Vacuum fresh corn production line data acquisition method and system

Abstract

This invention relates to the technical field of data acquisition methods for corn production lines, specifically: a data acquisition method and system for a vacuum fresh corn production line, comprising the following steps executed sequentially: Step 1, unified spatiotemporal tracking of individual corn: assigning a unique snowflake ID to each corn cob in the production line, achieving cross-process tracking of corn through industrial cameras and conveyor belt encoders, and establishing a unified spatiotemporal benchmark for the entire process using a two-layer scheme of process PTP synchronization and full-process NTP relative timestamp calibration; Step 2, raw material quality condition acquisition: predicting the intrinsic quality parameters of a single corn cob based on the RGB image of the corn appearance using a pre-trained lightweight convolutional neural network, synchronously acquiring sorting equipment condition data and binding it with the corresponding corn snowflake ID, completing data standardization preprocessing; This invention solves the core problems of existing vacuum fresh corn production lines, such as the imbalance between data acquisition cost and accuracy, the disconnect between sterilization control and quality, serious data silos, and insufficient traceability accuracy.

Description

Technical Field

[0001] This invention relates to the technical field of data acquisition methods for corn production lines, and in particular to a data acquisition method and system for a vacuum fresh corn production line. Background Technology

[0002] Vacuum-packed fresh corn, a core category in my country's deep-processed agricultural products sector, combines nutrition, convenience, and regional characteristics. In recent years, its market size has continued to expand, and consumers and regulatory authorities have increasingly stringent requirements regarding its food safety traceability, product quality consistency, and shelf-life stability. The production process of vacuum-packed fresh corn encompasses multiple core steps, including raw material sorting, washing and blanching, cooling and draining, vacuum packaging, high-temperature sterilization, and quality inspection and sorting. The process parameters, equipment operating status, and raw material quality characteristics at each stage directly determine the food safety and quality performance of the final product. Full-process, high-precision, and interconnected production line data collection is the core foundation for achieving intelligent control of fresh corn production, precise food safety traceability, and continuous optimization of product quality.

[0003] Currently, production line data acquisition solutions used in the industry generally suffer from a core contradiction between acquisition accuracy and implementation cost. Existing batch-level data acquisition solutions can only achieve a general record of production data for the entire batch, and cannot match the differences in raw material quality and processing history of different individual corns within the same batch. When quality problems such as bloated bags, excessive microorganisms, and abnormal taste occur, it is impossible to accurately locate the process and root cause of the problem, making it difficult to achieve refined control of the production process. On the other hand, solutions that can achieve individual-level data acquisition require equipping each product with RFID tags and inkjet coding equipment, and deploying hyperspectral detection equipment, array sensors, and a precision clock synchronization system at all nodes throughout the process. The hardware investment and continuous consumable costs are extremely high, making it difficult to implement in large-scale industrial production. Summary of the Invention

[0004] In response to the technical problems mentioned in the background art, the present invention provides a data acquisition method and system for a vacuum fresh corn production line.

[0005] The technical solution adopted in this invention is: a data acquisition method for a vacuum fresh corn production line, comprising the following steps performed sequentially: Step 1, Unified Spatiotemporal Tracking for Individuals: Assign a unique snowflake ID to each corn cob in the production line, and achieve cross-process tracking of corn through industrial cameras and conveyor belt encoders. Adopt a two-layer scheme of process PTP synchronization and full-process NTP relative timestamp calibration to establish a unified spatiotemporal benchmark for the entire process. Step 2, Raw material quality condition data collection: Using a pre-trained lightweight convolutional neural network, predict the intrinsic quality parameters of a single corn cob based on the RGB image of the corn appearance, simultaneously collect sorting equipment condition data and bind it to the corresponding corn snowflake ID, and complete the data standardization preprocessing. Step 3: Thermal history calculation and quality extraction: Collect process parameters for blanching and cooling processes, calculate the thermal processing history and center temperature change curve of a single corn cob based on the unsteady-state heat conduction equation, and predict the maturity and quality characteristics of the corn after thermal processing by combining the initial quality parameters. Step 4: Packaging process sealing data collection: High-precision collection of cavity vacuum degree, heat sealing temperature, heat sealing pressure and timing process parameters throughout the vacuum packaging process, binding them with the corresponding corn snowflake ID, and synchronously collecting packaging appearance defects and leakage rate sealing performance data; Step 5, F-value calculation and quality control: Based on the raw material quality and heat processing quality characteristics of a single corn cob, dynamically calculate the individual-specific optimal sterilization F0 target value; collect temperature field and pressure data of the entire sterilization process, establish a dynamic correction model to calculate the instantaneous center temperature of a single corn cob in real time, calculate the individual cumulative sterilization F-value in real time based on the Arrhenius formula, and simultaneously predict the quality indicators of the finished product after sterilization. Step 6: Environmental data collection for finished product quality: Collect environmental parameters such as temperature, humidity, and cleanliness throughout the entire production line workshop. Simultaneously complete full-scale appearance defect detection and sampling physicochemical, microbiological, and sensory quality testing of finished corn, and establish a feature set relating environmental parameters to finished product quality. Step 7: Data Alignment and Blockchain Traceability: The dynamic time warping algorithm is used to complete the spatiotemporal alignment of multi-source data for the entire process of a single corn cob. The SM3 cryptographic hash algorithm is used to generate individual feature hash values. Based on the batch Merkle tree, the blockchain is used to complete the tamper-proof evidence storage and build a traceability query system at the level of a single corn cob. Step 8: Closed-loop feedback model optimization: Based on the data collected throughout the entire process, the quality prediction and risk warning model is incrementally iteratively trained to predict the quality of finished products and the risk of defective products in real time. The process parameter optimization instructions are output and sent to the production line PLC control system to complete the adaptive closed-loop adjustment of the production process.

[0006] In one embodiment, the Snowflake ID adopts a 14-bit encoding rule consisting of a 6-bit date code, a 4-bit batch ID, and a 4-bit sequence code. The corn ID is bound across processes using the YOLOv8-nano target detection algorithm and the DeepSORT tracking algorithm. In the dual-layer synchronization scheme, the spatiotemporal alignment of cross-process data is completed based on the corn process trigger timestamp.

[0007] In one embodiment, the lightweight convolutional neural network uses the MobileNetV3 backbone network and outputs four quality parameters—soluble sugar content, moisture content, starch content, and maturity grade—based on dual-view RGB images of corn (front and side views). The intrinsic quality parameters have a prediction determination coefficient R0. 2 ≥0.9; The equipment operating condition data acquisition frequency is 10Hz, and the preprocessing is completed by moving average filtering, 3σ criterion outlier removal, and Min-Max normalization.

[0008] In one embodiment, the unsteady heat conduction equation introduces a corn ear taper correction coefficient, and the center temperature change curve during the corn blanching process is obtained by solving the finite difference method, with a center temperature calculation error ≤0.5℃; a thermal history and maturity quality model is constructed based on the random forest regression algorithm to predict the corn maturity grade, sugar loss rate and risk of epidermal damage.

[0009] In one embodiment, the individual-specific optimal sterilization F0 target value is calculated using a mapping model constructed by a gradient boosting tree algorithm. The model input features include the moisture content, soluble sugar content, starch content, maturity level, heat processing maturity level, and initial center temperature of the corn raw material. The instantaneous center temperature of a single corn cob is calculated using a dynamic correction model calibrated by multi-point fiber optic temperature sensors within the sterilization basket. The cumulative sterilization F value is calculated at a frequency of 20Hz, and the corn's texture, sugar retention rate, vitamin retention rate, and microbial safety redundancy are predicted in real time based on the cumulative F value.

[0010] In one embodiment, the vacuum packaging process parameters are acquired at a frequency of 20Hz, based on the feeding sequence of the multi-station packaging machine and the corn visual tracking time window; the packaging sealing performance is assessed by using a high-speed linear array camera to identify surface defects such as edge wrinkles, offsets, burns, and pinholes, combined with the vacuum attenuation method to detect the packaging leakage rate.

[0011] In one embodiment, the batch Merkle tree incorporates the characteristic hash value of a single corn cob, and only the batch Merkle root is written to the consortium blockchain main chain. The entire process data of a single corn cob is stored in IPFS distributed storage, and the hash verification value and IPFS address are retained on the chain. The traceability QR code enables the query and verification of the entire process production data, quality inspection data and blockchain evidence information of a single corn cob.

[0012] In one embodiment, the model incremental iteration cycle is once a week, and the quality prediction, F-value mapping, and risk warning models are updated based on the new production data; the defective product risk warning threshold is set to 80%, and when the warning is triggered, the optimized value of the process parameters is output based on the SHAP feature importance analysis and sent to the production line PLC control system through the OPCUA protocol to complete the adaptive adjustment of the production process.

