Rice online color sorting and quality grading system and method based on machine vision

Abstract

The application is suitable for the technical field of grain processing, and provides a rice online color selection and quality grading system based on machine vision, which comprises a multi-scale air film suspension fluidization and steady-state transportation module, a multi-modal endoplasmic perspective and polarized structured light imaging module, a suspended minimally invasive repair execution module, a physical-data dual-driven process knowledge decision engine, and a multi-stage flexible pneumatic shunt module. The multi-scale air film suspension fluidization and steady-state transportation module is used to generate laminar clean airflow to make the rice in a quasi-weightless suspended state, and to replace the solid pipe wall with a micro-air film guide layer, so that the single particle sequential feeding is contactless throughout the process. The multi-modal endoplasmic perspective and polarized structured light imaging module comprises a swept-frequency optical coherence tomography unit, a snapshot hyperspectral imaging unit, and a polarized structured light 3D imaging unit. The system realizes the whole process of rice from feeding, imaging, repair to shunting without physical contact. The problem of secondary damage of rice and increase of broken rice rate caused by mechanical friction, collision and extrusion in traditional equipment is eliminated.

Description

Technical Field

[0001] This invention relates to the field of grain processing technology, and in particular to a machine vision-based online color sorting and quality grading system and method for rice. Background Technology

[0002] As one of the world's most important staple crops, rice's processing precision and quality grade directly affect food safety, commercial value, and consumer experience. In the rice processing industry chain, color sorting and quality grading after hulling and milling are crucial refining stages. Their main task is to accurately separate discolored grains, chalky grains, broken rice, immature grains, and various foreign objects from the finished rice stream, and to grade qualified rice according to national standards (such as GB / T 1354). Traditional rice color sorting mainly relies on manual visual picking, which is inefficient and inconsistent. With the advancement of photoelectric sensing and automation technology, online color sorters have emerged. These machines can convert photoelectric signals into electrical signals during the high-speed flow of rice grains, allowing the machine to replace the human eye in identification, and high-pressure airflow to perform the rejection action. This achieves high-throughput, standardized online sorting operations and is gradually becoming standard equipment in modern rice mills.

[0003] Current mainstream online color sorting technology has evolved from early monochrome photoelectric sensors to multispectral detection systems centered on machine vision. These systems typically consist of a vibrating feeder, a conveying mechanism comprised of chutes or slides, a linear or area array industrial camera combining visible and near-infrared light, and a pneumatic rejection array composed of high-frequency electromagnetic spray valves. The working principle is as follows: rice grains slide down the chute at high speed under gravity, and the camera simultaneously scans them as they pass through the imaging zone, acquiring color and near-infrared images of each grain. Back-end image processing algorithms use color space thresholds or machine learning models to identify abnormally colored grains, moldy grains, and broken rice, classifying broken rice based on the grain's projected area and shape. Finally, the control system drives the corresponding spray valves, using compressed air to blow the detected defective grains into a waste bin, while qualified rice falls naturally into the finished product bin. Some high-end equipment has begun to incorporate multi-spectral combinations or AI deep learning algorithms to improve the detection rate of light-colored defects.

[0004] In traditional equipment, rice processing relies entirely on physical walls such as chutes and pipes for guidance and acceleration. High-speed friction, collision, and compression inevitably occur between rice grains and between rice grains and the equipment walls. This process, while sorting, itself creates new micro-cracks and broken rice, becoming one of the biggest sources of secondary damage in rice processing. Summary of the Invention

[0005] The purpose of this invention is to provide a machine vision-based online color sorting and quality grading system and method for rice, aiming to solve the problems in the background art.

[0006] Specifically: A machine vision-based online rice color sorting and quality grading system, including:

[0007] The multi-scale air film suspension fluidization and steady-state transport module is used to generate laminar clean airflow to put rice in a quasi-weightless suspension state, and uses a micro air film guiding layer to replace the solid pipe wall, enabling single-particle sequential feeding without contact throughout the process.

[0008] The multimodal internal structure imaging and polarized structured light imaging module includes a swept-frequency optical coherence tomography unit, a snapshot hyperspectral imaging unit, and a polarized structured light 3D imaging unit, which are used to generate a six-dimensional digital file for each suspended rice grain that integrates its internal structure, chemical composition, and three-dimensional morphology after de-reflection.

[0009] The suspended minimally invasive repair module consists of a coaxial micro-airflow attitude controller and an array of repair nozzles, and is used to perform non-contact plasma surface activation or nanomaterial filling repair on suspended repairable rice grains.

[0010] The physical-data dual-driven process knowledge decision engine is connected to the multimodal internal quality imaging and polarized structured light imaging module and the suspension minimally invasive repair execution module. It embeds the rice grain fracture mechanics equation and fluid dynamics physics equation, and constructs a process-defect-quality knowledge graph to calculate quality level, instability index, repairability index, and generate adaptive adjustment instructions for front-end processes.

[0011] The multi-stage flexible pneumatic diversion module uses a pneumatic network without physical baffles, separated by air curtains at specific angles, to guide rice grains to the finished product, repair, or by-product channels without contact.

[0012] The process digital twin and edge self-evolution module deploys a lightweight physical augmented neural network with offline incremental learning capabilities at the edge, and runs a process digital twin system in the cloud to maximize the overall yield rate of the production line through reinforcement learning.

[0013] In one embodiment, the rising airflow velocity of the multi-scale air film suspension fluidization and steady-state transport module is precisely controlled between the suspension velocity of each type of rice and the first blowing velocity, and the micro air film guiding layer is composed of micro jets tilted in opposite directions, forming a stable constraint potential field in the axial direction of the pipe.

[0014] In one embodiment, the swept frequency optical coherence tomography unit has a swept frequency light source with a center wavelength of 1280-1320 nm and an axial resolution better than 5 μm; the snapshot hyperspectral unit is based on a computational spectral imaging chip with a spectral range of 400-1000 nm, and can acquire a spectral cube in a single frame; the polarization structured light 3D imaging unit uses polarization-modulated infrared structured light to simultaneously acquire three-dimensional morphology and polarization degree information.

[0015] In one embodiment, the suspended minimally invasive repair execution module includes a low-temperature plasma upper body repair mode; the coaxial micro-airflow formed by the coaxial micro-airflow attitude controller and the array-type repair nozzle is used to wrap the rice grains so that their cracked surfaces are precisely exposed; the rice grains with exposed cracked surfaces are activated and cleaned by the short-time pulsed plasma jet emitted by the array-type repair nozzle.

