical extraction subprocess introduces the Arrhenius rate equation, the spray drying subprocess introduces Fick's second diffusion law, and the konjac glucomannan gelation subprocess introduces the cross-linking reaction kinetic equation.

10. The production process method according to claim 9, characterized in that, The thermodynamic-kinetic dual-engine driven neural network model is designed with a gelation state evaluation function. The gelation state evaluation value is calculated based on the concentration of konjac glucomannan, the pH value at each location and the system temperature, and then the low-speed anchor stirring speed is dynamically adjusted.

Images