Multi-objective optimization process intelligent agent control method and device based on food production

Abstract

The invention discloses a control method and device for a multi-objective optimization process agent based on food production. The method comprises the steps that raw material information is acquired and standardized, and raw material feature vectors are formed; constructing a three-dimensional process matrix containing equipment nodes, equipment types and process parameters, and performing embedded representation; training an agent model by using small sample data as approximation of high-fidelity simulation or experiment; performing multi-target iterative optimization by combining a global optimization algorithm, triggering computational fluid mechanics sample filling or experimental verification by combining an uncertainty evaluation strategy, and realizing an active learning closed loop of optimization-sample filling-updating; performing sequential logic and virtual debugging verification on the optimal process parameters through a digital twin system; and versioned issuing and closed-loop feedback of production data are realized through the manufacturing execution system. According to the method, the problems of strong parameter coupling, high experiment cost and poor cross-batch stability of a traditional extraction process are solved, and multi-target Pareto optimization of yield, energy consumption, time consumption and quality is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of food process engineering and intelligent manufacturing, in particular to a control method and device of multi-objective optimization process agent based on food production. BACKGROUND

[0002] The production of liquid food involves multiple unit operations, such as crushing, enzymatic hydrolysis, ultrasonic extraction, filtration, concentration, sterilization, homogenization, drying, filling, etc. The coupling relationship between these processes is complex, and the process parameters such as temperature, pressure, pH, shear, flow rate of each device need to be adjusted by manual experience, which is difficult to balance the extraction rate, energy consumption, processing time, equipment capacity and other multiple objectives. The existing technology relies on experience and point-by-point experiments, which is expensive and insufficient, and it is difficult to obtain the true multi-objective Pareto optimal solution within the affordable cost; the differences in raw materials are not quantified, the optimization-simulation-execution-feedback link is broken, resulting in poor cross-batch stability and long development cycle; at the same time, the existing scheme is usually single-objective or isolated simulation / MES management, lacking a unified mapping of "raw materials-process-target", agent acceleration and closed-loop self-evolution system. SUMMARY

[0003] The present application aims to solve the problems of the prior art, and provides a control method and device of multi-objective optimization process agent based on food production, which can realize one-key optimization of multiple objectives such as "maximum yield, minimum energy consumption and time consumption, and stable quality" under the conditions of multiple unit operations, strong coupling and multi-objective conflict within limited experimental / simulation cost, and can be quickly deployed and closed-loop self-evolved through MES / PLC.

[0004] The purpose of the present application is achieved by the following technical solutions: The present application provides a control method of multi-objective optimization process agent based on food production, comprising the following steps: S1 Obtain the physicochemical characteristic data of the target batch of raw materials and perform standardization processing; S2 Establish a three-dimensional process matrix representing the physical topology structure of the production line, generate a binary mask tensor according to the selection state of the device unit in the three-dimensional process matrix, and splice and fuse the three-dimensional process matrix and the raw material feature vector based on the mask tensor to generate a model input feature vector; S3 Input the input feature vector into a pre-trained agent model to obtain multi-dimensional process index prediction values and predicted uncertainty; the pre-trained agent model is an ensemble learning model or a multi-task deep neural network model obtained by training historical production data and an offline fluid mechanics simulation database; S4 if the prediction uncertainty is lower than a preset threshold, performing a global optimization algorithm based on the predicted value, and performing iterative optimization until convergence and outputting an optimal process configuration; S5 performing virtual verification on the optimal process configuration through a digital twin model; S6 issuing the verified process configuration scheme to a production system for execution, and collecting production data to update the agent model.

[0005] According to an embodiment of the present application, the step S1 is specifically: The physicochemical characteristic data of the target batch of raw materials is obtained, the physicochemical characteristic data includes continuous physicochemical characteristic data and discrete characteristic data; the Z-score standardization processing is performed on the continuous physicochemical characteristic data; the one-hot encoding processing is performed on the discrete characteristic data; the standardized continuous characteristic and the encoded discrete characteristic are spliced in a predetermined order to generate the standardized characteristic vector.

[0006] According to an embodiment of the present application, in the step S2, the three-dimensional process matrix is a tensor structure of N×M×P, where N corresponds to the physical process node, M corresponds to the selected equipment unit connected to each node, and P corresponds to the readable and writable process register address in the PLC controller of each equipment unit.

[0007] According to an embodiment of the present application, the step S2 includes the following sub-steps: S2.1 generating a binary mask tensor with the same structure as the three-dimensional process matrix according to the selection state of the equipment unit in the three-dimensional process matrix, wherein the value of the selected equipment unit and its associated entire process parameter space at the corresponding position in the mask tensor is 1, and the value of the unselected equipment unit and its associated parameter space at the corresponding position is 0; S2.2 splicing the standardized characteristic vector and the expanded three-dimensional process matrix to obtain a spliced vector; S2.3 calculating the Hadamard product of the spliced vector and the expanded binary mask tensor to filter out the features corresponding to the unselected equipment, and generating the model input vector.