[0016] In one embodiment, the physical-data dual-driven process knowledge decision engine's physical-driven path specifically combines the crack depth and morphology with the grain geometry measured by the swept-frequency optical coherence tomography unit. It calculates the stress intensity factor at the crack tip using an embedded grain fracture mechanics equation to quantify the risk of instability. The data on crack depth, morphology, and grain geometry drives the physical-data dual-driven process knowledge decision engine's physical-driven path and generates an appearance rating. The data on both crack depth and morphology and grain geometry are output to the fusion layer of the knowledge graph, generating process optimization suggestions through a causal reasoning chain.

[0017] In one embodiment, the various diversion channels of the multi-stage flexible pneumatic diversion module are separated by air curtains formed by clean air at specific flow rates and angles. Under the guidance of the airflow, rice grains cross the air curtains and enter different channels without any physical contact.

[0018] In one embodiment, the lightweight physical augmentation neural network deployed at the edge can perform instantaneous incremental updates of the model based on a small number of new samples under the constraints of fluid dynamics physical equations in an offline state, for rapid adaptation to new defects; the process digital twin system uses the maximization of the overall line yield as the reward function and autonomously generates the optimal control strategy for the upstream process through reinforcement learning.

[0019] In one embodiment, the multimodal endoplasmic perspective and polarized structured light imaging module further includes an acousto-optic co-excitation endoplasmic excitation unit, which includes a miniature acoustic radiation force pulse generator and a multi-angle laser speckle contrast imaging channel.

[0020] In one embodiment, the multimodal endoplasmic perspectral and polarized structured light imaging module further includes a fully polarized light field compound eye acquisition array, which consists of a multi-lens lens ring surrounding an air-bearing channel, with each lens integrating a micro-nano polarization filter array.

[0021] Another object of the present invention is to provide a method for an online color sorting and quality grading system for rice based on machine vision, comprising the following steps:

[0022] S1. Rice grains are transported non-contactly to the multimodal imaging area of ​​the multimodal internal matter perspective and polarized structured light imaging module through a multi-scale air flotation bed to form a quasi-weightless single-particle queue.

[0023] S2, synchronously triggered sweeping optical coherence tomography unit, snapshot hyperspectral imaging unit and polarized structured light 3D imaging unit, collect internal structure, spectrum and three-dimensional morphology data of rice grains, and generate six-dimensional digital archives;

[0024] S3, a physical-data dual-driven process knowledge decision engine, analyzes six-dimensional files, assesses quality levels, combines knowledge graphs to calculate repairability index and instability risk, and generates adaptive adjustment instructions for front-end processes.

[0025] S4. For repairable particles, the suspended minimally invasive repair module performs contactless surface repair; the multi-level flexible pneumatic diversion module diverts the rice particles into their respective channels without contact based on the final determination.

[0026] S5 and the edge model evolve in real time based on operator markings. The cloud-based digital twin runs reinforcement learning simulation based on full-line data and sends the optimal process parameters to the physical production line, forming a closed-loop intelligent manufacturing.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. This system integrates a multi-scale air-film suspension fluidization and steady-state transport module, a multi-modal internal quality imaging and polarized structured light imaging module, a suspension minimally invasive repair execution module, a physical-data dual-driven process knowledge decision engine, a multi-level flexible pneumatic diversion module, and a process digital twin and edge self-evolution module into a single unit. It proposes a fully air-float, zero-contact processing paradigm, replacing all physical chutes, tracks, and other mechanical contact components in traditional color sorting systems with a micro-air-film guide layer. This achieves zero physical contact throughout the entire process of rice sorting, from feeding, imaging, repair to diversion. It fundamentally eliminates the secondary damage and increased broken rice rate caused by mechanical friction, collision, and compression in traditional equipment. This represents a fundamental restructuring of the entire processing flow, bringing rice sorting from the era of destructive testing to the era of non-destructive intelligent manufacturing, providing an absolutely interference-free ideal working condition for all subsequent high-precision operations.

[0029] 2. The multimodal internal quality observation and polarized structured light imaging module integrates a swept-frequency optical coherence tomography unit, a snapshot-type hyperspectral imaging unit, and a polarized structured light 3D imaging unit. For the first time, it generates a six-dimensional digital profile for each suspended rice grain, including its three-dimensional morphology after de-reflection, internal microcrack / immature white core structure, and chemical composition. This endows the machine with the ability to see through and analyze chemically. On the one hand, polarization technology effectively eliminates the masking of subtle defects such as embryo spots and light chalkiness by the specular reflection on the rice grain surface. On the other hand, the fusion of swept-frequency optical coherence tomography and hyperspectral imaging can quantitatively assess the internal health status (such as whether the depth of microcracks exceeds the safety threshold and the proportion of white cores) and early mold growth without damaging the rice grain. This extends the grading standards from the surface to the interior, achieving a fundamental upgrade in quality judgment.

[0030] 3. The physics-data dual-driven process knowledge decision engine's physics-driven path is not purely data fitting. Specifically, this path combines the crack depth measured by the swept-frequency optical coherence tomography unit and the grain geometry obtained by the polarized structured light 3D imaging unit, using an embedded grain fracture mechanics equation (emphasizing the stress intensity factor criterion). The system calculates the stress intensity factor at the crack tip to quantify its dynamic instability risk; simultaneously, its data-driven pathway performs appearance grading; data on both crack depth and morphology and rice grain geometry are output to the fusion layer of the knowledge graph, generating process optimization suggestions through causal reasoning chains; this endows the system with decision robustness and interpretability beyond purely data-driven decision-making; the quantitative calculation of the instability index allows the system to predict whether a cracked grain of rice will break in subsequent processes, rather than relying solely on static images; this reasoning ability based on physical laws enables the model to maintain reliable judgment even when facing new rice varieties or extreme working conditions; at the same time, the causal reasoning of the knowledge graph traces the detected defects back to the specific process that produced them and autonomously generates adjustment instructions, transforming the sorting system from a passive quality inspector into an active "process engineer" for the first time, achieving a true closed loop between quality control and production;

[0031] 4. The multi-level flexible pneumatic diversion module uses air curtains at specific angles to replace physical baffles to achieve contactless diversion; a lightweight physical reinforcement neural network is deployed at the edge, which can complete real-time incremental updates of the model based on a small number of new samples in an offline state, under the constraints of fluid dynamics physical equations; at the same time, the process digital twin system running in the cloud uses the maximization of the overall line yield as the reward function, and autonomously explores and transmits the optimal control strategy of the previous process through reinforcement learning; forming a three-level intelligent evolution architecture of "autonomous execution of physical production line, real-time evolution of edge nodes, and global optimization of cloud brain"; it has the ability to learn offline and optimize globally without cloud dependence; the incremental learning at the edge under the constraints of physical equations completely solves the industry pain points of traditional AI models being slow to update when facing new defects and relying on massive annotations; while the reinforcement learning of the cloud "twin factory" with the overall line yield as the goal enables process optimization to move from the history of relying on human experience to a new stage of intelligent simulation exploration and the distribution of process formulas, realizing the leap from single-machine intelligence to system intelligence;

