A method, apparatus, and medium for quality control of production of a plant beverage

Abstract

This invention discloses a method, equipment, and medium for quality control in the production of plant-based beverages, belonging to the field of intelligent food manufacturing and reliable quality engineering technology. This invention pioneers a parallel control system integrating "biological-physical-chemical information flow." The method includes: deploying biomimetic olfactory / taste quantum dot array sensors to capture in real-time the spectra of volatile organic compounds and the spatial distribution of flavor substances during the beverage production process; constructing an inter-process "quality potential energy field" model to guide the optimal transmission path of quality flow using field strength gradients; introducing a digital twin based on evolutionary reinforcement learning, enabling it to autonomously evolve anti-disturbance control strategies in parallel interaction with the physical production line; and finally, transforming key quality evidence into standardized on-chain assets through verifiable credential technology. This invention achieves proactive anti-disturbance, self-evolution, and cross-chain verifiability in quality control.

Description

Technical Field

[0001] This invention relates to the fields of intelligent food manufacturing, industrial artificial intelligence and trusted computing technology, specifically to a method and system for quality control and traceability of plant-based beverages throughout their entire life cycle, which integrates biomimetic perception, field theory modeling, parallel intelligence and distributed trust technologies. Background Technology

[0002] Currently, the production of high-end plant-based beverages faces three technological gaps in terms of refined quality management. In terms of perception, traditional sensors can only acquire indirect physicochemical parameters such as temperature, pressure, and pH. They lack direct, online, and distributed monitoring methods for key biochemical substances that determine flavor, texture, and functionality (such as characteristic aroma components, flavor amino acids, and the conformation of active polysaccharides), resulting in a lack of precise basis for process adjustments. In terms of decision-making, existing advanced process control or model predictive control is mostly based on precise linear or simplified nonlinear models, which struggle to cope with the strong uncertainty, nonlinearity, and large hysteresis brought about by the biological complexity of plant raw materials, resulting in poor control robustness. In terms of trust, existing quality traceability is mostly limited to information recording and querying; data ownership and authenticity still rely on centralized institutions for endorsement, making it difficult to establish automated, machine-to-machine trust transfer of quality data in an open supply chain.

[0003] While some studies have attempted to incorporate spectroscopy or machine vision, their information dimensions are limited, failing to establish a deep correlation model between multimodal information and final quality. The application of artificial intelligence is also largely concentrated on offline analysis or single-stage prediction, lacking a parallel intelligent control system capable of real-time interaction with physical systems and autonomous evolution. Regarding data trustworthiness, simply putting data on the blockchain only addresses tamper-proofing, failing to solve the problems of data standardization and cross-entity mutual recognition. Therefore, a new quality control paradigm is urgently needed that can bridge the gap from molecular sensory perception to intelligent decision-making on the production line, and then to the trustworthy transmission throughout the supply chain. Summary of the Invention

[0004] The present invention aims to provide a concrete test block transportation and curing management system and method to solve the problems of low efficiency of manual handling, easy damage to test blocks, insufficient utilization of curing space, and chaotic management in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for quality control in the production of plant-based beverages, comprising the following steps: S1: Deploy a multimodal biomimetic sensor array to simultaneously collect volatile organic compound fingerprint spectra, flavor substance distribution cloud maps, and microstructure images from key workstations on the production line, generating a process-level "biological-physical-chemical" fused quality information flow; S2: Based on the information flow of step S1, construct a digital twin model of the "mass potential energy field" of the production line. This model maps the quality state of each process to the position and potential energy value in the potential energy field. The mass transfer process is simulated as motion driven by the field strength gradient. S3: Activate a virtual digital twin that runs synchronously with the physical production line. This twin is equipped with an evolutionary reinforcement learning algorithm. Through trial and error with the potential energy field model, it autonomously generates and outputs a real-time control strategy that can optimize the global potential energy of the final product quality. S4: By linking the execution evidence of the control strategy and the key feature values ​​of sensor data with decentralized identifiers and generating verifiable credentials, these credentials are anchored to the blockchain network to build quality credit assets that can be verified across platforms and systems.

[0006] Furthermore, the multimodal biomimetic sensing array in S1 specifically includes: Bionic olfactory unit: A cross-sensitive array composed of functionalized quantum dot gas sensors targeting characteristic flavor substances such as alcohols, aldehydes, and esters, outputting a fingerprint spectrum; Bionic taste unit: Based on a solid contact ion-selective electrode array, it measures the spatial distribution of concentrations of sweet, sour, umami, bitter and other taste substances in liquid materials in real time; Microstructure vision unit: integrates a high-speed microscope camera and a Raman spectroscopy probe to capture images of emulsion stability, particle size distribution and crystallization state online.