[0008] According to an embodiment of the present application, in the step S3, when the uncertainty of the prediction exceeds the preset threshold, the computational fluid dynamics simulation is automatically triggered to perform physical field sampling, and the sampling result is used as the true value to update the agent model. According to an embodiment of the present application, in the step S4, the global optimization algorithm is one of genetic algorithm, particle swarm optimization algorithm, differential evolution algorithm, reinforcement learning or Bayesian optimization.

[0009] The present application also provides a control device of a multi-objective optimization process intelligent agent based on food production, comprising: a data acquisition and standardization module configured to obtain characteristic data of a target batch of raw materials through a laboratory information management system, and to standardize and encode the characteristic data to generate a standardized characteristic vector of the target batch of raw materials; a three-dimensional process matrix construction module configured to establish a three-dimensional process matrix representing the physical topology of the production line; an intelligent decision-making core module in communication connection with the data acquisition and standardization module and the three-dimensional process matrix construction module, containing a global optimization algorithm engine, a feature mask encoder and a proxy model, and configured to realize process optimization based on mask mapping and active learning; a simulation verification module connected with the intelligent decision-making core module, and configured to perform virtual simulation and verification of control logic, timing and performance of the optimal process configuration scheme through a digital twin model; an execution and feedback module containing an MES terminal and a PLC control network, and configured to issue the verified process configuration scheme to the production system for execution, and to collect production data to update the proxy model.

[0010] The beneficial effects of the present application are: The three-dimensional process matrix realizes rapid migration across raw materials and automatic process path and parameter generation, the proxy model replaces expensive evaluation, the uncertainty drives the sampling, and the efficiency and accuracy are considered, the product development and process adjustment time is shortened, the optimization scheme can maximize the extraction rate while reducing energy consumption and time consumption, which helps to reduce production cost, the digital twin and simulation verification ensure the feasibility and quality stability of each optimization scheme, and reduce trial and error cost, the MES version realizes one-key issuance of the optimal scheme, reduces the risk of manual input errors, and the optimization scheme can be continuously updated and improved through data feedback of the actual production process, gradually improving the intelligentization and automation level of the production process. BRIEF DESCRIPTION OF DRAWINGS

[0011] The present application will be further described below in conjunction with the drawings and examples; Figure 1 The figure is a multi-objective optimization process intelligent agent system architecture of the present application; Figure 2 The figure is a main control method flow chart of the present application; Figure 3 The figure is a three-dimensional process matrix of the present application; Figure 4 The figure is an active learning flow chart driven by uncertainty; Figure 5 The figure is a control device schematic diagram of a multi-objective optimization process intelligent agent based on food production provided by the present application. DETAILED DESCRIPTION

[0012] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0013] like Figure 1 As shown, the control method of the multi-objective optimization process intelligence agent based on food production of the present invention includes: 1. Data Collection and Standardization: 1.1 Raw Material Data Acquisition: After the raw materials enter the production system, relevant data is first acquired through the Laboratory Information Management System (LIMS). This data includes: Physicochemical properties: continuous variables, such as moisture content (%), sugar content (Brix), total acid (TA), pH value, etc. Ingredient content: a continuous variable, such as the initial concentration of the target active ingredient (polyphenols, flavonoids); Source information: discrete variables, including raw material variety, place of origin, harvesting season and batch number.

[0014] 1.2 Data Cleaning and Standardization: The collected data is cleaned, including outlier detection, missing value imputation, and unit conversion, to generate raw material feature vectors. For continuous features, Z-score standardization is used to eliminate dimensional differences, making the parameters of different raw materials comparable. For discrete features, one-hot encoding is used for transformation, such as encoding discrete features like raw material origin and harvest season, ensuring that the features can be processed by intelligent algorithms.

[0015] 1.3 The standardized continuous features and the discrete sparse vectors encoded by one-hot encoding are concatenated in sequence to generate a standardized raw material feature vector, which includes both continuous and discrete raw material features.

[0016] 2. Definition of three-dimensional matrix: Establish a three-dimensional process matrix with dimensions N×M×P, which includes "equipment nodes, equipment types, and process parameters".

[0017] The matrix is not a simple data array, but a digital mapping container of the physical topology of the production line. Each element in the matrix is uniquely mapped to the underlying control network of the workshop (such as OPC UA address space). The X-axis (device node N) includes process sequence nodes for raw material preprocessing, extraction, filtration, purification, concentration, sterilization, drying, packaging, etc. The Y-axis (device type M) represents different device types available for each node, such as reverse flow extraction, ultrasonic extraction, supercritical extraction, etc. The Z-axis (process parameter P) refers to the adjustable control parameters of each device, including temperature, pH, pressure, ultrasonic power, flow rate, residence time, motor speed, etc.

[0018] The three-dimensional process matrix is shown in Figure 3 Through this mapping, any modification of the matrix by the algorithm is equivalent to a reconstruction of the physical production line control logic, thus solving the engineering problem of strong coupling between device selection and parameter control that traditional optimization algorithms cannot directly handle.

[0019] Sparse and constraint mechanism: mutual exclusion mask is preset inside the matrix. When a device m i is selected at a node n, the matrix automatically activates the Z-axis parameter space corresponding to the device, while locking the parameter space of other devices m j≠i at the same node (setting it to an invalid value or zero), forming a structured process path decision tree. When the global optimization algorithm tries to evaluate a process path, the mask tensor will force to shield all neuron connections of unselected devices.

