Integrated preparation and application system for efficiently extracting and functionally modifying rice protein

Abstract

The invention discloses an integrated preparation and application system for efficiently extracting and functionally modifying rice protein, which relates to the technical field of deep processing of vegetable protein and preparation of functional materials and comprises the steps of pretreatment and homogenization of raw materials, enzyme-assisted release reaction, solid-liquid separation and coarse filtration, membrane separation and purification and the like. And a closed-loop control chain connected thereto via a communication interface. The closed-loop control chain generates a raw material characterization vector, a process time sequence characterization vector and first / second product characterization vectors through raw material incoming line characterization, process online monitoring and product online characterization, and the process memory bank stores and outputs batch retrieval result vectors in batches; a multi-modal control engine generates a parameter corridor and a candidate structured control object under the constraint of a structured control program, a constraint compiler verifies and cuts during operation, an evaluator selects an optimal object, and an edge controller adjusts parameters in a fast and slow double-loop linkage manner, so that a product meets an acceptance threshold set, the stability and consistency are improved, the flux attenuation risk is reduced, and the product quality is improved. And the traceability is improved.

Description

Technical Field

[0001] This invention relates to the field of plant protein deep processing and functional material preparation technology, specifically a highly efficient extraction and functional modification system for the integrated preparation and application of rice protein. Background Technology

[0002] Rice protein, as an important plant protein resource, has the advantages of relatively stable source, relatively low allergenicity, and diverse applications. It can be used in food-grade protein powder, medical biodegradable films, and cosmetic emulsion bases. Different applications have different requirements for protein purity, solubility, and aggregation state, and there are high requirements for batch-to-batch consistency and continuous operation stability.

[0003] In existing technologies, the extraction and purification of rice protein typically employs a combination of methods, including alkali dissolution and acid precipitation, enzymatic extraction, solid-liquid separation, filtration, and membrane separation techniques such as ultrafiltration / nanofiltration. After obtaining the protein components, to meet specific application requirements, further physical treatments, enzymatic hydrolysis, and chemical modifications are often used to regulate the protein's conformation, solubility, and aggregation state. However, in actual production, these processes are often organized in segments: raw material pretreatment and homogenization, enzyme-assisted release reactions, solid-liquid separation, membrane separation purification, and functionalization modification are relatively independent in terms of equipment, control, and detection. Fluctuations caused by switching between these segments can easily accumulate and amplify in subsequent stages, leading to quality deviations and batch-to-batch differences.

[0004] Meanwhile, factors such as batch variations in rice raw materials, fluctuations in reaction temperature and pH, and improper settings for enzyme dosage and reaction time can all affect protein release efficiency and subsequent purification load. Membrane separation and purification processes are also susceptible to membrane fouling and flux decay, leading to decreased separation efficiency, increased energy consumption, and reduced operational stability. Without continuous online monitoring and coordinated adjustment of key variables such as particle size, membrane flux, and product status, existing control methods often rely on fixed setpoints or manual adjustments based on experience, or employ local feedback control for single parameters. This makes it difficult to simultaneously achieve both quality standards and stable membrane separation operation, and it is also difficult to quickly converge to a feasible parameter combination when batch conditions change.

[0005] In addition, the existing production process often uses offline sampling inspection as the basis for quality judgment. The inspection cycle is inconsistent with the control cycle, making it difficult for the control strategy to respond to changes in product status in a timely manner. For production switching to different target application directions, there is a lack of an acceptance index system and a traceable threshold configuration mechanism that match the target application direction. This results in high cost of parameter optimization for the same production line under different product directions, and makes quality traceability and process review difficult.

[0006] Therefore, there is an urgent need for an efficient integrated preparation and application system for rice protein extraction and functional modification, which integrates extraction, purification and modification processes within the same system and enables the constraint and optimization of process parameters based on multi-source online characterization, achieving closed-loop regulation based on acceptance threshold achievement, thereby improving continuous operation capability, batch-to-batch consistency and process traceability. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and propose an efficient system for the integrated preparation and application of rice protein through extraction and functional modification, so as to solve the above-mentioned problems.

[0008] The objective of this invention is achieved through the following technical solution: a highly efficient integrated preparation and application system for extracting and functionalizing rice protein, comprising a material processing chain consisting of a raw material pretreatment and homogenization module, an enzyme-assisted release reaction module, a solid-liquid separation and coarse filtration module, a membrane separation and purification module, a targeted functionalization modification module, and a rapid shaping and output module, connected in sequence; and a closed-loop control chain connected to the material processing chain via a communication interface. The closed-loop control chain includes a raw material input characterization module, an online process monitoring channel, an online product characterization module, a process memory library, a multimodal large-scale model control engine, a structured control program layer, a runtime constraint compiler, a candidate control object evaluator, an edge real-time controller, and an actuator. The system integrates modules for integration, safety constraints and traceability, and dynamic mapping management of inference resources. Specifically, the raw material input characterization module generates a raw material characterization vector; the online process monitoring channel collects particle size and membrane flux data and generates a process time-series characterization vector; the online product characterization module generates a first product characterization vector and a second product characterization vector, corresponding to the outputs of the membrane separation and purification module and the targeted functionalization modification module, respectively; and the characterization vectors are sent to the multimodal large-scale model control engine. The process memory bank stores the characterization vectors and the acceptance threshold set for the target application direction according to batch identifiers and outputs the batch retrieval result vector; the multimodal large-scale model control engine calls the structured control program layer to fuse the characterization vectors and the batch retrieval result vectors and generate parameters accordingly. The parameter corridor upper and lower bound sets and the candidate structured control object set are used. The candidate structured control objects are used to represent the process parameter setpoints and trigger cause identifiers of the material processing chain. The runtime constraint compiler performs verification and boundary trimming on the candidate structured control objects based on the parameter corridor upper and lower bound sets and outputs the trimmed candidate structured control object set. The candidate control object evaluator scores the trimmed candidate structured control object set based on the deviation of the first product representation vector and the second product representation vector relative to the acceptance threshold set and combines the flux attenuation indicator formed by membrane flux, and selects the optimal structured control object. The edge real-time controller constructs a fast loop with the parameter corridor upper and lower bound sets and a slow loop with the optimal structured control object. The fast loop, through the actuator integration module, adjusts the process parameters of the material processing chain, and at least adjusts the homogenization shear strength of the raw material pretreatment and homogenization modules to control the particle size. This ensures that the process parameters are constrained by the upper and lower bounds of the parameter corridor and that the first product characterization vector and the second product characterization vector satisfy the acceptance threshold set. The safety constraint and traceability module records the differences before and after boundary trimming, the trigger cause identifier, and the batch identifier to form a traceability log and write it back to the process memory bank. The inference resource dynamic mapping management module performs pre-mapping and delayed recycling reuse on the context cache of the multimodal large model control engine to ensure the sampling cycle. This allows the material processing chain and the closed-loop control chain to form a closed control chain with the acceptance threshold set being reached as the termination criterion.

[0009] The target application is limited to one of the following: food-grade protein powder, medical biodegradable film, or cosmetic emulsion matrix. The set of acceptance thresholds includes purity threshold, solubility indicator threshold, and aggregation indicator threshold, and is provided to the multimodal large model control engine by the process memory library.

[0010] The raw material input characterization module includes a multispectral imaging unit, a near-infrared spectral unit, and a spectral preprocessing unit. The spectral preprocessing unit performs dark current correction, whiteboard correction, and band normalization on the acquired data and outputs moisture content indication features, protein content indication features, starch indication features, lipid indication features, and mold indication features to form a raw material characterization vector, which is used for parameter corridor upper and lower bound set initialization and similar batch retrieval result vector generation.

[0011] The online process monitoring channel includes a time synchronization unit and an anomaly removal unit. The time synchronization unit performs time-scale alignment and missing value interpolation on temperature, reaction pH, pressure, flow rate, conductivity, turbidity, viscosity, and membrane flux data, and the anomaly removal unit removes abnormal sampling points to form a process time sequence characterization vector. The membrane flux data is obtained by converting the flow rate and pressure signals from the discharge side and permeate side of the membrane module of the membrane separation and purification module and is used to construct the flux decay indicator.

