Agricultural product storage intelligent management system based on big data

Abstract

The invention relates to the technical field of agricultural product storage management, in particular to an agricultural product storage intelligent management system based on big data, which comprises a multi-source data acquisition layer, an edge processing layer, a cloud processing layer and an execution control layer, compared with the prior art which only depends on manual recording of the picking date or simple visual sampling inspection for judging the freshness, the method has the advantages that the freshness key indexes such as the skin color degree and the texture structure of the agricultural product are synchronously and quantitatively extracted through the deep learning model in the warehousing link of the agricultural product; performing dual cross validation with the registration time to eliminate abnormal samples; the verified freshness reference data and parameters such as temperature, humidity, gas concentration and the like monitored in real time in a corresponding storage microenvironment are fused and input, and a self-adaptive shelf life prediction model based on a dynamic environmental stress factor is constructed; therefore, continuous and accurate quantification and early warning of decay processes of different types of agricultural products under various storage conditions are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural product storage management, and in particular to an intelligent agricultural product storage management system based on big data. Background Art

[0002] Intelligent warehousing management for agricultural products is a core component of the modern agricultural supply chain. Its core goal is to achieve efficient utilization of storage resources while ensuring the quality of agricultural products. With the penetration of IoT and AI technologies, traditional warehousing systems have gradually incorporated sensor monitoring and automated control modules, attempting to reduce waste rates by regulating environmental parameters. However, current technology systems still face significant bottlenecks in complex and diverse real-world application scenarios.

[0003] Existing technologies mainly face three contradictions: at the data acquisition layer, reliance on a single type of sensor or manual sampling leads to the lack of key information (such as the real-time physiological status of the product and external meteorological influences), and high-cost sensors cannot achieve full warehouse coverage; at the decision-making execution layer, the centralized cloud computing architecture leads to serious delays in responding to environmental anomalies and cannot meet the minute-level regulation and control needs of fresh agricultural products; at the optimization model layer, inventory allocation strategies are often out of touch with product quality and logistics needs, static empirical models are difficult to adapt to dynamic market fluctuations, and digital twin technology mostly uses pure mathematical interpolation, ignoring the real conduction laws of the physical field, resulting in significant deviations between the regulation strategy and the actual effect.

[0004] The present invention constructs a full-stack intelligent closed loop of "perception-decision-control": economical full-space coverage is achieved through heterogeneous networking of high-precision and low-cost sensors, and the freshness of agricultural products is double-checked by combining image recognition and registration systems; lightweight models are deployed on the edge side to achieve millisecond-level abnormal response, and a digital twin environment simulation engine that integrates physical laws (such as heat conduction equations) is simultaneously constructed; finally, a multi-objective optimization algorithm is used to dynamically integrate predicted orders, logistics timeliness and real-time quality parameters to generate a precise storage strategy that is strongly coupled with the physical field, fundamentally solving the three major problems of data one-sidedness, response hysteresis and strategy inaccuracy. Summary of the Invention

[0005] In order to overcome the problems raised in the above background technology, the present invention proposes an intelligent management system for agricultural product storage based on big data.

[0006] The technical solution of the present invention is: an intelligent management system for agricultural product storage based on big data, including a multi-source data acquisition layer, an edge processing layer, a cloud processing layer and an execution control layer, wherein: Multi-source data collection layer, used to collect storage environment parameters, agricultural product parameters, weather forecast data, historical logistics data and market demand data; The edge processing layer is deployed at the warehouse site and is used to analyze storage environment parameters and agricultural product parameters in real time through the deep learning model built into the edge computing node, and generate control instructions; The cloud processing layer integrates and processes multi-source data based on a distributed cloud platform, combines historical data with prediction models to generate warehouse control instructions, and sends them to the edge processing layer; The execution control layer is used to drive the temperature control device, intelligent sorting system and early warning device according to the control instructions of the edge processing layer, and to adjust the storage environment parameters and the storage status of agricultural products in real time. Specifically, it includes the temperature control device, intelligent sorting system and early warning device.

[0007] Preferably, the multi-source data acquisition layer specifically includes: A11: Environmental sensor network, consisting of sensor nodes and using the LoRa communication protocol. Each sensor node consists of a group of high-precision sensors and multiple groups of low-cost sensors. A12: Image acquisition equipment, used to collect image data of agricultural products when they are stored in the warehouse, and to collect images of agricultural products and the environment inside the warehouse in real time; A13: External data interface, used to connect to external systems to obtain weather forecast data, historical logistics data, and market demand data.

[0008] Preferably, the edge processing layer specifically includes: A21: Image analysis module, with a built-in lightweight deep learning model, is used to analyze collected agricultural product images, extract color, size, appearance, and texture features, and identify the quality, maturity, and freshness of agricultural products; A22: Environmental preprocessing module, used to filter, remove outliers, and synchronize time for the raw data collected by the sensor network to generate a local environmental data summary; A23: Control module, used to generate primary control instructions including local ventilation and temperature and humidity fine-tuning based on image analysis results and environmental preprocessing data.

[0009] Preferably, the image analysis module identifies the quality, maturity and freshness of agricultural products by: S11: Image acquisition and preprocessing. When agricultural products are put into storage, the original RGB agricultural product images are acquired through image acquisition equipment and preprocessed. The preprocessing methods used include size adjustment, image denoising, image enhancement, color space conversion, and standardization. S12: Object detection, using the built-in YOLOv5 model to perform object detection and classification to identify the specific categories of agricultural products; S13: Feature extraction based on the detection results: for each detected agricultural product, region cropping is performed and feature extraction is performed, wherein the extracted features include main hue value, average saturation, average brightness, size feature, appearance contour feature and texture feature; S14: Feature fusion and result judgment: combine the extracted features into a comprehensive feature vector, and use the trained classification model to judge the quality, maturity and freshness of agricultural products.

[0010] Preferably, the image environment preprocessing module filters, removes outliers, and synchronizes time on the raw data collected by the sensor network to generate a summary of local environmental data. It verifies the data of high-precision sensors and low-cost sensors in each group of sensor nodes, and when differences occur, it removes outliers and performs weighted fusion to obtain the final data of each group of sensor nodes.

[0011] Preferably, the cloud processing layer specifically includes: A31: Distributed data storage module, which uses a distributed consensus mechanism and includes an edge storage module and a cloud storage module. The edge storage module is distributed within the edge processing layer and is used to store data in the edge storage module and the cloud storage module. A32: Demand forecasting module, based on a fusion model of LSTM and XGBoost, combines historical orders, weather data, and market demand data to predict agricultural product inflows and outflows and regional sales for the next 30 days; A33: Inventory optimization module, which uses a genetic algorithm to combine predicted orders, logistics timeliness, and agricultural product quality grades to dynamically divide storage areas and match optimal temperature and humidity conditions. A34: Digital twin building module, used to fuse sensor data with physical models to generate a real-time virtual image of the warehouse environment. It uses Kriging interpolation to interpolate environmental parameters in sparse sensor areas and achieve millimeter-level spatial parameter simulation.

[0012] Preferably, the edge storage module is deployed at the warehouse site and includes multiple independent edge storage sub-modules. Each sub-module stores part of the warehouse data, and the storage data of all sub-modules has redundant coverage. Specifically, at any time, the total amount of data stored by all sub-modules is twice the total amount of warehouse data, that is, the same data item is repeatedly stored by at least two different sub-modules. The cloud storage module is deployed in a remote data center to store the full amount of warehouse data, that is, a complete set of all redundant data in the edge storage module. The edge storage module and the cloud storage module achieve data consistency through the following mechanism: A41: Timed synchronization. According to the set synchronization period, the edge storage module packages the redundant data of each submodule and synchronizes it to the cloud storage module. At the same time, the cloud storage module reversely synchronizes the latest global data to the edge storage module. A42: Confidence verification. During the synchronization process, a double verification strategy is adopted to ensure data accuracy: a hash value is generated for each set of synchronized data, and the cloud and edge respectively calculate and compare the hash values. If they are inconsistent, they are marked as suspicious data; for entries marked as suspicious data, the edge storage module retrieves all sub-module copies that store the data, and determines the correct version based on the principle of consistency of data in the majority of sub-modules; the cloud storage module synchronizes and uses the same rules to correct the data.

[0013] Preferably, the digital twin construction module interpolates the environmental parameters of the sparse sensor area through the Kriging interpolation method to achieve millimeter-level spatial parameter simulation, specifically including: S21: Collect historical temperature data of each temperature regulating device at different powers; S22: Based on physical simulation, establish a mapping relationship between the output power of each device and the change of ambient temperature; S23: In areas with sparse sensor networks, the spatial parameters of undeployed sensors are supplemented by the superposition effect of multiple groups of impact models combined with interpolation algorithms.

