Intelligent monitoring and early warning warehouse for agricultural products

Abstract

The invention provides an agricultural product intelligent monitoring and early warning warehouse, belongs to the technical field of agricultural product storage intelligent monitoring, and solves the problems of poor dynamic adaptability, weak early warning timeliness and insufficient control collaboration of an agricultural product monitoring warehouse. Comprising a sensor network module, an edge calculation module, an execution control module, an early warning module, a cloud platform module and a fault self-diagnosis module. The sensor network module integrates environmental parameters and agricultural product physiological state parameter two-dimensional acquisition; the edge calculation module is deployed in a local warehouse, the core bears real-time operation of three innovative algorithms, a multi-parameter coupling dynamic threshold optimization algorithm, a mildew risk prediction algorithm and a distributed cooperative control algorithm are designed, linkage control of execution units is realized based on parameter deviation priority ranking, environmental fluctuation is avoided, and the reliability of the system is improved. Through three algorithm closed loops and in combination with a cloud edge collaborative architecture, precise monitoring and efficient management and control of agricultural product storage are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for agricultural product storage, and in particular to an intelligent monitoring and early warning warehouse for agricultural products. Background Technology

[0002] Agricultural product storage is a crucial link in the agricultural industry chain. The stability of the warehouse storage environment and the evolution of the physiological state of agricultural products directly determine storage quality and loss rate. Although existing agricultural product warehouse monitoring systems can monitor basic parameters, they still face technical bottlenecks in large-scale, precise storage scenarios: First, the use of fixed threshold judgment mechanisms fails to consider the dynamic differences in agricultural product categories, maturity, and storage duration, and ignores the coupling relationship between environmental parameters, resulting in poor threshold adaptability and frequent misjudgments. Second, the early warning mode is limited to "real-time alarm after exceeding the standard," failing to predict the risk of quality deterioration in advance, making it difficult to avoid irreversible quality losses. Third, the execution unit uses independent control of a single parameter, lacking collaborative logic, which easily leads to environmental fluctuations and high energy consumption. Fourth, data analysis is limited to basic statistics, model training relies on local data from a single warehouse, has weak generalization ability, and requires manual recalibration when changing agricultural product categories or storage environments, which is cumbersome and has low adaptation efficiency.

[0003] To address the aforementioned issues, an intelligent monitoring system with dynamic adaptation, early warning, collaborative control, and wide adaptability is proposed to meet the high-precision storage needs of agricultural products of different categories and scales. Summary of the Invention

[0004] This invention aims to solve the technical problems of poor dynamic adaptability, weak early warning timeliness, insufficient control coordination, and weak model generalization ability in existing agricultural product monitoring warehouses. It provides an intelligent monitoring and early warning warehouse for agricultural products, with the specific technical solution as follows: A smart monitoring and early warning warehouse for agricultural products includes: a sensor network module, an edge computing module, an execution control module, an early warning module, a cloud platform module, and a fault self-diagnosis module. The sensor network module innovatively integrates the dual-dimensional acquisition of environmental parameters and physiological state parameters of agricultural products, addressing the shortcomings of existing systems that only monitor environmental parameters. Environmental parameters include temperature, humidity, oxygen concentration, carbon dioxide concentration, ethylene concentration, and dust concentration; physiological state parameters include surface humidity, hardness, and color difference of agricultural products. A distributed node deployment scheme is adopted, with one sensor node configured every 50-100 square meters. Each node integrates a group of environmental parameter sensors and a group of physiological state sensors. To improve data transmission stability and energy efficiency, the nodes adopt the LoRa wireless communication protocol, supporting on-demand wake-up and sleep switching. A built-in self-calibration mechanism automatically performs calibration every 24 hours to ensure the accuracy of collected data. The edge computing module is deployed locally in the warehouse, wirelessly communicating with the sensor network module. It corely supports the real-time operation of three innovative algorithms and simultaneously performs data preprocessing. The data preprocessing stage includes outlier removal, linear interpolation for missing values, and standardization to ensure data quality. The three core algorithms are: a multi-parameter coupled dynamic threshold optimization algorithm (solving the problem of poor adaptability of fixed thresholds), LSTM-attention for mold risk prediction, and... The system includes algorithms (for early warning) and distributed collaborative control algorithms (to improve the efficiency of execution unit linkage); it also features local data caching, storing key data (including parameters, warning records, and control commands) for the past 72 hours, and retaining data for 72 hours after power failure to ensure data continuity; after data processing and algorithm calculation, the results are synchronized to the cloud platform, and the system receives model updates and remote control commands from the cloud platform; the execution control module includes ventilation, dehumidification, refrigeration, atmosphere control, and pest control units, communicates with the edge computing module, and receives collaborative control commands; each unit is customized based on collaborative control requirements to ensure linkage stability; The aforementioned early warning module adopts a dual early warning architecture of "local audio-visual + remote multi-level push" to ensure that no early warning information is missed. The early warning module communicates with the edge computing module and outputs differentiated early warning signals according to the risk level. The cloud platform module communicates with the edge computing module through 5G or Ethernet, and its core functions include model iterative optimization and remote data management. Unlike existing cloud platforms that only have data storage functions, it introduces a federated learning framework to ensure data privacy while improving the model's generalization ability. The fault self-diagnosis module focuses on core equipment fault monitoring and communicates with the sensor network module, edge computing module, and execution control module to achieve accurate identification and early warning of critical equipment faults.

