Intelligent control system for food detection laboratory based on data analysis

Abstract

The application relates to the technical field of intelligent control and grain quality and safety detection, and particularly discloses a grain detection laboratory intelligent control system based on data analysis, which comprises a single-grain grain conveying device, a hyperspectral imaging unit, a medical image analysis and processing unit, a data fusion and intelligent discrimination server and a non-destructive grading execution mechanism; through single-grain detection, hyperspectral three-dimensional imaging and medical image segmentation model fusion analysis, accurate identification of worm-eaten, mildewed and toxin-enriched areas is realized, and non-destructive grading is carried out according to a multi-dimensional quality image. Through the above technical scheme, the internal and external characteristics of the grain can be synchronously obtained under the non-destructive premise, the discrimination accuracy and detection efficiency of early deterioration risks are improved, and an intelligent, traceable and clinical-level diagnosis means for grain quality and safety is provided.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control and grain quality and safety testing technology, specifically involving an intelligent control system for grain testing laboratories based on data analysis. Background Technology

[0002] With the continuous advancement of global food security standards and the construction of smart laboratories, high-precision grain testing technology has become a key support for ensuring the safety of the food supply chain at its source. Traditional grain testing mainly relies on batch sampling in physicochemical laboratories, aiming to assess the compliance of bulk grains through statistical sampling. Automated testing systems based on computer vision and data analysis have played a significant role in improving screening efficiency. Their basic principle is to use image sensors to acquire the appearance characteristics of grains and combine them with preset quality thresholds to achieve automatic classification and grading, reducing the labor intensity of manual sampling.

[0003] Existing detection technologies have limitations in handling heterogeneous quality deviations. In particular, the dilution effect of traditional batch mixing and grinding sampling methods can mask small amounts of highly toxic moldy grains with a large number of qualified samples, leading to potential food safety oversights. Traditional machine vision models can only extract external morphological features of grains, failing to reveal internal microstructural changes. Conventional chemical analysis methods are destructive and lack timeliness, making it difficult to achieve a simultaneous and accurate correlation between external characterization and internal biochemical quality in a non-destructive manner. Current analytical logic lacks in-depth analysis of the microscopic pathological characteristics of individual grains, failing to identify early-stage, hidden insect-eaten holes or deep mold hyphae networks. This results in insufficient accuracy in predicting potential grain spoilage risks, making it difficult to meet the stringent requirements of modern refined grain management.

[0004] Therefore, a data-driven intelligent control system for grain testing laboratories is needed. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent control system for grain testing laboratories based on data analysis, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: an intelligent control system for a grain testing laboratory based on data analysis, comprising a single grain conveying device, a hyperspectral imaging unit, a medical image analysis and processing unit, a data fusion and intelligent discrimination server, and a non-destructive grading execution mechanism, wherein:

[0007] The single grain conveying device is configured to separate the grains to be inspected one by one and convey them to the inspection station in a fixed posture, ensuring that each grain maintains a stable spatial position and does not obstruct each other during the imaging process.

[0008] The hyperspectral imaging unit is configured to perform full-band spectral scanning on a single grain to obtain reflectance spectral information at different depths on its surface and inside, and generate a three-dimensional data cube containing spatial coordinates and spectral dimensions to characterize the distribution characteristics of chemical components inside the grain.

[0009] The medical image analysis and processing unit is configured to receive the three-dimensional data cube output by the hyperspectral imaging unit, call the image segmentation model and lesion identification algorithm constructed based on medical pathological image analysis logic, perform pixel-level semantic segmentation of the internal microstructure of a single grain, and identify and mark potential insect-eaten holes, mold hyphae networks or toxin-rich areas.

[0010] The data fusion and intelligent discrimination server is configured to integrate the chemometric features provided by the hyperspectral imaging unit and the morphological pathological features output by the medical image analysis and processing unit to construct a multidimensional quality profile of a single grain, and generate a comprehensive quality judgment result for the constructed grain based on preset safety thresholds and risk level rules.

[0011] The non-destructive grading actuator is configured to perform non-destructive diversion operations on the single grains that have completed testing, based on the judgment results output by the data fusion and intelligent discrimination server, and accurately guide them to the corresponding qualified product channel, suspicious product temporary storage area or high-risk waste channel.

[0012] Preferably, the hyperspectral imaging unit adopts a continuous spectrum acquisition method from visible light to near-infrared bands. Its spectral resolution is sufficient to distinguish the characteristic absorption peaks of moisture, protein, fat and typical fungal toxins in grains. Furthermore, it suppresses surface reflection interference and improves the signal-to-noise ratio of internal microstructure imaging through multi-angle illumination and polarization filtering technology.

[0013] Preferably, the image segmentation model embedded in the medical image analysis and processing unit is trained based on the U-Net neural network architecture. Its training dataset comes from hyperspectral images of moldy grain samples verified by pathological sections. It can accurately identify internal abnormal structures at the millimeter or even sub-millimeter level and supports quantitative evaluation of the volume ratio, spatial connectivity and texture complexity of abnormal regions.