[0020] This method can significantly reduce the computational dimension of the model, accelerate the convergence, and more importantly, ensure the completeness and uniqueness of the feature vector of the input proxy model in the physical sense, which is not achievable by general machine learning methods.

[0021] 3. Global optimization algorithm generation and feature mask mapping 3.1 Solution vector encoding: map the device selection (Y-axis) and process parameters (Z-axis) in the three-dimensional process matrix to the solution vector structure that the global optimization algorithm can handle. If using genetic algorithm (GA), encode the matrix parameters as chromosomes; if using particle swarm optimization (PSO), encode the matrix parameters as particle position vectors; if using reinforcement learning (RL), encode the matrix state as state space. The system establishes a structure-aware parameter update mechanism to ensure that the discrete changes in device selection and the continuous changes in process parameters maintain the consistency of the physical topology during the algorithm iteration process.

[0022] 3.2 Feature flattening and exclusive mask processing: Set an encoder to splice and flatten the sparse "three-dimensional process matrix" and "raw material feature vector", generating a fixed-length input feature vector X input . Since the three-dimensional matrix contains a large number of mutually exclusive device paths (i.e. choosing device A cannot choose device B), the exclusive mask mechanism is introduced, the spliced vector and the mask tensor are subjected to Hadamard product (Element-wise Product) operation, and the Flatten operation is performed to generate a fixed-length input feature vector X with complete physical meaning and sparsity input . When the genetic algorithm tries to evaluate a process path, the mask tensor will force to shield all neuron connections of the unselected devices.

[0023] This method can significantly reduce the computational dimension of the model, accelerate the convergence, and more importantly, ensure the completeness and uniqueness of the feature vector input into the proxy model in the physical sense, which is not achievable by general machine learning methods.

[0024] 4, Proxy model prediction and active learning Use ensemble learning models (such as XGBoost) or multi-task deep neural networks (Multi-Task DNN) as proxy models. The input is X input , including raw material attributes and process attributes; the output is the predicted multi-dimensional vector Y (including yield, energy consumption, time consumption, and quality) and the uncertainty of the predicted value.

[0025] Proxy model pre-training: Select Kriging, radial basis function (RBF), etc. to generate training samples according to existing CFD or experimental data, perform 10-fold cross-validation to ensure the stability and accuracy of the model. Finally, the proxy model can give a fast and accurate prediction for unknown working conditions and provide the prediction variance of the model as an uncertainty measure.

[0026] Uncertainty evaluation: During optimization, when the prediction uncertainty of some parameter combinations exceeds the preset threshold, the system automatically calls the fluid dynamics simulation software (CFD) to generate a physical simulation task for the specific process matrix parameters. If the uncertainty ≤ preset threshold, the prediction result is considered reliable and the next step of fitness calculation is entered. The uncertainty threshold value ranges between 3% and 15%, preferably between 5% and 10%.

[0027] Sample calculation and self-correction: After each sample, the CFD simulation results are returned to the system, added to the training set, and the proxy model performs incremental learning. After model update, the current population is predicted again.

[0028] 5, Multi-objective evolutionary optimization and convergence criterion Objective function design: The optimization objectives include extraction rate (Y), energy consumption (E), time consumption (T), and quality (Q). A multi-objective fitness function is designed to comprehensively consider these objectives, using weighted coefficients (ω1, ω2, ω3, ω4) to balance the relationships between the objectives.

[0029] in, For extraction rate, For energy consumption, Because it takes time, For quality, The corresponding maximum or minimum boundary conditions, ω1, ω2, ω3, and ω4 are the corresponding weight coefficients, which can be dynamically adjusted according to the production strategy.

[0030] Intelligent optimization algorithms with global search capabilities (including but not limited to genetic algorithms, particle swarm optimization algorithms, differential evolution algorithms, or deep reinforcement learning) are used to traverse the high-dimensional mixed space of process parameters to find the parameter combination that optimizes the objective function.

[0031] Taking the genetic algorithm as an example, the convergence determination is as follows: Check whether the current population meets the convergence conditions (such as reaching the maximum number of iterations or fitness stagnation). If it has not converged, perform the selection, crossover, and mutation operations of the genetic algorithm to generate the next generation of the three-dimensional process matrix population, and return to step 3 to repeat the process. If it has converged, output the Pareto optimal solution set and proceed to the next verification stage.

[0032] Taking genetic algorithms as an example, the flowchart of uncertainty-driven active learning is shown below. Figure 4 .

[0033] 6. Digital twin simulation verification Digital Twin Construction: The digital twin system is built on a multiphysics simulation platform for joint simulation of fluid flow, heat transfer, and mass transfer behavior during the extraction process. The multiphysics simulation platform can be COMSOL Multiphysics, ANSYS Fluent, OpenFOAM, or an equivalent platform, and can work collaboratively with process-level and control-level simulation platforms (such as MATLAB / Simulink). A virtual production line model is established, and the optimal three-dimensional process matrix output by the GA is parsed into control commands for the digital twin for virtual testing. Testing includes: process cycle time and capacity assessment, energy and material consumption prediction, quality index simulation, and control timing and logic verification (such as valve opening and closing timing, PID control logic). Only solutions that pass digital twin verification are marked as "executable versions." If verification fails, constraints are fed back to the GA engine for re-optimization.