[0005] More preferably, the edge computing module generates dynamic thresholds adapted to the current agricultural product category, maturity, and real-time environment through a multi-parameter coupled dynamic threshold optimization algorithm; outputs the mold risk level 24-72 hours in advance through an LSTM-attention mold risk prediction algorithm; generates linkage control instructions for each execution unit through a distributed collaborative control algorithm to avoid environmental fluctuations caused by single control; uses non-volatile storage media for local data caching to ensure data security during power outages; and adopts incremental transmission for data synchronization to reduce bandwidth consumption.

[0006] Even better, the ventilation unit of the execution control module adopts a variable frequency centrifugal fan, supporting both timed and linked operation modes; the dehumidification unit adopts a rotary dehumidifier, linked with the ventilation unit for control, to avoid sudden changes in humidity; the refrigeration unit adopts a multi-head variable frequency compressor refrigeration system to ensure temperature uniformity within the warehouse; the controlled atmosphere unit integrates a nitrogen generator, a carbon dioxide generator, and an ethylene adsorber to achieve precise control of gas composition; and the pest control unit adopts a combination of ultrasonic insect repellent and low-toxicity trapping device, which is activated on a timed basis during non-high-priority control periods to avoid conflicts with the operation of other units.

[0007] More preferably, the local audible and visual early warning unit is installed at the warehouse entrance and control room, using a combination of red, yellow, and green indicator lights and a buzzer; the remote push unit pushes information to the administrator's APP and WeChat via 5G / 4G network, and supports telephone calls, enabling multi-level push to administrators, responsible persons, and management, ensuring timely response in unattended scenarios; specific grading standards: Level 1 Warning (Low Risk): Environmental parameters are close to the dynamic threshold, the probability of mold risk prediction is ≤30%, triggering a flashing green light on the local device (buzzer alarm at 5-second intervals), and a text reminder is pushed to the administrator's APP remotely; Level 2 Warning (Medium Risk): Parameters exceed the dynamic threshold by ±5%, the probability of mold risk is predicted to be 30%-60%, triggering a local yellow light to stay on (buzzer alarm at 2-second intervals), remotely pushing text and voice reminders to the administrator's APP and WeChat, and calling the administrator's phone. Level 3 warning (high risk): When parameters exceed the dynamic threshold by ±10% and the predicted probability of mold growth is >60%, the local red light flashes and stays on alternately (the buzzer sounds a continuous alarm). Warning information is pushed remotely to administrators, responsible persons, and management levels. When the dust concentration exceeds the standard, the fire protection system is activated.