[0014] Preferably, the data fusion and intelligent discrimination server adopts a multimodal feature fusion strategy. After spatially aligning the hyperspectral feature vector with the pathological segmentation mask, it uses an attention mechanism to weightedly fuse external morphological indicators and internal biochemical indicators to form a unified risk scoring system. The risk scoring system is dynamically calibrated based on historical test data and food safety standards to ensure that the discrimination results are both scientific and compliant.

[0015] Preferably, the single grain conveying device includes a vibrating feeder, a linear array of guide troughs, and a high-speed rotating positioning stage. The vibrating feeder controls the grain flow rate, the linear array of guide troughs is arranged in a single row, and the high-speed rotating positioning stage rotates each grain circumferentially at a uniform speed before detection, so that the hyperspectral imaging unit can acquire complete spectral information of its entire circumferential surface and near-surface layer.

[0016] Preferably, the non-destructive grading actuator consists of a pneumatic nozzle array and a multi-channel slide. The response time of each nozzle is less than a predetermined time. Based on the judgment result, it can trigger directional airflow within milliseconds to push a single grain of grain to the corresponding channel without contact, thus avoiding secondary damage caused by mechanical clamping.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. The intelligent control system for grain testing laboratories based on data analysis provided by this invention breaks through the problem of missed detection of highly toxic individuals caused by the "dilution effect" in traditional batch testing. By treating each grain as an independent testing unit, it realizes the transformation from homogeneous assessment to precise screening of micro-individual particles.

[0019] 2. The system integrates hyperspectral imaging and medical image analysis technologies to simultaneously acquire the external morphological characteristics and internal chemometric characteristics of grains without any destructive processes. This solves the technical bottleneck of traditional methods, which struggle to identify "hidden defects" such as intact exteriors but moldy interiors. With the help of a deep learning model derived from the field of pathological diagnosis, the system can perform three-dimensional reconstruction and qualitative analysis of the microstructures inside a single grain, such as insect-eaten holes and mold hyphae networks, thereby improving the ability to perceive and judge early spoilage risks.

[0020] 3. The entire system requires no grinding or chemical reagents, and the detection process is efficient, green, and traceable. It provides intelligent quality control methods with clinical-grade diagnostic capabilities for grain storage, processing, and distribution, and strongly supports the refined implementation of the national food security strategy. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention;

[0022] Figure 2 This is a schematic diagram of the core principle framework of intelligent discrimination that integrates hyperspectral chemical features and medical imaging morphological features in this invention;

[0023] Figure 3 This is a flowchart illustrating the main stages of pixel-level semantic segmentation and recognition of the internal microstructure of a single grain in this invention.

[0024] Figure 4This is a schematic diagram of the multi-level interaction relationship and data flow between the various front-end acquisition and execution units and the back-end data fusion server in this invention;

[0025] Figure 5 This is a schematic diagram of the logical framework for generating multimodal feature fusion scoring and triage instructions in this invention. Detailed Implementation

[0026] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0027] The intelligent control system for grain testing laboratories based on data analysis includes a single grain conveying device, a hyperspectral imaging unit, a medical image analysis and processing unit, a data fusion and intelligent discrimination server, and a non-destructive grading actuator.

[0028] The single grain conveying device is used to physically discretize the original bulk grain sample to be tested one by one. Through multi-level vibration and geometric constraint mechanisms, the randomly stacked grain particles are transformed into a single-row, equally spaced particle stream with specific spatial orientation. Each grain is then precisely conveyed to the optical path coverage area of ​​the hyperspectral imaging unit in a preset stable posture, ensuring that each grain is in a static or uniform motion state during the acquisition process, eliminating the interference of particle overlap, occlusion, or tumbling on imaging accuracy.

[0029] The hyperspectral imaging unit is used to acquire full-band spectral information of a single grain passing through the detection station under controlled illumination. The reflected light is decomposed into continuous spectral components by a spectroscopic element, and the two-dimensional spatial information and one-dimensional spectral information of the grain surface are captured simultaneously by a linear array or area array detector. A three-dimensional data cube containing spatial coordinates and spectral response intensity is constructed. The three-dimensional data cube contains the chemical composition characteristics of the grain from the epidermis to the near-surface interior at different depths, providing raw data support for subsequent qualitative and quantitative analysis of biochemical components.

[0030] The medical image analysis and processing unit is used to receive and parse the three-dimensional data cube transmitted by the hyperspectral imaging unit, treating it as digital image data similar to medical pathological slices. It calls the embedded deep learning model to preprocess, enhance features, and segment semantics of the image. By simulating the lesion identification logic in medical image diagnosis, it digitally reconstructs the microscopic physical structure inside a single grain of grain, identifies the boundary between normal tissue and abnormal structures (such as mycelial networks formed by mold infection, holes formed by insect lurking, or chemically abnormal areas formed by toxin accumulation), and generates high-precision pathological feature descriptors.

[0031] The data fusion and intelligent discrimination server is used to realize the deep fusion of multi-source heterogeneous data and decision logic judgment. On the one hand, it accesses the raw chemometric data provided by the hyperspectral imaging unit and extracts the spectral feature vectors reflecting the content of protein, fat, water and fungal toxins.

[0032] On the other hand, it accesses the morphological and pathological features output by the medical image analysis and processing unit, such as the geometric shape, texture complexity, and volume ratio of abnormal areas.