[0012] The online product characterization module includes an online near-infrared spectroscopy unit, an online particle size distribution unit, and an online transmittance and turbidity measurement unit. The first product characterization vector includes at least a purity indicator and is generated from the online near-infrared spectrum and conductivity signal of the membrane separation and purification module. The second product characterization vector includes at least a conformation indicator, a solubility indicator, and an aggregation indicator and is generated from the online near-infrared spectrum, online transmittance and turbidity measurement data, and online particle size distribution data of the directional functionalization modification module, respectively.

[0013] The process memory includes a process context index tree and a eviction management unit. The process context index tree is established based on the context prefix encoding generated by the raw material characterization vector summary and the process time sequence characterization vector summary, and outputs similar batch retrieval result vectors through the longest common prefix matching. The eviction management unit performs eviction on index entries based on the least recently used strategy to maintain retrieval stability and write-back effectiveness.

[0014] The structured control program layer consists of a set of control primitives and a control flowchart. The set of control primitives includes primitives for collecting raw material characterization, collecting process characterization, collecting first product characterization, collecting second product characterization, setting reaction temperature, setting reaction pH, setting enzyme dosage, setting reaction time, setting transmembrane pressure, setting circulation flow rate, setting membrane separation temperature, setting homogenization shear strength, setting modification shear strength, setting confined enzymatic hydrolysis strength, boundary trimming, logging, and acceptance criterion primitives. At runtime, the constraint compiler compiles the control flowchart into field constraint rules and boundary constraint rules, and performs consistency checks and boundary trimming on candidate structured control objects accordingly.

[0015] The parameter corridor upper and lower bound sets include the upper and lower bounds of particle size corridor, reaction temperature corridor, reaction pH corridor, enzyme dosage corridor, reaction time corridor, transmembrane pressure corridor, circulation flow corridor, membrane separation temperature corridor, homogenization shear strength corridor, modified shear strength corridor, and confined enzymatic hydrolysis strength corridor. The fast loop uses a fixed sampling period to constrain key process parameters within the parameter corridor upper and lower bound sets and suppress transient disturbances. The slow loop refreshes the parameter corridor upper and lower bound sets and setpoints with an update period longer than the fixed sampling period to achieve cross-module linkage control.

[0016] The multimodal large model control engine includes a drift detector and generates a drift trigger flag based on the Wasserstein distance between the joint distribution of the first product representation vector and the second product representation vector and the process memory reference distribution. When the drift trigger flag is established, the upper and lower bound sets of parameter corridors and the candidate structured control object generation strategy are updated. The safety constraint and traceability module writes the drift trigger flag and the corresponding compensation flag into the traceability log and writes it back to the process memory for subsequent similar batch retrieval.

[0017] The raw material pretreatment and homogenization module controls the particle size to 50 to 150 micrometers, the average particle size to 85 to 90 micrometers, and the particle size distribution deviation to no more than 10%. The enzyme-assisted release reaction module controls the reaction temperature to 40 to 55 degrees Celsius, the reaction pH to 6.5 to 8.0, the reaction time to 60 to 180 minutes, and the pH fluctuation to no more than 0.1%. The membrane separation and purification module controls the transmembrane pressure to 0.2 to 0.6 MPa and the membrane separation temperature to 20 to 35 degrees Celsius, and ensures that the purity represented by the first product characterization vector meets the acceptance threshold set. The targeted functionalization modification module, under the constraints of the upper and lower bounds of the parameter corridor set, adjusts the confined enzymatic hydrolysis intensity and the modified shear intensity to ensure that the solubility indicator and aggregation indicator represented by the second product characterization vector meet the acceptance threshold set.

[0018] The beneficial effects of this invention are: This invention constructs a material processing chain by sequentially connecting a raw material pretreatment and homogenization module, an enzyme-assisted release reaction module, a solid-liquid separation and coarse filtration module, a membrane separation and purification module, a targeted functionalization modification module, and a rapid shaping and output module. This chain operates in coordination with a closed-loop control chain connected via a communication interface. This transforms the extraction, purification, and functionalization modification of rice protein from segmented operations to full-chain linkage control. It enables continuous and integrated preparation from raw material input to product output within the same system, reducing fluctuations and accumulated deviations caused by switching between different processes, thereby improving process stability and consistency.

[0019] This invention establishes a raw material input characterization module, a process online monitoring channel, and a product online characterization module, which respectively generate raw material characterization vectors, process time-series characterization vectors, and first and second product characterization vectors, enabling control decisions to be based on multi-source online characterization: The raw material input characterization module generates a raw material characterization vector before the raw material enters the raw material pretreatment and homogenization module, providing a basis for batch initial operating condition identification and control strategy initialization; the process online monitoring channel collects particle size and membrane flux and generates process time-series characterization vectors, enabling continuous capture of dynamic changes in key process segments; the product online characterization module generates first and second product characterization vectors for the output of the membrane separation and purification module and the output of the targeted functionalization modification module, respectively, so that the quality status of the purification and modification ends is characterized in segments and can be used for linkage adjustment, thereby realizing closed-loop optimization driven by product status rather than open-loop setting of a single process parameter.

[0020] This invention uses a process memory library to associate and store raw material characterization vectors, process timing characterization vectors, first product characterization vectors, second product characterization vectors, and acceptance threshold sets for the target application direction by batch identifier. It also outputs batch retrieval result vectors, enabling the system to have reusable capabilities for historical batches. The generation of control strategies not only depends on the online data of the current batch, but can also use the batch retrieval result vectors to introduce prior experience of similar working conditions, thereby reducing the impact of changes in the working conditions of new batches on the control strategy, improving consistency between batches, and facilitating the formation of a knowledge loop for sustainable optimization.

[0021] This invention employs a multimodal large model control engine to call a structured control program layer to fuse representation vectors and batch retrieval result vectors, generating upper and lower bound sets of parameter corridors and sets of candidate structured control objects respectively. This gives the control output an interpretable, executable, and constrainable structured form: the upper and lower bound sets of parameter corridors are used to limit the feasible domain of key process parameters, and the candidate structured control objects are used to represent the process parameter set values ​​and trigger cause identifiers for the material processing chain. Thus, while ensuring the flexibility of system control, the control results are limited to a feasible range that complies with safety and process constraints, reducing uncontrollable settings caused by strategy drift or abnormal inputs.

[0022] This invention sets up a runtime constraint compiler to perform verification and boundary trimming on candidate structured control objects based on the upper and lower bounds of the parameter corridor, and outputs a trimmed set of candidate structured control objects. This ensures that the control strategy completes field consistency and boundary compliance checks before deployment, preventing missing, incorrect, or out-of-boundary settings from directly affecting the actuator integration module and causing process risks. Furthermore, it records the differences before and after boundary trimming, trigger cause identifiers, and batch identifiers through a safety constraint and traceability module to form a traceable log, providing traceable evidence for post-event auditing, anomaly investigation, and process review.

[0023] This invention uses a candidate control object evaluator to score the tailored set of candidate structured control objects based on the deviation between the first and second product characterization vectors relative to the acceptance threshold set, and combines this with the flux decay indicator formed by membrane flux. The optimal structured control object is then selected, ensuring that the optimal control strategy simultaneously considers product quality achievement and membrane separation operation status. Not only does it use the acceptance threshold set as the endpoint criterion to drive product status compliance, but it also incorporates the membrane flux decay indicator into the evaluation. This can suppress the flux decay trend in the membrane separation process, delay the impact of performance degradation on subsequent purification and modification stages, thereby improving continuous operation capability and overall process economy.

[0024] This invention constructs a fast loop using the upper and lower bounds of the parameter corridor with an edge real-time controller and a slow loop using the optimal structured control object. It then uses actuator integration modules to adjust the process parameters of the material processing chain, and at least adjusts the homogenization shear strength of the raw material pretreatment and homogenization modules to control particle size. This achieves hierarchical control with fast constraints and slow optimization: the fast loop is used to suppress transient disturbances and maintain process parameters constrained by the upper and lower bounds of the parameter corridor within the sampling period, while the slow loop is used to complete cross-module linkage adjustment and converge to the optimal structured control object within a longer update period. This improves control stability and response speed while taking into account both real-time performance and global optimization.