[0008] More preferably, the cloud platform module includes a data storage unit, a model iteration unit, a remote monitoring unit, and a data visualization unit. The data storage unit uses a distributed database to store raw sensor data, preprocessed data, algorithm parameters, and early warning records. The model iteration unit is a core innovative module that uses a federated learning framework to aggregate anonymized data from multiple warehouses to jointly train the LSTM-attention prediction algorithm, thus avoiding data privacy leaks and improving the model's adaptability to different agricultural product categories and storage environments. The remote monitoring unit supports access via web and mobile app, providing real-time parameter viewing, historical data querying, early warning record export, and remote control functions for the execution unit. The data visualization unit displays trends in environmental parameter changes, distribution of agricultural product physiological states, and early warning level statistics, and supports the generation of custom reports.

[0009] More preferably, the fault self-diagnosis module monitors the communication status and battery level of the sensor node. If communication fails three times consecutively or the battery level is below 10%, a sensor fault warning is triggered. It also monitors the operating current and fault codes of the execution unit. If the current is abnormal or the fault code is non-zero, an execution unit fault warning is triggered and the fault information is recorded. Furthermore, it monitors the core operating status of the edge computing module to ensure the stability of the algorithm operation.

[0010] Furthermore, the specific implementation logic of the three core algorithms of this invention is as follows, and their synergistic effect addresses the core defects of existing technologies: 1) A multi-parameter coupled dynamic threshold optimization algorithm addresses the shortcomings of existing fixed thresholds that cannot adapt to the dynamic storage needs of agricultural products. It generates adaptive thresholds through agricultural product category matching, multi-parameter weight coupling, and dynamic iterative updates. Specific steps include: S1. Data preprocessing: Outlier removal, missing value linear interpolation filling, and standardization are performed on the 8 core acquisition parameters (temperature, humidity, oxygen concentration, carbon dioxide concentration, ethylene concentration, hardness, surface humidity, and color difference). S2, Category-Maturity Feature Matching: The cloud platform pre-stores basic threshold datasets for different maturity levels of agricultural products. During system initialization, after the user selects the agricultural product category and maturity level, the edge computing module calls the corresponding basic threshold from the cloud platform. S3. Multi-parameter coupling weight calculation: The Analytic Hierarchy Process (AHP) is used to determine the weight of each parameter on the storage quality. A judgment matrix is ​​constructed based on the storage mechanism of agricultural products and experimental data to clarify the relative importance between parameters. Consistency test (consistency ratio < 0.1) is completed by calculating the maximum eigenvalue to ensure the rationality of weight allocation. The eigenvectors of the judgment matrix are normalized to obtain the final weight vector. S4. Dynamic threshold generation: Based on the preprocessed real-time parameter vector and weight vector, a dynamic threshold is generated through weighted iterative calculation. S5. Threshold validity verification: Calculate the deviation rate between the dynamic threshold and the actual parameters every hour. If the deviation rate is >5% for 3 consecutive times, re-execute S3-S4 to adjust the weight vector to ensure threshold adaptability.

[0011] The algorithm integrates multi-dimensional information such as category, maturity, and real-time parameters, breaking the static limitations of fixed thresholds; it quantifies the parameter coupling relationship through the AHP method to reduce the false judgment rate; and it introduces an effectiveness verification mechanism to ensure real-time adaptability, with the false judgment rate controlled within 3%.