[0032] 5. When a rice grain is suspended in a specific area, it is briefly fixed in position by a coaxial micro-airflow attitude controller. Then, an acoustic radiation force pulse generator emits a focused ultrasonic pulse with a very short duration (less than 1 ms) and extremely low energy (less than 0.1 mW) to the rice grain in a non-contact manner. The focused ultrasonic pulse generates a micrometer-level transient displacement response inside the rice grain. The viscoelastic modulus differs between healthy endosperm and immature white core, cracked areas, and moldy areas, and the response speed and recovery mode of the internal particles to the pulse are different. At the same time, the hyperspectral imaging channel or near-infrared channel switches to the multi-angle laser speckle contrast imaging channel mode to capture the changes in the internal speckle field caused by acoustic radiation force at high speed. The immature white core area with low elastic modulus will show a speckle decorrelation rate that is significantly different from that of normal endosperm. Through the active excitation method of "acoustic-optical" combination, non-contact detection of the internal mechanical properties of rice is achieved, which can effectively detect internal defects of the same refractive index that cannot be distinguished by traditional swept-frequency optical coherence tomography and spectroscopy.

[0033] 6. Multi-view synchronous acquisition using a fully polarized light field compound eye acquisition array abandons single-view structured light reconstruction and instead uses 16-32 miniature compound eye lenses surrounding the air-floating pipe to simultaneously acquire omnidirectional images of the rice grains. A single exposure can obtain its complete three-dimensional surface information. Each compound eye lens is covered with a micro-nano polarization filter array in front of its photosensitive element, which can simultaneously record the intensity, direction, and polarization state of the light. Based on the principle of light field imaging, the post-processing algorithm can perform virtual refocusing and depth-of-field expansion on the acquired light field data, completely eliminating local defocusing caused by the slight shift of the rice grains. At the same time, the acquisition of full polarization information can physically and accurately separate the surface specular reflection and the diffuse reflection of the chalky area, changing the chalky detection from residual analysis after de-reflection to direct calculation based on polarization characteristics, achieving a qualitative leap in accuracy and robustness. Attached Figure Description

[0034] Figure 1 This is a block diagram of the online color sorting and quality grading system for rice based on machine vision, as described in this invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The specific implementation of the invention will be described in detail below with reference to specific embodiments.

[0036] In one embodiment of the present invention, such as Figure 1 The aforementioned online color sorting and quality grading system for rice based on machine vision includes:

[0037] The multi-scale air film suspension fluidization and steady-state transport module is used to generate laminar clean airflow to put rice in a quasi-weightless suspension state, and uses a micro air film guiding layer to replace the solid pipe wall, enabling single-particle sequential feeding without contact throughout the process.

[0038] Further explanation is needed: the multi-scale air-film suspension fluidization and steady-state transport module includes a multi-scale laminar air flotation bed, which is composed of a porous sintered flow equalization plate and an airflow stabilizer. A vibrating feeder is installed on the multi-scale laminar air flotation bed. The upward clean airflow velocity of the multi-scale laminar air flotation bed is exactly between the suspension velocity of the rice grains and the first blowing velocity that can blow the rice grains out, so that the rice grains are balanced by forces in the vertical direction, forming a single-particle discrete queue in a quasi-weightless state. At the same time, micro-air film guiding layers are added on both sides of the pipe, replacing the solid pipe wall with a horizontal micro-airflow field, gently constraining the rice grains on the central axis of the air flotation queue, realizing absolutely contactless and stable sequential transport.

[0039] The multimodal internal structure imaging and polarized structured light imaging module includes a swept-frequency optical coherence tomography unit, a snapshot hyperspectral imaging unit, and a polarized structured light 3D imaging unit, which are used to generate a six-dimensional digital file for each suspended rice grain that integrates its internal structure, chemical composition, and three-dimensional morphology after de-reflection.

[0040] Further explanation is needed regarding the multimodal imaging region where the multimodal endoplasmic imaging and polarized structured light imaging module is located; synchronously triggering the frequency-sweeping optical coherence tomography unit, snapshot-type hyperspectral imaging unit, and polarized structured light 3D imaging unit to construct a six-dimensional digital twin profile for each grain of rice:

[0041] Sweep-frequency optical coherence tomography unit: center wavelength 1300 nm, axial resolution 5 μm, real-time scanning of the inside of rice grains to quantify the 3D morphology of microcracks and the volume and location of immature white nuclei;

[0042] Snapshot-type hyperspectral unit: Abandoning mechanical push-broom, it adopts a computational spectral imaging chip to acquire a spectral cube of 400-1000nm in a single frame, which is used to detect trace mold growth, fatty acid oxidation and other changes in chemical components;

[0043] Polarized structured light 3D imaging unit: Projects multi-angle polarized infrared structured light stripes, combined with polarization degree calculation, to accurately reconstruct the 3D morphology of rice grains and the volume of chalky areas, while eliminating specular reflections on rice skin and restoring detailed textures such as germ and light spots.

[0044] The suspended minimally invasive repair module consists of a coaxial micro-airflow attitude controller and an array of repair nozzles, and is used to perform non-contact plasma surface activation or nanomaterial filling repair on suspended repairable rice grains.

[0045] It needs further elaboration that: the suspended minimally invasive repair module is located after the multimodal imaging area and consists of an array of micro pneumatic repair nozzles; when a grain of rice is determined to have a repairable micro-crack (e.g., crack depth less than 60μm), the suspended minimally invasive repair module is activated.

[0046] The attitude of the rice grain in the air is precisely controlled by the coaxial micro-airflow attitude controller, so that the crack area is directly facing the micro-pneumatic repair nozzle; the micro-pneumatic repair nozzle releases a low-temperature plasma pulse or an aerosol jet carrying nano-repair materials including modified starch microparticles with precise control of duration.

[0047] The plasma treatment by low-temperature plasma pulse can clean and activate the crack surface. The nano-repair particles of aerosol jet then fill and bridge the crack under the action of van der Waals forces. The whole process is completed in a suspended state, and the rice grain itself is not subject to any mechanical stress, realizing online, non-contact, minimally invasive "surgical" repair of a single grain of rice.