[0034] 7. Closed-loop control and iteration MES system integration: Optimal formula and process parameters are issued to the PLC register of the production equipment through the Manufacturing Execution System (MES), driving the action of frequency converters, regulating valves and other actuators to realize real-time deployment of process parameters. Each process formula generates a version number and an effective time in the MES, and is subject to electronic signature audit.

[0035] Production data backflow: During the production process, the MES collects real-time process data and quality indicators and feeds these data back to the optimization system for subsequent model retraining and optimization.

[0036] Example 1: Multi-objective process optimization for flavonoid extraction from apple peel residue This example applies the above multi-objective optimization process agent for food production to optimize the extraction of flavonoid active substances from apple peel residue, a byproduct of apple juice processing. Apple peel residue is rich in flavonoids, but due to the different origins and ripeness of apples, the cellulose structure of the peel residue varies significantly, severely affecting the dissolution efficiency. Traditional fixed processes cannot balance high extraction rate and low energy consumption.

[0037] The specific implementation steps of this example are as follows: 1. Collection and standardization of raw material data: When the raw material of a certain batch number of apple peel residue enters the production line, the agent first obtains the physicochemical data of the batch of apple peel residue through the LIMS system interface and performs preprocessing, including the following steps: Continuous feature collection: Focus on collecting raw material particle size, moisture content, and peel residue cellulose content.

[0038] Discrete feature collection: Collect classification label information such as raw material origin (e.g. "Dangshan, Anhui") and harvest season (e.g. "late maturity") bound to this batch.

[0039] Data preprocessing: For continuous features (particle size, moisture content, fiber content), Z-score standardization is used, with the formula as follows:

[0040] Where x is a certain physicochemical index of the sample, μ is the average value of the physicochemical index of all samples, and σ is the standard deviation of the physicochemical index of all samples.

[0041] For discrete features (origin, season), One-Hot Encoding is used to generate a sparse vector. The standardized continuous features and the discrete feature sparse vector generated by One-Hot Encoding are concatenated in order to generate the standardized raw material feature vector V pearThe feature vector includes a raw material continuous type feature and a raw material discrete type feature.

[0042] 2. Three-dimensional matrix construction of the Sydney extraction process The system initializes a three-dimensional process matrix with dimensions N x M x P, where N corresponds to the physical section nodes of the production line, M corresponds to the optional equipment units physically connected under each node, and P corresponds to the readable and writable process register addresses in the PLC controller of each equipment unit. In this embodiment, the focus is on intelligent configuration for the "extraction section (Node 2)": Device node (X-axis): locked at Node 2 (extraction section).

[0043] Device type (Y-axis): the agent selects between "dynamic counter-current extraction tank" and "high-power ultrasonic extraction tank". Suppose the genetic algorithm (GA) activated "ultrasonic extraction tank (Device M=1)" in this round of optimization.

[0044] Process parameters (Z-axis): activate the 4 key adjustable parameter spaces corresponding to this device: ultrasonic power (P1): range [100, 800] W; ethanol concentration (P2): range [40, 100] %; extraction temperature (P3): range [30, 60] ℃; solid-liquid ratio (P4): range [1:5, 1:50].

[0045] Generate a binary mask tensor with the same structure as the three-dimensional process matrix, where the selected device units and their associated entire process parameter spaces have a value of 1 at the corresponding positions in the mask tensor, and the unselected device units and their associated parameter spaces have a value of 0 at the corresponding positions.

[0046] 3. Agent model prediction and active optimization Feature mapping: map the raw material feature vector V pear After vector splicing with the above three-dimensional matrix through the mutual exclusion mask mechanism, perform Element-wise Product operation to filter out the feature interference of invalid physical paths, and input the pre-trained XGBoost agent model.

[0047] Multi-objective optimization: the agent model outputs three predicted target values: extraction rate ( ), energy consumption ( ), and ethanol consumption ( ). Genetic algorithm performs iterative optimization based on the following fitness function:

[0048] where is the maximum extraction rate, is the minimum energy consumption, is the minimum ethanol consumption.

[0049] Intelligent Decision Making: Given the high cellulose content (>15%) detected in this batch of raw materials, the surrogate model predicts that the traditional "low power + high temperature" extraction method will not achieve sufficient extraction efficiency. In this embodiment, a genetic algorithm (GA) is preferably used: the algorithm iterates through selection, crossover, and mutation operations. With a population size of 100, a crossover probability of 0.8, and a mutation probability of 0.05, 120 generations of iterative searching are performed. The genetic algorithm converges to a new set of parameters: significantly increasing the ultrasonic power to 750W while simultaneously reducing the temperature to 45°C, controlling the ethanol concentration to 65%, and controlling the material-to-liquid mass-to-volume ratio to 1:20. At this point, the prediction uncertainty given by the surrogate model is below the threshold, indicating the feasibility of the proposed solution. The uncertainty threshold is set based on the allowable error of the process system, and in this embodiment, it is preferably approximately 10%. Ensuring both prediction reliability and avoiding CFD computational overload are key technical parameters for the intelligent process optimization efficiency of this invention.