[0012] 2) The LSTM-attention mold risk prediction algorithm addresses the shortcomings of existing technologies in providing delayed early warnings. It integrates environmental and physiological parameters to extract deep features, achieving long-leader mold risk prediction. Specific steps include: S1. Dataset Construction: Based on historical data collected by sensors, a three-dimensional dataset consisting of "environmental parameter sequence - physiological state parameter sequence - mold growth result label" is constructed. The time window length of the parameter sequence is set to 24 hours, the sampling interval is 1 hour, and the mold growth result label corresponds to the actual mold growth state 72 hours after the end of the time window. The dataset is divided into training set, validation set and test set in a 7:2:1 ratio. S2. Dual-domain feature extraction: Extract features from the parameter sequence in both the time domain (mean, variance, peak value) and the frequency domain (Fourier transform to extract the main frequency feature), construct a high-dimensional feature matrix, and comprehensively capture the evolution law of the parameters. S3, LSTM Feature Encoding: The feature matrix is ​​input into a 3-layer LSTM network with 128 hidden units in each layer. The tanh activation function is used, and the forget gate, input gate, and output gate use the sigmoid activation function. The deep correlation features of parameter evolution are extracted through time series modeling. S4. Attention Weighted Enhancement: Introducing a self-attention mechanism to assign weights to the hidden state sequence output by LSTM, enhancing the key time step features that significantly affect the risk of mold spoilage, and solving the problem of insufficient attention to important information by traditional LSTM. S5. Risk Level Prediction: Input the weighted hidden state into two fully connected layers. The first layer outputs a dimension of 64 (ReLU activation), and the second layer outputs a dimension of 3 (corresponding to three levels of mold risk, softmax activation). Output the probability distribution of the risk level. S6. Cloud-edge collaborative iterative optimization: Combining local fine-tuning with cloud-based joint training—The edge computing module fine-tunes the model every 7 days using the local data from the most recent 7 days to adapt to changes in the local environment of the warehouse; The cloud platform aggregates the anonymized model update parameters from multiple warehouses through a federated learning framework, generates a globally optimized model, and distributes it to each edge node to improve the model's generalization ability without the need for manual recalibration.

[0013] The algorithm improves the prediction accuracy through the fusion of two-dimensional parameters and two-domain features. The Attention mechanism strengthens key features, and the cloud-edge collaborative iteration ensures generalization ability. The early warning time can reach 24-72 hours.

[0014] 3) The distributed collaborative control algorithm aims at the defect of insufficient coordination of existing execution unit controls. Based on the parameter deviation priority, it realizes multi-unit linkage control. The specific steps are as follows: S1. Determination of risk-oriented control objectives: Combining the risk levels predicted by LSTM-attention and dynamic thresholds, the control objectives are set differentially - when the risk is low, the midpoint of the optimal interval of the dynamic threshold is taken as the target; when the risk is medium, the environmental parameters are targeted at the lower limit of the threshold, and the physiological state parameters are targeted at the upper limit of the threshold; when the risk is high, the environmental parameters are targeted at a value slightly lower than the midpoint of the optimal interval, and the physiological state parameters are targeted at a value slightly higher than the midpoint of the optimal interval. At the same time, the emergency processing unit is activated. S2. Perception of the state of the execution unit: Collect the operating status (running / stopped), real-time power, and fault codes of each execution unit, and construct a 5×3-dimensional state matrix (5 execution units, 3 state dimensions) to provide a basis for collaborative control. S3. Sorting of deviation priorities: Calculate the degree of deviation between each parameter and the target value, and sort them according to the deviation size to determine the control priority. The larger the deviation, the higher the priority. S4. Generation of linkage control instructions: Based on the priority, control tasks are assigned. The core linkage logic is as follows: When the temperature deviation is the largest, the refrigeration unit is started first, and the dehumidification unit is linked according to the humidity status (when the humidity is too high, it is linked to run, and when it is too low, the refrigeration unit runs alone); when the humidity deviation is the largest, the dehumidification unit is started first, and the operation duration of the ventilation unit is linked and regulated; when the gas concentration deviation is the largest, the gas conditioning unit is started and the ventilation unit is closed; the pest control unit is started during non-high-priority control periods to avoid operation conflicts. S5. Real-time feedback optimization: Parameters are collected every 15 minutes, and the deviation change trend is calculated - when the deviation shrinks, the current instruction is maintained; if it does not shrink, the execution parameters are adjusted and the priority is re-sorted; if it expands, S3-S4 are re-executed to ensure the control effect.

[0015] The risk-oriented target setting of the algorithm can improve the control pertinence. The priority sorting and linkage logic avoid control conflicts, and the real-time feedback ensures control accuracy. Edge computing local execution ensures that the response delay ≤ 1 second, and the time for the parameter to return to the target value ≤ 30 minutes. At the same time, through precise linkage, the frequent start-stop rate of the execution unit is reduced, and the energy consumption is reduced.