[0033] By establishing a multi-dimensional quality evaluation space, using a fusion algorithm to correlate and match the above features, a holographic quality profile of a single grain is constructed, and classification instructions for the currently detected grain are generated in real time according to preset safety level indicators.

[0034] The non-destructive grading actuator is used to perform the final physical sorting operation. It receives control commands from the data fusion and intelligent discrimination server. At the moment the grain particles flow through the sorting station, it applies a precise directional offset force to the target particles through a non-mechanical contact power medium, causing them to deviate from their original trajectory and enter the corresponding grading channel, thereby achieving automated separation of qualified products, suspicious products and waste products. The entire sorting process does not change the integrity and biological activity of the grain particles.

[0035] The single-grain conveying device includes a large-capacity vibrating feeder hopper, a flexible frequency control module, linear array guide troughs, and a high-speed rotating positioning worktable. The large-capacity vibrating feeder hopper generates adjustable-amplitude mechanical vibration through an electromagnetic vibration mechanism at its bottom, causing the accumulated grain to move towards the discharge port under the combined action of gravity and vibration force.

[0036] The flexible frequency control module is configured to dynamically adjust the vibration frequency based on the feedback speed of the back-end imaging unit, ensuring that the number of particles passing through per unit time remains constant. The linear array guide groove has a specific cross-sectional geometry, with its width slightly larger than the maximum transverse diameter of a single grain, forcing the grain particles to form a single column through physical restraint.

[0037] The high-speed rotating positioning stage is located below the imaging station and is equipped with micro negative pressure adsorption holes. At the moment of imaging, the grain is fixed by negative pressure and rotated circumferentially, so that the hyperspectral imaging unit can obtain complete information of the grain in the entire circumference, solving the problem of detection blind spots caused by single-sided imaging.

[0038] The hyperspectral imaging unit includes a broadband halogen tungsten lamp light source array, a multispectral beam splitter, a polarization filter adjustment module, and a high-sensitivity line scan camera. The broadband halogen tungsten lamp light source array adopts a ring-shaped symmetrical layout, providing uniform illumination covering the range of 400 nm to 2500 nm. Its power driver has a constant current and voltage regulation function to ensure that the fluctuation rate of luminous flux is lower than the preset extreme value.

[0039] The polarization filter adjustment module is installed at the front of the light source and camera lens. By adjusting the relative angle of the polarizer and analyzer, it suppresses specular reflection light caused by oil or moisture on the surface of grain particles and enhances the diffuse reflection signal entering the camera. The multispectral spectrometer uses a diffraction grating to disperse the reflected light to different rows of the photosensitive chip. The high-sensitivity line scan camera, synchronized with the pixel clock, continuously acquires the spectrum of the grain cross-section passing through the workstation, and finally stitches them together to form a complete three-dimensional data cube.

[0040] The medical image analysis and processing unit includes an image preprocessing module, a U-Net semantic segmentation engine, and a three-dimensional lesion reconstruction component. The image preprocessing module is configured to perform black-and-white correction, noise reduction, and pseudo-color enhancement on the original hyperspectral image to improve the visual saliency of low-contrast areas.

[0041] The U-Net semantic segmentation engine employs a deep encoder and symmetric decoder structure, preserving high-resolution spatial details through skip connection technology. Its embedded parameter matrix, trained on tens of thousands of grain samples annotated with pathological sections, can classify each pixel in a hyperspectral image as normal endosperm, damaged tissue, mycelium, or foreign impurities. The three-dimensional lesion reconstruction component integrates segmentation results acquired from multiple angles, calculates the spatial connectivity of abnormal regions within the grain, quantitatively assesses the depth distribution of insect-eaten holes and the volumetric density of mold diffusion, and outputs physically meaningful pathological quantitative parameters.

[0042] The data fusion and intelligent discrimination server includes a chemical feature extraction engine, a multimodal feature fusion matrix, and a dynamic risk assessment model. The chemical feature extraction engine uses partial least squares or principal component analysis to screen out characteristic wavelengths that are highly sensitive to specific fungal toxins (such as aflatoxin and vomitoxin) from hundreds of spectral bands and calculate their characteristic absorption intensity.

[0043] The multimodal feature fusion matrix employs an attention mechanism algorithm, assigning different weight coefficients to spectral and morphological features. If hyperspectral imaging detects abnormal chemical composition and medical images identify a clear hyphal network, the different weight coefficients assigned to spectral and morphological features will undergo nonlinear coupling enhancement, significantly improving the reliability of the discrimination results. The dynamic risk assessment model stores a database of food safety access standards from different regions globally and can automatically invoke corresponding logical thresholds based on the current testing task requirements, mapping the fused comprehensive features to the corresponding grading labels.

[0044] The non-destructive grading actuator includes a high-speed electromagnetic valve array, a pressure-adaptive compensated air source, and a multi-channel collection chute. The high-speed electromagnetic valve array consists of multiple micro-nozzles, and the response delay of each nozzle is strictly controlled within the millisecond range to ensure accurate hitting of target particles in a high-speed grain flow.