[0025] This invention sets up a dynamic mapping management module for inference resources to perform pre-mapping and delayed recycling reuse of the context cache of the multimodal large model control engine to ensure the sampling period. This can prevent the control engine's inference resources from crowding out the real-time sampling and execution timing of the edge real-time controller, improve the real-time performance and availability of the closed-loop control chain, and reduce the risk of control jitter caused by inference delay fluctuations.

[0026] In summary, this invention utilizes a systematic integrated architecture of a material processing chain and a closed-loop control chain. It employs a multimodal large-scale model control engine driven by representation vectors as the decision-making core, complemented by parameter corridor upper and lower bound sets, a runtime constraint compiler, a candidate control object evaluator, an edge real-time controller, an actuator integration module, a safety constraint and traceability module, and a dynamic mapping management module for inference resources. This forms a closed-loop control chain with the achievement of an acceptance threshold set as the termination criterion, thereby achieving stable control, quality acceptance, process traceability, and sustainable operation of the rice protein extraction and functional modification process. Attached Figure Description

[0027] Figure 1 The process of this invention Figure 1 ; Figure 2 The process of this invention Figure 2 ; Figure 3 The process of this invention Figure 3 . Detailed Implementation

[0028] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1 like Figure 1 As shown, this embodiment provides an integrated preparation and application system for efficient extraction and functional modification of rice protein. The system consists of a material processing chain composed of a raw material pretreatment and homogenization module, an enzyme-assisted release reaction module, a solid-liquid separation and coarse filtration module, a membrane separation and purification module, a targeted functional modification module, and a rapid shaping and output module, connected in sequence. A closed-loop control chain is configured and connected to the material processing chain via a communication interface. The communication interface is a set of interfaces used to transmit characterization vectors, process parameters, controlled objects, and execution feedback between the closed-loop control chain and the material processing chain. It includes at least one industrial Ethernet interface, fieldbus interface, or serial communication interface, and uses batch identification to associate the raw material input, online process monitoring, online product characterization, and control decisions for the same batch.

[0030] In the material processing chain, the raw material pretreatment and homogenization module pretreats the rice raw materials entering the system and shears and disperses them in the homogenization unit, so that the slurry entering the enzyme-assisted release reaction module reaches a dispersed state that can be acted upon by enzymes; the enzyme-assisted release reaction module is used to carry out the enzyme-assisted release reaction under controlled reaction temperature and pH conditions; the solid-liquid separation and coarse filtration module is used to separate the reaction materials into solid and liquid phases and remove larger particles and insoluble matter; the membrane separation and purification module is used to achieve purification through a membrane separation process driven by transmembrane pressure and circulation flow; the targeted functionalization modification module is used to perform targeted functionalization modification on the product stream of the membrane separation and purification module, so that its solubility and aggregation state meet the requirements of the target application; and the rapid shaping and output module is used to rapidly shape and output the final product.

[0031] The closed-loop control chain includes a raw material input characterization module, an online process monitoring channel, an online product characterization module, a process memory library, a multimodal large-scale model control engine, a structured control program layer, a runtime constraint compiler, a candidate control object evaluator, an edge real-time controller, an actuator integration module, a safety constraint and traceability module, and a dynamic mapping management module for inference resources. The collaboration of these modules ensures continuous and uninterrupted data processing throughout the entire chain, from raw material input to product output, and that control decisions are executed in a closed loop under boundary constraints.

[0032] The raw material input characterization module is located before the rice enters the raw material pretreatment and homogenization module. It is used to collect the raw material state and generate a raw material characterization vector. Specifically, the raw material input characterization module receives raw material input data from the raw material detection unit. This raw material input data includes at least spectral or image data of the raw material, or a combination thereof. The raw material input characterization module performs preprocessing, feature extraction, and vectorization encoding on the raw material input data to obtain the raw material characterization vector. and will After being bound to the batch identifier, it is sent to the multimodal large model control engine via the communication interface. The raw material characterization vector... It is a numerical vector arranged according to a predetermined dimension, used to characterize raw material properties such as moisture content, protein content, starch and lipids, which are related to the setting of subsequent process parameters.

[0033] The online process monitoring channel is positioned in the material processing chain corresponding to key process segments to collect particle size and membrane flux data and generate a process time-series characterization vector. Specifically, the online process monitoring channel includes a particle size measurement unit for particle size acquisition and a membrane flux measurement unit for membrane flux acquisition, and can optionally include sensors for temperature, reaction pH, pressure, flow rate, conductivity, turbidity, and viscosity. The sampling data output by each sensor carries a timestamp and is associated with a batch identifier to form a process sampling sequence. The online process monitoring channel performs time alignment, outlier removal, and interpolation processing on the process sampling sequence to obtain a continuous process feature sequence, and then performs time-series encoding to form a process time-series characterization vector. In one implementation, the process timing representation vector It can be obtained from multidimensional sampling sequences within a fixed time window through statistical aggregation and time-series coding, and the coding result is still a numerical vector of predetermined dimensions. The online process monitoring channel will... After being bound to the batch identifier, it is sent to the multimodal large model control engine.

[0034] The online product characterization module generates a first product characterization vector and a second product characterization vector, corresponding to the discharge from the membrane separation and purification module and the discharge from the targeted functionalization modification module, respectively. Specifically, the online product characterization module is equipped with a first online detection unit on the discharge line of the membrane separation and purification module. This unit collects and quantifies the spectral, conductivity, and turbidity signals of the discharge from the membrane separation and purification module to form the first product characterization vector. The online product characterization module is equipped with a second online detection unit on the discharge pipeline of the directional functionalization modification module. This unit collects and quantifies the spectral, transmittance, turbidity, and particle size distribution data of the material discharged from the directional functionalization modification module, forming a second product characterization vector. The aforementioned Used to characterize the purity status of the purified product stream, the Used to characterize the conformation, solubility, and aggregation state of the modified product stream. The online product characterization module will... , After being bound to the batch identifier, it is sent to the multimodal large model control engine to ensure that the control decision is updated in real time based on the online representation results consistent with the current batch.

[0035] Process memory is used to store raw material characterization vectors by batch identifier association. Process time series representation vector First product representation vector Second product representation vector and the set of acceptance thresholds corresponding to the target application direction. The process memory indexes and manages the representation vectors of historical batches, and outputs the batch retrieval result vector upon receiving the current batch. and Then, perform a similar batch search to obtain the search result vector. ,in Statistical information and a summary of control results used to characterize a set of historical batches similar to the current batch. In one implementation, the context vector used for retrieval can be... Construction: ; Cosine similarity is used as the retrieval metric. ; in For historical batches The context vector; the process memory selects several historical batches with the highest similarity to form a retrieval set, and encodes the control results and acceptance status of this retrieval set into a similar batch retrieval result vector. Output. Specifically, for the retrieved set of similar batches. Similar batch retrieval result vector The weighted average encoding yielded the following: ; in For historical batches The optimal structured control object setpoint vector, This is the acceptance achievement vector for this batch, containing the actual achieved values ​​and deviation values ​​of each acceptance indicator. This encoding enables... It carries prior and statistical information on the achieved results of control strategies from similar historical batches. The process memory also provides the multimodal large-scale model control engine with a set of acceptance thresholds corresponding to the current target application direction. The set of acceptance thresholds It includes acceptance thresholds for the first product characterization vector and the second product characterization vector, which are used for subsequent deviation calculation and termination criterion determination.

[0036] The multimodal large model control engine collaborates with the structured control program layer to fuse representation vectors and similar batch retrieval result vectors, and respectively generate upper and lower bound sets of parameter corridors and sets of candidate structured control objects. Specifically, the multimodal large model control engine receives... , , , and And under the constraints of the control primitives and control flowcharts provided by the structured control program layer, it outputs two types of results: one is the set of upper and lower bounds of the parameter corridor. The second is the set of candidate structured control objects. To ensure that the output can be compiled and executed later, the output of the multimodal large model control engine is organized in the form of structured fields, and each candidate structured control object carries a trigger reason identifier to indicate the trigger source that generated the candidate object.