[0013] In one embodiment, a data acquisition system for a vacuum fresh corn production line includes, in sequence, an individual tracking spatiotemporal unified module, a raw material quality and working condition acquisition module, a thermal history calculation and quality extraction module, a packaging process sealing acquisition module, an F-value calculation and quality control module, a finished product quality and environmental acquisition module, a data alignment blockchain traceability module, and a closed-loop feedback model optimization module. The individual tracking spatiotemporal unified module is used to assign a unique snowflake ID to a single corn stalk, realize cross-process ID binding of corn through industrial vision and encoder, and provide a unified spatiotemporal benchmark for data collection throughout the entire process. The raw material quality condition acquisition module is used to predict the internal quality parameters of a single corn cob based on the corn appearance image through a pre-trained neural network, synchronously collect sorting equipment condition data and bind it with the corresponding corn ID, and complete data preprocessing. The thermal history calculation and quality extraction module is used to collect thermal parameters of blanching and cooling processes, calculate the thermal processing history and core temperature of a single corn cob based on the heat conduction model adapted to corn, predict the quality characteristics after thermal processing, and bind and store them. The packaging process sealing acquisition module is used to collect vacuum packaging process time sequence data, accurately bind it with the ID of a single corn cob, and simultaneously collect packaging sealing performance data and mark defective products. The F-value calculation quality control module is used to collect sterilizer operating data, generate the optimal sterilization F0 target value for a single corn cob based on the quality characteristics of the entire corn process, calculate the corn center temperature and cumulative sterilization F value in real time, and simultaneously compare the F0 target value and predict the quality of the finished product after sterilization. The finished product quality environment acquisition module is used to collect environmental parameters of the entire production line and associate them with the corresponding corn ID, collect finished product quality inspection data to verify the previous prediction results, and extract the correlation features between the environment and finished product quality. The data alignment blockchain traceability module is used to complete the spatiotemporal alignment of multi-source data for a single corn cob throughout the entire process. It writes the corn data feature hash value into the consortium blockchain for storage through a batch Merkle tree, supporting full-chain traceability query for a single corn cob. The closed-loop feedback model optimization module is used to complete the periodic incremental iteration of each prediction model based on the new production data, predict the quality of finished products and the risk of defective products in real time, output process optimization instructions and send them to the production line PLC system to realize the adaptive closed-loop adjustment of the production process.

[0014] The beneficial effects of this invention are: Compared with existing technologies, this invention solves the core problems of existing vacuum fresh corn production lines, such as the imbalance between data acquisition cost and accuracy, the disconnect between sterilization control and quality, severe data silos, and insufficient traceability accuracy. By assigning a unique snowflake ID to each corn cob, combined with industrial camera visual tracking and a dual-layer clock synchronization scheme, it achieves precise tracking and data binding of individual corn cobs throughout the entire process at extremely low material costs, completely breaking down data silos between processes. By combining real-time center temperature and cumulative sterilization F-value calculation, it fundamentally resolves the industry dilemma of incomplete sterilization versus over-sterilization. At the same time, through full-dimensional data acquisition and model iterative optimization, it constructs an adaptive closed loop for the production process, coupled with a blockchain-based individual traceability system based on Merkle trees, which not only fully releases the value of production data but also meets the dual traceability needs of regulators and consumers.

[0015] Second, compared with the existing technology, the present invention uses a contactless visual tracking and master-slave encoder displacement correction scheme, which eliminates the need for consumables such as RFID tags and inkjet printing. It can achieve stable binding of the ID of a single corn cob throughout the entire process with only a small number of industrial cameras. Combined with a hierarchical clock synchronization strategy, it significantly reduces the hardware implementation cost while ensuring the accuracy and spatiotemporal consistency of individual-level data collection.

[0016] Third, compared with existing technologies, this invention fundamentally solves the dilemma of sterilization safety and product quality. It constructs a full-link linkage mechanism encompassing individual corn raw material quality, thermal processing process, dynamic optimal sterilization F-value target, real-time F-value calculation, and post-sterilization quality prediction. Based on the full-process quality characteristics of a single corn cob, a dedicated sterilization F-value target is customized. Simultaneously, a temperature field correction model within the sterilizer enables real-time calculation of the individual-level center temperature and F-value, replacing the traditional batch-uniform sterilization control mode.

[0017] Fourth, compared with existing technologies, this invention transforms collected production data into the core driver for finished product quality prediction, defective product risk warning, and process adaptive optimization by aligning and fusing multi-source data in time and space and combining it with continuous iterative optimization of the core model. At the same time, through the batch Merkle tree blockchain evidence storage architecture, it takes into account both the immutability of individual-level traceability data and the efficiency of on-chain, and realizes full-chain accurate traceability at the level of a single ear of corn. This not only meets the stringent traceability requirements of food safety supervision, but also allows consumers to conveniently verify the information of the entire production process of the product. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a content block diagram of the Snowflake ID in this invention; Figure 3 This is a system block diagram of the present invention; Figure 4This is a schematic diagram of the specific process in step 1 of the present invention; Figure 5 This is a schematic diagram of the specific process in step 2 of the present invention; Figure 6 This is a schematic diagram of the specific process in step 3 of the present invention; Figure 7 This is a schematic diagram of the specific process in step 4 of the present invention; Figure 8 This is a schematic diagram of the specific process in step 5 of the present invention; Figure 9 This is a schematic diagram of the specific process in step 6 of the present invention; Figure 10 This is a schematic diagram of the specific process in step 7 of the present invention; Figure 11 This is a schematic diagram of the specific process in step 8 of the present invention. Detailed Implementation

[0019] In the description of this invention, it should be noted that the terms "front", "up", "down", "left", "right", "vertical", "horizontal", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0020] refer to Figure 1-Figure 1 To address the problems existing in the background technology, this application proposes the following technical solution: a data acquisition method and system for a vacuum fresh corn production line, comprising the following steps performed sequentially: Step 1, Unified Spatiotemporal Tracking for Individuals: Assign a unique snowflake ID to each corn cob in the production line, and achieve cross-process tracking of corn through industrial cameras and conveyor belt encoders. Adopt a two-layer scheme of process PTP synchronization and full-process NTP relative timestamp calibration to establish a unified spatiotemporal benchmark for the entire process. This step provides a low-cost hardware foundation and a unified spatiotemporal reference for end-to-end individual-level data acquisition. The specific implementation method is as follows: The unique identification coding rule for each individual ear of corn is established by adopting a snowflake ID coding rule consisting of a 6-digit date code (YYMMDD), a 4-digit batch ID, and a 4-digit sequence code. Each ear of corn entering the production line is assigned a globally unique 14-digit digital ID. The date code is accurate to the production day, the batch ID corresponds to the production batch on that day, and the sequence code corresponds to the feeding order of the corn within a single batch. The coding capacity can meet the needs of large-scale production of up to 10,000 ears per batch, thus solving the risk of duplicate coding across years and days caused by the original 4-digit date code.

[0021] Non-contact individual cross-process tracking is achieved by deploying two 1920×1080 resolution, 30fps CMOS industrial RGB cameras at the feed inlet of the production line to capture frontal and side views of the corn. These cameras are paired with incremental encoders (1000P / R resolution) coaxially mounted on the drive roller of the conveyor belt. An auxiliary encoder is added to the driven roller, and the difference between the master and slave encoders corrects for displacement errors caused by conveyor belt slippage. The cameras capture real-time images of the corn at the feed inlet, and the YOLOv8-nano lightweight target detection algorithm is used to identify the corn target. A confidence threshold is set... Set to 0.7 and simultaneously assign snowflake IDs to the identified corn; based on the DeepSORT target tracking algorithm, and with the conveyor belt displacement data output by the encoder as an aid, the position of the corn on the conveyor belt is tracked in real time. At the feeding inlets of the four key processes of washing and blanching, cooling and draining, vacuum packaging, and sterilization pot feeding, one industrial RGB camera of the same specification is deployed as a trigger node. An auxiliary trigger camera is added at the turning point of the conveyor belt to avoid ID loss caused by corn stacking and disordered order. No physical RFID tags, inkjet printing and other consumables are required throughout the process. Individual tracking of the whole process can be achieved with only a small number of industrial cameras.

[0022] Lightweight Sensing Node Deployment Solution: This solution employs a tiered deployment strategy, focusing on precise deployment in core processes and data reuse in non-core processes. High-precision sensors are added only to the two core processes of vacuum packaging and high-temperature sterilization. For non-core processes such as raw material sorting, washing and blanching, cooling and draining, air drying, and sorting, the existing PLC control system of the production line is directly connected via the OPCUA protocol. The operating parameters and status data of the equipment are reused. Only a single low-cost Pt100 platinum resistance temperature sensor (Class A accuracy, measurement range -20℃~200℃) is added at the inlet and outlet of each process. This avoids the high cost of a full array of sensors, reducing the overall hardware deployment cost by more than 60%.