[0007] Furthermore, the construction of a digital twin model of the "mass potential energy field" in S2 is specifically as follows: Define the quality status of each process as a vector point in a multi-dimensional space; Based on historical high-quality batch data, a potential energy function is learned that maps the final product quality standard to a "low potential energy valley" in the potential energy field. The state point in real-time production has a specific potential energy value in this potential energy field. The goal of quality transfer optimization is to control the process parameters so that the state point moves along the gradient direction where the potential energy decreases the fastest.

[0008] Furthermore, the evolutionary reinforcement learning algorithm in S3 has the following operating mechanism: Population initialization: Initialize a set of virtual agents with different control policy parameters in the digital twin; Parallel trial and error: All virtual agents perform parallel simulations of the same production batch in the potential energy field model, and each agent outputs a set of control action sequences. Fitness evaluation: Evaluate the quality of the strategy based on the final potential energy value reached by the virtual agent at the guided state point; Evolutionary operation: Cross-pollinate and mutate the strategies of highly fit agents to generate a new generation of agent populations and iteratively optimize them; Policy delivery: Mapping and distributing the control policies of the best virtual agents to the physical production line for execution.

[0009] Furthermore, the construction of verifiable credentials in S4 involves the following steps: S41: Assign globally unique decentralized identifiers to production lines, equipment, batches, and critical data packets; S42: Encapsulate the characteristic hash of the sensor data, the digest of the control command, the timestamp, and other information according to the verifiable credential data model, and sign it with the producer's private key; S43: Publish the signed credential to a verifiable data registry and anchor its notarized fingerprint to the blockchain; S44: Downstream businesses or consumers can trust the authenticity of data by verifying the signature and storage status of the credentials without accessing the original database.

[0010] In a second aspect, the present invention provides an apparatus comprising: A multimodal biomimetic sensor array module, configured according to claim 2; Edge computing and field computing modules are used for real-time calculation of mass potential energy fields and operation of lightweight digital twins; A cloud-based evolutionary learning and parallel control platform for running large-scale virtual population evolutionary computations; The decentralized identifier and verifiable credential management module is used to generate, sign, and manage quality credentials that are linked on-chain and off-chain.

[0011] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method as described in any one of claims 1-5.

[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. Enhanced perception: The monitoring of "physicochemical conditions" has been elevated to the direct perception of "comprehensive quality elements" (flavor profile, taste distribution, physical structure), enabling direct, in-situ, and online evaluation of the sensory and functional quality of products, providing an unprecedentedly rich source of information for control.

[0013] 2. Qualitative Change in Decision-Making Mechanism: The quality control problem is transformed from the traditional "tracking the setpoint" to a global optimization problem of finding the "optimal descent path" in the "mass potential energy field." The introduced evolutionary reinforcement learning mechanism enables the control system to actively try and fail, learn autonomously, and continuously evolve in a parallel digital space, enabling it to cope with new raw materials or disturbances never seen before, thus achieving a leap from "automation" to "autonomy."

[0014] 3. Innovation in the Trust System: By employing decentralized identifiers and verifiable credentials, quality data is encapsulated into self-contained, independently verifiable credit atoms. This enables it to circulate freely among brands, retailers, testing institutions, and consumers without relying on a centralized database and to be automatically trusted, laying the foundation for building an open quality collaboration ecosystem.

[0015] 4. Foresight in System Architecture: The proposed three-layer parallel control architecture of "physical system - potential energy field model - parallel intelligent agent" provides a general framework for industrial intelligence. It is not only applicable to plant-based beverages, but also provides a paradigm for quality control in other complex biomanufacturing processes. Attached Figure Description

[0016] Figure 1 This is a diagram illustrating the overall architecture of the parallel quality control system described in this invention. Figure 2 This is a schematic diagram illustrating the principle of the "mass potential energy field" model in this invention. Figure 3 This is a flowchart of the quality data storage and verification process based on verifiable credentials in this invention. Detailed Implementation

[0017] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be emphasized that the embodiments described herein are only used to illustrate the working principle of the present invention and do not constitute any limitation on the scope of protection.