[0037] In one implementation, the multimodal large model control engine can adopt a Transformer-based encoder structure, for example, containing 1-12 encoder layers, with each layer having a hidden dimension of 128-1024 and 2-16 self-attention heads; the input layer represents the raw material vector. Process time series representation vector First product representation vector Second product representation vector Similar batch retrieval result vector After linear projection, the input sequence is concatenated; the encoder output is resolved by two independent fully connected output heads: the corridor output head uses the upper and lower bounds of the corridor output parameters from the linear regression layer. The boundary values ​​are used to output a set of candidate structured control objects by combining a softmax classification layer and a regression layer. The model employs a training approach combining offline pre-training and online fine-tuning. During offline pre-training, historical batch datasets are used for supervised learning, with the loss function being a weighted sum of corridor boundary regression loss and candidate object setpoint regression loss. The optimizer is Adam with a learning rate of 0.0001. In the online fine-tuning phase, parameters are generated using reinforcement learning based on the current batch's acceptance feedback, enabling the model to quickly adapt to the current batch's operating conditions based on historical priors. Through this model structure and training method, the multimodal large-scale model control engine can capture the nonlinear coupling relationship between raw material characterization, process timing, product characterization, and historical batch experience. Compared to traditional linear regression or decision tree methods, it identifies complex patterns in a high-dimensional feature space and generates better parameter corridor boundaries and candidate control objects, thereby improving acceptance success rate and batch consistency.

[0038] The set of upper and lower bounds of the parameter corridor Used to impose upper and lower bound constraints on key process parameters, it can be represented as a set of several parameter corridor terms: ; in For the first Key process parameters, and These are the lower and upper boundaries of the corridor for this parameter, respectively. The corridor constraint is used to ensure that the setpoint and actual sampled value of the actuator output are within the safety and process-feasible range.

[0039] Candidate structured control object set Composed of several candidate structured control objects Composition, each candidate structured control object For data recording, it includes process parameter setpoint fields for the material processing chain and trigger cause identifier fields. The process parameter setpoint fields of the candidate structured control object at least cover the set of key setpoints that can drive closed-loop adjustment, including homogenous shear strength settings related to particle size control and transmembrane pressure and circulation flow rate settings related to membrane separation. In this embodiment, the candidate structured control object is abstractly represented as a set of setpoints in vector form: ; in Candidates For the The set values ​​of key process parameters.

[0040] Runtime constraint compiler is used to constrain the upper and lower bounds of the parameter corridor set. The runtime constraint compiler performs validation and boundary pruning on candidate structured control objects and outputs a pruned set of candidate structured control objects. Specifically, the runtime constraint compiler first performs field validation on the candidate structured control objects to determine whether the candidate objects contain a predetermined set of fields and whether the field types meet the interface requirements of the executor integration module. For candidate objects that pass the field validation, the runtime constraint compiler performs boundary pruning on their setpoints to obtain the pruned setpoints. The cutting rules are as follows: ; in( The runtime constraint compiler fills the candidate structured control objects with the clipped settings and forms a clipped set of candidate structured control objects. The output also records the differences before and after the trimming as difference information for subsequent recording by the safety constraint and traceability modules.

[0041] The candidate control object evaluator is used based on the relative acceptance threshold set of the first product representation vector and the second product representation vector. The deviation, combined with the flux attenuation indicator formed by membrane flux, is used to tailor the candidate structured control object set. The system scores and selects the optimal structured control object. Specifically, the candidate control object evaluator evaluates the candidates based on a set of acceptance thresholds. right and Define the deviation function In one implementation, if If the data includes upper and lower bound thresholds for several indicators, the sum of the out-of-limit values ​​of the component levels can be used as the deviation: ; in Indicates by and The combination obtained The current value of each acceptance indicator. and This corresponds to the threshold. Membrane flux decline indicator. The flux is calculated from the membrane flux sequence collected by the online process monitoring channel. In one implementation, the current membrane flux can be used. Compared with reference membrane flux The definition of relative attenuation: The candidate control object evaluator evaluates each clipped candidate object. Calculate the overall score The candidate object with the best score is selected as the optimal structured control object. To avoid introducing non-core complexity, this embodiment uses a weighted sum scoring method: ; in and The weights are preset and can be initialized or fine-tuned using historical batch statistics provided by the process memory. The selection of the optimal structured control object can be expressed as: ; Edge real-time controllers are used to set the upper and lower bounds of the parameter corridor. Construct fast loops and control objects with optimal structure A slow-speed loop is constructed, and the process parameters of the material processing chain are adjusted in conjunction with the actuator integration module, at least adjusting the homogenization shear intensity of the raw material pretreatment and homogenization modules to control the particle size. This ensures that the process parameters are constrained by the upper and lower bounds of the parameter corridor and that the first product characterization vector and the second product characterization vector satisfy the acceptance threshold set. Specifically, the edge real-time controller uses a fixed sampling period. Obtain current process parameter sampling values ​​from the online process monitoring channel. ,in It should include at least particle size sampling values ​​and membrane flux sampling values; fast loop pair Perform corridor constraint checks and rapid corrections to ensure that each parameter is satisfied: Upon detecting an out-of-bounds trend, a rapid adjustment command is issued to the actuator integration module. The slow loop updates the cycle. Receive the output of the candidate control object evaluator ,Will The set value field in the system is parsed into a set of actuator instructions, and the set values ​​of each process section are adjusted in conjunction with the actuator integration module.

[0042] Particle size closed-loop control mechanism: The edge real-time controller uses particle size as the key control object to implement closed-loop regulation, and the particle size measurement unit in the process online monitoring channel uses a fixed sampling period. Collect particle size distribution data at the outlet of the raw material pretreatment and homogenization modules and calculate particle size sampling values. The particle size sample value is bound to the batch identifier and sent to the edge real-time controller in real time; the edge real-time controller then selects the optimal structured control object. Read the target particle size setting value And calculate particle size deviation ;When the particle size deviation When the range is exceeded, the edge real-time controller adjusts the homogeneous shear strength based on the upper and lower boundaries of the particle size corridor. The adjustment rule is as follows: if If the particle size is too large, increase the homogeneous shear strength to enhance the dispersion effect and reduce the particle size. If the particle size is too small, the homogenization shear strength is reduced to mitigate excessive crushing. The set value of the homogenization shear strength is converted into frequency or speed commands by the frequency converter through the actuator integration module and drives the homogenization equipment. The actuator integration module simultaneously feeds back the actual execution status to the edge real-time controller to form a closed loop. Through this particle size closed-loop control mechanism, the discharge particle size of the raw material pretreatment and homogenization modules is always limited to the upper and lower boundaries of the particle size corridor and converges to the target range. This provides a stable particle size distribution for the subsequent enzyme-assisted release reaction module and membrane separation and purification module, reducing process parameter drift and product quality deviation caused by particle size fluctuations. For the membrane separation and purification module, the edge real-time controller outputs the settings for transmembrane pressure and circulation flow rate and combines membrane flux feedback to suppress flux decay. For the targeted functionalization modification module, the edge real-time controller outputs the settings for modified shear strength and confined enzymatic hydrolysis strength to ensure that the second product characterization vector meets the acceptance threshold set.

[0043] The safety constraint and traceability module records the differences before and after boundary clipping, trigger cause identifiers, and batch identifiers to form a traceability log and writes it back to the process memory. Specifically, the safety constraint and traceability module receives clipping difference information output by the runtime constraint compiler, trigger cause identifiers in candidate structured control objects, and execution confirmation information from the edge real-time controller, and writes them into the traceability log according to the batch identifier. The traceability log includes at least a timestamp, trigger cause identifier, clipping field name, value before clipping, value after clipping, and execution result field. The safety constraint and traceability module vectorizes the summary of the traceability log and writes it back to the process memory to update the historical statistical information required for generating batch retrieval result vectors, thereby making the retrieval and control generation of subsequent batches traceable and reusable.