[0023] A hierarchical spatiotemporal reference unification is achieved by employing a dual-layer synchronization scheme: precise PTP synchronization for core processes and relative timestamp calibration via NTP throughout the entire process. For temperature acquisition nodes in the high-temperature sterilization process, an IEEE 1588 PTP slave clock module is installed, connected to the PTP master clock of the production line main controller via industrial Ethernet to achieve sub-microsecond clock synchronization with a synchronization accuracy of ≤100ns, ensuring the time accuracy of F-value calculation. Acquisition nodes, cameras, and encoders in other processes are synchronized at the millisecond level via the local area network NTP network time protocol with a synchronization accuracy of ≤1ms. Using the visual recognition timestamp of corn at each process trigger node as a relative reference, a linear interpolation algorithm is used to complete the spatiotemporal alignment of data from different acquisition frequencies and nodes, eliminating time offsets in cross-process data. While ensuring the accuracy of core data, the implementation cost of the clock synchronization system is reduced by more than 70%.

[0024] Basic data storage rules: All collected raw data is appended with a triple tag consisting of a Snowflake ID, a UTC timestamp, and a parameter name, and written in real time to the Redis real-time database of the edge computing node in key-value pair format, providing a foundation for individual data association in subsequent steps.

[0025] I. Corn Target Recognition and Snowflake ID Assignment Based on YOLOv8-nano, with an example as follows: YOLOv8-nano is the lightweight model with the smallest number of parameters (approximately 3.2M) and the fastest inference speed in the YOLOv8 series. It is very suitable for real-time deployment on edge devices in production lines. Its specific implementation process is as follows: Model preparation and training; Dataset Construction: Collect RGB images of corn from different locations on the production line, such as the feed inlet and sorting process. The samples need to cover: Different varieties (mainstream varieties such as sweet corn and waxy corn); Different states (with husks, without husks, with broken kernels, stacked); Different environments (natural light, workshop lighting, steam and water mist interference).

[0026] The sample size is no less than 5,000 images. Use tools such as LabelImg to annotate the bounding boxes (bboxes) of corn targets. The annotation format is YOLO format (category ID + center point x / y + width / height). Only one category is set: "corn".

[0027] Model training: Fine-tuning was performed on a self-built dataset based on the official YOLOv8-nano pre-trained model: The input size is set to 640×640 (balancing accuracy and speed). Data augmentation employs Mosaic, MixUp, random flipping, and brightness adjustment to enhance the model's robustness to the complex environment of the production line; The loss function uses a combination of CIoU (bounding box regression), BCE (classification), and DFL (distribution focus loss); The training rounds were set to 100, the batch size to 16, the initial learning rate to 0.001, and cosine annealing decay was used.

[0028] Model Quantization and Deployment: After training, the model is exported to ONNX format and quantized using TensorRT with INT8. After quantization, the model inference speed is increased by 2-3 times. When deployed on edge computing nodes of the production line (such as NVIDIA Jetson Nano or industrial control computers), the inference time for a single 1920×1080 image is ≤15ms, meeting the 30fps real-time requirement.

[0029] Real-time inference and target recognition; Image preprocessing: The 1920×1080 image captured by the industrial RGB camera at the feed inlet is first resized to 640×640, then normalized (pixel value / 255), and converted into a tensor input model in NCHW format.

[0030] Forward propagation and post-processing: After the model outputs the prediction results through forward propagation, the following post-processing is performed: Confidence filtering: Only retain prediction boxes with a category confidence score ≥ 0.7 (i.e., the 0.7 threshold set in the claims), and remove false positives with low confidence scores (such as background clutter and shadows); Non-maximum suppression (NMS): Using NMS with an IoU threshold of 0.5, duplicate detection boxes for the same corn bud are removed, resulting in a unique bounding box for each corn bud.

[0031] Snowflake ID assignment; ID encoding rules implemented: The Snowflake ID is generated according to the rule of "6-digit date code (YYMMDD) + 4-digit batch ID + 4-digit sequence code". The date encoding is obtained from the system time of the edge node (e.g., March 25, 2026 is recorded as 260325). Batch IDs are issued daily by the production line PLC system according to the production sequence (e.g., the 3rd batch of the day is recorded as 0003). The sequence code is the order in which corn is fed into a single batch, starting from 0001 and incrementing, with a maximum of 9999 ears per batch.

[0032] ID is bound to the detection box: For each corn bounding box with a confidence level ≥ 0.7 identified, the latest Snowflake ID is immediately assigned to it, the sequence code is automatically incremented by 1, and the "Snowflake ID + bounding box coordinates and detection timestamp" is stored in the edge node Redis cache to provide initial data for subsequent tracking.

[0033] II. Real-time tracking of corn based on DeepSORT+ encoder displacement, illustrated in the following example: DeepSORT is a multi-target tracking algorithm that adds appearance feature (ReID) to SORT. Combined with encoder displacement data from the production line conveyor belt as an aid, it can significantly improve the tracking stability of corn on the conveyor belt. The specific implementation process is as follows: DeepSORT basic workflow adaptation for appearance feature extraction: After YOLOv8-nano detection, a lightweight ReID network (such as MobileNetV2) is added to extract a 128-dimensional appearance feature vector for each corn bounding box, which is used to distinguish different individual corns and avoid ID switching when stacking.

[0034] Kalman Filter State Design: Considering the motion characteristics of corn on the conveyor belt, the state vector of the Kalman filter is designed as follows: ; in, The coordinates of the center point of the bounding box. The bounding box width and height, The velocity of the center point in the image coordinate system (initially 0). Hungarian matching association: For the detection box in the current frame and the tracking trajectory in the previous frame, three cost matrices are calculated: bounding box IoU cost matrix; appearance feature cosine distance cost matrix; motion Mahalanobis distance cost matrix; weighted fusion (IoU weight 0.4, appearance weight 0.4, motion weight 0.2) is used to obtain the final cost matrix, and the association between the detection box and the tracking trajectory is completed by the Hungarian algorithm.

[0035] Encoder displacement data-assisted tracking (core innovation): Encoder data acquisition and conversion. An incremental encoder (1000P / R resolution) is coaxially mounted on the drive roller of the conveyor belt, which outputs pulse signals in real time. The edge nodes collect the number of pulses through a counter and convert it into the physical displacement of the conveyor belt according to the following formula: ; in, The pulse increment between two tracking operations. The diameter of the drive roller. The encoder resolution. Combined with camera calibration parameters (intrinsic and extrinsic parameters), the physical displacement is... Convert to pixel displacement in image coordinate system (Along the direction of the conveyor belt).

[0036] Kalman filter state correction: Pixel shifts converted by the encoder As external observations, corrections are added to the Kalman filter's prediction and update process: Prediction phase: First, predict the state using the conventional Kalman filter. ; Displacement correction: Adjust the predicted center point x-coordinate Add encoder pixel displacement The corrected predicted x-coordinates are obtained. ; Update phase: The corrected predicted state is associated with the current detection box to update the tracking trajectory. This correction effectively solves the tracking loss of corn on the conveyor belt caused by occlusion (such as stacking before entering the blanching tank) and image blurring (such as steam interference), with a tracking ID retention rate of ≥99.9%.

[0037] Cross-process ID verification and rebinding are implemented at the inlet of four key processes: washing and blanching, cooling and draining, vacuum packaging, and sterilizer loading. Industrial RGB cameras of the same specifications are deployed as trigger nodes. When corn enters the trigger node's field of view, the trigger node's YOLOv8-nano identifies the corn and extracts its appearance features; Perform cosine distance matching between the appearance features and the appearance features of the tracking trajectory cached in the previous process (threshold ≤ 0.3). After a successful match, the snowflake ID from the previous process is rebound to the detection box of the current trigger node to complete the cross-process ID transfer and ensure that the snowflake ID of the same corn is unique throughout the entire process.

[0038] The above technical solution utilizes a dual-view industrial camera paired with a master-slave encoder to achieve contactless, end-to-end tracking of individual corn ears. This eliminates the need for additional consumables such as RFID tags and inkjet printing equipment. A hierarchical clock synchronization strategy is employed, with precision clock synchronization used in the core sterilization process and network time protocol used for calibration in other processes. Furthermore, lightweight sensing nodes are deployed to directly interface with the existing PLC system on the production line, reusing equipment operation data. This solution significantly reduces hardware investment and ongoing consumable costs, solving industry-wide problems such as tracking inaccuracies caused by conveyor belt slippage and misaligned ID bindings across processes. It also establishes a unified spatiotemporal benchmark for the entire process, ensuring all collected data accurately corresponds to a single corn ear, completely breaking down data silos between processes. This lays a solid hardware and data foundation for subsequent end-to-end individual-level data collection, correlation analysis, and quality traceability.