[0018] (I) Detailed description of the technical solution of the present invention The overall architecture of the technical solution of this invention is as follows: Figure 1 As shown, the core lies in building an intelligent quality control ecosystem in which a physical production system and a manual control system execute in parallel and interact in two directions.

[0019] 1. Construction and data fusion of multimodal biomimetic sensor arrays The perceptual basis of this invention is a hardware-software fusion system that simulates biological senses.

[0020] Hardware layer: Integrated sensor chambers are installed at key workstations (such as fermenter exhaust ports, homogenizer outlets, and maturation tanks). The chambers integrate: ① a 16-channel gas sensor array, with each channel's quantum dot material sensitizing a specific type of VOC (such as hexanal - grassy smell, ethyl acetate - fruity aroma); ② an 8-electrode taste sensor array covering multiple ions including H+ (acid), Na+ (salty), and glutamate (fresh); ③ an online microfluidic imaging unit capable of capturing the microscopic shear thinning behavior and fat globule distribution of materials.

[0021] Data fusion layer: Multidimensional data collected every second is fed into a deep convolutional fusion network. This network first extracts "aroma profile features" from the gas fingerprint spectrum, "flavor balance features" from the ion concentration, and "texture structure features" from the image. Subsequently, these heterogeneous features are concatenated and cross-attention is calculated at higher layers of the network, ultimately outputting a unified "real-time comprehensive quality vector", such as [fruit aroma intensity: 0.87, acidity-freshness ratio: 1.2, delicacy index: 0.95].

[0022] 2. Construction and Visualization of the "Mass Potential Energy Field" Model To intuitively understand and optimize quality transfer, this invention introduces the concept of a "field." For example... Figure 2 As shown, a two-dimensional potential energy plane (actually a high-dimensional space) is constructed using two key quality attributes (such as "stability" and "flavor intensity") as coordinate axes.

[0023] Potential energy surface generation: Based on massive historical high-quality batch data, a potential energy function E=f(x, y) is learned through methods such as kernel density estimation. The higher the quality of the final product, the lower the potential energy value E, which is displayed as a "low potential energy valley" in the graph; the lower the quality, the higher the potential energy, which is displayed as a "high potential energy mound".

[0024] Real-time positioning and navigation: During the production process, the real-time comprehensive quality vector at the end of each process is mapped to a point in the field. The system can display the position and potential energy value of this point on the potential energy surface in real time. The control objective becomes extremely intuitive: by adjusting the process, this point is driven to move along the steepest potential energy descent gradient towards the nearest "low potential energy valley." This provides intuitive geometric navigation for complex optimization problems.

[0025] 3. Parallel Control Policy Generation Based on Evolutionary Reinforcement Learning like Figure 1 As shown, the control core is a cluster of digital twins running in the cloud, parallel to the physical production line.

[0026] Virtual clone: ​​Whenever a batch is started on the physical production line, N (e.g., 100) virtual clones of that batch are generated synchronously in the cloud. Each clone has a built-in intelligent agent with a random initial strategy.

[0027] Parallel Evolution: These 100 virtual agents, each in their own digital twin "mass potential energy field," simultaneously begin exploring control strategies for this virtual batch. They experiment with different temperature profiles, stirring speeds, and the timing of additive additions.

[0028] Survival of the fittest: After the virtual batch is completed, the virtual products are sorted according to the final potential value (i.e., predicted quality) reached by each agent. A genetic algorithm is used: the bottom 30% of inferior strategies are eliminated; the top 30% of superior strategies are paired to generate offspring strategies; and all strategies undergo a small-probability mutation to introduce new changes.

[0029] Strategy Transmission: After several rounds (e.g., 10 rounds) of rapid evolution, the optimal strategy is selected. This strategy is "translated" into a specific, time-sequential sequence of PLC control instructions and sent to the physical production line for execution. The results of the physical production line's execution (new sensor data) are fed back to fine-tune the potential energy field model and evaluate the effectiveness of the strategy, forming a parallel and reinforcing closed loop of "learning-execution-feedback-relearning".

[0030] 4. Assetization of Quality Credit Based on Verifiable Evidence To address cross-enterprise data trust issues, this invention abandons simple data on-chaining and adopts more advanced verifiable credential technology, as follows: Figure 3 As shown.

[0031] Certificate Issuance: When a critical quality event occurs (such as "UHT sterilization temperature confirmation"), the system automatically generates a verifiable certificate compliant with W3C standards. This certificate includes: a statement ("temperature set to 137.5℃"), the issuer (the manufacturer's DID), the holder (the batch's DID), a digital signature, and optional evidence anchors.