[0045] The pressure-adaptive compensation air source is equipped with a precision pressure reducing valve and an air storage tank, which can automatically adjust the jet pressure according to the quality of the grain to be inspected (such as the gravity difference between corn and rice), preventing pressure overload from causing particle breakage or insufficient pressure from causing sorting failure. The multi-channel collection chute is lined with a polymer buffer material to reduce the impact force when the grain particles fall, ensuring that the sorted grain still maintains a good appearance and marketability.

[0046] Furthermore, in the single-grain conveying device, the surface of the linear array of guide channels is coated with a Teflon anti-friction layer. This layer has an extremely low coefficient of friction, preventing grain particles from adhering during conveying or undergoing changes in biochemical properties due to frictional heat. The adsorption pressure of the high-speed rotary positioning worktable is monitored in real time by a proportional pressure sensor. When the sensor detects that the adsorption pressure is lower than a set threshold, the control system automatically triggers a cleaning pulse, using reverse airflow to remove dust from the adsorption pores, ensuring the reliability of continuous operation.

[0047] Furthermore, in the hyperspectral imaging unit, the acquisition frequency of the line scan camera and the conveying speed of the single-grain grain conveying device are linked in real time via a hardware trigger bus. When the encoder of the conveying device detects that a grain particle has entered a predetermined trigger position, it immediately sends an exposure pulse to the camera to ensure that the physical displacement corresponding to each frame of the image is completely consistent, eliminating image stretching or compression distortion caused by motion jitter. In addition, the hyperspectral imaging unit also includes a standard reflectance calibration plate. Before the start of each detection task, the system automatically pushes the standard reflectance calibration plate into the imaging area for automatic calibration to compensate for the effects of light source attenuation and ambient light drift.

[0048] Furthermore, within the medical image analysis and processing unit, the U-Net semantic segmentation engine integrates a transfer learning module, allowing the system to rapidly fine-tune the model with only a small number of labeled samples when detecting different types of grains (e.g., switching from wheat to soybeans). The U-Net semantic segmentation engine also includes a robust enhancement layer, which, by simulating bias field correction techniques in medical images, automatically compensates for uneven illumination distribution caused by the extremely irregular shape of the grains, ensuring that the segmentation accuracy in the edge regions remains consistent with that in the central region.

[0049] Furthermore, in the data fusion and intelligent discrimination server, the dynamic risk assessment model possesses self-evolution capabilities. During operation, the system stores the data of each grain identified as "suspicious" and its corresponding measured physicochemical indicators (if subsequent laboratory verification data are available) in a cloud database. Through deep reinforcement learning algorithms, the discrimination boundary is continuously optimized, enabling the system to adapt to changes in grain production areas and the drift of basic components caused by climate fluctuations during long-term operation, thereby improving the generalization performance of the discrimination system.

[0050] Example 2: As a further optimization and supplement to Example 1, this example provides an intelligent control system for a grain testing laboratory based on a distributed architecture, which aims to solve the computational load bottleneck and system real-time requirements in large-scale, ultra-high-speed grain screening tasks.

[0051] The intelligent control system for grain testing laboratories based on data analysis includes a front-end distributed acquisition array, an edge computing node cluster, a centralized decision server, and a distributed execution array.

[0052] The front-end distributed acquisition array comprises multiple parallel single-grain grain conveying branches and hyperspectral acquisition modules. Each branch independently performs grain discretization and spectral imaging. Through multi-path concurrent processing, the system's total throughput is significantly improved. Each hyperspectral acquisition module is equipped with an independent local cache for temporarily storing high-bandwidth raw spectral data, which is then distributed to the corresponding edge computing nodes via a high-speed Ethernet interface.

[0053] The edge computing node cluster consists of multiple high-performance graphics processing units (GPUs), with each node corresponding to one or more front-end acquisition branches. The edge computing nodes are responsible for performing real-time image preprocessing, spatial-spectral calibration, and basic feature extraction tasks. Since medical image processing involves a large number of convolutional operations, deploying U-Net segmentation logic at the edge allows the processed semantic mask, rather than the original massive image, to be transmitted to the central server, reducing the transmission pressure on the backbone network. The edge nodes are also responsible for real-time attitude monitoring of the conveying device, using visual feedback algorithms to fine-tune the rotation speed of the rotary positioning stage to ensure the integrity of the image for each grain of grain.

[0054] The centralized decision server is the logical brain of the entire system. It operates in a highly available computing environment and is responsible for receiving processing results from various edge nodes. The central server maintains a global index of single-grain characteristics and uses a multimodal fusion algorithm to cross-verify spectral consistency and pathological morphological similarity.

[0055] The central server also integrates a knowledge graph-based risk inference engine, which can dynamically adjust the risk assessment weight of individual grains based on the overall statistical distribution patterns of the current batches being tested. For example, when it detects that grains from a certain region generally have a risk of exceeding the standard for a certain type of mold, the system will automatically increase the sensitivity coefficient of that mold exceeding the standard feature in the discrimination logic.