[0016] The present invention provides an intelligent monitoring and warning warehouse for agricultural products, which has the following beneficial effects: When in use, this invention reduces the false judgment rate to below 3% through a multi-parameter coupled dynamic threshold optimization algorithm, and achieves early warning of mold risk in 24-72 hours by combining the LSTM-attention algorithm, allowing sufficient time for quality control and effectively reducing the loss rate of agricultural products. The distributed collaborative control algorithm enables the linkage of each execution unit, and the parameter regression target value time is ≤30 minutes, reducing energy consumption and balancing control effect and energy saving requirements.

[0017] In addition, the cloud-edge collaborative federated learning iteration mode enables the model to adapt to different types of agricultural products and warehouses of various sizes without manual calibration, significantly improving its generalization ability; remote monitoring by the cloud platform enables unattended operation and reduces labor costs.

[0018] In addition, local edge computing can operate independently, ensuring normal monitoring, early warning and control functions even when the network is down. Sensor self-calibration and fault self-diagnosis mechanisms further enhance system stability. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below.

[0020] The accompanying drawings described below are only related to some embodiments of the invention and are not intended to limit the invention.

[0021] In the accompanying drawings of the instruction manual: Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is the system logic diagram of the present invention.

[0022] Figure Labels 1. Sensor Network Module; 101. Sensor Node; 2. Edge Computing Module; 3. Execution Control Module; 301. Ventilation Unit; 302. Dehumidification Unit; 303. Cooling Unit; 304. Controlled Atmosphere Unit; 305. Pest Control Unit; 4. Early Warning Module; 5. Cloud Platform Module; 501. Data Storage Unit; 502. Model Iteration Unit; 503. Remote Monitoring Unit; 504. Data Visualization Unit; 6. Fault Self-Diagnosis Module. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the described embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example: Please refer to the appendix. Figure 1 To be continued Figure 2 : This invention proposes an intelligent monitoring and early warning warehouse for agricultural products, comprising: a sensor network module 1, an edge computing module 2, an execution control module 3, an early warning module 4, a cloud platform module 5, and a fault self-diagnosis module 6. The sensor network module 1 collects environmental parameters of the agricultural products' storage environment and physiological state parameters within the warehouse. The edge computing module 2 is deployed locally in the warehouse and wirelessly connected to the sensor network module 1. It preprocesses the collected parameters in real time and runs a multi-parameter coupled dynamic threshold optimization algorithm, an LSTM-attention mold risk prediction algorithm, and a distributed collaborative control algorithm. The execution control module 3 includes... The system includes a ventilation unit 301, a dehumidification unit 302, a cooling unit 303, a controlled atmosphere unit 304, and a pest control unit 305. The execution control module 3 is communicatively connected to the edge computing module 2. The early warning module 4 includes a local audible and visual early warning unit 401 and a remote push unit 402. The early warning module 4 is communicatively connected to the edge computing module 2. The cloud platform module 5 communicates with the edge computing module 2 via 5G or Ethernet and is used for parameter storage, algorithm model iteration, remote monitoring, and data visualization. The fault self-diagnosis module 6 is communicatively connected to the sensor network module 1, the edge computing module 2, and the execution control module 3.

[0025] In this embodiment, environmental parameters include temperature, humidity, oxygen concentration, carbon dioxide concentration, ethylene concentration, and dust concentration. Physiological state parameters of agricultural products include surface humidity, hardness, and color difference. Sensor network module 1 sets up one sensor node 101 every 50-100 square meters. Each sensor node 101 integrates an environmental parameter sensor group and a physiological state sensor group. The environmental parameter sensor group includes a temperature sensor, a humidity sensor, a gas sensor, and a dust sensor, wherein the gas sensors include an oxygen concentration sensor, a carbon dioxide concentration sensor, and an ethylene concentration sensor. The agricultural product physiological state sensor group includes a surface humidity sensor, a hardness sensor, and a color difference sensor. Furthermore, the sensor node 101 uses the LoRa wireless communication protocol and supports automatic wake-up and sleep switching.