[0048] The micro-airflow attitude controller includes an attitude sensing unit, a flow generation and control unit, and a precise attitude control strategy. Before the rice grain is repaired, the attitude sensing unit reuses the high-precision three-dimensional data of the rice grain captured by the preceding multimodal endoplasmic perspective and polarized structured light imaging module to analyze the current position, outline, and spatial orientation of the crack surface of the rice grain in real time, providing target instructions for attitude adjustment.

[0049] The flow generation and control unit is configured with a micro-nozzle array independently controlled by a high-frequency micro-solenoid valve (response frequency greater than 1kHz) around multiple planes of the suspension channel; each nozzle in the micro-nozzle array can independently adjust the pressure and pulse duration of the ejected airflow through an algorithm; the micro-nozzle array corresponds to the array-type repair nozzle;

[0050] A precise attitude control strategy is used to generate coordinated commands for four micro-airflows, enabling multi-dimensional attitude adjustment of the rice grains. By sequentially exciting one or more pairs of opposing nozzles, a controllable high-speed vortex airflow is generated, driving the rice grains to rotate precisely around a specific axis, aligning the crack area with the repair nozzle. While rotating and positioning, the remaining nozzles continuously release low-pressure, stable coaxial enveloping laminar flow. This airflow acts like an invisible air bearing, not only counteracting the centrifugal drift that may be caused by rotation and ensuring that the rice grains are completely and stably suspended at the moment of repair, but also providing crucial contactless buffering between the rice grains and the external repair environment during the repair process.

[0051] The physical-data dual-driven process knowledge decision engine is connected to the multimodal internal quality imaging and polarized structured light imaging module and the suspension minimally invasive repair execution module. It embeds the rice grain fracture mechanics equation and fluid dynamics physics equation, and constructs a process-defect-quality knowledge graph to calculate quality level, instability index, repairability index, and generate adaptive adjustment instructions for front-end processes.

[0052] It needs to be further explained that the core of the physics-data dual-drive process knowledge decision engine is a dual-drive model that includes a physical simulation model, a deep neural network, and a process knowledge graph.

[0053] The data-driven path of the physical-data dual-driven process knowledge decision engine consists of a high-performance convolutional neural network that integrates multimodal features, responsible for quickly identifying common defects and evaluating appearance levels;

[0054] The physical-data dual-driven process knowledge decision engine's physical-driven path: In the training and inference of the dual-driven model of the process knowledge graph, rice grain fracture mechanics, fluid dynamics, and electromagnetic scattering equations are embedded as constraints; this gives the dual-driven model of the process knowledge graph the ability to calculate the stress intensity factor of internal cracks observed by the swept-frequency optical coherence tomography unit and predict their dynamic instability risk.

[0055] Knowledge Graph and Causal Reasoning: A knowledge graph for rice manufacturing, consisting of "process-defect-quality", has been constructed. When the dual-drive model of the process knowledge graph determines a defect, it will activate the causal reasoning chain in the graph. For example, the defect is likely caused by excessive pressure in the preceding rice milling machine, and it is recommended to execute parameter A.

[0056] The physical-data dual-driven process knowledge decision engine ultimately outputs a comprehensive solution that includes quality grade, instability risk index, repairability index, and specific optimization parameter instructions for upstream processes; it transforms the role of the machine vision-based online rice color sorting and quality grading system from a passive quality inspector to an active process engineer.

[0057] I. Equations of Mesh Grain Fracture

[0058] The goal is to quantitatively assess the dynamic instability risk of a rice grain during subsequent processing (such as polishing and transportation) based on internal microcrack data obtained from nondestructive testing using a swept-frequency optical coherence tomography unit; this enables the system to move from simply observing cracks to predicting cracks.

[0059] 1. Core Mechanism: Stress Intensity Factor Criterion

[0060] This system uses the stress intensity factor (K) from linear elastic fracture mechanics as the criterion for crack instability. Basic equation form:

[0061]

[0062] in, Crack depth or characteristic size is obtained directly and non-destructively by the sweep frequency OCT unit in the multimodal internal quality perspective module;

[0063] σ: Stress acting on the crack; This system assumes a standard load stress for subsequent processing, or accurately calculates the stress state of the rice grain under a specific polishing pressure based on a digital twin model.

[0064] Y: Geometric configuration factor, which is dynamically determined in real time by the three-dimensional morphology and crack location data of the rice grains obtained by the polarized structured light 3D imaging unit through finite element fitting or table lookup method.

[0065] Instability risk quantification: The system calculates the stress intensity factor K of the current rice grain in real time. calc And compared with the fracture toughness (K) of this rice variety as determined by experiments. IC By comparing these values, the Crack Instability Index (CSI) is generated:

[0066]

[0067] When the CSI is much smaller than the threshold, the crack is stable and the rice grains can withstand conventional processing.

[0068] When the CSI approaches or exceeds the threshold, the crack is very likely to propagate and cause the rice grain to break, which is judged as a high risk of instability.

[0069] 2. Embedding method in the decision engine (physical augmentation neural network)

[0070] This equation is transformed into the regularization term of the loss function for a physical information neural network:

[0071] Loss=Loss data (Grading error) + λ⋅Loss phy

[0072] Among them, the physical loss term Loss phy The design is as follows: when the CSI value predicted by the model based on the swept-frequency optical coherence tomography unit image is inconsistent with the reference value calculated by the above fracture mechanics equation, a penalty is applied:

[0073]

[0074] II. Fluid Mechanics Physical Equations

[0075] The objective is to accurately describe the dynamic behavior of rice grains in the air flotation flow field and the repair microjet, so as to provide a precise control basis for actuators such as coaxial micro airflow attitude controllers;

[0076] 1. Core Equations: Navier-Stokes Equations and Particle Motion Equations

[0077] The system uses a simplified fluid dynamics model to describe the flow field (air can be regarded as an incompressible Newtonian fluid) and couples it with the particle trajectory equation to form a "fluid-structure interaction" model.

[0078] 1) Fluid control equations:

[0079] Describe the airflow field generated by the repair nozzle and attitude controller, and solve it using computational fluid dynamics methods:

[0080]

[0081] Where u is the airflow velocity vector and p is the pressure. and Given air density and viscosity; the accurate solution to the equations must satisfy boundary condition constraints.