[0032] Voucher Holding and Transfer: This voucher is linked to the batch and can be transferred to downstream distributors along with product logistics metadata. Distributors can then store it in their own digital wallets.

[0033] Zero-knowledge proof: When it is necessary to prove to a third party (such as an e-commerce platform) that a specific process has been used, there is no need to provide the original data. A short proof can be generated through zero-knowledge proof to verify that the statement "sterilization temperature is above 137°C" is true and that the certificate has not been revoked, thereby protecting business details.

[0034] On-chain anchoring and auditing: The summaries (Merkle Root) of all issued credentials are periodically written to the blockchain in batches, providing a global timestamp and tamper-proof audit trail. Regulatory agencies only need to verify the credential signature and on-chain anchor to confirm the credibility of the data, without needing to interface with heterogeneous databases of various enterprises.

[0035] (II) Specific Implementation Example: Parallel Quality Control of Premium Oat and Plant Protein Yogurt Taking the production of an oat-based plant protein yogurt with "zero additives and a rich taste" as an example, the core challenges are stabilizing the plant protein, simulating the taste of milk fat, and inhibiting the generation of undesirable flavors.

[0036] 1. System Deployment and Initialization Sensor arrays are deployed in the enzymatic hydrolysis tank, fermentation tank, homogenizer, and post-ripening tank. During initialization, data from the past 1000 batches of high-quality products are imported to train a quality potential field for this product, characterized by "rich taste and pure flavor" (see principle). Figure 2 The potential energy field shows that the optimal quality region corresponds to a narrow range of "viscosity index > 0.8, acetaldehyde concentration < 5 ppm, and diacetyl concentration between 2 and 3 ppm".

[0037] 2. Parallel control operation process Enzymatic hydrolysis and flavor precursor control: The physical system begins enzymatic hydrolysis of the oat milk. Simultaneously, 100 virtual agents in the cloud initiate parallel evolution (architecture see...). Figure 1 This study explored the effects of different enzymatic hydrolysis temperatures and times on the degree of protein hydrolysis and the formation of potential off-odor precursors. The physical system operated according to the default process. At the 15-minute mark, the parallel system issued an early warning: based on the virtual evolution results, excessive hexanal (cabbage odor) would be produced after 30 minutes at the current temperature. The physical system immediately received instructions from the optimal virtual strategy and lowered the enzymatic hydrolysis temperature from 55°C to 50°C.

[0038] Fermentation and Texture Evolution: Entering the fermentation stage, the core processes are acid production and aroma production. A biomimetic olfactory array monitors the concentration ratio of diacetyl (creamy aroma) and acetaldehyde (green aroma) in real time. Potential field model (see...) Figure 2 The system indicated a risk of sliding towards a "high acetaldehyde" state. The virtual agents in the parallel system competed fiercely, and the winning strategy was not to adjust the temperature, but to propose briefly initiating low-speed stirring for 10 seconds at a specific acidity level to alter the microenvironment of the bacterial community. After the physical system executed this, the sensors confirmed that the diacetyl / acetaldehyde ratio had rebounded, and the state point had returned to a safe valley.

[0039] Post-ripening and quality control: During the post-ripening stage, microscopic imaging of the physical system showed a tendency for fat globules to aggregate (resulting in a decrease in the fineness index). Upon receiving this data, the parallel system instantly simulated the effects of adding different hydrophilic colloids in virtual space and recommended a fine-tuning scheme of "adding 0.015% gellan gum" within one minute. After implementing this scheme, the fineness index recovered.

[0040] Voucherization and Supply Chain Display: This batch of products is assigned DID:did:example:yogurt-batch-2023-001. The system automatically generated multiple verifiable vouchers for it, such as: "Declaration of No Added Preservatives Voucher", "Fermentation Strain Activity > 1e9 CFU / mL Voucher", and "Critical Process Parameter Compliance Voucher" (see process details). Figure 3 When products enter high-end supermarkets, the supermarket's smart receiving system automatically verifies the authenticity and validity of these vouchers by scanning RFID tags, completing the acceptance and warehousing process in seconds. Consumers who scan the QR code on the bottle can not only see traceability information, but also verify the digital signature of the brand's claim of "no additives other than sterile strains added throughout the entire process," greatly enhancing their sense of trust.