[0050] 4. Digital Twin Verification and Closed-Loop Execution Virtual verification: The optimal parameter combination (750W, 65%, 45℃, 1:20) was input into a digital twin constructed in the COMSOL Multiphysics platform. Simulation results show that at this power, the tank vibration frequency is within a safe range, and there is no safety hazard caused by excessively rapid ethanol evaporation.

[0051] Physical execution: The MES system generates the recipe version. Parameters are written to the PLC address via the OPC UA protocol.

[0052] 5. Comparison of Implementation Results Compared with traditional fixed empirical parameters (500W, 80%, 60℃, 1:40), the extraction rate is improved: the total flavonoid yield in pear peel residue increases from 1.2% to 1.45%; energy consumption is reduced: due to the lower extraction temperature and targeted shortening of time, the energy consumption per batch decreases by 18%; solvent is saved: the concentration of ethanol used is reduced, and the solvent consumption is reduced by 12% by optimizing the material-liquid ratio.

[0053] Example 2: Multi-objective dynamic optimization of Polygonatum polysaccharide extraction process Polygonatum ( Polygonatum sibiricum Polysaccharides are the core active ingredients in functional beverages and oral liquids. Traditional extraction methods typically use fixed parameters (such as constant high temperature and long-term decoction), which not only consumes a lot of energy but also makes it difficult to cope with batch-to-batch differences in raw materials from different origins, often resulting in large fluctuations in extraction rates and severe degradation of heat-sensitive components.

[0054] This embodiment utilizes the aforementioned multi-objective optimization process intelligence agent to perform "one-click" dynamic optimization for the Polygonatum polysaccharide extraction process. The specific steps are as follows: 1. Collection and standardization of multi-source raw material data Dried rhizomes of Polygonatum were selected as raw materials, and samples were collected from 10 batches from major domestic producing areas, including southeastern Guizhou, Zhaotong in Yunnan, Loudi in Hunan, and Huangshan in Anhui. Upon warehousing of the raw materials, key physicochemical data were entered into the LIMS system: Continuous characteristics: moisture content (%), crude polysaccharide content (%), initial pH value, reducing sugar content, etc. Discrete characteristics: origin of raw materials (e.g., "southern Guizhou"), harvest season (e.g., "autumn harvest").

[0055] Data preprocessing: For continuous features, the Z-score standardization formula is used to eliminate the influence of dimensions.

[0056] Where x is a certain physicochemical index of the sample, μ is the average value of the physicochemical index of all samples, and σ is the standard deviation of the physicochemical index of all samples.

[0057] For discrete features (origin, season), one-hot encoding is used to generate sparse feature vectors.

[0058] The standardized continuous features and the discrete sparse vectors encoded by one-hot encoding are concatenated sequentially to generate the standardized raw material feature vector V. polygonatum .

[0059] 2. Three-dimensional matrix construction and initialization of the extraction process: The system establishes an N×M×P three-dimensional process matrix. This embodiment focuses on the intelligent configuration of the "polysaccharide extraction section". Matrix structure definition: X-axis (equipment node): locked at the extraction section.

[0060] Y-axis (equipment type): Activate the "ultrasonic assisted extraction tank" according to the production line configuration.

[0061] Z-axis (process parameters): Maps the four key adjustable control parameters of the equipment: ultrasonic power (P1): range [100, 800] W; extraction time (P2): range [5, 30] min; ethanol concentration (P3): range [40, 100]%; extraction temperature (P4): range [30, 60] ℃.

[0062] Matrix initialization strategy: An initial matrix library is constructed based on historical experimental datasets (containing the extraction effects of raw materials from different origins under different processes). For missing process points, Design of Experiments (DOE) methods (such as Box-Behnken design) are used to fill in the initial data for pre-training the agent model, ensuring that the agent has basic predictive capabilities in the initial stage.

[0063] 3. Agent model prediction with multi-objective evolutionary optimization The raw material feature vector V polygonatum Spliced with a three-dimensional process matrix, input the agent model.

[0064] Multi-objective optimization setting: Set three optimization objectives: Maximize extraction rate (Y): Improve the utilization rate of core raw materials. Minimize energy consumption (E): Reduce power and steam consumption. Minimize time consumption (T): Improve production line turnover efficiency. At the same time, introduce a penalty function to prevent component degradation: When the temperature > 80℃ and the time > 20 min, the fitness function gives heavy penalties to simulate the heat-sensitive protection mechanism.

[0065] Intelligent optimization process: A batch of "Zhaotong, Yunnan" raw materials, LIMS data shows that it has tight texture and low water content. Traditional pain points: Traditional fixed process (60℃, 60min) will cause the outer layer of polysaccharides to be gelatinized, hindering the internal dissolution, and long time high temperature will cause partial polysaccharide hydrolysis. Intelligent agent decision: After receiving the characteristics of the raw material, the process intelligent agent uses the agent model as the objective function evaluator, and performs global search in the process parameter space through genetic algorithm. In the search process, the agent model outputs the prediction values of extraction rate, energy consumption, etc., and calculates the corresponding uncertainty index. The uncertainty threshold is set according to the allowable error of the process system, which is preferably about 8% in this embodiment. For parameter combinations with uncertainty higher than the preset threshold, the system automatically triggers CFD simulation for high-precision recalculation, and returns the results to the agent model. Genetic algorithm converges to a new set of parameters under the guidance of the agent model: high ultrasonic power (700W): use strong mechanical waves to break the tight fiber structure and speed up dissolution; low temperature (45℃): avoid polysaccharide thermal degradation; short time (18min): achieve target extraction rate with significantly reduced energy consumption.