[0026] In this embodiment, the edge computing module 2 generates an adaptive dynamic threshold through a multi-parameter coupled dynamic threshold optimization algorithm, predicts the risk of mold growth in agricultural products 24-72 hours in advance through an LSTM-attention mold risk prediction algorithm, and realizes the linkage and collaborative control of each execution unit through a distributed collaborative control algorithm.

[0027] In this embodiment, the specific configuration of each unit of the execution control module 3 is as follows: the ventilation unit 301 adopts a variable frequency centrifugal fan; the dehumidification unit 302 adopts a rotary dehumidifier; the refrigeration unit 303 adopts a variable frequency compressor refrigeration system; the controlled atmosphere unit 304 includes a nitrogen generator, a carbon dioxide generator, and an ethylene adsorber; and the pest control unit 305 adopts an ultrasonic insect repellent combined with a low-toxicity trapping device.

[0028] In this embodiment of the disclosure, the logical steps of the multi-parameter coupled dynamic threshold optimization algorithm include: S1. Data preprocessing: Outlier removal, missing value filling and standardization of environmental and physiological parameters collected by sensors; S2. Agricultural product category feature matching: Based on the category and maturity level of the stored agricultural products, the corresponding basic threshold dataset is retrieved from cloud platform module 5; S3. Multi-parameter coupling weight calculation: The analytic hierarchy process is used to determine the weight of each parameter on the storage quality of agricultural products, a judgment matrix is ​​constructed, and the weight vector is obtained after passing the consistency test. S4. Dynamic threshold iterative update: Based on the standardized parameter vector collected in real time, the dynamic threshold correction coefficient is calculated by combining the weight vector, and a dynamic threshold vector is generated. S5. Threshold validity verification: Calculate the deviation rate between the dynamic threshold and the actual parameters every hour. If the deviation rate exceeds 5% for 3 consecutive times, repeat steps S3 to S4 to adjust the weight vector and correction coefficient.

[0029] In this embodiment of the disclosure, the logical steps of the LSTM-attention mold risk prediction algorithm include: S1. Dataset Construction: Based on the historical data collected by sensor network module 1, including environmental parameter sequences, physiological state parameter sequences and corresponding moldy result labels, a training dataset is constructed. The time window length of the environmental parameter sequences and physiological state parameter sequences is set to 24 hours, and the sampling interval is 1 hour. S2. Feature Extraction: Perform time-domain and frequency-domain feature extraction on the parameter sequence to construct a feature matrix; S3, LSTM Feature Encoding: Input the feature matrix into the LSTM network. The LSTM network contains 3 hidden layers, each with 128 hidden units. The activation function is tanh, and the activation functions for the forget gate, input gate, and output gate are sigmoid. Calculate the hidden state. S4, Attention Mechanism Weighting: Introducing a self-attention mechanism to the hidden state sequence of the LSTM output; S5. Risk Level Prediction: Input the weighted hidden state into a fully connected layer. The fully connected layer contains two layers. The first layer has an output dimension of 64 and uses ReLU as the activation function. The second layer has an output dimension of 3 (corresponding to 3 mold risk levels) and uses softmax as the activation function. Calculate the probability distribution of the risk level. S6. Model Iteration and Optimization: The prediction error is calculated using the cross-entropy loss function, the network parameters are updated using the Adam optimizer, and the model is fine-tuned every 7 days using historical data from cloud platform module 5.