[0082] 2) Equations of particle motion:

[0083] Describe the forces and motion of rice grains as discrete phases in a flow field:

[0084]

[0085] The key component, drag force, originates from:

[0086]

[0087] The drag coefficient depends on the particle shape and Reynolds number. The windward area of ​​the particles is provided by real-time imaging data;

[0088] 2. Embedding method in decision engine and execution module

[0089] 1) Forward solution and inverse problem optimization:

[0090] For attitude adjustment in minimally invasive suspension repair, the control algorithm solves the inverse problem in real time: given the current and desired attitude of the rice grain, the required airflow pulse force is calculated based on the particle motion equation. (Precise control), and then calculate the required air supply pressure and control timing for each micro-nozzle through the fluid control equation;

[0091] 2) Virtual-Real Mapping and Self-Calibration:

[0092] The physical core of the cloud-based digital twin is the high-precision offline simulation of the aforementioned fluid-structure interaction model; it runs continuously, comparing the simulation results with actual sensor data and correcting the drag coefficient online. These parameters drive the physical model to continuously self-calibrate, achieving a precise mapping between the digital world and the physical production line;

[0093] The multi-stage flexible pneumatic diversion module uses a pneumatic network without physical baffles, separated by air curtains at specific angles, to guide rice grains to the finished product, repair, or by-product channels without contact.

[0094] The process digital twin and edge self-evolution module deploys a lightweight physical augmentation neural network with offline incremental learning capabilities at the edge, and runs a process digital twin system in the cloud to maximize the overall yield rate of the production line through reinforcement learning.

[0095] What needs further elaboration is:

[0096] At the edge: Deploy lightweight physical augmentation neural networks with offline incremental learning capabilities; operators can label new defects at any time, and the model can quickly complete online evolution based on physical constraints and a small number of samples, without waiting for major version updates in the cloud;

[0097] Cloud-based: A 1:1 digital twin system running the production line receives real-time data on production line status and rice quality. It can be used not only for predictive maintenance, but also, based on reinforcement learning, autonomously explore and recommend the optimal process formula with the goal of maximizing the overall yield of the production line, achieving a leap from individual intelligence to system intelligence.

[0098] This system integrates a multi-scale air-film suspension fluidization and steady-state transport module, a multi-modal endogenous imaging and polarized structured light imaging module, a suspension minimally invasive repair execution module, a physical-data dual-driven process knowledge decision engine, a multi-level flexible pneumatic diversion module, and a process digital twin and edge self-evolution module. It proposes a fully air-float, zero-contact processing paradigm, replacing all physical chutes, tracks, and other mechanical contact components in traditional color sorting systems with a micro-air-film guiding layer. This achieves zero physical contact throughout the entire process of rice sorting, from feeding, imaging, repair to diversion. It fundamentally eliminates the secondary damage and increased broken rice rate caused by mechanical friction, collision, and compression in traditional equipment. This represents a fundamental restructuring of the entire processing flow, bringing rice sorting from the era of destructive testing to the era of non-destructive intelligent manufacturing, providing an absolutely interference-free ideal working condition for all subsequent high-precision operations.

[0099] In another embodiment of the present invention, the rising airflow velocity of the multi-scale air film suspension fluidization and steady-state transport module is precisely controlled between the suspension velocity of each type of rice and the first blowing velocity, and the micro air film guiding layer is composed of micro jets tilted in opposite directions, forming a stable constraint potential field in the axial direction of the pipe.

[0100] The swept-frequency optical coherence tomography unit has a swept-frequency light source with a center wavelength of 1280-1320 nm and an axial resolution better than 5 μm; the snapshot-type hyperspectral unit is based on a computational spectral imaging chip with a spectral range of 400-1000 nm, and can acquire a spectral cube in a single frame; the polarization structured light 3D imaging unit uses polarization-modulated infrared structured light to simultaneously acquire three-dimensional morphology and polarization degree information.

[0101] The multimodal internal quality imaging and polarized structured light imaging module integrates a swept-frequency optical coherence tomography unit, a snapshot-type hyperspectral imaging unit, and a polarized structured light 3D imaging unit. For the first time, it generates a six-dimensional digital profile for each suspended rice grain, including its three-dimensional morphology after de-reflection, internal microcrack / immature white nucleus structure, and chemical composition. This endows the machine with the ability to see through and analyze chemically. On the one hand, polarization technology effectively eliminates the masking of subtle defects such as embryo spots and light chalkiness by specular reflection on the rice grain surface. On the other hand, the fusion of swept-frequency optical coherence tomography and hyperspectral imaging can quantitatively assess the internal health status (such as whether the depth of microcracks exceeds the safety threshold and the proportion of white nucleus) and early mold growth without damaging the rice grain. This extends the grading standards from the surface to the interior, achieving a fundamental upgrade in quality judgment.

[0102] The suspended minimally invasive repair execution module includes a low-temperature plasma upper body repair mode; a coaxial micro-airflow formed by a coaxial micro-airflow attitude controller and an array-type repair nozzle is used to envelop rice grains and precisely expose their cracked surfaces; the rice grains with exposed cracked surfaces are activated and cleaned by short-time pulsed plasma jets emitted by the array-type repair nozzles.

[0103] It needs to be further explained that: the low-temperature plasma top body repair mode is a technology that, under normal pressure and open environment, uses a coaxial micro-airflow to encapsulate rice grains and then directionally emits a low-temperature plasma jet with precise control over time and energy into the microcrack area. The purpose of the low-temperature plasma top body repair mode is to perform atomic-level cleaning and activation of the inner surface of the crack, greatly enhancing its surface energy, and creating ideal interfacial chemical conditions for the subsequent spraying of nano-repair particles to firmly adsorb and bridge under the action of van der Waals forces.

[0104] The low-temperature plasma upper body repair mode mainly consists of a micro plasma jet generator and a coordinated gas supply system, specifically:

[0105] Miniature plasma jet generator: It adopts a dielectric barrier discharge structure; the central electrode is a tungsten needle, which is wrapped with a quartz glass dielectric tube on the outside, and a grounded copper ring is provided at the outlet; the working gas is ionized under the action of the high voltage electric field between the electrodes and is ejected from the nozzle to form a plasma jet;

[0106] Collaborative gas supply system: It mainly uses high-purity argon (Ar, purity ≥99.99%) as the main component, and carries a small amount of active gas. Under specific ratios, it can generate different particles such as active oxygen and active nitrogen to meet the cleaning needs of different rice types and pollution types.

[0107] Working mechanism:

[0108] 1) Ionization process: Under pulsed high voltage (amplitude 5-15kV, frequency 10-50kHz), the working gas is broken down and ionized to generate a plasma composed of high-energy electrons, ions, excited-state atoms and free radicals;

[0109] 2) Jet formation: Driven by the airflow, the plasma is ejected from the quartz tube nozzle and forms a stable jet in the air with a length of about 5-15 mm and a diameter of about 0.5-2 mm.

[0110] 3) Surface treatment: The active particles in the jet react chemically with the organic contaminants on the crack surface, and at the same time, the physically adsorbed impurities are removed by ion bombardment, thus achieving cleaning; at the same time, the high-energy particles break the surface chemical bonds and introduce polar groups, thus achieving activation.