[0056] The distributed execution array comprises multiple sets of sorting actuators, each corresponding to a front-end acquisition branch. To ensure precise synchronization of execution, each actuator is directly driven by a local logic controller. The local logic controller receives delayed trigger commands from the central server and accurately calculates the jetting timing based on the precise physical distance required for the grain to travel from the imaging point to the execution point and the real-time belt speed. The system also includes a sorting effect feedback loop, which monitors the material flow rate of each channel in real time through weight sensors or photoelectric gates installed at the end of the diversion channels. If an abnormal sorting ratio is detected (such as a large number of particles that should be discarded appearing in the qualified product channel), an alarm is automatically triggered and the system enters self-test mode.

[0057] In terms of hardware configuration, the single-grain conveying device in this embodiment adopts airflow suspension guidance technology to replace the physical contact guide trough. By setting a micro-pore matrix at the bottom of the conveying track, a stable air film is formed under the grain particles using controlled clean compressed air, so that the grain is in a near frictionless suspension state. This not only eliminates physical wear, but also makes ultra-high-speed conveying (the number of grains processed per second is increased to more than 3 times the original level) possible.

[0058] In this embodiment, the hyperspectral imaging unit is upgraded to a dual-sided omnidirectional imaging structure. A set of hyperspectral cameras and a light source are arranged above and below the conveying path, respectively, to achieve simultaneous scanning of the top and bottom surfaces of the grain particles using optical synchronization technology. For the detection of internal microstructures, this embodiment introduces time-domain gating imaging technology. Through the combination of an extremely short pulse light source and an ultra-fast shutter, multiple scattering light from the surface is filtered out, directly acquiring the pathological characteristics of deep grain tissues and improving the detection rate of occult lesions.

[0059] In this embodiment, the medical image analysis and processing unit incorporates a three-dimensional generative adversarial network (GAN). When hyperspectral scan data is locally missing due to occlusion or other reasons, the GAN can physically fill in the missing areas based on the principles of grain symmetry and historical anatomical features, thereby generating a more complete three-dimensional pathological model. The medical image analysis and processing unit integrates a multi-scale feature extractor, capable of simultaneously capturing sub-millimeter-level micro-mold spots and centimeter-level structural damage, ensuring that detection accuracy is unaffected by changes in defect scale.

[0060] In this embodiment, the data fusion and intelligent judgment server employs a blockchain-based quality traceability framework. For each grain of grain tested, its holographic quality profile, testing time, geographical coordinates, and judgment result are encapsulated into data blocks and encrypted for storage, ensuring the authenticity and immutability of the test data. For high-risk batches, the system can automatically identify all related grains from the same origin and transport vehicle through the traceability chain, achieving precise risk mitigation.

[0061] In this embodiment, the non-destructive grading actuator is upgraded to a multi-stage flexible paddle array. Compared to pneumatic sorting, the flexible paddles offer higher energy conversion efficiency and lower noise when processing large grain particles (such as corn and soybeans). The paddles are driven by piezoelectric ceramics, enabling action switching on a microsecond scale. Their surfaces are coated with medical-grade silicone material to ensure no physical indentations are created when in contact with the grain.

[0062] Furthermore, this embodiment also includes a laboratory environment intelligent sensing subsystem. This subsystem utilizes temperature and humidity sensors, dust concentration sensors, and vibration sensors distributed throughout the experimental space to monitor the impact of environmental conditions on detection accuracy in real time. When environmental humidity exceeds a preset range, causing a shift in the hyperspectral absorption peak, the system automatically activates a compensation algorithm to perform deconvolution repair on the spectral curve based on the humidity value. When excessive ground vibration is detected, the system automatically adjusts the camera's exposure compensation and filtering parameters to ensure clinical-grade detection accuracy is maintained even in complex laboratory environments.

[0063] Example 3: This example focuses on describing an intelligent control system optimized for extremely heterogeneous grain samples (such as highly irregular grains with husks or attachments). Based on Examples 1 and 2, the intelligent control system enhances the physical feature compensation and deep semantic parsing capabilities.

[0064] The intelligent control system for grain testing laboratories, based on data analysis, incorporates an adaptive morphology recognition mechanism in its single-grain conveying device. Before the grain enters the hyperspectral imaging zone, a low-power structured light 3D scanner acquires a rough outline of the grain. Based on this outline data, the flexible mechanical fingers in the conveying device automatically adjust the clamping force and angle of the grain, ensuring that the longest axis of each grain remains parallel to the scanning line of the imaging camera. This preprocessing mechanism simplifies the geometric correction required for subsequent image analysis.

[0065] In this embodiment, the hyperspectral imaging unit employs acousto-optic tunable filter (AOTF) technology. Compared to traditional diffraction gratings, AOTF allows the system to randomly access specific spectral bands at extremely high frequencies. During detection, the system first performs a rapid full-band scan. Once the medical imaging unit identifies a suspected lesion area, the imaging unit immediately switches to a narrow characteristic band with extremely high sensitivity for lesions (such as Fusarium head blight and smut) for a fine secondary scan. This "coarse-to-fine" detection strategy ensures both detection speed and the ultimate accuracy in capturing key risk indicators.