[0014] Preferably, the inventory optimization module is also used to predict the shelf life of agricultural products, including: S31: When agricultural products are put into storage, the picking time of the agricultural products is recorded; S32: extracting freshness features of the agricultural products based on the image recognition results, and comparing the freshness features with the registered picking time; S33: After the comparison is successful, the environmental data of the corresponding storage area is integrated and a Weibull distribution model is established to predict the remaining shelf life.

[0015] As a preferred option, the demand forecasting module predicts the agricultural product inflow and outflow volume and regional sales volume for the next 30 days through the following steps: S41: Data cleaning and feature engineering: performing time series aggregation processing on historical order data and removing outliers. At the same time, performing missing value filling and generating derived features on the meteorological monitoring data. The derived features include the number of consecutive high-temperature days and accumulated precipitation. External platform indicators are integrated with the market demand data to generate a quantitative value of regional demand heat. S42: Dual-model collaborative forecasting uses an LSTM neural network to process time series dependencies. It inputs historical inbound and outbound sequences and meteorological time series data, extracts dynamic features through three LSTM hidden layers, and outputs primary inbound and outbound forecasts for the next 30 days. It also uses the XGBoost algorithm to process structured features. It inputs a feature vector containing category attributes, regional identifiers, and promotion intensity, and outputs secondary inbound and outbound forecasts. It also converts meteorological factors into agricultural product corruption coefficients, which are then combined in the LSTM hidden layers for joint training. S43: Dynamic weighted fusion and decision output, dynamically adjusting the fusion weight coefficient according to the environmental fluctuation state , and according to the formula: , generate the final prediction value, where is the final predicted value, is the predicted value of the LSTM neural network, is the predicted value of the XGBoost algorithm; S44: Real-time error correction: daily comparison of the predicted value with the actual inventory in and out. If the regional sales error rate exceeds 15%, the following actions will be performed: analyzing feature contribution through SHAP value; expanding the training sample set; and fine-tuning the weight parameters of the last layer of the LSTM network. S45: Output application, generate the inventory in and out time series curve for the next 7 days; generate the regional sales geographic heat map for the next 30 days; trigger inventory optimization and logistics scheduling instructions based on the forecast results.

[0016] As a preferred method, when converting meteorological factors into agricultural product corruption coefficients, the method specifically includes: S51: Establishing temperature-respiration entropy mapping function , where T is the daily average temperature, and is the category experience parameter, e is the natural constant; S52: Defining the Corruption Coefficient ,in, is the corruption coefficient, d is the prediction date, N is the number of days in the historical window, is the precipitation attenuation factor.

[0017] As a preference, dynamically adjust the fusion weight coefficient according to the environmental fluctuation state When , the dynamic weight adjustment mechanism adopted is: A51: When the weather volatility exceeds the preset threshold, the weight coefficient is set. , where the quantitative formula for meteorological volatility is: ,in, is the quantitative value of meteorological volatility, is the standard deviation function, is the daily maximum temperature difference series, is the cumulative precipitation series, is the regional adjustment factor, and the fluctuation threshold is ,in, is the mean function; A52: When a sudden promotion event is detected, set the weight coefficient ; A53: Default weight coefficient .

[0018] Preferably, the inventory optimization module comprises the following steps when operating: S51: Data input and preprocessing, obtaining demand forecast data, agricultural product quality grading data, logistics time constraints and environmental strategies; S52: Multi-objective optimization modeling, establish a multi-objective optimization model, the objective function is: ,in, is the weighted distance from the shelf to the platform and emergency area, is the deviation between shelf temperature and humidity and the optimal value of the category, is the penalty item for exceeding the regional floor area ratio limit, 、 and is the weight coefficient; S53: Select an algorithm based on the optimization scale: NSGA-II genetic algorithm is used for global layout updates; DQN reinforcement learning model is used for local storage location adjustments; S54: Instruction execution and closed-loop monitoring, outputting environmental control parameters and product positioning strategies to the execution control layer, and triggering re-optimization when the deviation exceeds the threshold in real time. The re-optimization mechanism is as follows: when the environmental sensor detects that the difference between the actual temperature and the set temperature is greater than 1°C for more than 5 minutes, local re-optimization is automatically triggered; the re-optimization process retains 80% of the valid solutions of the original strategy as the initial population to accelerate convergence.

[0019] As a preferred method, when selecting an algorithm based on the optimization scale, the configuration of the genetic algorithm is as follows: the crossover operation uses a two-point crossover operator with a crossover probability of 0.85; the mutation operation uses a Gaussian mutation operator with a mutation probability of 0.02, and the adaptability evaluation function is ; The DQN reinforcement learning model includes: the state space is the shelf occupancy rate, environmental parameters and real-time order queue; the action space is the category location migration vector; the reward function is , =0.3 is the utilization gain bonus coefficient, is the change in transportation cost, is the change in environmental cost, is the space utilization gain value.

[0020] Beneficial effects of the present invention: 1. Compared with existing technologies that rely solely on manual recording of picking dates or simple visual spot checks to determine freshness, which are subject to strong subjectivity, ignore the impact of the actual storage environment, and have large prediction errors, this solution simultaneously adopts an automatic picking time registration system and multispectral image feature analysis technology during the agricultural product storage process. Through deep learning models, key freshness indicators such as the skin color and texture structure of agricultural products are quantitatively extracted, and double cross-validation is performed with the registration time to eliminate abnormal samples. The verified freshness benchmark data is then integrated with real-time monitoring of temperature, humidity, gas concentration and other parameters in the corresponding storage microenvironment to construct an adaptive shelf life prediction model based on dynamic environmental stress factors. This achieves continuous and accurate quantification and early warning of the decay process of different types of agricultural products under various storage conditions. 2. Compared with existing technologies that use traditional storage architectures that lack redundancy or one-way synchronization mechanisms, resulting in poor data consistency, weak disaster recovery capabilities, and high network dependence, leading to a high risk of real-time control interruption, this solution adopts a redundant coverage design of distributed data storage modules. Through timed synchronization and confidence verification, it realizes bidirectional data flow and majority voting mechanisms. It has the advantages of high reliability and consistency, forming a robust system of edge-cloud collaborative storage, effectively preventing transmission errors and hardware failures, ensuring strong data consistency and high availability, and supporting continuous service capabilities under extreme conditions. 3. Compared to existing technologies that use uniform deployment of single high-precision sensors or pure mathematical interpolation to fill blind spots, which have the disadvantages of high hardware costs and low data physical credibility, this solution uses a unique hierarchical deployment architecture for heterogeneous sensors: a small number of high-precision sensors are set as benchmark anchors at key monitoring nodes, combined with low-cost sensors to form a dense observation network. The edge layer fusion verification mechanism is used to automatically verify data differences between devices and correct drift errors. At the same time, a thermodynamic physical model is integrated into the cloud-based digital twin construction module, and the environmental parameter transfer function is established based on the historical power-temperature curve of the temperature control equipment. The Kriging interpolation algorithm under physical constraints is used to reconstruct sparse area data. This reduces hardware costs by 60% while ensuring that the environmental simulation values ​​in areas without sensor deployments are more than 95% consistent with the real physical field dynamics. 4. Compared to existing technologies that simplify inventory allocation to static partitioning or mechanical order sorting, ignoring the shortcomings of coordinated optimization of product characteristics, logistics efficiency, and storage costs, this solution deeply integrates three key elements through a full-link data integration mechanism: Based on the multi-dimensional order spatiotemporal distribution heat map output by the demand forecasting module, combined with the real-time agricultural product quality grading map and logistics time constraint matrix, a multi-objective genetic algorithm is used to dynamically calculate the value weight of storage locations. An environmental strategy generator automatically matches the optimal temperature and humidity combination for each category, and the physical coordinates of the shelves are dynamically divided in milliseconds based on the category priority coefficient. This achieves intelligent three-dimensional storage planning of "perishable goods near the export, storable goods on high floors, and promotional goods on fast access", achieving an optimal balance between overall warehouse turnover efficiency and energy consumption control. 5. Compared to existing technologies that use static prediction models and suffer from poor environmental adaptability, this solution innovatively builds an LSTM-XGBoost dual-model collaborative prediction system, incorporating a dynamic weight adjustment mechanism. Through meteorological-biological coupled modeling and real-time error correction, the prediction system can intelligently perceive environmental fluctuations and market changes, maintaining stable output even in extreme weather such as typhoons, significantly improving the foresight and accuracy of long-term warehousing strategies. 6. Compared with the existing technology that uses a single type of sensor or manually enters data, which has the disadvantages of one-sided data, poor real-time performance, and large coverage blind spots, this solution adopts a "high-precision + low-cost" hybrid sensor network, multi-dimensional image acquisition and external data interface collaborative acquisition to build a multi-source heterogeneous data system covering micro-environmental status, agricultural product characteristics and macro-market meteorological elements, realizing deep fusion perception of physical and information space inside and outside the warehouse, and providing a solid and reliable data foundation for intelligent decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Shown is a schematic diagram of the structure of the big data-based agricultural product storage intelligent management system of the present invention; Figure 2 Shown is a second three-dimensional structural diagram of the big data-based agricultural product storage intelligent management system of the present invention. DETAILED DESCRIPTION

[0022] The present invention will be further described below with reference to the accompanying drawings and examples.