[0030] In this embodiment of the disclosure, the logical steps of the distributed cooperative control algorithm include: S1. Control target determination: Based on the risk level output by the LSTM-attention prediction algorithm and the multi-parameter coupled dynamic threshold, the control target vector is determined, where each target parameter is the midpoint of the optimal interval of the dynamic threshold; S2. Execution Unit Status Assessment: Collect the current operating status of each execution unit and construct a status matrix, which includes 5 execution units (ventilation, dehumidification, cooling, controlled atmosphere, and pest control) and 3 status dimensions (operating status, power, and fault status). S3. Control Priority Ranking: Determine the control priority based on the degree of deviation of each parameter from the target value and the storage requirements of agricultural products; S4. Cooperative Control Instruction Generation: Based on control priority, control tasks are assigned to each execution unit. The specific logic is as follows: If the temperature deviation is the largest: the cooling unit 303 will be activated first, and the dehumidification unit 302 will be adjusted according to the humidity parameters (if the humidity is higher than the target value, the dehumidification unit 302 and the cooling unit 303 will operate in conjunction; if the humidity is lower than the target value, the cooling unit 303 will operate alone, and the power of the dehumidification unit 302 will be reduced). If the humidity deviation is the largest: Dehumidification unit 302 will be activated first, while the operating time of ventilation unit 301 will be controlled at the same time; If the gas concentration deviation is the largest: activate the gas conditioning unit 304 to adjust the ratio of oxygen and carbon dioxide, and at the same time shut down the ventilation unit 301; The pest control unit 305 operates independently and is activated during periods when non-high-priority control is not in effect, based on dust concentration and pest monitoring data. S5. Control effect feedback and adjustment: Parameters are collected every 15 minutes, the deviation between the parameters and the target value is calculated, and if the deviation is not reduced, the running parameters of the execution unit are adjusted, the control priority is recalculated, and the control instructions are iteratively optimized.

[0031] In this embodiment, the fault self-diagnosis module 6 periodically detects the communication status and battery level of the sensor node 101. When communication fails three times consecutively or the battery level is below 10%, a sensor fault warning is triggered. The module 6 also detects the CPU utilization and memory usage of the edge computing module 2. When the CPU utilization exceeds 80% for one hour or the memory usage exceeds 90%, non-core processes are automatically shut down. The module 6 also detects the operating current and fault codes of the execution unit. When the current is abnormal or the fault code is not zero, an execution unit fault warning is triggered, and fault information is recorded.

[0032] In this embodiment of the disclosure, the warning level of the warning module 4 is divided into three levels, corresponding to different warning methods: Level 1 Warning (Low Risk): When the storage environment parameters of agricultural products are close to the upper / lower limit of the dynamic threshold, the probability of mold risk is ≤30%, triggering the local audible and visual warning unit (green light flashing, buzzer alarm at 5-second intervals), and the remote push unit sends a text reminder to the administrator APP. Level 2 warning (medium risk): When the parameter exceeds the dynamic threshold ±5%, the predicted probability of mold risk is >30% but ≤60%, the local sound and light warning unit is triggered, and the remote push unit sends text and voice reminders to the administrator's APP and WeChat, and at the same time dials the administrator's phone. Level 3 Warning (High Risk): When parameters exceed the dynamic threshold by ±10% and the predicted probability of mold risk is >60%, the local audible and visual warning unit is triggered, and the remote push unit pushes the warning information to the administrator, warehouse manager, and enterprise management at multiple levels, while simultaneously linking the fire protection system.

[0033] In this embodiment of the disclosure, the cloud platform module 5 is connected to the user terminal via either a web application or an app. The cloud platform module 5 includes: Data storage unit 501 uses a distributed database to store raw data collected by the sensor, preprocessed data, algorithm model parameters, and early warning records. Model Iteration Unit 502 adopts a federated learning framework, which combines anonymized data from multiple repositories to jointly train the LSTM-attention prediction algorithm, avoiding data privacy leaks and improving the model's generalization ability. The remote monitoring unit 503 supports access via Web and APP, and provides functions such as real-time parameter viewing, historical data query, early warning record export, and execution unit control. Data visualization unit 504 displays trends in environmental parameters, distribution of physiological states of agricultural products, and statistics on early warning levels, and supports the generation of custom reports.

[0034] The working principle, and the specific steps for monitoring and early warning in this embodiment, are as follows: After system initialization, environmental and physiological parameters are collected through sensor node 101. The data is preprocessed, and the core algorithm is run through edge computing module 2. A dynamic threshold is generated using a multi-parameter coupled dynamic threshold algorithm, and a risk level is output using an LSTM-attention prediction algorithm. Then, an early warning judgment is made. If the parameters exceed the dynamic threshold and risk level, an early warning is triggered, and the warning information is pushed to the user through remote monitoring unit 503. At the same time, abnormal parameters and corresponding areas are highlighted in data visualization unit 504. Users can view detailed warning content through the web or APP, retrieve historical data for comparative analysis, and remotely control execution control module 3 to implement environmental adjustments. Parameters are fed back after 15 minutes. If there are no abnormalities, the system repeats the above steps for continuous monitoring.