[0111] The operation of the suspended minimally invasive repair module in the suspended minimally invasive repair process: After the physical-data dual-driven decision engine determines that a grain of rice is "repairable", the system executes the following steps:

[0112] 1) Suspension positioning: The coaxial micro-airflow attitude controller responds first, driving the rice grain to rotate in suspension, so that its crack surface faces the repair nozzle;

[0113] 2) Jet pretreatment: The plasma jet generator in the array-type repair nozzle is activated, and 1-3 pulses of low-temperature plasma jet are emitted into the crack area to complete surface cleaning and activation;

[0114] 3) Repair execution: Immediately afterwards, the same nozzle switches working mode and sprays a nano-repair aerosol jet carrying modified starch particles to fill the activated cracks with repair particles, completing the bridging repair.

[0115] The suspended minimally invasive repair execution module consists of a coaxial micro-airflow attitude controller and an array of repair nozzles. The coaxial micro-airflow first envelops the rice grain, and the vortex airflow generated by the multi-axis micro-jet array makes it rotate precisely in a suspended state, so that the crack surface faces the repair nozzle. Subsequently, the array of repair nozzles emits short-time pulsed low-temperature plasma jets to activate and clean the inner surface of the crack. This realizes surgical-level online repair of single grains on the production line. It completely subverts the wasteful logic of discarding defective rice in traditional sorting, and instead establishes a generative chain of "assessment-suspension positioning-minimally invasive repair-value reshaping". The cleaned / activated crack surface creates an ideal interface for subsequent nanomaterial filling, making it possible for micro-cracked rice that would otherwise become substandard to be repaired into superior whole rice, directly creating huge economic value.

[0116] The physical-data dual-driven process knowledge decision engine's physical-driven path specifically combines the crack depth and morphology with the grain geometry measured by the swept-frequency optical coherence tomography unit. It calculates the stress intensity factor at the crack tip using an embedded grain fracture mechanics equation to quantify instability risk. The data on crack depth, morphology, and grain geometry drives the physical-data dual-driven process knowledge decision engine's physical-driven path and generates an appearance rating. The data on both crack depth and morphology and grain geometry are output to the knowledge graph's fusion layer, generating process optimization suggestions through a causal reasoning chain.

[0117] The physics-data dual-driven process knowledge decision engine's physics-driven path is not purely data fitting. Specifically, this path combines the crack depth measured by the swept-frequency optical coherence tomography unit and the grain geometry obtained by the polarized structured light 3D imaging unit, and uses an embedded grain fracture mechanics equation (emphasizing the stress intensity factor criterion). The system calculates the stress intensity factor at the crack tip to quantify its dynamic instability risk; simultaneously, its data-driven pathway performs appearance grading; data on both crack depth and morphology and rice grain geometry are output to the fusion layer of the knowledge graph, generating process optimization suggestions through causal reasoning chains; this endows the system with decision robustness and interpretability beyond purely data-driven approaches; the quantitative calculation of the instability index allows the system to predict whether a cracked grain of rice will break in subsequent processes, rather than relying solely on static images; this reasoning ability based on physical laws enables the model to maintain reliable judgments even when facing new rice varieties or extreme working conditions; furthermore, the causal reasoning of the knowledge graph traces detected defects back to the specific process that caused them and autonomously generates adjustment instructions, transforming the sorting system from a passive quality inspector into an active "process engineer" for the first time, achieving a true closed loop between quality control and production.

[0118] The multi-stage flexible pneumatic diversion module is separated into different diversion channels by air curtains formed by clean air at specific flow rates and angles. Under the guidance of the airflow, rice grains cross the air curtains and enter different channels without any physical contact.

[0119] The lightweight physical augmentation neural network deployed at the edge can perform instantaneous incremental updates of the model based on a small number of new samples under the constraints of fluid dynamics physical equations in an offline state, which is used to quickly adapt to new defects; the process digital twin system uses the maximization of the overall line yield as the reward function and autonomously generates the optimal control strategy for the upstream process through reinforcement learning.

[0120] Specifically, the multi-level flexible pneumatic diversion module uses air curtains at specific angles to replace physical baffles to achieve contactless diversion; a lightweight physical reinforcement neural network is deployed at the edge, which can complete real-time incremental updates of the model based on a small number of new samples in an offline state, under the constraints of fluid dynamics physical equations; at the same time, the process digital twin system running in the cloud uses the maximization of the overall line yield as the reward function, and autonomously explores and transmits the optimal control strategy of the previous process through reinforcement learning; thus forming a three-level intelligent evolution architecture of "autonomous execution of physical production line, real-time evolution of edge nodes, and global optimization of cloud brain"; it has the ability to learn offline and optimize globally without cloud dependence; the incremental learning at the edge under the constraints of physical equations completely solves the industry pain points of traditional AI models being slow to update when facing new defects and relying on massive annotations; and the reinforcement learning of the cloud "twin factory" with the overall line yield as the goal enables process optimization to move from the history of relying on human experience to a new stage of intelligent simulation exploration and the distribution of process formulas, realizing the leap from single-machine intelligence to system intelligence.

[0121] In another embodiment of the present invention, addressing the problem that passive sensing relying solely on frequency-sweeping optical coherence tomography and snapshot-type hyperspectral imaging has a physical limit in terms of sensitivity for early-stage diseases that have not yet formed a significant optical density difference (such as very early-stage mold mycelium that has not yet produced pigment) or transparent immature white nuclei with a very small difference in refractive index from normal endosperm, the present invention further proposes that: the multimodal endoplasmic perspectral and polarized structured light imaging module also includes an acousto-optic co-excitation unit for endoplasmic excitation, which includes a miniature acoustic radiation force pulse generator and a multi-angle laser speckle contrast imaging channel.

[0122] As the rice grains levitate through a specific area, a coaxial micro-airflow attitude controller briefly fixes their posture. Then, an acoustic radiation force pulse generator non-contactly emits a focused ultrasonic pulse with extremely short durations (less than 1 ms) and very low energy (less than 0.1 mW) towards the grain. This focused ultrasonic pulse generates a micrometer-level transient displacement response within the rice grain. The viscoelastic modulus differs between healthy endosperm and immature white cores, cracked areas, and moldy areas, resulting in different response speeds and recovery mechanisms of the internal particles to the pulse. Simultaneously, the hyperspectral imaging channel or near-infrared channel switches to a multi-angle laser speckle contrast imaging channel, rapidly capturing changes in the internal speckle field caused by the acoustic radiation force. The immature white core region, with its low elastic modulus, exhibits a speckle decorrelation rate significantly different from that of normal endosperm. Through this combined acoustic-optical active excitation method, non-contact detection of the internal mechanical properties of rice is achieved, effectively detecting heterogeneous internal defects with the same refractive index that cannot be distinguished by traditional swept-frequency optical coherence tomography and spectroscopy.