[0023] See also Figure 1-2 The present invention provides an embodiment: an intelligent management system for agricultural product storage based on big data, comprising a multi-source data acquisition layer, an edge processing layer, a cloud processing layer, and an execution control layer, wherein: Multi-source data collection layer, used to collect storage environment parameters, agricultural product parameters, weather forecast data, historical logistics data and market demand data; The edge processing layer is deployed at the warehouse site and is used to analyze storage environment parameters and agricultural product parameters in real time through the deep learning model built into the edge computing node, and generate control instructions; The cloud processing layer integrates and processes multi-source data based on a distributed cloud platform, combines historical data with prediction models to generate warehouse control instructions, and sends them to the edge processing layer; The execution control layer is used to drive the temperature control device, intelligent sorting system and early warning device according to the control instructions of the edge processing layer, and to adjust the storage environment parameters and the storage status of agricultural products in real time. Specifically, it includes the temperature control device, intelligent sorting system and early warning device.

[0024] As mentioned above, refined management and control of the entire chain is achieved through the "edge-cloud" collaborative intelligent architecture: the edge computing layer uses deep learning models to achieve millisecond-level response to the storage environment and agricultural product status (such as temperature mutation warning), ensuring real-time execution of control instructions; the cloud integrates multi-source heterogeneous data such as meteorology, market, and historical logistics, and uses prediction models to dynamically optimize long-term storage strategies (such as adjusting temperature zone allocation in advance), forming a "real-time control + long-term prediction" dual guarantee, effectively reducing the loss rate of agricultural products (about 15% to 30%), reducing energy waste (intelligent temperature control saves about 20% energy), and improving market response speed (sorting efficiency increased by 40%+). Ultimately, through data-driven, the dual improvement of storage cost optimization and agricultural product quality is achieved.

[0025] Preferably, the multi-source data acquisition layer specifically includes: A11: Environmental sensor network, consisting of sensor nodes and using the LoRa communication protocol. Each sensor node consists of a group of high-precision sensors and multiple groups of low-cost sensors. A12: Image acquisition equipment, used to collect image data of agricultural products when they are stored in the warehouse, and to collect images of agricultural products and the environment inside the warehouse in real time; A13: External data interface, used to connect to external systems to obtain weather forecast data, historical logistics data, and market demand data.

[0026] As mentioned above, the multi-source data acquisition layer of this technical solution significantly improves the breadth and accuracy of data coverage through heterogeneous sensor fusion and internal and external data collaboration: the environmental sensor network adopts a "high-precision + low-cost" hybrid networking mode, which reduces deployment costs (saving hardware costs by approximately 30%) while ensuring the measurement accuracy of key parameters such as temperature and humidity, and achieves low-power wide-area coverage (communication distance of 5km+) through the LoRa protocol; image acquisition equipment combines visual AI technology to achieve non-destructive testing of the appearance quality of agricultural products (such as mold and damage) (accuracy >95%), and simultaneously monitors abnormal storage environments (such as water leaks); external data interfaces open up meteorological, logistics and market data links, providing cross-domain decision-making support for cloud-based prediction models (such as combining weather forecasts to predict cold chain demand), and ultimately forming a three-dimensional data acquisition system of "micro-environmental perception + macro trend insight", providing a highly reliable data foundation for intelligent regulation.

[0027] Preferably, the edge processing layer specifically includes: A21: Image analysis module, with a built-in lightweight deep learning model, is used to analyze collected agricultural product images, extract color, size, appearance, and texture features, and identify the quality, maturity, and freshness of agricultural products; A22: Environmental preprocessing module, used to filter, remove outliers, and synchronize time for the raw data collected by the sensor network to generate a local environmental data summary; A23: Control module, used to generate primary control instructions including local ventilation and temperature and humidity fine-tuning based on image analysis results and environmental preprocessing data.

[0028] As described above, lightweight AI combined with real-time data processing enables intelligent autonomy and rapid response at the warehouse site. The image analysis module utilizes lightweight deep learning models (such as MobileNet or YOLOTiny) to achieve millisecond-level detection of agricultural product quality (e.g., mold and damage) and maturity (latency <50ms) with over 90% accuracy, even in low-computing environments. The environmental preprocessing module uses adaptive filtering and anomaly detection algorithms (such as DBSCAN) to improve sensor data reliability and reduce noise interference (data error is reduced by 40%). The control module dynamically generates local precision control commands (e.g., adjusting temperature and humidity for specific shelves) based on real-time analysis results. Compared to traditional centralized control, this reduces cloud reliance (response speed increases by 5 times) and reduces communication bandwidth pressure (data upload volume is reduced by 60%). This design forms a closed "perception-analysis-execution" loop at the edge, significantly improving the real-time and robustness of warehouse management while reducing the computing burden on the cloud.

[0029] The control module generates primary control instructions, including local ventilation and temperature and humidity fine-tuning, based on image analysis results and environmental pre-processing data. Specifically, the following steps are involved: S61: Receive data from the image analysis module in real time, including agricultural product quality grade, maturity index, abnormality flag, and freshness decay rate; and simultaneously obtain environmental parameters output by the environmental preprocessing module, including temperature distribution map, humidity distribution map, gas concentration index, and environmental status of the adjacent area; S62: Performing weighted fusion of the image data and environmental parameters based on a preset decision matrix to generate a control demand intensity value, wherein the decision matrix includes: a temperature deviation factor and its weight coefficient, a humidity deviation factor and its weight coefficient, a maturity index factor and its weight coefficient, an abnormality mark factor and its weight coefficient, and a regional coordination factor and its weight coefficient; S63: Generate an initial control instruction based on the intensity value of the control demand, including: when an abnormal sign is detected, generate an emergency ventilation instruction, whose intensity is positively correlated with the severity of the abnormality; when the absolute value of the temperature deviation is greater than 0.5°C or the absolute value of the humidity deviation is greater than 3%, calculate the temperature and humidity fine-tuning value according to the formula: temperature adjustment amount , humidity adjustment , where P is the priority coefficient, which is determined by the quality grade of agricultural products. and are the basic quantities of temperature and humidity, and is the adjustment coefficient; when the maturity index exceeds the threshold, ripening and slow ripening strategy instructions are generated; S64: Optimizing the power consumption of the initial control instruction, and obtaining optimized instruction parameters while satisfying the control accuracy constraint by solving an energy consumption minimization model; S65: Encapsulate the optimization instruction parameters into a standard control protocol data packet, and send the instruction to the execution device. The data packet includes the control area identifier, action type, adjustment value, duration and energy consumption budget.

[0030] Among them, the dynamic adjustment method of the decision matrix weight coefficient is as follows: the temperature and humidity weights are periodically updated according to the sensor reliability index; the maturity factor weight is automatically increased by 20% during the agricultural product harvest period; the anomaly factor weight is increased by 50% when an anomaly is detected in the adjacent area.

[0031] When generating the initial control instruction according to the regulation demand intensity value, a nonlinear correction of the instruction calculation is included: When the absolute value of the temperature deviation exceeds 2°C, an attenuation coefficient of 0.7 is applied to the temperature adjustment amount; When the absolute value of the humidity deviation exceeds 10%, an attenuation coefficient of 0.6 is applied to the humidity adjustment amount; The ventilation strategy decision adopts a three-level logic tree: Level 1: Natural convection (CO2 concentration < 1500ppm and no abnormalities); Level 2: Moderate ventilation (1500ppm≤CO2 concentration<3000ppm); Level 3: Strong ventilation (CO2 concentration ≥ 3000ppm or abnormal corruption occurs).