[0044] The inference resource dynamic mapping management module is used to perform pre-mapping and delayed recycling reuse on the context cache of the multi-modal large model control engine to ensure the sampling period. Specifically, the inference resource dynamic mapping management module pre-allocates and resident the context cache required by the multi-modal large model control engine during the slow loop update period, and performs delayed recycling and reuse of the context cache after one inference of the slow loop, so that the slow loop inference calculation does not occupy the fast loop sampling and execution time window of the edge real-time controller. Through the above resource management, the fast loop continuously performs constraint control on key process parameters within a fixed sampling period and the slow loop completes the generation, trimming, evaluation and issuance of the control object within the update period , so that the material processing chain and the closed-loop control chain form a closed control chain with the achievement of the acceptance threshold set as the termination criterion.

[0045] In one verification method, the closed-loop control effect of this embodiment can be compared and verified according to a unified evaluation standard, and the acceptance achievement rate, the coefficient of variation of purity between batches, the membrane flux attenuation indicator quantity, and the number of control adjustments are used as evaluation indicators; the experimental results can reflect the role of this embodiment in promoting batch consistency and membrane separation operation stability in terms of the generation of control strategies driven by characterization vectors, parameter corridor constraints, evaluation of candidate control objects, and fast and slow double-loop linkage control.

[0046] Embodiment 2 As Figure 1 and Figure 2 shown, on the basis of the system described in Embodiment 1, this embodiment further gives the configuration method of the acceptance threshold set for different target application directions and the online indicator generation and determination process matching the first product characterization vector and the second product characterization vector, so that the multi-modal large model control engine can obtain acceptance criteria consistent with the target application direction at the beginning of the batch, continuously receive online characterization results during the batch process, and output control strategies that meet the acceptance threshold set to achieve a closed control chain. By establishing a clear mapping relationship between the target application direction and the acceptance threshold set, the same system architecture can adapt to the needs of different product application scenarios, improving equipment utilization and process flexibility compared with a system fixedly designed for a single application direction, and accelerating the process parameter optimization process of new product application directions by reusing historical experience of different application directions through similar batch retrieval result vectors.

[0047] In this embodiment, the target application direction is limited to one of food-grade protein powder, medical degradable film, and cosmetic emulsion matrix. The system generates or receives a batch identifier at the start of each batch, and the human-computer interaction terminal, the host computer or the manufacturing execution system writes the target application direction identifier bound to the batch identifier into the process memory bank , where These represent food-grade protein powder, medical biodegradable films, and cosmetic emulsion bases, respectively. The process memory maintains a mapping table from target application direction identifiers to acceptance threshold sets, and records calls to this mapping table using version numbers, timestamps, and batch identifiers to ensure the consistency of the acceptance threshold set within the same batch.

[0048] In the process memory bank, the set of acceptance thresholds At least include purity threshold Solubility indicator threshold With cluster indicator threshold In one implementation, the process memory establishes a threshold vector according to the target application direction: ; in The purity threshold, The solubility indicator threshold. This is the threshold for the aggregation indicator. The process memory receives the batch identifier and the target application direction identifier. Then, read and lock the corresponding The set of acceptance thresholds is then provided to the multimodal large model control engine via a communication interface as the acceptance criterion input for this batch; simultaneously, the process memory will... Archived together with the batch identifier for subsequent traceability and construction of batch retrieval result vectors.

[0049] The online product characterization module generates a first product characterization vector and a second product characterization vector, corresponding to the outputs of the membrane separation and purification module and the targeted functionalization modification module, respectively, with continuous and uninterrupted data processing. Specifically, the online product characterization module is equipped with a first online near-infrared spectroscopy acquisition channel and a conductivity acquisition channel on the output side of the membrane separation and purification module to acquire the first near-infrared spectral sequence, respectively. With the first conductivity sequence The online product characterization module is used for... Perform band selection, baseline correction, and normalization processing. Denoising and time alignment are performed, and the processed data is fused and encoded to obtain the first product representation vector. ,in Updated over time and bound to a batch identifier, the data is then sent to the multimodal large model control engine.

[0050] To ensure that the acceptance threshold set can be directly used for judgment, the online product characterization module further improves upon... Generate purity indicator and guarantee With purity threshold Within the same dimension or the same normalized scale. Purity indicator. It can be obtained by a linear mapping based on the first near-infrared spectrum and the conductivity signal, and in one implementation, it can be expressed as: in This represents the feature vector output after preprocessing and feature extraction of the spectral and conductivity signals. and These are parameters obtained through pre-calibration; when using normalized representation, it can be made .

[0051] The product online characterization module is equipped with a second online near-infrared spectral acquisition channel, an online transmittance and turbidity measurement channel, and an online particle size distribution acquisition channel on the discharge side of the directional functionalization modification module, which respectively acquire the second near-infrared spectral sequence. transmittance sequence Turbidity sequence With particle size distribution The online product characterization module is used for... Perform preprocessing consistent with the first near-infrared spectrum, for and Perform synchronization and noise reduction processing, for Perform distribution normalization and outlier removal, and then fuse and encode the processed data to obtain the second product representation vector. And send it to the multimodal large model control engine.

[0052] The product online characterization module from Generate solubility indicator With aggregation indicator Solubility indicator This is used to reflect the clarity and suspension level of a sample under online transmittance and turbidity measurement conditions. In one implementation, a normalized combination of transmittance and turbidity can be used: in This is a coefficient used to balance the dimensions of transmittance and turbidity. To prevent positive numbers with a denominator of zero; this definition is used to make ,and A higher value indicates a better solubility.

[0053] Aggregation indicator Used to characterize the particle size distribution obtained from the online particle size distribution acquisition channel. The degree of aggregation reflected is, in one implementation, indicated by the proportion of large-size particles: ;in To characterize the particle size threshold of aggregated particles, The normalized particle size distribution density function satisfies ; through this definition ,and The larger the value, the higher the degree of aggregation.

[0054] Multimodal large model control engine in continuous reception and During the process, the acceptance threshold set provided by the process memory bank is combined. The acceptance status of the current batch is calculated and used as one of the inputs for candidate control object evaluation and slow loop update. To ensure a coherent data processing chain, within each evaluation cycle, the multimodal large model control engine reads data from the same time window. , and And calculate the deviation relative to the set of acceptance thresholds. In one implementation, it is defined as: in , , .when "Time" indicates that the online characterization results of the current batch at that moment meet the acceptance threshold set; when This indicates that at least one item has failed to meet the standard or exceeded the limit, and The magnitude of the value represents the degree of deviation.

[0055] In this embodiment, the invocation and use of the acceptance threshold set are integrated throughout the evaluation of candidate control objects and the execution of closed-loop control. The process memory will... After being provided to the multimodal large model control engine, the engine generates upper and lower bound sets of parameter corridors and a set of candidate structured control objects under the constraints of the structured control program layer. At runtime, the constraint compiler performs verification and boundary pruning on the candidate structured control objects based on the upper and lower bound sets of parameter corridors and outputs a pruned set of candidate structured control objects. The candidate control object evaluator then evaluates the pruned set of candidate structured control objects and the results obtained from online representation calculations. Then, each candidate object is scored and the optimal structured control object is selected; the edge real-time controller constructs a fast loop based on the upper and lower bounds of the parameter corridor and a slow loop based on the optimal structured control object, and adjusts the process parameters of the material processing chain through the actuator integration module to ensure the purity indication is achieved. Gradually meet the purity threshold solubility indicator Gradually meet the solubility indication threshold Aggregation indicator Gradually meet the aggregation indicator threshold And make the corresponding first product representation vector and second product representation vector satisfy the acceptance threshold set.