[0066] In this embodiment, the medical image analysis and processing unit integrates knowledge graph-driven auxiliary diagnostic logic. This logic compares real-time segmented microstructural features with an expert-level grain pathology knowledge base, enabling "multi-labeling" of complex lesion types. For example, it can not only identify the presence of pores inside the grain but also accurately determine, based on the edge roughness of the pores and the spectral characteristics of residues on the inner wall, whether the pores were caused by specific types of storage pests or by secondary mold growth following mechanical damage. This in-depth qualitative analysis provides scientific decision support for subsequent grain storage management.

[0067] In this embodiment, the data fusion and intelligent discrimination server employs a multimodal alignment technique based on contrastive learning. Due to the extremely high biodiversity of grain samples, traditional fixed threshold methods are prone to misjudgment. This server uses a contrastive learning framework to learn the relative distance between "healthy grains" and "pathological grains" in a multidimensional feature space. Even when different grain varieties are mixed, the system can accurately identify high-risk individuals by finding outliers in the feature space (i.e., those individuals whose distribution is significantly different from healthy samples), achieving true single-grain-level anomaly detection.

[0068] In this embodiment, a visual verification module is added to the non-destructive grading actuator. A high-speed, high-definition camera is installed at the actuator's exit to capture real-time images of the sorted waste particles. If a large number of misclassified healthy particles are detected among the sorted particles, the system will immediately adjust the front-end discrimination logic or the actuator's action parameters through closed-loop control signals. This "execution-feedback-optimization" closed-loop mechanism ensures the long-term operational stability of the system in an unattended state.

[0069] In addition, the system includes an intelligent maintenance reminder module. By analyzing the cumulative illumination time of the hyperspectral light source, the spectral energy decay rate, and the response time trend of the pneumatic actuator, a predictive maintenance algorithm is used to assess the remaining lifespan of each core component and automatically prioritize maintenance according to the importance level of the components, ensuring that the laboratory system is always in optimal working condition.

[0070] Example 4: This example describes an intelligent control system for a grain testing laboratory with edge self-healing and cloud-based collaborative capabilities. Its core feature lies in the deep integration of the precision of single-grain testing with the macroscopic predictive capabilities of a big data platform.

[0071] In this architecture, the single-grain conveying device is equipped with an online self-cleaning unit. This unit uses a high-pressure ion air gun to electrostatically de-staticate and clean the stage after each conveying cycle, preventing residual fine powder (such as mold spore dust) from causing cross-contamination of the next grain or obstructing the optical window. The conveying device incorporates a precision weighing sensor, enabling simultaneous measurement of the mass and density of each grain, providing a crucial physical benchmark for biochemical analysis.

[0072] The hyperspectral imaging unit employs multi-excitation source fluorescence spectral compensation technology. In addition to conventional reflectance spectral acquisition, the system utilizes ultraviolet lasers of specific wavelengths to excite fluorescent substances within the grain. Many mycotoxins exhibit significant fluorescence characteristics; by overlaying fluorescence emission spectra with hyperspectral reflectance data, the system can detect trace toxin-rich areas with significantly higher sensitivity, achieving a detection limit more than 10 times that of conventional methods.

[0073] The medical image analysis and processing unit employs an automatically optimized model based on Neural Architecture Search (NAS). As detection data accumulates, the system automatically searches for the most suitable deep learning network topology for the current grain type and detection task. When processing grains with extremely complex textures (such as peanuts in their shells), the system automatically increases the depth of convolutional layers and the number of residual connections to extract deeper structural features; while when processing grains with simple textures (such as refined rice), the model is automatically simplified to improve processing speed.

[0074] The data fusion and intelligent discrimination server integrates a federated learning protocol. This protocol allows food testing laboratories in different geographical locations to jointly train and optimize discrimination models without sharing raw, sensitive data. By exchanging encrypted gradient information, each laboratory system can obtain the latest spoilage characteristics from other regions in real time, building a global collaborative defense network against new mold risks.

[0075] In this embodiment, the non-destructive grading actuator is configured with multi-level recycling logic. The system can not only distinguish between qualified and unqualified grains, but also perform more detailed classification based on the degree of grain deterioration (such as mild insect damage, deep mold, physical breakage, etc.). Grains with mild defects can be directed to the feed-grade processing area, while deeply moldy grains enter the biofuel conversion area, maximizing the remaining value of the grain samples.

[0076] Finally, the system is equipped with a remote control terminal based on augmented reality (AR). Laboratory managers can use tablets or dedicated head-mounted devices to observe the real-time "3D physical examination" dynamic image of each grain of grain inside the detection system, and watch the model delineate the lesion area in real time. When the system encounters an extreme edge sample that cannot be determined, it will automatically push the image to an expert terminal for remote manual assistance in judgment. The judgment result will then be immediately fed back to the system for online learning, realizing a closed-loop evolution of human-machine collaboration.

[0077] In summary, this invention combines the rigorous logic of medical image analysis with the rich biochemical dimensions of hyperspectral imaging to construct a single-grain-level "holographic physical examination" system for grains. It not only solves the dilution effect problem of traditional sampling inspection from a physical structure perspective, but also achieves digital insight into hidden defects within grains from a logical level. The entire system provides extremely high detection accuracy and processing efficiency while ensuring non-destructive operation, offering a revolutionary technological means for modern food security management.