[0032] As mentioned above, multi-factor dynamic decision-making + intelligent optimization are used to achieve precise and energy-saving warehouse control: based on the preset decision matrix (S62), the quality of agricultural products and environmental parameters are weighted and integrated (for example, the maturity weight is automatically increased by 20% during the harvest period), and nonlinear correction is combined (for example, an attenuation coefficient is applied when there are large deviations in temperature and humidity) to avoid over-control; a three-level ventilation logic tree (S63) is used to achieve a gradient response to CO2 concentration and abnormal conditions, and the emergency ventilation response speed is increased by 60%; through the energy consumption minimization model (S64), while ensuring the control accuracy (temperature and humidity fluctuations ±0.3℃ / ±2%), energy consumption is reduced by 15% to 25%; the dynamic weight mechanism (for example, the weight of the abnormal factor is increased by 50% when there is an adjacent abnormality) enhances the robustness of the system, so that the local control instructions (S65) are both real-time (delay <100ms) and globally coordinated, ultimately achieving the dual goals of "abnormality handling in seconds + optimal maintenance of normal conditions" in the warehouse environment.

[0033] Preferably, the image analysis module identifies the quality, maturity and freshness of agricultural products by: S11: Image acquisition and preprocessing. When agricultural products are put into storage, the original RGB agricultural product images are acquired through image acquisition equipment and preprocessed. The preprocessing methods used include size adjustment, image denoising, image enhancement, color space conversion, and standardization. S12: Object detection, using the built-in YOLOv5 model to perform object detection and classification to identify the specific categories of agricultural products; S13: Feature extraction based on the detection results: for each detected agricultural product, region cropping is performed and feature extraction is performed, wherein the extracted features include main hue value, average saturation, average brightness, size feature, appearance contour feature and texture feature; S14: Feature fusion and result judgment, combining the extracted features into a comprehensive feature vector, and using the trained classification model to judge the quality, maturity and freshness of agricultural products. Among them, multiple trained base models are used to judge the quality, maturity and freshness of agricultural products, and the prediction results of multiple base models are integrated through the voting method to obtain the final result. The base models used include VM, RF and DNN.

[0034] As described above, the image analysis module achieves high-precision nondestructive testing of agricultural product quality through multi-stage intelligent analysis. YOLOv5-based object detection rapidly identifies agricultural product categories (with >98% accuracy). Multi-dimensional feature extraction quantifies key metrics such as hue and texture (for example, a hue deviation >15° in moldy areas automatically triggers an alert). An ensemble learning strategy fuses predictions from support vector machines (SVMs), radio frequency (RF), and dense neural networks (using a voting method to improve generalization), resulting in a maturity determination error of <5% and a freshness grading accuracy of 93%. In the preprocessing stage, adaptive image enhancement (such as the low-light CLAHE algorithm) ensures analysis stability under complex lighting conditions. The entire process completes on the edge device in just 200ms, a 50-fold increase in efficiency compared to traditional manual sampling. This also reduces contact damage to agricultural products, providing real-time, reliable visual data support for intelligent warehouse management.

[0035] Among them, when using the built-in YOLOv5 model to perform target detection and classification and identify the specific categories of agricultural products, it specifically includes: S121: Collect multi-source agricultural product image samples to construct a training dataset covering multiple varieties of target agricultural products, all maturity stages, typical surface defects, differentiated ambient lighting, and stacked occlusion scenarios; S122: Analyzing the width and height distribution of the real bounding boxes in the training data set based on the K-Means clustering algorithm to generate an adaptive anchor box set that matches the morphological characteristics of the target agricultural product; S123: Construct a lightweight deep neural network based on a single-stage detection framework. The backbone network uses a cross-stage local fusion structure to optimize gradient flow, the neck network uses a bidirectional path aggregation structure (BiFPN) to achieve multi-scale feature fusion, and the output layer of the detection head is configured with feature maps of at least three different scales to adapt to the detection requirements of agricultural products of different sizes. S124: Performing environmental simulation enhancement operations on the training data set, including: random perturbation of brightness and saturation within a controllable range to simulate natural light fluctuations; Gaussian noise injection and motion blur simulation to reproduce imaging noise in a high-speed sorting environment; Mosaic-based image stitching and occlusion generation to improve the detection robustness of small targets and stacked targets; S125: Introducing the α-IoU loss function during network training to optimize bounding box regression accuracy, and combining it with the difficult sample mining strategy to dynamically adjust the confidence loss weight to suppress complex background interference; S126: Using knowledge distillation technology to guide the learning of the lightweight student model with a pre-trained high-precision teacher model, thereby improving the representation ability of the model with a limited number of parameters, wherein a generalized feature alignment loss is used; S127: Implement quantization-aware training to simulate the FP32 to INT8 numerical conversion process in the network forward propagation to generate hardware-friendly low-bit deployable models. Specifically, it includes: inserting pseudo-quantization nodes in the training graph to simulate INT8 inference behavior during deployment; using a pass-through estimator to bypass the non-differentiability of quantization operations; and exporting to a TensorRT or OpenVINO-compatible INT8 engine during deployment.

[0036] In step S123, the bidirectional path aggregation structure (BiFPN) performs top-down and bottom-up bidirectional information interaction on the high-level, mid-level, and low-level feature maps output by the backbone network through a weighted feature fusion mechanism. The formula is expressed as: ; in, is the output feature map, is the top-down path feature, is the upsampling function, is the bottom-up path feature, is the downsampling function, and are learnable weights, Represents an element-wise addition operation.

[0037] As mentioned above, the accuracy and efficiency of agricultural product target detection are significantly improved through full-process optimization design: based on multi-source sample construction and environmental simulation enhancement, a training set covering complex scenarios is formed (for example, the detection rate of stacked occlusion is increased by 40%), and combined with adaptive anchor boxes, the bounding box matching degree is improved by 25%; the lightweight network architecture uses BiFPN to achieve multi-scale feature fusion, and the three-scale output of the detection head is adapted to targets of different sizes (the missed detection rate of small targets is reduced to 3%); the positioning accuracy of the model in complex backgrounds is improved through α-IoU loss and knowledge distillation (mAP@0.5 reaches 96.5%), and quantitative training reduces the model size by 4 times and the inference speed reaches 120FPS (2.3 times faster than the FP32 model), ultimately achieving high-precision real-time detection on edge devices (latency <8ms), meeting the dual stringent requirements of speed and accuracy in warehousing and sorting scenarios.

[0038] Preferably, the image environment preprocessing module filters, removes outliers, and synchronizes time on the raw data collected by the sensor network to generate a summary of local environmental data. It verifies the data of high-precision sensors and low-cost sensors in each group of sensor nodes, and when differences occur, it removes outliers and performs weighted fusion to obtain the final data of each group of sensor nodes.

[0039] As mentioned above, the environmental preprocessing module significantly improves the reliability and data accuracy of environmental monitoring through a heterogeneous sensor data fusion and verification mechanism: by comparing the data differences between high-precision sensors and low-cost sensors in real time (for example, verification is triggered when the temperature deviation is >0.5°C), a dynamic weighted fusion algorithm (with high-precision sensors accounting for more than 70% of the weight) is used to automatically eliminate outliers (abnormal data identification accuracy >98%), effectively overcoming the problem of single sensor failure or drift (system fault tolerance is improved by 60%); combined with time synchronization technology (error <10ms), the spatiotemporal consistency of multi-node data is ensured, and the comprehensive error of the final generated local environmental data summary (such as the temperature and humidity distribution map) is controlled within ±0.2°C / ±1.5%, which is three times more reliable than the single-sensor solution, providing a high-precision and highly robust environmental status benchmark for subsequent intelligent control.

[0040] Preferably, the cloud processing layer specifically includes: A31: Distributed data storage module, which uses a distributed consensus mechanism and includes an edge storage module and a cloud storage module. The edge storage module is distributed within the edge processing layer and is used to store data in the edge storage module and the cloud storage module. A32: Demand forecasting module, based on a fusion model of LSTM and XGBoost, combines historical orders, weather data, and market demand data to predict agricultural product inflows and outflows and regional sales for the next 30 days; A33: Inventory optimization module, which uses a genetic algorithm to combine predicted orders, logistics timeliness, and agricultural product quality grades to dynamically divide storage areas and match optimal temperature and humidity conditions. A34: Digital twin building module, used to fuse sensor data with physical models to generate a real-time virtual image of the warehouse environment. It uses Kriging interpolation to interpolate environmental parameters in sparse sensor areas and achieve millimeter-level spatial parameter simulation.