[0039] Step 2, Raw material quality condition data collection: Using a pre-trained lightweight convolutional neural network, predict the intrinsic quality parameters of a single corn cob based on the RGB image of the corn appearance, simultaneously collect sorting equipment condition data and bind it to the corresponding corn snowflake ID, and complete the data standardization preprocessing. This step targets the raw material sorting process, achieving accurate prediction of the intrinsic quality of all individual corn kernels at extremely low hardware costs. Simultaneously, it completes the linked acquisition and individual binding of equipment operating data, providing core foundational data for subsequent quality F-value linkage control. This solves the problems of the original solution where a single camera cannot acquire dual-view images and the model's insufficient adaptability for different varieties. The specific implementation method is as follows: Construction of a pre-trained prediction model for appearance features and intrinsic qualities: Dataset Construction: Before the production line goes into operation, select no fewer than 1,000 sets of corn samples covering two major categories, sweet corn and waxy corn, no fewer than 10 mainstream varieties, different maturity levels, and different origins. The sample size for each mainstream variety is no fewer than 100 sets to ensure the model's adaptability to different varieties. For each set of samples, RGB dual-view images and laboratory calibration data are collected simultaneously. Laboratory calibration uses a hyperspectral imager to detect three core intrinsic quality parameters of the samples: soluble sugar content, moisture content, and starch content. The maturity level of the samples is determined by manual grading (levels 1-5), forming a labeled dataset with one-to-one correspondence between appearance images and intrinsic quality parameters. 80% of the dataset is used as the training set, 10% as the validation set, and 10% as the test set.

[0040] Model Training and Optimization: A lightweight convolutional neural network, MobileNetV3, was used as the backbone network. RGB dual-view images of corn were used as input, and soluble sugar content, moisture content, starch content, and maturity level were used as outputs. Joint training was performed using the MSE loss function (for regression tasks) and the cross-entropy loss function (for classification tasks). The Adam optimizer was used to update the network weights. The learning rate was set to 0.001, and the training epochs were set to 100. The early stopping strategy was set to stop training if the validation set loss did not decrease for 10 consecutive epochs. After model training, the model was validated using a test set. The prediction determination coefficient R of the intrinsic quality parameters was measured. 2 ≥0.9, maturity level accuracy ≥95%, meeting the precision requirements of industrial production, the model is deployed on the edge computing node of the production line after quantization, and the inference time for a single image is ≤10ms.

[0041] Low-cost real-time prediction of the quality of all individual corn plants: During the entire production process, the dual-view industrial RGB camera at the feed inlet deployed in step 1 collects front and side view images of each corn plant. These images are then input into a pre-trained quality prediction model, which outputs four core quality parameters in real time: soluble sugar content, moisture content, starch content, and maturity grade. This eliminates the need for a full-process deployment of hyperspectral detection equipment; only two cameras are required to achieve full-scale detection of the intrinsic quality of all individual corn plants, reducing hardware detection costs by more than 80%. The prediction results are bound to the individual corn plant's snowflake ID in real time and written into a real-time database.

[0042] Data reuse and individual binding of sorting equipment operating conditions: Communication is established with the existing PLC control system of the sorting equipment through the OPCUA protocol to collect four core operating condition data in real time: conveyor belt running speed, vibrating screen vibration frequency, air separator air pressure parameters, and grading equipment opening degree. The acquisition frequency is set to 10Hz, and each set of data is appended with a UTC timestamp. Based on the individual tracking timestamp in step 1, the equipment operating condition data within the same time window are batch-bound with the snowflake ID of the corresponding corn individual. The binding time window matches the conveyor belt stepping cycle to ensure that the operating condition data corresponds one-to-one with the corn individual.

[0043] Data standardization preprocessing: The collected raw material quality data and equipment operating condition data are preprocessed in three steps: First, random noise in the sensor-collected data is removed using a moving average filtering algorithm with a window length of 5. Second, outliers exceeding a preset threshold range are removed using the 3σ criterion, and the outliers are filled with the mean of the three valid data points before and after them in the same sequence. Third, the Min-Max normalization method is used to map all parameter data to the [0,1] interval to eliminate the dimensional differences between different parameters. The preprocessed data is stored in the real-time database and the InfluxDB time-series database with the Snowflake ID as the primary key.

[0044] In the above technical solution: a pre-trained lightweight deep learning model is used to establish a mapping relationship between the visual features of corn appearance and its internal quality. During formal production, the internal quality of all corn can be predicted in real time using only an industrial camera at the feed inlet. Simultaneously, the operating data of the existing PLC system of the sorting equipment is reused to accurately bind the equipment operating parameters with individual corn plants within the corresponding time window. The collected data is then standardized and pre-processed. This solution completely solves the high cost problem caused by the deployment of hyperspectral equipment for traditional full-volume quality inspection. It achieves full-volume quality inspection of each corn raw material with extremely low hardware investment. Furthermore, binding raw material quality data and equipment operating data to the individual dimension provides core foundational data for subsequent customized heat processing and sterilization processes. It also allows for early identification of raw material quality anomalies, controlling product quality consistency from the source of production and avoiding significant fluctuations in finished product quality due to raw material differences.

[0045] Step 3: Thermal history calculation and quality extraction: Collect process parameters for blanching and cooling processes, calculate the thermal processing history and center temperature change curve of a single corn cob based on the unsteady-state heat conduction equation, and predict the maturity and quality characteristics of the corn after thermal processing by combining the initial quality parameters. This step targets the two core heat processing steps of washing and blanching, and cooling and draining. It achieves precise quantitative calculation of the heat processing process for each individual corn kernel with extremely low hardware costs, while simultaneously predicting the quality characteristics after heat processing in real time. This provides a second core input parameter for subsequent calculation of the individual-specific sterilization F-value target. The specific implementation method is as follows: Low-cost acquisition of core parameters for the hot processing procedure: One set of Class A Pt100 platinum resistance temperature sensors (measurement accuracy ±0.1℃) is deployed at the inlet and outlet of the blanching water tank, with a sampling frequency of 10Hz, to acquire real-time water temperature distribution data. One set of the same Pt100 temperature sensors is deployed at the inlet and outlet of the cooling water tank, with a sampling frequency of 10Hz, to acquire real-time cooling water temperature data. Simultaneously, via the OPCUA protocol, four equipment operating parameters—conveyor belt speed, steam valve opening, water circulation flow rate, and fan pressure—are acquired in real-time from the existing PLC systems of the blanching and cooling equipment, with a sampling frequency of 10Hz. This eliminates the need for array-type sensor deployment, significantly reducing hardware costs.

[0046] Precise calculation of the individual thermal processing process of corn: Accurate acquisition of heat processing time: Based on the individual visual tracking data in step 1, obtain the accurate timestamps of each corn entering and leaving the blanching water tank, and calculate the actual blanching time of a single corn; similarly, obtain the accurate timestamps of each corn entering and leaving the cooling water tank, and calculate the actual cooling time, with a time calculation accuracy of ≤10ms.

[0047] Calculation of individual core temperature change curve: Based on the Fourier unsteady-state heat conduction equation, combined with real-time data of blanching water temperature, corn blanching time, conveyor belt speed, initial corn moisture content, and ear diameter, and introducing an ear taper correction coefficient, an ear heat conduction model adapted to different corn varieties is established. The equation expression is as follows: ; in, For corn ear density, For the specific heat capacity of corn, For the thermal conductivity of corn, The core temperature of the corn. Heating time, The radial coordinates of the ear of fruit The ear taper correction coefficient (calibrated based on measured values ​​for the variety, ranging from 0.85 to 1.05) is used. The above equation is solved using the finite difference method to calculate the center temperature change curve and cumulative thermal history value for each ear of corn during the blanching process. Similarly, based on the cooling water temperature and cooling time, the temperature drop history and final center temperature of the corn during the cooling process are calculated. This allows for accurate calculation of the individual-level thermal processing history without the need for implanting sensors in the corn, minimizing the error in center temperature calculation. .

[0048] Real-time extraction of individual heat-processing quality characteristics: Based on a pre-trained thermal history and ripeness quality correlation model, combined with the initial raw material quality data and blanching thermal history data of corn, the model predicts in real time three core quality characteristics of each corn cob after blanching and cooling: ripeness grade, sugar loss rate, and risk of skin damage. This model uses a random forest regression algorithm, trained on no fewer than 5000 sets of heat-processing test data, achieving a ripeness grade prediction accuracy of ≥93% and a sugar loss rate prediction accuracy of R0.05. 2 ≥0.88; The prediction result is bound to the individual corn snowflake ID in real time and stored in the real-time database, serving as the core input for subsequent individual-specific sterilization F-value target calculation.