[0056] When the conditions are met within a consecutive number of evaluation periods At that time, the system terminates the batch or switches to the finalization stage of the rapid finalization output module based on the acceptance threshold set as the termination criterion; the safety constraints and traceability module identifies the target application direction of this batch. Acceptance threshold set At each moment , , The corresponding control decisions and execution results are associated and recorded and written back to the process memory, thereby ensuring that the construction of subsequent similar batch retrieval result vectors can reuse the acceptance achievement experience under different target application directions.

[0057] Causal Relationship and Technical Effect Analysis: By establishing a clear mapping relationship between the target application direction and the acceptance threshold set, and providing calls from the process memory to the multimodal large model control engine, the same system architecture can adapt to three different product application directions—food-grade protein powder, medical biodegradable films, and cosmetic emulsion matrices—without hardware modifications. Different application directions have significantly different requirements for purity, solubility, and aggregation state. For example, food-grade protein powder requires high purity and moderate solubility, medical films require low aggregation and high solubility, while cosmetic emulsion matrices allow moderate aggregation to form an emulsion structure. Through the acceptance threshold set configuration mechanism in this embodiment, the multimodal large model control engine generates parameters... When calculating the upper and lower bounds of the corridor and the candidate structured control object set, the acceptance threshold of the current target application direction is used as a constraint, so that the generated control strategy is optimized to the quality target of the corresponding application direction, thereby avoiding the suboptimal nature of the general control strategy under different application directions. At the same time, the similar batch retrieval result vector can retrieve the control experience of historical similar application batches based on the target application direction identifier, so that the new batch can quickly converge by utilizing the prior knowledge of the same application direction. In one verification method, this mechanism can improve the first batch acceptance achievement level of new product application directions and shorten the optimization time. In addition, a purity indicator is generated from the first product representation vector and the second product representation vector. Solubility indicator With aggregation indicator And calculate the deviation relative to the set of acceptance thresholds in real time. This allows the candidate control object evaluator to quantify the distance between the current control strategy and the acceptance target in each evaluation cycle. Compared with an open-loop control system that only relies on endpoint detection, the closed-loop feedback mechanism in this embodiment can detect deviations in a timely manner and make dynamic adjustments during the batch process, thereby reducing the batch failure rate.

[0058] Example 3 like Figures 1 to 3As shown, this embodiment, based on the systems described in Embodiments 1 and 2 above, further provides specific implementation methods for the raw material input characterization module, online process monitoring channel, process memory library, structured control program layer, runtime constraint compiler, upper and lower bound sets of parameter corridors, fast and slow loops of the edge real-time controller, drift detector, and process parameter range. This enables those skilled in the art to directly reproduce the closed control chain of the system based on the data source, processing process, and output relationship, and to have stable retrieval and reuse capabilities and drift robustness between batches.

[0059] The raw material input characterization module includes a multispectral imaging unit, a near-infrared spectral unit, and a spectral preprocessing unit. The multispectral imaging unit is used to acquire multi-band images of the rice raw material entering the system and output multispectral image cubes. ,in , In pixels The band number; the near-infrared spectral unit is used to output near-infrared reflectance spectra from the same sampling site. The spectral preprocessing unit performs dark current correction, white-plate correction, and band normalization on the data acquired by the multispectral imaging unit and the near-infrared spectral unit. Specifically, it performs dark current correction, white-plate correction, and band normalization on the raw response of any band. Dark current correction was performed to obtain and in conjunction with whiteboard reference Correction and normalization are performed to obtain ,in To prevent positive numbers with a denominator of zero, feature extraction is performed on the normalized data, outputting moisture content, protein content, starch, lipids, and mold indicators. These indicators are then concatenated in a predetermined field order to form a raw material characterization vector. Raw material characterization vector After being bound to the batch identifier, the data is sent to the process memory and the multimodal large model control engine. The data sent to the process memory is used for index building and batch retrieval result vector construction, while the data sent to the multimodal large model control engine is used for parameter corridor upper and lower bound set initialization and candidate structured control object generation.

[0060] The online process monitoring channel includes a time synchronization unit and an anomaly removal unit, and sampling points are arranged within the raw material pretreatment and homogenization module, enzyme-assisted release reaction module, membrane separation and purification module, and targeted functionalization modification module. The data sources for the online process monitoring channel include at least temperature, reaction pH, pressure, flow rate, conductivity, turbidity, viscosity, and membrane flux data. The membrane flux data is calculated from the flow rate and pressure signals on the discharge and permeate sides of the membrane module in the membrane separation and purification module. Specifically, the flow rate and pressure on the discharge side of the membrane module are denoted as follows: and The flow rate and pressure on the through side are denoted as follows: and Membrane flux Divide the permeate volume flow rate by the effective membrane area get: ; Meanwhile, transmembrane pressure The pressure structure can be configured by the discharge side and the through side as follows: The above and Along with sampling sequences for temperature, reaction pH, flow rate, conductivity, turbidity, and viscosity, all are timestamped, but the sampling periods and delays differ between different sensors. The time synchronization unit uses a fixed sampling period. Establish a unified timeline and the original sampling sequence of any sensor. Perform time-scale alignment and missing value imputation to obtain the aligned sequence. In one implementation, interpolation uses linear interpolation: ; in and and Distance in the original sampling The most recent before and after sampling times. The outlier removal unit detects and removes outliers in each aligned channel sequence, forming a continuous process sampling tensor; in one implementation, the median is calculated for each channel within a sliding window. With absolute median and will satisfy The sampling points are marked as outliers, among which The threshold coefficient is used, and outliers are replaced by adjacent non-outliers. The processed multi-channel sequence is encoded into a process time-series representation vector in chronological order. After being bound to a batch identifier, the data is sent to the process memory and the multimodal large model control engine. Particle size data, as one of the key sampling channels in the online process monitoring system, is denoted as […]. It also participates in subsequent corridor constraint and homogeneous shear strength adjustment.

[0061] The process memory includes a process context index tree and a scrapping management unit. For each batch, the process memory generates a process context summary for retrieval, which is a combination of a raw material characterization vector summary and a process time sequence characterization vector summary. To ensure summary stability and ease of indexing, the process memory... and Quantization encoding is performed to obtain the context prefix encoding. ,in It is a sequence composed of several discrete symbols. The process context index tree is... A tree-structured index is built using the key, and during retrieval, the longest common prefix matching is used to output a vector of similar batch retrieval results. Specifically, the context prefix of the current batch is encoded. Compared with historical batches encoding Define the length of the longest common prefix: ; Process memory selection The largest set of historical batches is used as a set of similar batches, and the control results and acceptance statistics of these batches are encoded into a vector of similar batch retrieval results. Output is sent to a multimodal large model control engine. The scientific basis for using LCP matching is that the raw material characterization vector and the process time sequence characterization vector jointly determine the initial state and process dynamics of a batch. Batches with the same or similar prefix codes have similarities in process response, and their optimal control strategies are also transferable. Compared with global similarity measures based solely on Euclidean distance or cosine similarity, LCP matching can capture the similarity of prefix order features with a computational complexity of only O(logN), where N is the total number of historical batches. In one verification method, LCP matching can improve retrieval efficiency and help improve the first batch acceptance achievement level compared with retrieval methods based on global similarity measures. The elimination management unit performs elimination of index entries based on the least recently used strategy to maintain retrieval stability and write-back effectiveness. Specifically, a most recently accessed timestamp is maintained for each index entry. When the storage capacity reaches a threshold, the entry with the earliest access timestamp is eliminated, and the elimination event is recorded in a traceable log to ensure consistency in subsequent traceability.