[0041] As mentioned above, the cloud processing layer achieves global optimization and precise simulation of warehouse management through multimodal intelligent fusion: the distributed data storage module adopts an edge-cloud collaborative storage mechanism (data synchronization delay <50ms), which ensures data security while reducing network load (bandwidth usage is reduced by 40%); the demand forecasting module combines LSTM time series modeling and XGBoost feature analysis to achieve a 30-day sales forecast error rate of <8% (35% improvement in accuracy compared to traditional methods); the inventory optimization module uses an adaptive genetic algorithm to dynamically plan storage locations (such as prioritizing perishables in low-temperature areas), thereby increasing storage space utilization by 25% and extending the shelf life of agricultural products by 15%; the digital twin module uses Kriging interpolation technology to construct millimeter-level environmental images (sparse area data filling accuracy >92%), supports virtual control rehearsals (trial and error costs are reduced by 60%), and forms a closed-loop management paradigm of "data-driven decision-making + virtual verification and optimization", which overall reduces warehouse operating costs by 18% to 30%.

[0042] Preferably, the edge storage module is deployed at the warehouse site and includes multiple independent edge storage sub-modules. Each sub-module stores part of the warehouse data, and the storage data of all sub-modules has redundant coverage. Specifically, at any time, the total amount of data stored by all sub-modules is twice the total amount of warehouse data, that is, the same data item is repeatedly stored by at least two different sub-modules. The cloud storage module is deployed in a remote data center to store the full amount of warehouse data, that is, a complete set of all redundant data in the edge storage module. The edge storage module and the cloud storage module achieve data consistency through the following mechanism: A41: Timed synchronization. According to the set synchronization period, the edge storage module packages the redundant data of each submodule and synchronizes it to the cloud storage module. At the same time, the cloud storage module reversely synchronizes the latest global data to the edge storage module. A42: Confidence verification. During the synchronization process, a double verification strategy is adopted to ensure data accuracy: a hash value is generated for each set of synchronized data, and the cloud and edge respectively calculate and compare the hash values. If they are inconsistent, they are marked as suspicious data; for entries marked as suspicious data, the edge storage module retrieves all sub-module copies that store the data, and determines the correct version based on the principle of consistency of data in the majority of sub-modules; the cloud storage module synchronizes and uses the same rules to correct the data.

[0043] As described above, the edge-cloud collaborative storage mechanism achieves high data reliability and consistency through redundant coverage and dual verification. The edge storage module utilizes a distributed redundant architecture (double-redundant data storage), ensuring rapid recovery from other submodules even if a single node fails (data availability reaches 99.99%). A scheduled synchronization mechanism, combined with bidirectional data flow (synchronization period configurable down to minutes), keeps data latency between the edge and the cloud under 5 seconds. Confidence verification is ensured through a combination of hash checksums and majority voting (with 100% accuracy in correcting suspicious data), effectively protecting against transmission errors and hardware failures (data error rate <0.001%). This design maintains full edge service capabilities for 72 hours (relying on locally redundant data) even in extreme situations such as network outages. After synchronization is restored, data differences are automatically aligned, reducing network dependency by 60% compared to traditional single-cloud storage solutions. This simultaneously meets the dual requirements of strong data consistency and high disaster recovery for real-time warehouse management.

[0044] Preferably, the digital twin construction module interpolates the environmental parameters of the sparse sensor area through the Kriging interpolation method to achieve millimeter-level spatial parameter simulation, specifically including: S21: Collect historical temperature data of each temperature regulating device at different powers; S22: Based on physical simulation, establish a mapping relationship between the output power of each device and the change of ambient temperature; S23: In areas with sparse sensor networks, the spatial parameters of undeployed sensors are supplemented by the superposition effect of multiple groups of impact models combined with interpolation algorithms.

[0045] As described above, high-precision three-dimensional reconstruction of the warehouse environment is achieved through the fusion of physical models and data-driven methods: based on historical data and physical simulation modeling, the power-temperature response curve of the temperature control device is established (fitting error <0.1°C), and the Kriging interpolation algorithm is used to perform millimeter-level spatial parameter estimation of sparsely sensed areas (interpolation accuracy reaches ±0.3°C), which increases the accuracy of temperature field simulation in areas without deployed sensors to over 92%. By superimposing multiple groups of impact models (such as thermal radiation coupling calculations of adjacent devices), the airflow disturbances and heat conduction effects in the real environment are effectively restored (dynamic simulation delay <200ms). Compared with traditional linear interpolation methods, the spatial resolution is improved by 10 times, providing a full-space, high-fidelity virtual mirror for intelligent warehouse control, supporting simulation rehearsals for decisions such as temperature zone optimization (reducing trial and error costs by 70%).

[0046] Preferably, the inventory optimization module is also used to predict the shelf life of agricultural products, including: S31: When agricultural products are put into storage, the picking time of the agricultural products is recorded; S32: extracting freshness features of the agricultural products based on the image recognition results, and comparing the freshness features with the registered picking time; S33: After the comparison is successful, the environmental data of the corresponding storage area is integrated and a Weibull distribution model is established to predict the remaining shelf life.

[0047] As mentioned above, the inventory optimization module achieves accurate shelf life prediction of agricultural products through multimodal data fusion and dynamic modeling: at the warehousing stage, benchmark data is established through picking time verification and image freshness feature extraction (recognition accuracy >95%), and a Weibull distribution prediction model is constructed in combination with real-time environmental monitoring data (such as temperature and humidity fluctuation curves). This fully quantifies the impact of storage conditions on the decay rate of agricultural products (prediction error <8 hours); this model can dynamically adapt to the characteristics of different categories (such as automatic adjustment of the difference coefficient between leafy vegetables and root vegetables), thereby improving the remaining shelf life prediction accuracy of perishable agricultural products such as strawberries by 40%. At the same time, it supports priority outbound warning (automatic marking of products approaching their expiration date), reducing the storage loss rate from 15% of traditional methods to below 5%, significantly improving inventory turnover efficiency and economic benefits.

[0048] As a preferred option, the demand forecasting module predicts the agricultural product inflow and outflow volume and regional sales volume for the next 30 days through the following steps: S41: Data cleaning and feature engineering: performing time series aggregation processing on historical order data and removing outliers. At the same time, performing missing value filling and generating derived features on the meteorological monitoring data. The derived features include the number of consecutive high-temperature days and accumulated precipitation. External platform indicators are integrated with the market demand data to generate a quantitative value of regional demand heat. S42: Dual-model collaborative forecasting uses an LSTM neural network to process time series dependencies. It inputs historical inbound and outbound sequences and meteorological time series data, extracts dynamic features through three LSTM hidden layers, and outputs primary inbound and outbound forecasts for the next 30 days. It also uses the XGBoost algorithm to process structured features. It inputs a feature vector containing category attributes, regional identifiers, and promotion intensity, and outputs secondary inbound and outbound forecasts. It also converts meteorological factors into agricultural product corruption coefficients, which are then combined in the LSTM hidden layers for joint training. S43: Dynamic weighted fusion and decision output, dynamically adjusting the fusion weight coefficient according to the environmental fluctuation state , and according to the formula: , generate the final predicted value, where is the final predicted value, is the predicted value of the LSTM neural network, is the predicted value of the XGBoost algorithm; S44: Real-time error correction: daily comparison of the predicted value with the actual inventory in and out. If the regional sales error rate exceeds 15%, the following actions will be performed: analyzing feature contribution through SHAP value; expanding the training sample set; and fine-tuning the weight parameters of the last layer of the LSTM network. S45: Output application, generate the inventory in and out time series curve for the next 7 days; generate the regional sales geographic heat map for the next 30 days; trigger inventory optimization and logistics scheduling instructions based on the forecast results.

[0049] As described above, the demand forecasting module achieves accurate predictions of agricultural product supply and demand through multimodal data fusion and dynamic optimization. Based on data cleaning and feature engineering, a high-dimensional feature system is constructed, including meteorological influencing factors (such as the number of consecutive high-temperature days) and market heat indicators. The LSTM-XGBoost dual-model collaborative mechanism simultaneously captures temporal dependencies (LSTM's three hidden layers extract dynamic features) and structured feature associations (XGBoost analyzes the intensity of promotional activities, etc.), reducing prediction errors by 35% compared to a single model. A dynamic weighted fusion algorithm is used to automatically adjust weights based on environmental fluctuations (for example, the weight of meteorological factors is increased by 40% during heavy rain), stabilizing the 30-day sales forecast error rate to within 8%. Real-time error correction combined with SHAP interpretability analysis enables model self-optimization (an average of 1.2 iterations per day). The final output geographic heat map (S45) and in-and-out inventory curves guide inventory scheduling, reducing the unsold rate by 50% and increasing cold chain logistics resource utilization by 30%, forming a closed-loop optimization system of "prediction-decision-feedback."