[0078] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A data-analysis-based intelligent control system for grain testing laboratories, characterized in that, The system includes a single grain conveying device, a hyperspectral imaging unit, a medical image analysis and processing unit, a data fusion and intelligent discrimination server, and a non-destructive grading execution mechanism. The single grain conveying device is configured to perform physical-level discretization processing on bulk grain samples to be inspected and to steadily convey the grains to be inspected to the inspection station in a fixed posture, ensuring that each grain maintains a stable spatial position during the imaging process. The hyperspectral imaging unit is configured to perform full-band spectral scanning on a single grain of grain passing through the detection station, obtain reflectance spectral information of its surface and different internal depths, and generate a three-dimensional data cube containing spatial coordinates and spectral dimensions. The medical image analysis and processing unit is connected to the hyperspectral imaging unit and is configured to receive the three-dimensional data cube. Through the embedded image segmentation model and lesion recognition algorithm, it performs pixel-level semantic segmentation of the internal microstructure of a single grain of grain, and identifies and marks the internal abnormal structural regions. The medical image analysis and processing unit includes an image preprocessing module, a semantic segmentation engine, and a three-dimensional lesion reconstruction component. The image preprocessing module is configured to perform black-and-white correction, noise reduction, and pseudo-color enhancement on the original hyperspectral image to improve the visual saliency of low-contrast areas. The semantic segmentation engine adopts a deep encoder and symmetric decoder structure, and retains high-resolution spatial details through skip connection technology. Its embedded parameter matrix is ​​trained with pre-labeled grain samples and configured to classify pixels in hyperspectral images into normal tissue, damaged tissue, mycelium or foreign impurities. The three-dimensional lesion reconstruction component is configured to integrate the segmentation results acquired from multiple angles, calculate the spatial connectivity of abnormal areas within the grain, quantitatively assess the depth distribution of insect-eaten holes and the volume density of mold spread, and output physically meaningful pathological quantitative parameters. The semantic segmentation engine also integrates a transfer learning module, configured to enable rapid fine-tuning of the model by inputting a small number of labeled samples when detecting different types of grains; The data fusion and intelligent discrimination server is connected to the hyperspectral imaging unit and the medical image analysis and processing unit, respectively. It is configured to integrate the chemometric features provided by the hyperspectral imaging unit and the morphological pathological features output by the medical image analysis and processing unit to construct a multi-dimensional quality profile of a single grain of food, and generate a quality judgment result for the constructed single grain of food based on a preset safety threshold. The non-destructive grading actuator is connected to the data fusion and intelligent discrimination server and is configured to perform non-destructive diversion operations on the single grains that have completed the detection based on the judgment result.

2. The intelligent control system for grain testing laboratories based on data analysis according to claim 1, characterized in that: The single grain conveying device includes a large-capacity vibrating feeder hopper, a flexible frequency control module, a linear array of guide troughs, and a high-speed rotating positioning worktable; the large-capacity vibrating feeder hopper generates mechanical vibration through an electromagnetic vibration mechanism at its bottom, causing the accumulated grain to move toward the discharge port. The flexible frequency control module is configured to dynamically adjust the vibration frequency according to the feedback speed of the back-end imaging unit to ensure that the number of particles passing through per unit time remains constant. The width of the linear array guide groove is greater than the maximum transverse diameter of a single grain. Through physical limiting, the grain particles are forced to form a single column. The surface of the linear array guide groove is coated with a friction-reducing layer to prevent the grain particles from adhering during the conveying process. The high-speed rotating positioning stage is located below the imaging station. Its surface is provided with micro negative pressure adsorption holes, which are configured to fix the grain with negative pressure at the moment of imaging and drive the grain to rotate circumferentially, so that the hyperspectral imaging unit can acquire the spectral information of the grain in the entire circumference.

3. The intelligent control system for grain testing laboratories based on data analysis according to claim 1, characterized in that: The hyperspectral imaging unit includes a broadband light source array, a multispectral beam splitter, a polarization filter adjustment module, and a high-sensitivity line scan camera. The broadband light source array adopts a ring-shaped symmetrical layout to provide uniform illumination covering the visible to near-infrared bands. Its power driver has a constant current and voltage regulation function to ensure that the fluctuation rate of light flux is lower than the preset extreme value. The polarization filter adjustment module is installed at the front of the light source and the camera lens. It is configured to suppress specular reflection light formed on the surface of grain particles due to oil or moisture by adjusting the relative angle between the polarizer and the analyzer, and enhance the diffuse reflection signal entering the camera. The multispectral dispersion component uses a diffraction grating to disperse the reflected light to different rows of the photosensitive chip; The high-sensitivity line scan camera and the single grain conveying device are linked in real time through a hardware trigger bus. Under the synchronization of the pixel clock, the cross-sectional spectrum of the grain passing through the workstation is continuously collected and finally pieced together to form a complete three-dimensional data cube.