[0049] Data preprocessing and associated storage: For the collected thermal processing parameters, calculated thermal history data, and quality prediction data, perform filtering, outlier removal, and normalization preprocessing consistent with step 2. The preprocessed data is associated across processes using the snowflake ID as the primary key and synchronously written to the time-series database.

[0050] In the above technical solution: a small number of temperature sensors are deployed at the inlet and outlet of the blanching and cooling processes. The operating parameters of existing equipment are reused, and combined with heat conduction models adapted to different varieties, the heat processing history and core temperature changes of a single corn cob are accurately calculated. Then, a pre-trained model is used to predict core quality characteristics of the corn after heat processing, such as ripeness and sugar loss. Finally, data preprocessing and cross-process associated storage are completed. This solution eliminates the need to implant detection sensors in the corn, achieving precise quantification of the individual-level heat processing history. It solves the problem that traditional solutions cannot match the differences in heat processing of individual corn cobs. Simultaneously, it accurately predicts the quality characteristics after heat processing, providing crucial input for customized F-value calculations in the subsequent sterilization process. It also enables real-time identification of process anomalies in the heat processing stage, avoiding over-blanching or under-ripening, and ensuring the stability of product quality during the heat processing stage.

[0051] Step 4: Packaging process sealing data collection: High-precision collection of cavity vacuum degree, heat sealing temperature, heat sealing pressure and timing process parameters throughout the vacuum packaging process, binding them with the corresponding corn snowflake ID, and synchronously collecting packaging appearance defects and leakage rate sealing performance data; This step targets key processes in vacuum packaging, enabling precise collection of packaging process parameters, one-to-one binding with individual corn kernels, and linked collection of packaging sealing performance test data. This provides data support for risk prevention and control of finished product packaging defects. The specific implementation method is as follows: High-precision timing data acquisition of vacuum packaging process parameters: A high-precision piezoresistive vacuum sensor (measurement range 0~101.3kPa, measurement accuracy ±0.1kPa) is deployed in the vacuum chamber of the vacuum packaging machine, with a sampling frequency of 20Hz, to collect real-time timing data of the chamber vacuum level during the vacuuming, pressure holding, and heat sealing stages. At the upper and lower heat sealing blade positions of the packaging machine, two sets of patch-type PT1000 temperature sensors (measurement accuracy ±0.2℃) and film pressure sensors (measurement range 0~100N, accuracy ±1%FS) are deployed, with a sampling frequency of 20Hz, to collect real-time timing data of the sealing blade temperature distribution and heat sealing pressure during the heat sealing process. Simultaneously, the system communicates with the packaging machine PLC system via the Modbus RTU protocol to collect three core time parameters in real-time: vacuuming time, pressure holding time, and heat sealing time, with a sampling frequency of 20Hz. All process parameters are accompanied by a PTP-synchronized UTC timestamp.

[0052] Precise binding of packaging process parameters to individual corn: For multi-station vacuum packaging machines, the feeding sequence of each station is synchronized by encoders. Based on the conveyor belt stepping cycle and the visual tracking trigger timestamp of step 1, the precise time window of each corn entering the corresponding station of the packaging machine and completing packaging is obtained. The packaging process sequence data such as vacuum degree, heat sealing temperature, heat sealing pressure, and time parameters of the corresponding station within the time window are bound one by one with the snowflake ID of the corresponding individual corn. The binding time accuracy is ≤10ms, ensuring that the packaging process data of each corn is traceable and associative, and completely solving the problem of parameter binding misalignment in multi-station equipment.

[0053] Data acquisition for packaging sealing performance testing is linked: At the exit of the vacuum packaging process, a high-speed linear scan camera with a resolution of 2560×1440 and a frame rate of 60fps is deployed sequentially, along with a differential pressure method airtightness testing device. The high-speed linear scan camera acquires high-definition images of the packaging sealing edge. Through an improved Faster R-CNN target detection model, it identifies four packaging appearance defects: sealing edge wrinkles, sealing edge offset, heat seal burns, and pinholes, with a defect identification accuracy of ≥98%. The airtightness testing device uses the vacuum attenuation method, with a detection accuracy of ±1Pa and a detection cycle of ≤2s / piece. It collects leakage rate data of the packaging in real time to determine whether there are micro-leakage defects in the packaging. All sealing performance test results are bound in real time to the snowflake ID of the corresponding corn packaging, marking defective products and simultaneously outputting rejection signals to the production line sorting mechanism.

[0054] Data preprocessing and dataset construction: The collected packaging process parameters and sealing test data are preprocessed by moving average filtering and outlier removal to construct an individual-level correlation dataset of packaging process parameters and sealing defects. The preprocessed data is stored synchronously in the real-time database and historical database with the snowflake ID as the primary key, providing data support for subsequent packaging process optimization.

[0055] In the above technical solution: high-precision sensors are deployed in the vacuum packaging machine to collect process parameters such as vacuum degree, heat sealing temperature, and heat sealing pressure in real time throughout the packaging process. The feeding sequence of the multi-station packaging machine is synchronized via encoders, accurately binding the packaging process parameters to a single ear of corn. Then, high-speed visual inspection and airtightness testing equipment complete the linked acquisition of packaging sealing performance and mark defective products. This solution solves the pain point of traditional multi-station packaging machines not being able to accurately map process parameters to individual products. It achieves individual-level linkage and binding of packaging process parameters and sealing performance test data, not only accurately tracing the packaging process details of each bag of product but also identifying problems such as micro-leakage and sealing defects in advance, reducing the risk of bloated bags and spoilage of finished products from the source. Simultaneously, the accumulated process and defect correlation data provides reliable data support for continuous optimization of the packaging process.

[0056] Step 5, F-value calculation and quality control: Based on the raw material quality and heat processing quality characteristics of a single corn cob, dynamically calculate the individual-specific optimal sterilization F0 target value; collect temperature field and pressure data of the entire sterilization process, establish a dynamic correction model to calculate the instantaneous center temperature of a single corn cob in real time, calculate the individual cumulative sterilization F-value in real time based on the Arrhenius formula, and simultaneously predict the quality indicators of the finished product after sterilization. This step achieves deep linkage between the quality characteristics of individual corn throughout the entire process and the real-time calculation of sterilization F-values, replacing the traditional batch-wide sterilization control model. It resolves the core contradiction in the industry between insufficient and excessive sterilization, and also corrects the problems of mismatched corn placement and insufficient temperature sampling points in the original solution. The specific implementation method is as follows: Dynamic calculation of the optimal sterilization F-value target for each individual corn plant: Mapping Model Construction: Based on the thermal lethality characteristics of the target bacteria (Clostridium botulinum) for sterilizing fresh corn, and combined with no less than 100,000 sets of production test data, finished product microbial test data, and sensory quality test data, the Gradient Boosting Tree (GBDT) algorithm was used to construct a mapping model between individual quality characteristics and the optimal sterilization F-value. The model input features are six core parameters of the individual corn: raw material moisture content, soluble sugar content, starch content, maturity grade, heat processing maturity grade, and initial core temperature. The model output is the optimal sterilization F0 target value specific to that corn. The model was cross-validated with a 5-fold fold, and the microbial safety assurance rate was 100%.

[0057] Real-time calculation of individual F0 target: For each ear of corn entering the sterilization process, the full-process quality characteristic data bound to its snowflake ID is retrieved from the real-time database, input into the pre-trained mapping model, and the optimal sterilization F0 target value for that ear of corn is calculated in real time. This achieves "one sterilization target for each ear of corn" control, replacing the traditional batch-wide fixed F0 value. For sweet corn with high moisture content, the F0 target value is dynamically lowered to avoid over-sterilization and resulting in a soft and mushy texture. For high-starch glutinous corn and individuals with low ripeness, the F0 target value is dynamically increased to ensure thorough sterilization, balancing food safety and product quality.

[0058] High-frequency synchronous acquisition of sterilization process parameters: Two sets of Class A Pt100 platinum resistance temperature sensors (measurement accuracy ±0.1℃) and high-precision pressure sensors (measurement accuracy ±0.1kPa) are deployed on the upper, middle and lower layers of the water bath / steam sterilizer. The acquisition frequency is set to 20Hz, and sub-microsecond time synchronization with the main controller is achieved through the IEEE1588PTP protocol to collect real-time data on the temperature field distribution and pressure time sequence data inside the sterilizer throughout the sterilization process. At the same time, it communicates with the sterilizer control system through the OPCUA protocol to collect parameters such as steam valve opening, circulating pump operating frequency, heating time, constant temperature time and cooling time in real time, with an acquisition frequency of 20Hz.