[0050] As a preferred method, when converting meteorological factors into agricultural product corruption coefficients, the method specifically includes: S51: Establishing temperature-respiration entropy mapping function , where T is the daily average temperature, and is the category experience parameter, e is the natural constant; S52: Defining the Corruption Coefficient ,in, is the corruption coefficient, d is the prediction date, N is the number of days in the historical window, is the precipitation attenuation factor.

[0051] As mentioned above, the accuracy of agricultural product spoilage prediction is significantly improved through meteorological-biological coupling modeling: the metabolic activity intensity of different types of agricultural products at specific temperatures is quantified based on the temperature-respiratory entropy mapping function (for example, the respiratory entropy of leafy vegetables increases by 300% at 25°C), and the accelerating effect of high humidity on spoilage is dynamically corrected by the precipitation attenuation factor (error rate <5%), so that the generated spoilage coefficient can accurately reflect the level of environmental stress (for example, the strawberry spoilage coefficient automatically increases to 0.85 after 3 consecutive days of 30°C + heavy rain); the model captures the cumulative effect of meteorological influences through historical window sliding calculation (N=7 days), and the prediction sensitivity is increased by 40% compared with the traditional empirical threshold method, providing interpretable biological feature inputs for the LSTM prediction network, ultimately achieving a 92% match between inventory turnover plans and actual spoilage progress, reducing agricultural product waste due to prediction bias by more than 23%.

[0052] As a preference, dynamically adjust the fusion weight coefficient according to the environmental fluctuation state When , the dynamic weight adjustment mechanism adopted is: A51: When the weather volatility exceeds the preset threshold, the weight coefficient is set. , where the quantitative formula for meteorological volatility is: ,in, is the quantitative value of meteorological volatility, is the standard deviation function, is the daily maximum temperature difference series, is the cumulative precipitation series, is the regional adjustment factor, and the fluctuation threshold is ,in, is the mean function; A52: When a sudden promotion event is detected, set the weight coefficient ; A53: Default weight coefficient .

[0053] As described above, adaptive weight allocation across multiple scenarios significantly improves the forecasting system's environmental responsiveness. Based on a quantitative model of meteorological volatility, the system calculates the combined variability of temperature and precipitation in real time (for example, when the delta value exceeds the threshold γ during heavy rain), dynamically increasing the LSTM weight to above 0.7 (improving forecast accuracy by 25% for weather-sensitive categories). During unexpected promotions, the system quickly switches to XGBoost-dominant mode (λ2 = 0.8), accurately capturing the structural characteristics of market demand (e.g., sales forecast error < 5% during promotional periods). By default, the system maintains a balanced fusion of the two models (λ3 = 0.5), ensuring basic forecast stability (30-day error rate < 8%). This mechanism improves the model's forecast robustness by 40% in extreme weather conditions such as typhoons. Furthermore, it supports sub-second weight switching (latency < 10ms), creating a forecasting system that intelligently switches between "weather-sensitive" and "market-oriented" modes. This reduces inventory decision errors caused by environmental shocks by over 60%.

[0054] Preferably, the inventory optimization module comprises the following steps when operating: S51: Data input and preprocessing, obtaining demand forecast data, agricultural product quality grading data, logistics time constraints and environmental strategies; S52: Multi-objective optimization modeling, establish a multi-objective optimization model, the objective function is: ,in, is the weighted distance from the shelf to the platform and emergency area, is the deviation between shelf temperature and humidity and the optimal value of the category, is the penalty item for exceeding the regional floor area ratio limit, 、 and is the weight coefficient; S53: Select an algorithm based on the optimization scale: NSGA-II genetic algorithm is used for global layout updates; DQN reinforcement learning model is used for local storage location adjustments; S54: Instruction execution and closed-loop monitoring, outputting environmental control parameters and product positioning strategies to the execution control layer, and triggering re-optimization when the deviation exceeds the threshold in real time. The re-optimization mechanism is as follows: when the environmental sensor detects that the difference between the actual temperature and the set temperature is greater than 1°C for more than 5 minutes, local re-optimization is automatically triggered; the re-optimization process retains 80% of the valid solutions of the original strategy as the initial population to accelerate convergence.

[0055] As described above, the inventory optimization module achieves optimal allocation of warehouse resources through multi-objective dynamic optimization and intelligent algorithm switching. Based on multi-objective modeling, it simultaneously balances transportation efficiency (minimizing weighted distance), storage quality (optimizing temperature and humidity deviations), and space utilization (volume ratio penalty). It uses a hybrid strategy of NSGA-II and DQN to achieve global-local dual-scale optimization (global layout optimization speed increased by 3 times, and local adjustment response delay decreased to less than 30 seconds). The closed-loop monitoring system automatically initiates thermal gravimetric optimization (retaining 80% of valid solutions to accelerate convergence) through a threshold trigger mechanism (e.g., temperature drift >1°C for 5 minutes), improving the stability of warehouse environmental parameters by 40% and shelf turnover efficiency by 25%. This solution supports dynamic weight adjustment (e.g., prioritizing shortening logistics distances during peak season and focusing on energy conservation during off-season), ultimately reducing warehouse operating costs by 18% and agricultural product loss to below 5%, forming an intelligent closed-loop management and control system based on "decision-making, execution, and feedback."

[0056] As a preferred method, when selecting an algorithm based on the optimization scale, the configuration of the genetic algorithm is as follows: the crossover operation uses a two-point crossover operator with a crossover probability of 0.85; the mutation operation uses a Gaussian mutation operator with a mutation probability of 0.02, and the adaptability evaluation function is ; The DQN reinforcement learning model includes: the state space is the shelf occupancy rate, environmental parameters and real-time order queue; the action space is the category location migration vector; the reward function is , =0.3 is the utilization gain bonus coefficient, is the change in transportation cost, is the change in environmental cost, is the space utilization gain value.

[0057] As described above, efficient dynamic allocation of warehouse resources is achieved through collaborative optimization using hybrid intelligent algorithms. The genetic algorithm utilizes a two-point crossover (with a probability of 0.85) and Gaussian mutation (with a probability of 0.02) configuration to maintain population diversity during global optimization (reducing convergence generations by 40%). Combined with an adaptive evaluation function (including temperature and humidity deviations), the optimal solution set's environmental compatibility improves by 35%. The DQN reinforcement learning model uses three-dimensional state-space modeling (shelf occupancy, environmental parameters, and order queues) to enable real-time localized storage location decisions. Its reward function (including a 0.3 utilization gain coefficient) reduces handling costs by 15% while ensuring space utilization remains above 92%. This algorithm combination reduces the time required for large-scale warehouse layout adjustments from six hours compared to traditional methods to 90 minutes, and enables local dynamic relocation response speeds of 10 seconds per transaction, improving overall logistics efficiency by 30%. This creates a paradigm for intelligent warehouse optimization that combines robust global optimization with agile local response.

[0058] Among them, the digital twin construction module includes the following steps when working: S71: collects temperature, humidity, and CO2 concentration data from physical sensors in the warehouse in real time and associates the data points with their three-dimensional coordinates and timestamps; S72: Dynamically select an interpolation algorithm based on the spatial distribution density of sensors: Kriging interpolation is used to construct the semivariogram in areas where the sensor spacing is less than 2 meters, and anisotropic inverse distance weighting is used in areas near the boundary. S73: Couple the computational fluid dynamics model and use the interpolation results as boundary conditions to solve the energy conservation equation and material transport equation; S74: Compare the measured value with the twin prediction value through the residual compensation mechanism. When the residual exceeds the threshold, the Gaussian process regression model is activated to correct the interpolation parameters. S75: Outputs a virtual image of the environment with a resolution of ≤5mm grid, including temperature field, humidity field, CO2 concentration field and three-dimensional airflow vector.

[0059] As described above, the digital twin construction module achieves ultra-high-precision simulation of the warehouse environment through multimodal fusion and dynamic correction: based on the three-dimensional spatiotemporal correlation of sensor data and an adaptive interpolation algorithm (intelligent switching between Kriging and anisotropic weighted methods), the data filling error in sparse areas is controlled to ±0.3°C / ±2%RH; coupled with the CFD model to solve the energy-matter transfer equation and restore the actual airflow organization (wind speed prediction error <0.1m / s); the residual compensation mechanism uses Gaussian process regression to correct model parameters in real time (response time when residual exceeds the limit <500ms), and finally generates a 5mm grid-level virtual image, which includes temperature and humidity fields, CO2 distribution, and three-dimensional airflow vectors. Compared with traditional methods, the spatial resolution is improved by 10 times, providing millimeter-level decision-making basis for precise environmental control, increasing the identification speed of abnormal areas by 80%, and reducing warehouse energy consumption by 15% to 25%.