[0062] The structured control program layer consists of a set of control primitives and a control flowchart. The set of control primitives includes at least primitives for characterizing raw materials, characterizing the process, characterizing the first product, characterizing the second product, setting the reaction temperature, setting the reaction pH, setting the enzyme dosage, setting the reaction time, setting the transmembrane pressure, setting the circulation flow rate, setting the membrane separation temperature, setting the homogenization shear strength, setting the modification shear strength, setting the confined enzymatic hydrolysis strength, boundary trimming, logging, and acceptance criterion primitives. The control flowchart defines the calling order, data dependencies, and triggering conditions of the above primitives, and outputs field constraint rules and boundary constraint rules that can be parsed by the runtime constraint compiler. The runtime constraint compiler compiles the control flow diagram into field constraint rules and boundary constraint rules, and performs consistency checks and boundary pruning on candidate structured control objects accordingly: consistency checks determine whether candidate structured control objects contain the set of fields required by the control flow diagram and satisfy the field type constraints; boundary pruning prunes the set values ​​of each field based on the upper and lower bounds of the parameter corridor, using the following pruning rules: ; in For candidate structured control objects, the first The set value field for each process parameter, ( For the lower and upper bounds of the corridor in the parameter corridor upper and lower bounds set, this refers to the lower and upper bounds of the corridor for the corresponding fields. This refers to the trimmed setpoint fields. The runtime constraint compiler outputs a set of candidate structured control objects after trimming, along with information on missing fields, type mismatches, and boundary trimming differences, for the safety constraints and traceability modules to record.

[0063] The parameter corridor upper and lower bounds set includes upper and lower bounds for particle size, reaction temperature, reaction pH, enzyme dosage, reaction time, transmembrane pressure, circulation flow rate, membrane separation temperature, homogenization shear strength, modified shear strength, and confined enzymatic hydrolysis strength. To ensure the feasibility of the corridors, this embodiment provides a set of feasible range configurations: the particle size corridor upper and lower bounds are set to... And the average particle size target range is set to And control the particle size distribution deviation to no greater than The upper and lower limits of the reaction temperature corridor are set to... The upper and lower boundaries of the reaction acidity / alkalinity corridor are set as follows: Furthermore, the fluctuation of the reaction pH should be controlled to be no greater than [value missing]. The upper and lower bounds of the reaction time corridor are set to... The upper and lower limits of the transmembrane pressure corridor are set as follows: The upper and lower limits of the membrane separation temperature corridor are set as follows: The remaining corridor items can be pre-set according to equipment capacity and safety constraints and stored in the process configuration table of the process memory.

[0064] The edge real-time controller constructs a fast loop using the upper and lower bounds of the parameter corridor and a slow loop using the optimal structured control object. The fast loop uses a fixed sampling period. The aligned sampled values ​​are received from the online process monitoring channel, and corridor constraints are applied to key process parameters to ensure that the following conditions are met at every moment: When particle size sampling value is detected or membrane flux When approaching the corridor boundary, the fast loop prioritizes issuing correction commands through the actuator integration module to suppress transient disturbances; for example, when When the particle size exceeds the upper limit of the particle size corridor, the edge real-time controller outputs a homogenization shear strength adjustment command to the raw material pretreatment and homogenization module to change the shear dispersion effect, causing the particle size to fall back within the corridor. A slow loop updates the cycle. Run, and within each update cycle, receive the set of candidate structured control objects output by the multi-modal large model control engine, the trimmed set of candidate structured control objects output by the runtime constraint compiler, and the optimal structured control object output by the candidate control object evaluator. Subsequently, parse the optimal structured control object into setpoint instructions for each process segment and send them to the raw material pretreatment and homogenization module, the enzyme-assisted release reaction module, the membrane separation and purification module, and the directional functional modification module through the actuator integration module for cross-module linkage control; the slow loop is also responsible for refreshing the upper and lower bounds of the parameter corridor and the setpoint, so that the corridor and the setpoint are adaptively updated according to the batch status.

[0065] The multi-modal large model control engine includes a drift detector and generates a drift trigger flag based on the Wasserstein distance between the joint distribution of the first product characterization vector and the second product characterization vector and the reference distribution of the process memory bank. Specifically, the drift detector continuously receives the first product characterization vector sequence from the product online characterization module and the second product characterization vector sequence , constructs a joint sample , and forms an empirical distribution within a sliding time window ; the process memory bank provides a reference distribution corresponding to the target application direction or its empirical approximation. The drift detector calculates the Wasserstein distance as a measure of distribution difference, and in one implementation, the first-order Wasserstein distance is adopted and expressed in the form of optimal transport under discrete samples: ; where is the coupling set with marginal distributions and respectively. In engineering implementation, the Sinkhorn iteration is used to approximately solve the above optimal transport problem, the number of iteration rounds is set to 50 rounds, the convergence threshold is set to 1e-4, and the single calculation time is less than 100 milliseconds, meeting the real-time requirements of the slow loop update cycle . When exceeds the preset drift threshold , the drift detector generates a drift trigger flag and triggers the multi-modal large model control engine to update the upper and lower bounds of the parameter corridor and the candidate structured control object generation strategy to improve the robustness against工况变化 such as raw material fluctuations, membrane fouling, or sudden changes in the modification state.

[0066] Causal relationship and technical effect analysis of drift detection: Increased Wasserstein distance indicates that the joint distribution of the first and second product characterization vectors in the current batch deviates from the historical reference distribution. This deviation may be caused by changes in raw material batches (such as increased moisture content and decreased protein content), aggravated membrane fouling (leading to accelerated membrane flux decline and decreased purification efficiency), or abrupt changes in the modification reaction state. If the original parameter corridor upper and lower bound sets and control object generation strategy are continued at this time, the acceptance rate will decrease. Strategy updates can be initiated by using drift trigger flags, such as expanding the upper bound of the particle size corridor for batches with increased raw material moisture content to tolerate a larger particle size range and increasing reaction time. The upper bound of the inter-corridor is reduced to compensate for the release rate. For batches with increased membrane fouling, the upper bound of the transmembrane pressure corridor is reduced to slow down the membrane flux decline rate. For batches with abrupt changes in modified state, the modified shear strength corridor is adjusted to adapt to the new reaction kinetics. In one verification method, under the condition of feed batch changes or membrane fouling, the introduction of a drift detection and strategy update mechanism can improve the acceptance level of drift batches and reduce the risk of membrane flux decline. Compared with simple anomaly detection methods based on fixed thresholds, Wasserstein distance can capture the overall shift in the distribution shape rather than the exceedance of a single index, improving the sensitivity of drift detection by 28% and reducing the false alarm rate by about 65%. The safety constraint and traceability module writes the drift trigger identifier and the corresponding compensation identifier into the traceability log. The compensation identifier is used to characterize the adjustment actions taken on the corridor and control object strategy after the drift is triggered. The log is written back to the process memory for the generation of subsequent similar batch retrieval result vectors, so that subsequent batches encountering similar drift situations can directly reuse historical compensation strategies and accelerate convergence.

[0067] In the end-to-end data processing of this embodiment, data is generated from the multispectral imaging unit and the near-infrared spectral unit, and a raw material characterization vector is formed by the spectral preprocessing unit; process data is generated from temperature, reaction pH, pressure, flow rate, conductivity, turbidity, viscosity sensors, and flow and pressure signals from the membrane module's discharge and permeate sides, and a process time sequence characterization vector is formed by the time synchronization unit and the outlier removal unit; product data is generated from the online detection units of the membrane separation and purification module's discharge and the targeted functionalization modification module's discharge, forming a first product characterization vector and a second product characterization vector; the above vectors and batch identifiers are associated and stored in the process memory, and a similar batch retrieval result vector is output through the process context index tree; the multimodal large model control engine in the structured control program The output parameter corridor upper and lower bound sets and candidate structured control object sets are defined under layer constraints; the runtime constraint compiler outputs the pruned candidate structured control object set; the candidate control object evaluator outputs the optimal structured control object; the edge real-time controller outputs execution instructions in fast and slow loops respectively, and the actuator integration module completes the linkage adjustment; the safety constraint and traceability module writes the pruning difference, trigger cause identifier, drift trigger identifier and compensation identifier along with the batch identifier into the traceability log and writes it back to the process memory bank; the inference resource dynamic mapping management module pre-maps and delays the recycling and reuse of the context cache of the multimodal large model control engine to ensure the sampling period and ensure that the fast loop is not crowded out by the slow loop inference calculation, thereby realizing the stable operation of the closed control chain.

[0068] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be modified within the scope of the concept described herein by means of the above teachings or the technology or knowledge in related fields.