[0123] In another embodiment of the present invention, although single-polarization structured light 3D imaging can de-reflect light and reconstruct three-dimensional morphology, the viewing angle is fixed. When rice grains undergo slight and irregular deflection in the air-float due to their irregular shape, it may cause single-frame point cloud occlusion or edge blurring, making it impossible to achieve complete 360-degree surface quality detection. Furthermore, it is proposed that the multimodal endoplasmic vision and polarization structured light imaging module also includes a fully polarized light field compound eye acquisition array. The fully polarized light field compound eye acquisition array is composed of a multi-eye compound eye lens ring surrounding the air-float channel, and each lens integrates a micro-nano polarization filter array.

[0124] The multi-view synchronous acquisition of the all-polarized light field compound eye acquisition array abandons the single-view structured light reconstruction and instead uses 16-32 miniature compound eye lenses surrounding the air-floating pipe to simultaneously acquire omnidirectional images of the rice grains. A single exposure can obtain its complete three-dimensional surface information. Each compound eye lens is covered with a micro-nano polarization filter array in front of its photosensitive element, which can simultaneously record the intensity, direction and polarization state of the light. Based on the principle of light field imaging, the post-processing algorithm can perform virtual refocusing and depth-of-field expansion on the acquired light field data, completely eliminating local defocusing caused by the slight displacement of the rice grains. At the same time, the acquisition of all-polarization information can physically and accurately separate the surface specular reflection and the diffuse reflection of the chalky area, changing the chalky detection from residual analysis after de-reflection to direct calculation based on polarization characteristics, achieving a qualitative leap in accuracy and robustness.

[0125] In another embodiment of the present invention, in response to the problem that each imaging unit only acquires data "independently and synchronously" without utilizing the air-bearing module's ability to control the grain attitude to actively improve data quality, the following is further proposed: In the multimodal endoplasmic perspectral and polarized structured light imaging module, before the grain enters the imaging area, the coaxial micro-airflow attitude controller actively drives the grain to enter the imaging area with the optimal display attitude (e.g., standardizing the germ orientation and long axis direction) based on the preliminary contour of the previous imaging, rather than passively and randomly.

[0126] While the air flotation transport propels the rice grains in a straight line, a coaxial micro-airflow attitude controller drives the rice grains to slowly rotate around their long axis at a controllable angular velocity. The swept-frequency optical coherence tomography unit uses this controlled rotation to perform a spiral scan on the rice grains, rapidly reconstructing the entire internal three-dimensional volume data of the rice grains with a completeness far exceeding that of traditional B-scan, completely eliminating "missed detection layers". The improved edge computing engine uses multi-frame, multi-view data to perform temporal domain super-resolution reconstruction, generating a final archive with spatial resolution exceeding the physical limits of a single sensor.

[0127] The combination of a fully polarized light field compound eye acquisition array and air-bearing attitude control enables the system to acquire complete, unobstructed light field data containing precise polarization information. After combining frequency-sweep optical coherence tomography and acousto-optic data, the system performs deep fusion under a unified and self-consistent physical model. The resulting six-dimensional digital archive has achieved a leap in accuracy, robustness, and information richness.

[0128] In another embodiment of the present invention, another objective of the present invention is to provide a method for an online color sorting and quality grading system for rice based on machine vision, comprising the following steps:

[0129] S1. Rice grains are transported non-contactly to the multimodal imaging area of ​​the multimodal internal matter perspective and polarized structured light imaging module through a multi-scale air flotation bed to form a quasi-weightless single-particle queue.

[0130] S2, synchronously triggered sweeping optical coherence tomography unit, snapshot hyperspectral imaging unit and polarized structured light 3D imaging unit, collect internal structure, spectrum and three-dimensional morphology data of rice grains, and generate six-dimensional digital archives;

[0131] S3, a physical-data dual-driven process knowledge decision engine, analyzes six-dimensional files, assesses quality levels, combines knowledge graphs to calculate repairability index and instability risk, and generates adaptive adjustment instructions for front-end processes.

[0132] S4. For repairable particles, the suspended minimally invasive repair module performs contactless surface repair; the multi-level flexible pneumatic diversion module diverts the rice particles into their respective channels without contact based on the final determination.

[0133] S5 and the edge model evolve in real time based on operator markings. The cloud-based digital twin runs reinforcement learning simulation based on full-line data and sends the optimal process parameters to the physical production line, forming a closed-loop intelligent manufacturing.

[0134] Furthermore, in step S3, when the causal reasoning chain of the knowledge graph determines that the root cause of a certain type of defect is excessive pressure in the preceding rice milling process, the decision engine automatically generates control instructions containing the pressure reduction ratio and the target ampere number, and sends them to the rice milling machine PLC via industrial Ethernet.

[0135] In step S4, for rice grains designated as repairable, the online color sorting and quality grading system for rice based on machine vision controls the flow rate and direction of the coaxial micro-airflow through a suspended minimally invasive repair execution module consisting of a coaxial micro-airflow attitude controller and an array of repair nozzles. This allows the rice grains to rotate slowly in a suspended state, aligning their cracked areas with the array of repair nozzles. Plasma pulse activation and nano-repair particle aerosol spraying repair are then performed sequentially, all while the rice grains remain suspended and in a non-contact state.

[0136] In another embodiment of the present invention, another objective of the present invention is to provide a method for an online color sorting and quality grading system for rice based on machine vision, comprising the following steps:

[0137] S1. Quasi-weightless air flotation transport and holographic perception: Rice grains enter a multi-scale air flotation bed to form a contactless single-particle queue; in the multi-modal imaging area, the frequency sweep optical coherence tomography unit, snapshot hyperspectral and polarization structured light unit are triggered simultaneously to establish a six-dimensional digital archive.

[0138] S2, Physical-Data Dual-Driven Intelligent Assessment: Six-dimensional data input physical-data dual-driven process knowledge decision engine calculates its quality level, crack instability index and repairability index, and traces the root process cause of defects through knowledge graph;

[0139] S3, Suspension Minimally Invasive Repair and Flexible Diversion: For rice grains determined to be repairable, the suspension minimally invasive module performs plasma or nanomaterial filling repair; the repaired rice grains are inspected and merged into the finished product stream, while the remaining rice grains are guided to their destination without contact by a multi-stage pneumatic diversion network;

[0140] S4. Digital Twin Closed Loop and Self-Evolution: The edge model learns in real time, the cloud-based "twin factory" simulates synchronously and continuously downloads the globally optimal process adjustment scheme, driving the rice milling machine, polishing machine and other equipment to adaptively adjust parameters, forming a manufacturing closed loop of autonomous optimization throughout the entire process.