[0059] Real-time prediction of individual corn core temperature: One fiber optic temperature sensor (measurement accuracy ±0.2℃, response time ≤0.5s) is implanted in each of the three layers (top, middle, and bottom) of each sterilization basket. Data is collected synchronously during the sterilization of multiple baskets in a single batch at a frequency of 20Hz to obtain time-series data of the core temperature of corn at standard positions. Based on visual tracking data during loading into the sterilizer, the row and column positions of each corn cob in the sterilization basket are recorded. The spatial coordinates of the three layers inside the sterilizer are matched. Combined with the collected core temperature data at the standard positions, the temperature field distribution data inside the sterilizer, and the initial core temperature of the corn, a dynamic correction model of "sterilizer ambient temperature, sterilizer position, and corn core temperature" is established using a multiple linear regression algorithm. Based on this model, the instantaneous core temperature of each corn cob is calculated in real time without the need to implant sensors in each corn cob. The prediction error of individual core temperature is ≤0.3℃, providing core data for the calculation of individual F-values.

[0060] Real-time calculation of individual-level F-values ​​and quality-linked prediction: Real-time calculation of instantaneous F-value: Based on the Arrhenius thermal lethality formula, the instantaneous sterilization F-value of each corn cob is calculated in real time at a frequency of 20Hz. The formula expression is as follows: ; in, Let be the instantaneous center temperature of the corn at time t, and let Z be the thermal lethality parameter for Clostridium botulinum, taken as 10°C; simultaneously calculate the cumulative F value for each corn cob, and compare the cumulative F value with the specific F value of that corn. The target value is compared in real time.

[0061] Sterilization quality linkage prediction: While calculating the F value in real time, based on the pre-trained "cumulative F value - finished product quality" correlation model, the four core quality indicators of the corn after sterilization are predicted in real time: taste and hardness, sugar retention rate, vitamin retention rate, and microbial safety redundancy. When the cumulative F value of an individual reaches the exclusive F0 target value, the corn is marked as sterilized and the final quality prediction result is locked at the same time. All calculation data are bound and stored one by one with the individual snowflake ID of the corn.

[0062] Batch-level sterilization process auxiliary control: For the entire batch sterilization process, the uniformity of the temperature field inside the oven is monitored in real time. When the temperature difference within the same layer of the oven exceeds 1°C, an early warning signal for uneven temperature field is immediately generated and pushed to the central control system. Simultaneously, based on the individual F-value calculation results of the entire batch of corn, the sterilization constant temperature time is dynamically adjusted to ensure that all individual corn in the batch reaches the exclusive F0 target value, while avoiding overall over-sterilization, thus achieving precise control of the batch sterilization process.

[0063] The above technical solution involves: firstly, combining raw material quality and heat processing quality data from the entire corn production process, a customized optimal sterilization F-value target is determined for each ear of corn; then, through distributed sensors within the sterilizer and fiber optic temperature measurement data at standard points, a temperature field correction model is established within the sterilizer to calculate the center temperature and instantaneous and cumulative F-value of each ear of corn in real time, simultaneously predicting the quality of the finished product after sterilization, and providing auxiliary control over the batch sterilization process. This solution completely breaks away from the traditional industry model of uniform sterilization F-values ​​for each batch, achieving precise control of "one sterilization target per ear of corn." While thoroughly ensuring the microbial safety of the product, it maximizes the preservation of the corn's taste, nutrition, and freshness, solving the long-standing industry dilemma of incomplete or excessive sterilization. Furthermore, the deep integration of sterilization F-value calculation and finished product quality prediction enables real-time control and quality prediction of the sterilization process, significantly improving the consistency of finished product quality and effectively reducing the product defect rate.

[0064] Step 6: Environmental data collection for finished product quality: Collect environmental parameters such as temperature, humidity, and cleanliness throughout the entire production line workshop. Simultaneously complete full-scale appearance defect detection and sampling physicochemical, microbiological, and sensory quality testing of finished corn, and establish a feature set relating environmental parameters to finished product quality. This step enables continuous collection of environmental parameters throughout the entire production process, as well as the collection of final inspection quality data for finished products. It also verifies individual quality prediction results and performs correlation analysis between the environment and quality, providing data support for model iteration and production environment control. The specific implementation method is as follows: Low-cost collection of environmental parameters throughout the entire production process: In the workshop areas of the eight core processes of the production line, one set of temperature and humidity sensors (measurement accuracy ±0.3℃, ±2%RH) is deployed, with a collection frequency set to 1Hz, to collect real-time environmental temperature and humidity data in the workshop; in the three key cleanliness areas of raw material pretreatment, packaging, and post-sterilization cooling, one laser dust particle counter is deployed in each area, automatically collecting workshop cleanliness data every 30 minutes; one airborne bacteria sampler is deployed at the workshop entrance and the finished product warehouse, collecting the total number of airborne bacteria colonies twice a day; all environmental parameters are appended with a UTC timestamp and linked to the individual corn ID range produced within the corresponding time period.

[0065] Finished Product Final Inspection Quality Data Collection and Prediction Verification: In the finished product quality inspection and sorting process, a combination of "full-scale visual inspection and sampling laboratory testing" is adopted to complete the collection of finished product quality data. A full-scale visual inspection of the finished product packaging is conducted using an industrial camera to collect data on packaging appearance defects, corn ear mold, and kernel damage. The test results are linked to the individual corn snowflake ID. At a sampling rate of no less than 5% per batch, finished corn samples are randomly selected. Using a hyperspectral imager, texture analyzer, and rapid microbial detection equipment, five core physicochemical and microbiological indicators are tested: soluble sugar content, moisture content, hardness, total bacterial count, and coliform count. Simultaneously, sensory evaluation (taste, color, and odor) is completed by professional quality inspectors. All test data are linked to the corresponding sample's snowflake ID. The actual measured quality data of the finished product is compared with the quality prediction results after sterilization in step 5 to generate prediction deviation data, providing real label data for subsequent model iterations.

[0066] Environmental and quality correlation feature extraction: Based on the collected full-process environmental parameters, finished product microbial test data, and shelf-life test data, the correlation coefficients between workshop temperature and humidity, cleanliness, finished product microbial indicators, and shelf-life are calculated using the Pearson correlation coefficient analysis method. An environmental-quality correlation feature set is established to identify the threshold values ​​of key environmental parameters affecting the shelf life of finished products. The correlation feature set and the full-process data of individual corn are stored together in the historical database to provide data support for the refined management and control of the production environment.

[0067] The above technical solution involves deploying environmental sensors in key process areas of the production line to continuously collect environmental parameters such as workshop temperature, humidity, and cleanliness. These parameters are then linked to individual corn kernels produced during the corresponding time periods. Simultaneously, through full-scale visual inspection and sampling laboratory testing, final quality data for the finished product is collected, validating the quality prediction results of each stage. Ultimately, the correlation characteristics between the production environment and finished product quality are extracted. This solution fills the technological gap in traditional solutions that neglect the impact of the production environment on finished product quality. It achieves the linked collection and correlation analysis of environmental parameters and finished product quality data throughout the entire production process. This not only verifies the accuracy of the quality prediction models for each stage, providing real label data for model iteration and optimization, but also identifies key environmental thresholds affecting the shelf life and microbiological indicators of the finished product. This provides data support for refined control of the production environment, further ensuring the food safety and shelf-life stability of the finished product.

[0068] Step 7: Data Alignment and Blockchain Traceability: The dynamic time warping algorithm is used to complete the spatiotemporal alignment of multi-source data for the entire process of a single corn cob. The SM3 cryptographic hash algorithm is used to generate individual feature hash values. Based on the batch Merkle tree, the blockchain is used to complete the tamper-proof evidence storage and build a traceability query system at the level of a single corn cob. This step completes the spatiotemporal alignment and fusion of multi-source data throughout the entire process. Based on consortium blockchain technology, it constructs an immutable traceability chain at the level of a single ear of corn, solving the problems of block congestion and low on-chain efficiency caused by uploading all single ears of corn to the chain in the original solution. This achieves accurate traceability of the entire product chain. The specific implementation method is as follows: Spatiotemporal alignment and fusion of multi-source data throughout the entire process: The entire process data corresponding to a single corn snowflake ID is retrieved from real-time and time-series databases. This includes raw material quality data, process parameters for each step, thermal processing history data, packaging and sealing data, individual-specific F0 targets and actual cumulative F values, sterilization quality prediction data, finished product measured data, and production environment data. Based on the relative timestamp benchmark from step 1, the Dynamic Time Warping (DTW) algorithm is used to perform spatiotemporal alignment of multi-source time-series data from different acquisition frequencies and nodes, eliminating time offsets between data and generating a standardized time-series dataset for the entire corn process. The dataset uses the snowflake ID as the primary key to complete the fusion and association of all-dimensional data and is stored in a MySQL relational database.