[0041] 3. Long-term self-evolution of the system After a year of operation, the system has undergone more than 300 batches of parallel evolution. Its virtual agent population has evolved a "strategy library" to cope with complex scenarios such as seasonal fluctuations in oat quality and slow decline in equipment efficiency. When a new production line is connected to the system, it can directly download this constantly evolving "strategy library" from the cloud as initial knowledge, enabling instant replication and transfer of expert experience, shortening the quality compliance cycle of the new production line by more than 60%.

Claims

1. A method for quality control in the production of plant-based beverages, characterized in that, Includes the following steps: S1: Deploy a multimodal biomimetic sensor array to simultaneously collect volatile organic compound fingerprint spectra, flavor substance distribution cloud maps, and microstructure images from key workstations on the production line, generating a process-level "biological-physical-chemical" fused quality information flow; S2: Based on the information flow of step S1, construct a digital twin model of the "mass potential energy field" of the production line. This model maps the quality state of each process to the position and potential energy value in the potential energy field. The mass transfer process is simulated as motion driven by the field strength gradient. S3: Activate a virtual digital twin that runs synchronously with the physical production line. This twin is equipped with an evolutionary reinforcement learning algorithm. Through trial and error with the potential energy field model, it autonomously generates and outputs a real-time control strategy that optimizes the global potential energy of the final product quality. S4: By linking the execution evidence of the control strategy and the key feature values ​​of sensor data with decentralized identifiers and generating verifiable credentials, these credentials are anchored to the blockchain network to build quality credit assets that can be verified across platforms and systems.

2. The method according to claim 1, characterized in that, The multimodal biomimetic sensing array in S1 specifically includes: Bionic olfactory unit: A cross-sensitive array composed of functionalized quantum dot gas sensors targeting characteristic flavor substances such as alcohols, aldehydes, and esters, outputting a fingerprint spectrum; Bionic taste unit: Based on a solid contact ion-selective electrode array, it measures the spatial distribution of concentrations of sweet, sour, umami, bitter and other taste substances in liquid materials in real time; Microstructure vision unit: integrates a high-speed microscope camera and a Raman spectroscopy probe to capture images of emulsion stability, particle size distribution and crystallization state online.

3. The method according to claim 1, characterized in that, The digital twin model of the "mass potential energy field" constructed in S2 is as follows: Define the quality status of each process as a vector point in a multi-dimensional space; Based on historical high-quality batch data, a potential energy function is learned that maps the final product quality standard to a "low potential energy valley" in the potential energy field. The state point in real-time production has a specific potential energy value in this potential energy field. The goal of quality transfer optimization is to control the process parameters so that the state point moves along the gradient direction where the potential energy decreases the fastest.

4. The method according to claim 1, characterized in that, The evolutionary reinforcement learning algorithm in S3 has the following operating mechanism: Population initialization: Initialize a set of virtual agents with different control policy parameters in the digital twin; Parallel trial and error: All virtual agents perform parallel simulations of the same production batch in the potential energy field model, and each agent outputs a set of control action sequences. Fitness evaluation: Evaluate the quality of the strategy based on the final potential energy value reached by the virtual agent at the guided state point; Evolutionary operation: Cross-pollinate and mutate the strategies of highly fit agents to generate a new generation of agent populations and iteratively optimize them; Policy delivery: Mapping and distributing the control policies of the best virtual agents to the physical production line for execution.

5. The method according to claim 1, characterized in that, The specific steps for constructing verifiable credentials in step S4 are as follows: S41: Assign globally unique decentralized identifiers to production lines, equipment, batches, and critical data packets; S42: Encapsulate the characteristic hash of the sensor data, the digest of the control command, the timestamp, and other information according to the verifiable credential data model, and sign it with the producer's private key; S43: Publish the signed credential to a verifiable data registry and anchor its notarized fingerprint to the blockchain; S44: Downstream businesses or consumers can trust the authenticity of data by verifying the signature and storage status of the credentials without accessing the original database.

6. An apparatus for implementing the method according to any one of claims 1-5, characterized in that, include: A multimodal biomimetic sensor array module, configured according to claim 2; Edge computing and field computing modules are used for real-time calculation of mass potential energy fields and operation of lightweight digital twins; A cloud-based evolutionary learning and parallel control platform for running large-scale virtual population evolutionary computations; The decentralized identifier and verifiable credential management module is used to generate, sign, and manage quality credentials that are linked on-chain and off-chain.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-5.

Images