[0066] 4. Digital twin and MES closed-loop execution logic verification: The optimal parameter combination (700W, 18min, 45℃, 60% ethanol) is verified by digital twin. The system confirms that the high-power ultrasonic will not cause the tank body resonance frequency to exceed the standard, and the 18min beat can be perfectly matched with the downstream concentration section without the risk of blockage. Closed-loop issuance: The MES system generates a batch formula. Parameters are automatically written into the PLC register.

[0067] 5. The implementation effect solves the three major problems of polysaccharide extraction from Rhizoma Polygonati: improving quality and stable yield: the standard deviation of extraction rate between batches of raw materials from different places is reduced by 65%, realizing "thousand materials and thousand skills, quality normalization". Cost reduction and efficiency improvement: compared with the traditional fixed process, the extraction time of a single batch is shortened by 40% (from 30 min to about 18 min), and the comprehensive energy consumption is reduced by 25%. Active retention: due to the avoidance of long-term high temperature, the average molecular weight distribution of polysaccharide in the extract is more concentrated, and the retention rate of heat-sensitive functional ingredients is increased by 15%.

[0068] Example 3: Multi-objective dynamic regulation of fresh milk UHT ultra-high temperature instantaneous sterilization The ultra-high temperature instantaneous sterilization (UHT) of fresh milk is the core link of liquid milk production. The traditional production usually adopts the "over-sterilization" strategy (such as fixedly setting 137℃ for 4 seconds) to ensure the shelf life safety, but this often leads to the increase of heat-sensitive indicators such as furosine, produces "cooking flavor", and the energy waste is serious. In addition, as the production proceeds, milk scale (Fouling) will form on the surface of the heat exchanger, leading to the dynamic attenuation of the heat transfer coefficient, and it is difficult to maintain constant sterilization intensity with fixed parameters.

[0069] This embodiment applies the above intelligent agent system to optimize the dynamic balance of "safety-quality-energy consumption" for the UHT section. The specific steps are as follows: 1. Real-time acquisition and standardization of raw milk data After each batch of raw milk enters the milk storage tank, the system automatically obtains the physicochemical indicators of this batch from LIMS: Key physicochemical indicators: fat content (Fat%), protein content (Protein%), non-fat milk solids (SNF), initial total number of colonies (CFU / mL), acidity (°T).

[0070] Data preprocessing: Z-score standardization is performed on the above continuous variables. "Production running time" is introduced as a special input feature, because as the running time increases, the fouling layer thickens, and the thermal efficiency will decrease nonlinearly.

[0071] 2. Construction of three-dimensional matrix of UHT sterilization section The system initializes a three-dimensional process matrix with dimensions N×M×P, and locks the "sterilization section (Node X=5)".

[0072] X-axis (device node): sterilization section.

[0073] Y-axis (device type): the system activates "tubular sterilization machine B group (Device Y=2)" according to the shift production plan.

[0074] Z-axis (process parameters): Map the key adjustable parameters in the PLC of the device: sterilization temperature (P1): range [135.0, 142.0] ℃; material flow (P2): range [5000, 12000] L / h; homogenization pressure (P3): range [15, 25] MPa; back pressure (P4): range [2.0, 4.0] bar.

[0075] 3. Predict the proxy model considering the fouling effect, sequentially splice "raw milk physicochemical data features + running time" and "three-dimensional process matrix", and input the proxy model.

[0076] Multi-objective optimization settings: Maximize sterility assurance (F0 value): ensure F0 > 5.0 to meet commercial sterility requirements; minimize heat damage (furosine value): ensure furosine < 12 mg / 100g protein to maintain fresh taste; minimize steam consumption (E): reduce steam consumption per ton of milk.

[0077] Uncertainty-driven active sampling: After the device has been running continuously for 4 hours, due to fouling on the heat exchange tube wall, the prediction uncertainty of the proxy model for the "actual material heating temperature" exceeds the threshold value. The uncertainty threshold is set according to the allowable error of the process system, and in this embodiment, it is preferably about 5%. The system automatically triggers CFD simulation to simulate the "pipe wall fouling thermal resistance model" under the current protein content. The simulation results show that the current fouling layer has caused a 5% decrease in heat transfer efficiency. The system feeds this "true value" back to the model, which immediately adjusts the prediction logic and no longer blindly trusts the set temperature, but predicts the actual temperature obtained by the fluid.

[0078] 4. Evolutionary optimization and decision-making Based on the corrected model, the genetic algorithm (GA) generates the optimal process parameter combination: Original strategy: set 137℃, flow rate 8000 L / h.

[0079] New strategy: detect that the microbial background of this batch of raw milk is low and has good heat stability; the agent decides to lower the sterilization temperature to 135.5℃ and increase the flow rate to 8500 L / h.