[0069] Individual-level blockchain immutable evidence storage implementation: Consortium blockchain architecture: A consortium blockchain is built, including five types of nodes: manufacturers, market regulators, third-party testing agencies, distributors, and retailers. It adopts the PBFT consensus algorithm, with a block generation time of 10 seconds. Each node synchronously stores a complete blockchain ledger to ensure that the data on the chain is immutable and traceable.

[0070] On-chain data processing and storage: Merkle roots are generated in batches, and the feature hash values ​​of individual corn cobs are incorporated into the batch Merkle trees. For each corn cob's standardized time-series dataset, a unique 256-bit feature hash value is generated using the SM3 cryptographic hash algorithm. All individual hash values ​​are used to construct the batch Merkle trees. Only the batch Merkle roots are written to the blockchain main chain, while the detailed data of individual corn cobs is stored in IPFS distributed storage. Only the hash verification value and IPFS address are retained on the chain, which ensures that the data is immutable and greatly improves the efficiency of on-chain storage, avoids block congestion, and ensures that the on-chain data is immutable, cannot be deleted, and is permanently retained.

[0071] Traceability system setup: A unique traceability QR code is generated for each batch of corn. The QR code contains a built-in blockchain query portal and batch retrieval address. Consumers and regulators can scan the code with their mobile phones and retrieve the full-process production data, quality inspection data, and blockchain evidence information of the corresponding single corn cob through Snowflake ID. At the same time, they can verify the consistency between the hash value on the chain and the hash value of the local data to verify the authenticity of the data and achieve precise traceability of the entire chain at the single corn cob level.

[0072] The above technical solution utilizes a dynamic time warping algorithm to achieve spatiotemporal alignment of multi-source data from different collection frequencies for a single ear of corn throughout the entire process, generating a standardized full-process dataset. Then, through a batch Merkle tree structure, the feature hash value of each ear of corn is written into the blockchain for notarization. Detailed data is stored in a distributed storage system, and a QR code traceability query system for consumers and regulators is established. This solution solves the problems of spatiotemporal misalignment of multi-source data and the inability to accurately link to individual products in traditional solutions. It achieves complete integration of individual-level full-process data, while blockchain technology ensures the immutability and permanent retention of traceability data, balancing on-chain efficiency and data security. This meets the stringent requirements of food safety supervision for full-chain traceability and allows consumers to easily query the entire production process and quality information of a single ear of corn, significantly improving consumer trust and brand credibility.

[0073] Step 8: Closed-loop feedback model optimization: Based on the data collected throughout the entire process, the quality prediction and risk warning model is incrementally iteratively trained to predict the quality of finished products and the risk of defective products in real time. The process parameter optimization instructions are output and sent to the production line PLC control system to complete the adaptive closed-loop adjustment of the production process.

[0074] This step, based on the dataset collected throughout the entire process, constructs and continuously optimizes a core model linking quality and F-value, outputs process optimization instructions, and feeds them back to the production line control system, forming a complete data closed loop of "collection, analysis, optimization, and re-collection." The specific implementation method is as follows: Continuous iterative optimization of the core linkage model: A fixed model iteration cycle of once a week is set, automatically retrieving no less than 10,000 new valid production data points added this week, including corn individual full-process characteristic data, finished product measured quality data, F-value prediction deviation data, and sterilization quality prediction deviation data; an incremental learning method is used to incrementally train the prediction models of "appearance characteristics and internal quality", "thermal history and maturity quality", "individual quality and optimal F-value" mapping model, and "cumulative F-value and finished product quality" prediction model, and update the model weight parameters; after each iteration, the model accuracy is verified through the test set. If the model accuracy decreases, it is rolled back to the previous stable model, ensuring that the model's prediction accuracy continues to improve with the accumulation of production data, and can adapt to the production needs of fresh corn of different varieties, seasons, and production areas.

[0075] Real-time quality risk early warning and adaptive process optimization: The system inputs real-time, full-process characteristic data of individual corn products collected from the production line into the trained quality prediction model and defective product risk early warning model to predict the finished product quality indicators and defective product risk probability of the corn being produced in real time. When the predicted defective product risk probability exceeds a preset threshold (80%), an audible and visual early warning signal is immediately generated and pushed to the production line central control system, simultaneously marking the risk cause and corresponding process. Based on the SHAP feature importance analysis results of the model, the system outputs optimized values ​​of process parameters for the corresponding process, including core parameters such as blanching water temperature, blanching time, vacuum packaging heat sealing temperature, and sterilization constant temperature time. The optimization instructions are sent to the production line PLC control system in real time via the OPCUA protocol to complete the adaptive closed-loop adjustment of the production process and achieve early prevention and control of product quality risks.

[0076] Deep mining of full-process data value: Based on the full production data of the historical database, Pearson correlation analysis and random forest feature importance ranking methods are used to mine the quantitative mapping relationship between raw material quality, heat processing parameters, packaging process, sterilization F value and finished product quality and shelf life, and output the optimal production process parameter package for different varieties of fresh corn; at the same time, based on the accumulated sterilization data, a sterilization process database of corn with different quality characteristics is established, providing complete data support for product formula optimization, production process upgrade and shelf life extension, realizing the full-dimensional value release of the collected data.

[0077] In the above technical solution: based on the production and testing data accumulated throughout the process, the core models such as quality prediction and F-value customization are incrementally trained and iteratively optimized on a regular basis. At the same time, the real-time collected production data is input into the model to complete the finished product quality prediction and defective product risk warning, output process optimization instructions and send them to the production line PLC system to realize the adaptive adjustment of the production process. In addition, the optimal production process parameter package for different varieties of corn is output through data mining.

[0078] Another possible embodiment is a data acquisition system for a vacuum fresh corn production line, comprising an individual tracking spatiotemporal unified module, a raw material quality and working condition acquisition module, a thermal history calculation and quality extraction module, a packaging process sealing acquisition module, an F-value calculation and quality control module, a finished product quality and environment acquisition module, a data alignment blockchain traceability module, and a closed-loop feedback model optimization module, all connected in sequence. The individual tracking spatiotemporal unified module is used to assign a unique snowflake ID to a single corn stalk, realize cross-process ID binding of corn through industrial vision and encoder, and provide a unified spatiotemporal benchmark for data collection throughout the entire process. The raw material quality condition acquisition module is used to predict the internal quality parameters of a single corn cob based on the corn appearance image through a pre-trained neural network, synchronously collect sorting equipment condition data and bind it with the corresponding corn ID, and complete data preprocessing. The thermal history calculation and quality extraction module is used to collect thermal parameters of blanching and cooling processes, calculate the thermal processing history and core temperature of a single corn cob based on the heat conduction model adapted to corn, predict the quality characteristics after thermal processing, and bind and store them. The packaging process sealing acquisition module is used to collect vacuum packaging process time sequence data, accurately bind it with the ID of a single corn cob, and simultaneously collect packaging sealing performance data and mark defective products. The F-value calculation quality control module is used to collect sterilizer operating data, generate the optimal sterilization F0 target value for a single corn cob based on the quality characteristics of the entire corn process, calculate the corn center temperature and cumulative sterilization F value in real time, and simultaneously compare the F0 target value and predict the quality of the finished product after sterilization. The finished product quality environment acquisition module is used to collect environmental parameters of the entire production line and associate them with the corresponding corn ID, collect finished product quality inspection data to verify the previous prediction results, and extract the correlation features between the environment and finished product quality. The data alignment blockchain traceability module is used to complete the spatiotemporal alignment of multi-source data for a single corn cob throughout the entire process. It writes the corn data feature hash value into the consortium blockchain for storage through a batch Merkle tree, supporting full-chain traceability query for a single corn cob. The closed-loop feedback model optimization module is used to complete the periodic incremental iteration of each prediction model based on the new production data, predict the quality of finished products and the risk of defective products in real time, output process optimization instructions and send them to the production line PLC system to realize the adaptive closed-loop adjustment of the production process.

[0079] In summary, this invention addresses the core issues of existing vacuum fresh corn production lines, including the imbalance between data acquisition cost and accuracy, the disconnect between sterilization control and quality, severe data silos, and insufficient traceability accuracy. By assigning a unique snowflake ID to each individual corn cob, combined with industrial camera visual tracking and a dual-layer clock synchronization scheme, it achieves precise tracking and data binding of individual corn cobs throughout the entire process at extremely low material costs, completely breaking down data silos between processes. Furthermore, by combining real-time center temperature and cumulative sterilization F-value calculations, it fundamentally resolves the industry dilemma of incomplete sterilization versus over-sterilization. Simultaneously, through multi-dimensional data acquisition and iterative model optimization, it constructs an adaptive closed-loop production process, coupled with a Merkle tree-based blockchain-based individual traceability system, fully releasing the value of production data while meeting the dual traceability needs of regulators and consumers.