[0060] Example 1: Collaborative Management of Intelligent Warehousing and Distribution for Fresh Fruit (In-depth Application in High-Loss Strawberries) Scenario: A cold chain logistics center stored 300 tons of strawberries, facing the dual pressures of rotting caused by high humidity during a rainy season and a sudden surge in market orders. The traditional system, due to delayed environmental control (>5 minutes response time) and inefficient manual sampling (covering only 5% of the goods), resulted in a loss rate of up to 30%.

[0061] Implementation details: 1. Multi-source perception enhancement Heterogeneous sensor network: Hybrid nodes consisting of one high-precision temperature and humidity sensor (±0.1°C) and four low-cost sensors were deployed on 200 shelves. Data was transmitted in real time via the LoRa self-organizing network, reducing the coverage blind spot rate to 0.5%.

[0062] Visually enhanced acquisition: A multispectral camera captures RGB / near-infrared images of the strawberry surface (resolution 4096×2160) when the strawberries are put into storage. A ring light source is used to eliminate shadows in the stacked state.

[0063] External data integration: Connect to the Meteorological Bureau API to obtain red alerts for heavy rain (humidity > 95% in the next 72 hours) and integrate e-commerce platform promotion data (predicting a 200% increase in order volume).

[0064] 2. Edge Intelligent Autonomy Lightweight model reasoning: After knowledge distillation and compression, the YOLOv5s model achieved mAP@0.5 of 98.2% for the teacher model and 96.5% for the student model, and was tested at 120 FPS on a Jetson Nano. In the feature extraction stage, the HSV hue value (deviation ΔH>25°) and texture entropy increment (>0.3) of the moldy area are analyzed, and after fusion, a corruption probability value of 87.5% is output.

[0065] Environmental data refinement: Remove drift nodes based on DBSCAN clustering (sensor #173 is considered abnormal due to a sudden temperature change of 3°C); After spatiotemporal alignment, a shelf-level temperature and humidity cloud map (0.5m×0.5m grid accuracy) was generated.

[0066] Dynamic decision engine: The decision matrix integrates the mildew flag (weight 0.6), humidity deviation +8% (weight 0.3), and the neighboring area CO2 concentration of 2800ppm (weight 0.1) to generate a control intensity value of 0.82 (S62); According to the nonlinear correction rule (applying a 0.6 attenuation coefficient when the humidity deviation is greater than 10%), the humidity adjustment amount is calculated as -12%RH; The power consumption optimization model reduces ventilation time from 30 minutes to 18 minutes, saving 41% energy.

[0067] 3. Cloud-based global optimization Corruption coefficient injection prediction: The rainstorm data were converted to a corruption coefficient of ω = 0.85 (T = 26 °C, parameter β_strawberry = 1.32 in the temperature-respiration entropy function); After splicing the ω value in the LSTM hidden layer, the prediction error rate dropped from 12% to 6.8%.

[0068] Digital Twin Preview: Kriging interpolation was used to complete the cold bridge area data on the west wall (sensor spacing was 3.5 meters), and coupled with CFD simulation to simulate condensation risks (areas with relative humidity >97% were marked with a red alert). The residual compensation mechanism detected a deviation of 0.8°C in the simulation value of area 3, triggering a Gaussian process regression to correct the heat transfer coefficient.

[0069] Joint Inventory Strategy: The NSGA-II algorithm aims to minimize the transport distance (f1) and the spoilage rate (f2). The output solution is: strawberry shelves with a spoilage risk greater than 60% are moved to the low-temperature quick-freeze area (4°C); Weibull model predicts shelf life: combining the picking timestamp (2023-05-2006:00) with surface freshness characteristics (ΔL* value <2.5), an early warning is issued for batches that are about to expire within 18 hours.

[0070] 4. Execution layer closed-loop control The intelligent sorting system sorts fruit based on image analysis results: Category I fruit (corruption rate <5%) is directly supplied to supermarkets, while Category II fruit (corruption rate 5% to 20%) is processed into jam; The temperature control device adjusts the air supply angle (15° depression angle) according to the digital twin airflow vector diagram to prevent condensation from spraying directly onto the cargo.

[0071] Technical achievements: The loss rate during heavy rain cycles dropped from 30% to 5.2%, the sorting efficiency reached 800 boxes / hour (a three-fold increase), and the cold chain energy consumption decreased by 35%.

[0072] Example 2: Intelligent Control of Bulk Grain Reserves (Rice Pest Control and Energy-Saving Storage) Scenario: Storing 100,000 tons of rice requires addressing two major challenges: 1) Early detection of pests during their incubation period is difficult; 2) Long-term temperature control energy consumption accounts for 60% of operating costs. Traditional phosphine fumigation methods have a slow response time (>48 hours) and high energy consumption.

[0073] Implementation details 1. Multimodal Perception Network Deep Vision Detection: 200 infrared cameras capture grain pile thermal images (resolution 2560×1920) and identify grain borer activity areas based on residual temperature differences (temperature difference > 0.5°C); 6-meter-long grain stacks are imaged using penetrating X-rays, combined with YOLOv5 to detect grain infestations (small target detection rate of 95.7%).

[0074] Gas sensor array: CO2 / O2 multi-channel sensors are deployed every 50 square meters to detect pest metabolic gases (CO2 concentration > 3000ppm for 2 hours triggers the threshold).

[0075] 2. Real-time response at the edge layer Pest fusion judgment: Lightweight MobileNetV3 analyzes the texture features of wood grains (entropy value > 7.2) and simultaneously verifies the coordinates of infrared thermal anomaly areas; Output pest density index (level 3: mild <5 insects / ㎡, severe >20 insects / ㎡).

[0076] Oxygen control strategy generation: The decision matrix weights the O2 concentration deviation (weight 0.4), pest density (weight 0.5), and neighboring area status (weight 0.1), and the demand intensity value reaches 0.91; Start level 3 strong ventilation (CO2=3200ppm) according to the ventilation logic tree, and calculate the nitrogen injection amount (target O2<12%) in combination with the oxygen reduction model.

[0077] 3. Cloud-based Intelligent Hub Twin space reconstruction: Generate a 3D grain pile model based on sparse sensors (500 points) using anisotropic inverse distance weighted method: CFD simulation of nitrogen diffusion: solving the NS equations to restore the gas flow boundary layer (eddy strength error <0.05%).

[0078] Hybrid optimization decision: Global planning: The NSGA-II algorithm (crossover probability 0.85) prioritizes "pest risk factor × cargo value" and allocates high-quality rice to core low-oxygen areas (O2 < 10%). Local real-time adjustment: The DQN model receives the order queue (state space: {shelf occupancy rate 82%, O2 concentration 14%, 3 urgent orders}), outputs a shelf migration vector (action space: movement direction + distance), and the reward function calculates the space utilization gain ΔU = +7.3%.

[0079] Self-healing data storage: The edge storage module has double redundancy (six nodes store the same data), and when node #4 fails, it is automatically repaired through a majority voting mechanism; Execute oxygen reduction instructions locally during network outage (72 hours of offline operation).

[0080] 4. Precision Intervention at the Executive Level The early warning device triggers an audible and visual alarm (the pest coordinates are projected onto the twin platforms); The intelligent ventilation system optimizes the air supply trajectory according to the airflow vector simulated by CFD (velocity field 0.8m / s, wind direction angle 35°); The nitrogen injection device is linked to the temperature control system to maintain a constant temperature of 15°C to avoid condensation.

[0081] Technical achievements: Pests are discovered in the incubation period (<24 hours) ahead of schedule, energy consumption for temperature control is reduced from 58kWh / ton to 35kWh / ton, and the stored grain loss rate is <0.1% (national standard requirement is <0.3%).

[0082] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims

1. An intelligent management system for agricultural product storage based on big data, characterized by: It includes multi-source data collection layer, edge processing layer, cloud processing layer and execution control layer, among which: Multi-source data collection layer, used to collect storage environment parameters, agricultural product parameters, weather forecast data, historical logistics data and market demand data; The edge processing layer is deployed at the warehouse site and is used to analyze storage environment parameters and agricultural product parameters in real time through the deep learning model built into the edge computing node, and generate control instructions; The cloud processing layer integrates and processes multi-source data based on a distributed cloud platform, combines historical data with prediction models to generate warehouse control instructions, and sends them to the edge processing layer; The execution control layer is used to drive the temperature control device, intelligent sorting system and early warning device according to the control instructions of the edge processing layer, and to adjust the storage environment parameters and the storage status of agricultural products in real time. Specifically, it includes the temperature control device, intelligent sorting system and early warning device.