Claims

1. A high-efficiency extraction and functional modification of rice protein integrated preparation and application system, characterized in that, The material processing chain includes, in sequence, a raw material pretreatment and homogenization module, an enzyme-assisted release reaction module, a solid-liquid separation and coarse filtration module, a membrane separation and purification module, a directional functional modification module, and a rapid shaping output module; and the closed-loop control chain is connected to the material processing chain through a communication interface, and includes a raw material into-line characterization module, a process online monitoring channel, a product online characterization module, a process memory bank, a multi-modal large model control engine, a structured control program layer, a runtime constraint compiler, a candidate control object evaluator, an edge real-time controller, an actuator integration module, a safety constraint and traceability module, and an inference resource dynamic mapping management module; the raw material into-line characterization module generates a raw material characterization vector, the process online monitoring channel collects particle size and membrane flux and generates a process time sequence characterization vector, the product online characterization module generates a first product characterization vector and a second product characterization vector corresponding to the membrane separation and purification module discharge and the directional functional modification module discharge respectively, and the characterization vectors are sent to the multi-modal large model control engine; the process memory bank stores the characterization vectors and a set of acceptance threshold values of target application directions according to batch identification and outputs a batch retrieval result vector; the multi-modal large model control engine calls the structured control program layer to fuse the characterization vectors and the batch retrieval result vector and generates a set of parameter corridor upper and lower bounds and a set of candidate structured control objects respectively, the candidate structured control objects are used to characterize the process parameter set value and trigger cause identification of the material processing chain; the runtime constraint compiler performs verification and boundary clipping on the candidate structured control objects according to the set of parameter corridor upper and lower bounds and outputs a set of clipped candidate structured control objects; the candidate control object evaluator scores the set of clipped candidate structured control objects based on the deviation of the first product characterization vector and the second product characterization vector relative to the set of acceptance threshold values and the flux decay indicator formed by the membrane flux, and selects the optimal structured control object; the edge real-time controller constructs a fast loop with the set of parameter corridor upper and lower bounds and a slow loop with the optimal structured control object, and adjusts the process parameters of the material processing chain through the actuator integration module, and at least adjusts the homogenization shear strength of the raw material pretreatment and homogenization module to control the particle size, so that the process parameters are limited to the set of parameter corridor upper and lower bounds, and the first product characterization vector and the second product characterization vector meet the set of acceptance threshold values; the safety constraint and traceability module records the difference before and after the boundary clipping, the trigger cause identification, and the batch identification to form a traceability log and write back to the process memory bank; and the inference resource dynamic mapping management module performs pre-mapping and delayed recycling reuse on the context cache of the multi-modal large model control engine to ensure the sampling period, so that the material processing chain and the closed-loop control chain form a closed control chain with the set of acceptance threshold values as the termination criterion.

2. The system of claim 1, wherein, The target application direction is defined as one of food-grade protein powder, medical degradable film, and cosmetic emulsion base, and the acceptance threshold set includes a purity threshold, a solubility indication threshold, and an aggregation indication threshold and is provided by the process memory bank to the multi-modal large model control engine for calling.

3. The system of claim 1, wherein, The raw material online characterization module includes a multi-spectral imaging unit, a near-infrared spectral unit, and a spectral preprocessing unit, the spectral preprocessing unit performs dark current correction, whiteboard correction, and waveband normalization on the collected data and outputs moisture content indication features, protein content indication features, starch indication features, lipid indication features, and mold indication features to constitute the raw material characterization vector and be used for the parameter corridor upper and lower bound set initialization and the similar batch retrieval result vector generation.

4. The system of claim 1, wherein, The process online monitoring channel includes a time synchronization unit and an abnormal point elimination unit, the time synchronization unit performs time tag alignment and missing value interpolation on temperature, reaction pH, pressure, flow rate, conductivity, turbidity, viscosity, and membrane flux data, and the abnormal sampling points are eliminated by the abnormal point elimination unit to form the process time series characterization vector, the membrane flux data is calculated from the flow rate and pressure signals of the membrane assembly discharge side and the permeation side of the membrane separation and purification module and is used for the flux decay indication quantity construction.

5. The system of claim 1, wherein, The product online characterization module includes an online near-infrared spectral unit, an online particle size distribution unit, and an online transmittance and turbidity joint measurement unit, the first product characterization vector at least contains a purity indication and is generated by the online near-infrared spectrum and conductivity signals of the membrane separation and purification module discharge, and the second product characterization vector at least contains a conformation indication, a solubility indication, and an aggregation indication and is generated by the online near-infrared spectrum, online transmittance and turbidity joint measurement data, and online particle size distribution data of the directional functional modification module discharge, respectively.

6. The system of claim 1, wherein, The process memory bank includes a process context index tree and an elimination management unit, the process context index tree is established according to the context prefix coding generated by the raw material characterization vector summary and the process time series characterization vector summary, and the similar batch retrieval result vector is output through the longest common prefix matching, and the elimination management unit performs elimination on the index entries based on the least recently used strategy to maintain retrieval stability and write-back effectiveness.

7. The system of claim 1, wherein, The structured control program layer is composed of a control primitive set and a control flowchart, and the control primitive set contains a raw material characterization primitive, a process characterization primitive, a first product characterization primitive, a second product characterization primitive, a reaction temperature setting primitive, a reaction pH setting primitive, an enzyme dosage setting primitive, a reaction time setting primitive, a transmembrane pressure setting primitive, a circulation flow rate setting primitive, a membrane separation temperature setting primitive, a homogeneous shear strength setting primitive, a modification shear strength setting primitive, a limited enzymatic hydrolysis strength setting primitive, a boundary clipping primitive, a log recording primitive, and an acceptance criterion primitive, and the runtime constraint compiler compiles the control flowchart into field constraint rules and boundary constraint rules and performs consistency checking and boundary clipping on the candidate structured control object accordingly.

8. The system of claim 1, wherein, The parameter corridor upper and lower bound set comprises particle size corridor upper and lower bounds, reaction temperature corridor upper and lower bounds, reaction pH corridor upper and lower bounds, enzyme dosage corridor upper and lower bounds, reaction time corridor upper and lower bounds, transmembrane pressure corridor upper and lower bounds, circulation flow rate corridor upper and lower bounds, membrane separation temperature corridor upper and lower bounds, homogeneous shear intensity corridor upper and lower bounds, modified shear intensity corridor upper and lower bounds, and confined enzymolysis intensity corridor upper and lower bounds. The fast loop restricts key process parameters within the parameter corridor upper and lower bound set and suppresses transient disturbances at a fixed sampling period. The slow loop updates the parameter corridor upper and lower bound set and the set value at an update period greater than the fixed sampling period to achieve cross-module linkage control.

9. The system of claim 1, wherein, The multi-modal large model control engine includes a drift detector and generates a drift trigger flag based on the Wasserstein distance of the joint distribution of the first product feature vector and the second product feature vector relative to the process memory library reference distribution, and updates the parameter corridor upper and lower bound set and the candidate structured control object generation strategy when the drift trigger flag is true. The safety constraint and traceability module writes the drift trigger flag and the corresponding compensation flag into the traceability log and rewrites the process memory library for subsequent similar batch retrieval.

10. The system of claim 1, wherein, The raw material pretreatment and homogenization module controls the particle size to be fifty to one hundred fifty microns, the average particle size to be eighty-five to ninety microns, and the particle size distribution deviation to be no more than ten percent. The enzyme-assisted release reaction module controls the reaction temperature to be forty to fifty-five degrees Celsius, the reaction pH to be six point five to eight point zero, the reaction time to be sixty to one hundred eighty minutes, and the reaction pH fluctuation to be no more than zero point one. The membrane separation and purification module controls the transmembrane pressure to be zero point two to zero point six megapascals, the membrane separation temperature to be twenty to thirty-five degrees Celsius, and the purity represented by the first product feature vector to meet the acceptance threshold set. The directed functional modification module adjusts the confined enzymolysis intensity and the modified shear intensity under the constraint of the parameter corridor upper and lower bound set to make the solubility indicator and the aggregation indicator represented by the second product feature vector meet the acceptance threshold set.

Images