[0080] Although embodiments of the invention have been shown and described, the scope of the invention will be defined by the appended claims and their equivalents by those skilled in the art.

Claims

1. A data acquisition method for a vacuum fresh corn production line, characterized in that, This includes the following steps performed sequentially: Step 1, Unified Spatiotemporal Tracking for Individuals: Assign a unique snowflake ID to each corn cob in the production line, and achieve cross-process tracking of corn through industrial cameras and conveyor belt encoders. Adopt a two-layer scheme of process PTP synchronization and full-process NTP relative timestamp calibration to establish a unified spatiotemporal benchmark for the entire process. Step 2, Raw material quality condition data collection: Using a pre-trained lightweight convolutional neural network, predict the intrinsic quality parameters of a single corn cob based on the RGB image of the corn appearance, simultaneously collect sorting equipment condition data and bind it to the corresponding corn snowflake ID, and complete the data standardization preprocessing. Step 3: Thermal history calculation and quality extraction: Collect process parameters for blanching and cooling processes, calculate the thermal processing history and center temperature change curve of a single corn cob based on the unsteady-state heat conduction equation, and predict the maturity and quality characteristics of the corn after thermal processing by combining the initial quality parameters. Step 4: Packaging process sealing data collection: High-precision collection of cavity vacuum degree, heat sealing temperature, heat sealing pressure and timing process parameters throughout the vacuum packaging process, binding them with the corresponding corn snowflake ID, and synchronously collecting packaging appearance defects and leakage rate sealing performance data; Step 5, F-value calculation and quality control: Based on the raw material quality and heat processing quality characteristics of a single corn cob, dynamically calculate the individual-specific optimal sterilization F0 target value; collect temperature field and pressure data of the entire sterilization process, establish a dynamic correction model to calculate the instantaneous center temperature of a single corn cob in real time, calculate the individual cumulative sterilization F-value in real time based on the Arrhenius formula, and simultaneously predict the quality indicators of the finished product after sterilization. Step 6: Environmental data collection for finished product quality: Collect environmental parameters such as temperature, humidity, and cleanliness throughout the entire production line workshop. Simultaneously complete full-scale appearance defect detection and sampling physicochemical, microbiological, and sensory quality testing of finished corn, and establish a feature set relating environmental parameters to finished product quality. Step 7: Data Alignment and Blockchain Traceability: The dynamic time warping algorithm is used to complete the spatiotemporal alignment of multi-source data for the entire process of a single corn cob. The SM3 cryptographic hash algorithm is used to generate individual feature hash values. Based on the batch Merkle tree, the blockchain is used to complete the tamper-proof evidence storage and build a traceability query system at the level of a single corn cob. Step 8: Closed-loop feedback model optimization: Based on the data collected throughout the entire process, the quality prediction and risk warning model is incrementally iteratively trained to predict the quality of finished products and the risk of defective products in real time. The process parameter optimization instructions are output and sent to the production line PLC control system to complete the adaptive closed-loop adjustment of the production process.

2. The data acquisition method for a vacuum fresh corn production line according to claim 1, characterized in that, In step 1, the Snowflake ID adopts a 14-bit encoding rule consisting of a 6-bit date code, a 4-bit batch ID, and a 4-bit sequence code. The corn ID is bound across processes using the YOLOv8-nano target detection algorithm and the DeepSORT tracking algorithm. In the dual-layer synchronization scheme, the spatiotemporal alignment of cross-process data is completed based on the corn process trigger timestamp.

3. The method of claim 1, wherein, In step 2, the lightweight convolutional neural network uses the MobileNetV3 backbone network and outputs four quality parameters—soluble sugar content, moisture content, starch content, and maturity grade—based on dual-view RGB images of corn (front and side views). The intrinsic quality parameters have a prediction determination coefficient R0. 2 ≥0.9; The equipment operating condition data acquisition frequency is 10Hz, and the preprocessing is completed by moving average filtering, 3σ criterion outlier removal, and Min-Max normalization.

4. The method of claim 1, wherein, In step 3, the unsteady heat conduction equation introduces the ear taper correction coefficient, and the center temperature change curve during the corn blanching process is obtained by solving the finite difference method, with a center temperature calculation error ≤0.5℃; a thermal history and maturity quality model are constructed based on the random forest regression algorithm to predict the corn maturity grade, sugar loss rate and risk of skin damage.

5. The data acquisition method for a vacuum fresh corn production line according to claim 1, characterized in that, In step 5, the individual-specific optimal sterilization F0 target value is calculated through a mapping model constructed by the gradient boosting tree algorithm. The model input features include the moisture content of corn raw materials, soluble sugar content, starch content, maturity level, heat processing maturity level, and initial core temperature. The instantaneous center temperature of a single corn cob is calculated using a dynamic correction model calibrated by multi-point fiber optic temperature sensors inside the sterilization basket; the cumulative sterilization F-value is calculated at a frequency of 20Hz, and the corn's texture, sugar retention rate, vitamin retention rate, and microbial safety redundancy are predicted in real time based on the cumulative F-value.

6. The method of claim 1, wherein, In step 4, the vacuum packaging process parameters are collected at a frequency of 20Hz, based on the feeding sequence of the multi-station packaging machine and the corn visual tracking time window; the packaging sealing performance is assessed by using a high-speed linear array camera to identify defects such as edge wrinkles, offsets, burns, and pinholes, and by using the vacuum attenuation method to detect the packaging leakage rate.

7. The method of claim 1, wherein, In step 7, the batch Merkle tree incorporates the characteristic hash value of a single corn cob, and only the batch Merkle root is written into the consortium blockchain main chain. The entire process data of a single corn cob is stored in IPFS distributed storage, and the hash verification value and IPFS address are retained on the chain. The traceability QR code enables the query and verification of the entire process production data, quality inspection data and blockchain evidence information of a single corn cob.

8. The method of claim 1, wherein, In step 8, the model incremental iteration cycle is once a week. The quality prediction, F-value mapping, and risk warning models are updated based on the new production data. The defective product risk warning threshold is set to 80%. When the warning is triggered, the optimized process parameter value is output based on the SHAP feature importance analysis and sent to the production line PLC control system through the OPCUA protocol to complete the adaptive adjustment of the production process.

9. A data acquisition system for a vacuum fresh corn production line, characterized in that, It includes a unified spatiotemporal tracking module for individuals, a raw material quality and working condition acquisition module, a thermal history calculation and quality extraction module, a packaging process and sealing acquisition module, an F-value calculation and quality control module, a finished product quality and environmental acquisition module, a data alignment and blockchain traceability module, and a closed-loop feedback model optimization module, all connected in sequence. The individual tracking spatiotemporal unified module is used to assign a unique snowflake ID to a single corn stalk, realize cross-process ID binding of corn through industrial vision and encoder, and provide a unified spatiotemporal benchmark for data collection throughout the entire process. The raw material quality condition acquisition module is used to predict the internal quality parameters of a single corn cob based on the corn appearance image through a pre-trained neural network, synchronously collect sorting equipment condition data and bind it with the corresponding corn ID, and complete data preprocessing. The thermal history calculation and quality extraction module is used to collect thermal parameters of blanching and cooling processes, calculate the thermal processing history and core temperature of a single corn cob based on the heat conduction model adapted to corn, predict the quality characteristics after thermal processing, and bind and store them. The packaging process sealing acquisition module is used to collect vacuum packaging process time sequence data, accurately bind it with the ID of a single corn cob, and simultaneously collect packaging sealing performance data and mark defective products. The F-value calculation quality control module is used to collect sterilizer operating data, generate the optimal sterilization F0 target value for a single ear of corn based on the quality characteristics of the entire corn process, calculate the corn center temperature and cumulative sterilization F value in real time, and simultaneously compare the F0 target value and predict the quality of the finished product after sterilization. The finished product quality environment acquisition module is used to collect environmental parameters of the entire production line and associate them with the corresponding corn ID, collect finished product quality inspection data to verify the previous prediction results, and extract the correlation features between the environment and finished product quality. The data alignment blockchain traceability module is used to complete the spatiotemporal alignment of multi-source data for a single corn cob throughout the entire process. It writes the corn data feature hash value into the consortium blockchain for storage through a batch Merkle tree, supporting full-chain traceability query for a single corn cob. The closed-loop feedback model optimization module is used to complete the periodic incremental iteration of each prediction model based on the new production data, predict the quality of finished products and the risk of defective products in real time, output process optimization instructions and send them to the production line PLC system to realize the adaptive closed-loop adjustment of the production process.

Images