2. The big data-based intelligent management system for agricultural product storage according to claim 1, characterized in that: The multi-source data collection layer specifically includes: A11: Environmental sensor network, consisting of sensor nodes and using the LoRa communication protocol. Each sensor node consists of a set of high-precision sensors and multiple sets of low-cost sensors. A12: Image acquisition equipment, used to collect image data of agricultural products when they are stored in the warehouse, and to collect images of agricultural products and the environment inside the warehouse in real time; A13: External data interface, used to connect to external systems to obtain weather forecast data, historical logistics data, and market demand data.

3. The big data-based intelligent agricultural product storage management system according to claim 2, characterized in that: The edge processing layer specifically includes: A21: Image analysis module, with a built-in lightweight deep learning model, is used to analyze collected agricultural product images, extract color, size, appearance, and texture features, and identify the quality, maturity, and freshness of agricultural products; A22: Environmental preprocessing module, used to filter, remove outliers, and synchronize time on the raw data collected by the sensor network to generate a local environmental data summary. When filtering, removing outliers, and synchronizing time on the raw data collected by the sensor network to generate the local environmental data summary, the environmental preprocessing module verifies the data from the high-precision sensors and low-cost sensors in each group of sensor nodes. If discrepancies arise, the module removes outliers and performs weighted fusion to obtain the final data for each group of sensor nodes. A23: Control module, used to generate primary control instructions including local ventilation and temperature and humidity fine-tuning based on image analysis results and environmental preprocessing data.

4. The big data-based intelligent management system for agricultural product storage according to claim 3 is characterized by: The image analysis module identifies the quality, maturity, and freshness of agricultural products through the following steps: S11: Image acquisition and preprocessing: when agricultural products are put into storage, the original RGB agricultural product images are acquired through the image acquisition device and the original RGB agricultural product images are preprocessed; S12: Object detection, using the built-in YOLOv5 model to perform object detection and classification to identify the specific categories of agricultural products; S13: Based on the feature extraction of the detection results, for each detected agricultural product individual, region cropping is performed and feature extraction is performed; S14: Feature fusion and result judgment: combine the extracted features into a comprehensive feature vector, and use the trained classification model to judge the quality, maturity and freshness of agricultural products.

5. The big data-based intelligent management system for agricultural product storage according to claim 4, characterized in that: The cloud processing layer specifically includes: A31: Distributed data storage module, which uses a distributed consensus mechanism and includes an edge storage module and a cloud storage module. The edge storage module is distributed within the edge processing layer and is used to store data in the edge storage module and the cloud storage module. A32: Demand forecasting module, based on a fusion model of LSTM and XGBoost, combines historical orders, weather data, and market demand data to predict agricultural product inflows and outflows and regional sales for the next 30 days; A33: Inventory optimization module, which uses a genetic algorithm to combine predicted orders, logistics timeliness, and agricultural product quality grades to dynamically divide storage areas and match optimal temperature and humidity conditions. A34: Digital twin building module, used to fuse sensor data with physical models to generate a real-time virtual image of the warehouse environment. It uses Kriging interpolation to interpolate environmental parameters in sparse sensor areas and achieve millimeter-level spatial parameter simulation.

6. The big data-based intelligent management system for agricultural product storage according to claim 5, characterized in that: The edge storage module is deployed at the warehouse site and contains multiple independent edge storage sub-modules. Each sub-module stores part of the warehouse data, and the storage data of all sub-modules has redundant coverage. Specifically, at any time, the total amount of data stored by all sub-modules is twice the total amount of warehouse data, that is, the same data item is stored repeatedly by at least two different sub-modules. The cloud storage module is deployed in a remote data center and stores the full amount of warehouse data, that is, a complete set of all redundant data in the edge storage module. The edge storage module and the cloud storage module achieve data consistency through the following mechanisms: A41: Timed synchronization. According to the set synchronization period, the edge storage module packages the redundant data of each submodule and synchronizes it to the cloud storage module. At the same time, the cloud storage module reversely synchronizes the latest global data to the edge storage module. A42: Confidence Verification: During the synchronization process, a dual verification strategy is used to ensure data accuracy: a hash value is generated for each set of synchronized data. The cloud and edge nodes calculate and compare the hash values. If they are inconsistent, the data is marked as suspicious. For any item marked as suspicious, the edge storage module retrieves copies of all submodules storing the data and determines the correct version based on the principle of data consistency among the majority of submodules. The cloud storage module synchronously uses the same rules to correct data.

7. The big data-based intelligent management system for agricultural product storage according to claim 6, characterized in that: The digital twin construction module uses the Kriging interpolation method to interpolate data on environmental parameters in sparse sensor areas to achieve millimeter-level spatial parameter simulation. Specifically, it includes: S21: Collect historical temperature data of each temperature regulating device at different powers; S22: Based on physical simulation, establish a mapping relationship between the output power of each device and the change of ambient temperature; S23: In areas with sparse sensor networks, the spatial parameters of undeployed sensors are supplemented by the superposition effect of multiple groups of impact models combined with interpolation algorithms.

8. The big data-based intelligent management system for agricultural product storage according to claim 7, characterized in that: The inventory optimization module is also used to predict the shelf life of agricultural products, including: S31: When agricultural products are put into storage, the picking time of the agricultural products is recorded; S32: extracting freshness features of the agricultural products based on the image recognition results, and comparing the freshness features with the registered picking time; S33: After the comparison is successful, the environmental data of the corresponding storage area is integrated and a Weibull distribution model is established to predict the remaining shelf life.

9. The big data-based intelligent management system for agricultural product storage according to claim 8, characterized in that: The demand forecasting module predicts the agricultural product inflow and outflow volume and regional sales volume for the next 30 days through the following steps: S41: Data cleaning and feature engineering: performing time series aggregation processing on historical order data and removing outliers, performing missing value filling and derivative feature generation on meteorological monitoring data, and integrating external platform indicators with the market demand data to generate quantitative values ​​of regional demand heat; S42: Dual-model collaborative forecasting uses an LSTM neural network to process time series dependencies. It inputs historical inbound and outbound sequences and meteorological time series data, extracts dynamic features through three LSTM hidden layers, and outputs primary inbound and outbound forecasts for the next 30 days. It also uses the XGBoost algorithm to process structured features. It inputs a feature vector containing category attributes, regional identifiers, and promotion intensity, and outputs secondary inbound and outbound forecasts. It also converts meteorological factors into agricultural product corruption coefficients, which are then combined in the LSTM hidden layers for joint training. S43: Dynamic weighted fusion and decision output, dynamically adjusting the fusion weight coefficient according to the environmental fluctuation state , and according to the formula: , generate the final prediction value; in, is the final predicted value, is the predicted value of the LSTM neural network, is the predicted value of the XGBoost algorithm; S44: Real-time error correction: daily comparison of the predicted value with the actual inventory in and out. If the regional sales error rate exceeds 15%, the following actions will be performed: analyzing feature contribution through SHAP value; expanding the training sample set; and fine-tuning the weight parameters of the last layer of the LSTM network. S45: Output application, generate the inventory in and out time series curve for the next 7 days; generate the regional sales geographic heat map for the next 30 days; trigger inventory optimization and logistics scheduling instructions based on the forecast results.

10. The big data-based intelligent management system for agricultural product storage according to claim 9, characterized in that: The inventory optimization module works by: S51: Data input and preprocessing, obtaining demand forecast data, agricultural product quality grading data, logistics time constraints and environmental strategies; S52: Multi-objective optimization modeling, establish a multi-objective optimization model, the objective function is: ,in, is the weighted distance from the shelf to the platform and the emergency area, is the deviation between shelf temperature and humidity and the optimal value of the category, is the penalty item for exceeding the regional floor area ratio limit, 、 and is the weight coefficient; S53: Select an algorithm based on the optimization scale: NSGA-II genetic algorithm is used for global layout updates; DQN reinforcement learning model is used for local storage location adjustments; S54: Instruction execution and closed-loop monitoring, outputting environmental control parameters and product positioning strategies to the execution control layer, and triggering re-optimization when the deviation exceeds the threshold in real time. The re-optimization mechanism is as follows: when the environmental sensor detects that the difference between the actual temperature and the set temperature is greater than 1°C for more than 5 minutes, local re-optimization is automatically triggered; the re-optimization process retains 80% of the valid solutions of the original strategy as the initial population to accelerate convergence.

Images