4. The intelligent control system for grain testing laboratories based on data analysis according to claim 1, characterized in that: The data fusion and intelligent discrimination server includes a chemical feature extraction engine, a multimodal feature fusion matrix, and a dynamic risk assessment model; The chemical feature extraction engine utilizes a feature selection algorithm to screen out characteristic wavelengths highly sensitive to mycotoxins from the spectral bands, and calculates... Its characteristic absorption intensity; The multimodal feature fusion matrix employs an attention mechanism algorithm to assign different weight coefficients to spectral and morphological features. When hyperspectral imaging detects abnormal chemical composition and medical imaging identifies significant internal structural damage, the weight coefficients are nonlinearly coupled and enhanced. The dynamic risk assessment model stores a database of food safety access standards at different levels and is configured to call the corresponding logical thresholds according to the current detection task requirements, mapping the fused comprehensive features to the corresponding grade labels. The dynamic risk assessment model has self-evolution capability and is configured to store the judgment results and their corresponding physicochemical indicators into a database, and continuously optimize the judgment boundary through deep reinforcement learning algorithms.

5. The intelligent control system for grain testing laboratories based on data analysis according to claim 1, characterized in that: The non-destructive graded actuator includes a high-speed electromagnetic valve array, a pressure adaptive compensation air source, and a multi-channel collection slide. The high-speed electromagnetic valve array consists of multiple micro-nozzles, each with a response delay in the millisecond range, configured to accurately hit target particles in a high-speed grain flow. The pressure adaptive compensation air source is equipped with a precision pressure reducing valve and an air storage tank, which is configured to automatically adjust the jet pressure according to the quality of the grain to be inspected, so as to prevent pressure overload from causing particle breakage or insufficient pressure from causing sorting failure. The multi-channel collection chute is lined with a polymer buffer material, which is configured to reduce the impact force when grain particles fall, and ensure that the sorted grains maintain their appearance integrity. The non-destructive grading actuator is equipped with a visual verification module at its exit, which is configured to capture images of the sorted particles using a high-definition camera. When the misjudgment ratio is found to exceed a preset range, a closed-loop adjustment signal is fed back to the data fusion and intelligent discrimination server.

6. The intelligent control system for grain testing laboratories based on data analysis according to claim 1, characterized in that: The single grain conveying device also includes an online self-cleaning unit and a precision weighing sensor; the online self-cleaning unit uses a high-pressure ion air gun to eliminate static electricity and clean dust from the platform after each conveying cycle to prevent residual powder from causing cross-contamination of subsequent detected particles. The precision weighing sensor is configured to measure the mass of a single grain and send the mass data to the data fusion and intelligent discrimination server as a physical reference parameter for biochemical analysis. The single grain conveying device adopts airflow suspension guidance technology, which sets a micro-pore matrix at the bottom of the conveying track and uses controlled compressed air to form an air film under the grain particles, so that the grain is in a frictionless suspension state.

7. The intelligent control system for grain testing laboratories based on data analysis according to claim 1, characterized in that: The system adopts a distributed architecture, including a front-end distributed acquisition array, an edge computing node cluster, a centralized decision server, and a distributed execution array; The front-end distributed acquisition array includes multiple sets of parallel single-grain grain conveying branches and hyperspectral acquisition modules. The edge computing node cluster consists of multiple high-performance graphics processing units, configured to perform real-time image preprocessing, spatial spectral calibration, and primary feature extraction tasks, and transmit the processed semantic mask to the centralized decision server. The centralized decision server is configured to receive processing results from each edge node, use a multimodal fusion algorithm to cross-verify spectral consistency and pathological morphological similarity, and integrate a risk inference engine based on a knowledge graph. It is configured to dynamically adjust the risk judgment weight of individual particles by combining the overall statistical distribution pattern of the test batch.

8. The intelligent control system for grain testing laboratories based on data analysis according to claim 1, characterized in that: The hyperspectral imaging unit has a dual-sided omnidirectional imaging structure, with a set of hyperspectral cameras and a light source arranged above and below the conveying path, respectively, and synchronous scanning of the top and bottom surfaces of the grain particles is achieved through optical synchronization technology. The hyperspectral imaging unit also incorporates time-domain gating imaging technology, configured to filter out multiple scattered light from the surface and obtain pathological characteristics of deep grain tissues through the combination of pulsed light source and ultra-fast shutter. The medical image analysis and processing unit introduces a three-dimensional generative adversarial network, which is configured to physically fill in the missing areas when there are local gaps in the hyperspectral scan data, based on the principle of grain symmetry and historical anatomical features, and generate a complete three-dimensional pathological model.

9. The intelligent control system for grain testing laboratories based on data analysis according to claim 1, characterized in that: The data fusion and intelligent discrimination server integrates a blockchain-based quality traceability framework, configured to encapsulate the quality profile, detection time, geographical coordinates, and judgment results of each detected grain into encrypted data blocks for storage. The system also includes a laboratory environment intelligent sensing subsystem, which uses temperature and humidity sensors, dust concentration sensors and vibration sensors distributed in the experimental space to monitor the environmental status in real time, and when the environmental humidity exceeds the preset range, causing the hyperspectral absorption peak to shift, it activates a compensation algorithm to perform deconvolution repair on the spectral curve. The system is also equipped with an augmented reality-based remote control terminal, configured to observe the three-dimensional dynamic image of each grain of grain inside the detection system in real time, and push the image to an expert terminal for remote manual assistance when encountering edge samples that cannot be determined.

Images