[0080] Expected results: under the premise of ensuring F0 compliance, furosine generation is reduced by 15% and production capacity is increased by 6%.

[0081] 5. Digital twin verification and PLC issuance Virtual verification: the digital twin simulates the whole process of fluid passing through the holding tube.

[0082] Verification point 1: confirm that the F0 value calculated by the integral of the fluid centerline temperature meets the safety standard.

[0083] Verification point 2: Verify that the pressure drop of the pipeline under high flow rate does not exceed the lift limit of the pump. The maximum allowable lift of the delivery pump is about 30 m of water column, corresponding to a pressure of about 0.30 MPa. Considering the safety redundancy of industrial operation, the system sets the upper limit of the effective allowable pressure drop of the pump to 80% of the maximum lift, i.e. about 0.24 MPa.

[0084] Closed-loop execution: After verification, the MES generates a formula and writes it to the Siemens S7-1500 PLC through OPC UA, updating the PID set value of the steam regulating valve to 135.5℃; the frequency of the feed variable frequency pump is updated, corresponding to a flow rate of 8500 L / h.

[0085] 6. Implementation effect After introducing the method, the product quality is improved: the content of furanosine in the finished product is reduced by an average of 12%-18%, significantly improving the "freshness" taste of the product and eliminating the burnt taste caused by overheating; the running period is extended: by dynamically adjusting the homogenization pressure and temperature difference, the fouling rate of the heat exchanger is slowed down, and the single continuous production time (CIP cycle) is extended by 1.5 hours; energy saving and consumption reduction: avoid over-heating, reduce steam consumption per unit product by 8%.

[0086] Example 4: Process optimization alternative based on particle swarm algorithm (PSO) In order to verify the universality of the "three-dimensional process matrix + surrogate model" architecture proposed in the present application, the multi-objective particle swarm optimization algorithm (MOPSO) is used instead of the genetic algorithm in the previous examples to optimize the same snow pear skin residue extraction process.

[0087] The difference between the specific implementation methods is: Encoding method: encode the equipment selection (Y-axis discrete variable) and process parameters (Z-axis continuous variable) in the three-dimensional process matrix as the position vector of "particles". Among them, the discrete variable adopts the binary particle swarm (BPSO) logic, and the speed is mapped to the selection probability of [0,1] through the Sigmoid function; the continuous variable adopts the standard PSO logic.

[0088] Update mechanism: instead of performing crossover and mutation operations, the speed and position of the particles are updated according to the individual historical optimal solution (Pbest) and the global optimal solution (Gbest).

[0089] Speed update formula:

[0090] Position update formula:

[0091] In the above speed update formula and position update formula, and respectively the velocity component and the position component of the i-th particle in the t-th iteration in the d-th dimension; is the inertia weight, used to balance the global search and local exploitation ability of the algorithm; and is the learning factor (acceleration constant), representing the step size of the particle towards and the learning step size weight; and is a random number subject to a uniform distribution in [0, 1].

[0092] Optimization result: under the same agent model and mutual exclusion mask mechanism, the PSO algorithm can also converge to the Pareto front. Tests show that compared with the genetic algorithm, the convergence speed of PSO in handling continuous parameters (such as temperature, power) is improved by about 15%, proving that the system architecture proposed by the present application can be compatible with multiple global optimization algorithms.

[0093] Those skilled in the art should understand that although the multi-objective optimization process is described in detail in the present embodiment taking the genetic algorithm (Genetic Algorithm, GA) as an example, the present application is not limited thereto. Without departing from the core idea of the present application (i.e. closed loop based on three-dimensional matrix structure and agent model), any heuristic algorithm or machine learning algorithm with global optimization ability can replace the genetic algorithm.

[0094] Specifically, the alternative algorithms include but are not limited to: Swarm intelligence algorithms: such as particle swarm optimization algorithm (PSO), ant colony algorithm (ACO), artificial bee colony algorithm (ABC), grey wolf optimization algorithm (GWO); Evolution algorithm variants: such as differential evolution algorithm (DE), evolutionary strategy (ES), simulated annealing algorithm (SA); Learning-based algorithms: such as deep reinforcement learning (DRL), deep Q network (DQN), proximal policy optimization (PPO); Model-based algorithms: such as Bayesian optimization (BO), tree-structured Pareto estimation (TPE).

[0095] Any technical solution that uses the above algorithms to iteratively update the three-dimensional process matrix described in the present application, and combines the mutual exclusion mask mechanism and the agent model for evaluation, should be covered within the protection scope of the present application.

[0096] The present application embodiment also provides a control device of a multi-objective optimization process intelligent agent based on food production, such as Figure 5As shown, the device comprises a data acquisition and standardization module 301, a three-dimensional process matrix construction module 302, an intelligent decision-making core module 303, a simulation verification module 304, and an execution and feedback module 305.