[0141] In the description of this invention, although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine vision-based online color sorting and quality grading system for rice, characterized in that, include: The multi-scale air film suspension fluidization and steady-state transport module is used to generate laminar clean airflow to put rice in a quasi-weightless suspension state, and uses a micro air film guiding layer to replace the solid pipe wall, enabling single-particle sequential feeding without contact throughout the process. The multimodal internal structure imaging and polarized structured light imaging module includes a swept-frequency optical coherence tomography unit, a snapshot hyperspectral imaging unit, and a polarized structured light 3D imaging unit, which are used to generate a six-dimensional digital file for each suspended rice grain that integrates its internal structure, chemical composition, and three-dimensional morphology after de-reflection. The suspended minimally invasive repair module consists of a coaxial micro-airflow attitude controller and an array of repair nozzles, and is used to perform non-contact plasma surface activation or nanomaterial filling repair on suspended repairable rice grains. The physical-data dual-driven process knowledge decision engine is connected to the multimodal internal quality imaging and polarized structured light imaging module and the suspension minimally invasive repair execution module. It embeds the rice grain fracture mechanics equation and fluid dynamics physics equation, and constructs a process-defect-quality knowledge graph to calculate quality level, instability index, repairability index, and generate adaptive adjustment instructions for front-end processes. The multi-stage flexible pneumatic diversion module uses a pneumatic network without physical baffles, separated by air curtains at specific angles, to guide rice grains to the finished product, repair, or by-product channels without contact. The process digital twin and edge self-evolution module deploys a lightweight physical augmented neural network with offline incremental learning capabilities at the edge, and runs a process digital twin system in the cloud to maximize the overall yield rate of the production line through reinforcement learning.

2. The online color sorting and quality grading system for rice based on machine vision according to claim 1, characterized in that, The upward airflow velocity of the multi-scale air film suspension fluidization and steady-state transport module is precisely controlled between the suspension velocity of each type of rice and the first blowing velocity, and the micro air film guiding layer is composed of micro jets tilted in opposite directions, forming a stable constraint potential field in the axial direction of the pipeline.

3. The online color sorting and quality grading system for rice based on machine vision according to claim 1, characterized in that, The swept frequency optical coherence tomography unit has a swept frequency light source with a center wavelength of 1280-1320 nm and an axial resolution better than 5 μm. The snapshot-type hyperspectral unit is based on a computational spectral imaging chip with a spectral range of 400-1000 nm, and can acquire a spectral cube in a single frame; The polarized structured light 3D imaging unit modulates the infrared structured light to simultaneously acquire three-dimensional morphology and polarization degree information.

4. The online color sorting and quality grading system for rice based on machine vision according to claim 1, characterized in that, The suspended minimally invasive repair execution module includes a low-temperature plasma upper body repair mode; a coaxial micro-airflow formed by a coaxial micro-airflow attitude controller and an array-type repair nozzle is used to envelop rice grains and precisely expose their cracked surfaces; the rice grains with exposed cracked surfaces are activated and cleaned by short-time pulsed plasma jets emitted by the array-type repair nozzles.

5. The online color sorting and quality grading system for rice based on machine vision according to claim 1, characterized in that, The physical-data dual-driven process knowledge decision engine's physical-driven path specifically combines the crack depth and morphology measured by the swept-frequency optical coherence tomography unit with the grain geometry, and calculates the stress intensity factor at the crack tip through the embedded grain fracture mechanics equation to quantify the instability risk; the data-driven physical-data dual-driven process knowledge decision engine's physical-driven path of crack depth, morphology and grain geometry generates appearance rating. The data on crack depth and morphology, along with the geometric dimensions of the rice grains, are output to the fusion layer of the knowledge graph and used to generate process optimization suggestions through causal reasoning chains.

6. The online color sorting and quality grading system for rice based on machine vision according to claim 1, characterized in that, The multi-stage flexible pneumatic diversion module is separated into different diversion channels by air curtains formed by clean air at specific flow rates and angles. Under the guidance of the airflow, rice grains cross the air curtains and enter different channels without any physical contact.

7. The online color sorting and quality grading system for rice based on machine vision according to claim 1, characterized in that, The lightweight physical augmentation neural network deployed at the edge can perform instantaneous incremental updates of the model based on a small number of new samples under the constraints of the fluid dynamics physical equations in an offline state, enabling rapid adaptation to new defects. The process digital twin system uses the maximization of the overall line yield as the reward function and autonomously generates the optimal control strategy for the upstream process through reinforcement learning.

8. The online color sorting and quality grading system for rice based on machine vision according to claim 1, characterized in that, The multimodal endoplasmic perspective and polarized structured light imaging module also includes an acoustic-optical co-excitation endoplasmic excitation unit, which includes a miniature acoustic radiation force pulse generator and a multi-angle laser speckle contrast imaging channel.

9. The online color sorting and quality grading system for rice based on machine vision according to claim 1, characterized in that, The multimodal endoplasmic vision and polarized structured light imaging module also includes a fully polarized light field compound eye acquisition array, which consists of a multi-lens lens ring surrounding the air-bearing channel, with each lens integrating a micro-nano polarization filter array.

10. A method for an online color sorting and quality grading system for rice based on machine vision according to any one of claims 1-7, characterized in that, Includes the following steps: S1. Rice grains are transported non-contactly to the multimodal imaging area of ​​the multimodal internal matter perspective and polarized structured light imaging module through a multi-scale air flotation bed to form a quasi-weightless single-particle queue. S2, synchronously triggered sweeping optical coherence tomography unit, snapshot hyperspectral imaging unit and polarized structured light 3D imaging unit, collect internal structure, spectrum and three-dimensional morphology data of rice grains, and generate six-dimensional digital archives; S3, a physical-data dual-driven process knowledge decision engine, analyzes six-dimensional files, assesses quality levels, combines knowledge graphs to calculate repairability index and instability risk, and generates adaptive adjustment instructions for front-end processes. S4. For repairable particles, the suspended minimally invasive repair module performs contactless surface repair; the multi-level flexible pneumatic diversion module diverts the rice particles into their respective channels without contact based on the final determination. S5 and the edge model evolve in real time based on operator markings. The cloud-based digital twin runs reinforcement learning simulation based on full-line data and sends the optimal process parameters to the physical production line, forming a closed-loop intelligent manufacturing.

Images