[0097] The data acquisition and standardization module 301 is configured to obtain characteristic data of a target batch of raw materials through a laboratory information management system, and to perform standardization and coding processing on the characteristic data to generate a standardized characteristic vector of the target batch of raw materials. The three-dimensional process matrix construction module 302 is configured to establish a three-dimensional process matrix representing the physical topology of the production line. The intelligent decision-making core module 303 is in communication connection with the data acquisition and standardization module and the three-dimensional process matrix construction module, and comprises a global optimization algorithm engine, a feature mask encoder, and a proxy model, and is configured to implement process optimization based on mask mapping and active learning. The simulation verification module 304 is connected with the intelligent decision-making core module, and is configured to perform virtual simulation and verification of control logic, timing, and performance of the optimal process configuration scheme through a digital twin model. The execution and feedback module 305 comprises an MES terminal and a PLC control network, and is configured to issue the verified process configuration scheme to the production system for execution, and to collect production data to update the proxy model.

[0098] Although the embodiments of the present application have been shown and described above, it should be understood that these embodiments are only exemplary and do not limit the scope of protection of the present application. Those skilled in the art can make various changes, modifications, replacements or adjustments to the above embodiments within the scope defined by the present application. It should be noted that although the embodiments are described in the field of food raw materials, the ideas and methods are also applicable to other process optimization fields with similar characteristics.

Claims

1. A control method for a multi-objective optimization process intelligent agent based on food production, characterized in that, Includes the following steps: S1. Obtain the physicochemical characteristic data of the target batch of raw materials and perform standardization processing; S2. Establish a three-dimensional process matrix representing the physical topology of the production line. Generate a binary mask tensor based on the selection state of equipment units in the three-dimensional process matrix. Based on the mask tensor, concatenate and fuse the three-dimensional process matrix with the raw material feature vector to generate the model input feature vector. S3 Input the input feature vector into the pre-trained surrogate model to obtain the predicted values ​​of multi-dimensional process indicators and the uncertainty of the prediction; the pre-trained surrogate model is an ensemble learning model or a multi-task deep neural network model obtained by training through historical production data and offline fluid dynamics simulation database. S4 If the prediction uncertainty is lower than a preset threshold, then a global optimization algorithm is executed based on the prediction value, and iterative optimization is performed until convergence and the optimal process configuration is output. S5 uses a digital twin model to virtually verify the optimal process configuration; S6 will send the verified process configuration scheme to the production system for execution and collect production data to update the proxy model.

2. The control method for a multi-objective optimization process intelligent agent based on food production according to claim 1, characterized in that, Step S1 specifically involves: acquiring the physicochemical characteristic data of the target batch of raw materials, the physicochemical characteristic data including continuous physicochemical characteristic data and discrete characteristic data; performing Z-score standardization on the continuous physicochemical characteristic data; performing one-heat encoding on the discrete characteristic data; and concatenating the standardized continuous features and the encoded discrete features in a predetermined order to generate the standardized feature vector.

3. The control method for a multi-objective optimization process intelligent agent based on food production according to claim 1, characterized in that, In step S2, the three-dimensional process matrix is ​​an N×M×P tensor structure, where N corresponds to the physical section node of the production line, M corresponds to the optional equipment unit physically connected under each node, and P corresponds to the readable and writable process register address in the PLC controller of each equipment unit.

4. The control method for a multi-objective optimization process intelligent agent based on food production according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.1 Generate a binary mask tensor with the same structure as the three-dimensional process matrix based on the selection status of the equipment units in the three-dimensional process matrix, wherein the value of the selected equipment unit and its associated entire process parameter space at the corresponding position in the mask tensor is 1, and the value of the unselected equipment unit and its associated parameter space at the corresponding position is 0. S2.2 The standardized feature vector is concatenated with the expanded three-dimensional process matrix to obtain the concatenated vector; S2.3 Calculate the Hadamard product of the concatenated vector and the expanded binary mask tensor, filter out the features corresponding to the unselected devices, and generate the model input vector.

5. The control method for a multi-objective optimization process intelligent agent based on food production according to claim 1, characterized in that, In step S3, when the uncertainty of the prediction exceeds a preset threshold, the computational fluid dynamics simulation is automatically triggered to perform physical field resampling, and the resampling result is used as the true value to update the surrogate model.

6. The control method for a multi-objective optimization process intelligent agent based on food production according to claim 1, characterized in that, In step S4, the global optimization algorithm is one of the following: genetic algorithm, particle swarm optimization algorithm, differential evolution algorithm, reinforcement learning, or Bayesian optimization.

7. A control device for a multi-objective optimization process intelligent agent based on food production, characterized in that, include: The data acquisition and standardization module is configured to acquire feature data of the target batch of raw materials, and to standardize and encode the feature data to generate a standardized feature vector of the target batch of raw materials. The 3D process matrix construction module configures a 3D process matrix to characterize the physical topology of the production line. The intelligent decision-making core module is communicatively connected to the data acquisition and standardization module and the three-dimensional process matrix construction module. It includes a global optimization algorithm engine, a feature mask encoder, and a proxy model, and is configured to realize process optimization based on mask mapping and active learning. The simulation verification module is connected to the intelligent decision-making core module and is configured to perform virtual simulation and verification of the control logic, timing and performance of the optimal process configuration scheme through a digital twin model. The execution and feedback module includes an MES terminal and a PLC control network, configured to send the verified process configuration scheme to the production system for execution, and collect production data to update the agent model.

Images