[0035] The edge computing module disclosed in this embodiment adopts an industrial-grade edge gateway, configured with an ARM Cortex-A53 CPU (4 cores, 1.5GHz), 2GB of memory, 16GB of storage, supports local data caching with a cache capacity of ≥10GB, retains data for 72 hours after power failure, and has electromagnetic interference protection, dustproof, and moisture-proof functions. Its operating temperature range is -40℃ to 85℃, and it meets the industrial-grade IP65 protection standard. It also features a variable frequency centrifugal fan with an air volume of 1000-5000 m³ / h. 3 / h, power 0.75-3kW, supports timed ventilation and linked ventilation; rotary dehumidifier, dehumidification capacity 5-20kg / h, humidity control accuracy ±2%RH, linked with ventilation unit to avoid humidity fluctuations; variable frequency compressor refrigeration system, cooling capacity 5-30kW, temperature control accuracy ±0.5℃, multi-head linkage ensures uniform cooling; nitrogen generator: nitrogen purity ≥99.5%, carbon dioxide generator: output concentration 0-10%, ethylene adsorber: adsorption efficiency ≥90%; ultrasonic insect repellent, frequency 20-80kHz adjustable, but not limited to the above models.

Claims

1. An intelligent monitoring and early warning warehouse for agricultural products, characterized in that, The application relates to a warehouse environment and physiological state parameter monitoring and control system. The warehouse environment and physiological state parameter monitoring and control system comprises a sensor network module (1) which collects agricultural product storage environment parameters and agricultural product physiological state parameters in a warehouse; an edge computing module (2) which is arranged in the warehouse and is in wireless communication connection with the sensor network module (1), performs real-time preprocessing on the collected parameters, and runs a multi-parameter coupling dynamic threshold optimization algorithm, an LSTM-attention mildew risk prediction algorithm and a distributed collaborative control algorithm; an execution control module (3) which comprises a ventilation unit (301), a dehumidification unit (302), a refrigeration unit (303), an air conditioning unit (304) and a pest control unit (305) and is in communication connection with the edge computing module (2); an early warning module (4) which comprises a local sound-light early warning unit (401) and a remote pushing unit (402) and is in communication connection with the edge computing module (2); a cloud platform module (5) which is in communication with the edge computing module (2) through one of 5G or Ethernet and is used for parameter storage, algorithm model iteration, remote monitoring and data visualization; and a fault self-diagnosis module (6) which is in communication connection with the sensor network module (1), the edge computing module (2) and the execution control module (3). The environment parameters comprise temperature, humidity, oxygen concentration, carbon dioxide concentration, ethylene concentration and dust concentration, the agricultural product physiological state parameters comprise surface humidity, hardness and color difference, one sensor node (101) is arranged every 50-100 square meters, each sensor node (101) is integrated with an environment parameter sensor group and a physiological state sensor group, the environment parameter sensor group comprises a temperature sensor, a humidity sensor, a gas sensor and a dust sensor, wherein the gas sensor comprises an oxygen concentration sensor, a carbon dioxide concentration sensor and an ethylene concentration sensor; the agricultural product physiological state sensor group comprises a surface humidity sensor, a hardness sensor and a color difference sensor; and the sensor node (101) adopts a LoRa wireless communication protocol and supports automatic wake-up and sleep switching. The edge computing module (2) generates self-adaptive dynamic thresholds through the multi-parameter coupling dynamic threshold optimization algorithm, predicts the agricultural product mildew risk 24-72 hours in advance through the LSTM-attention mildew risk prediction algorithm, and realizes linkage and collaborative control of each execution unit through the distributed collaborative control algorithm. The units of the execution control module (3) are specifically configured as follows: the ventilation unit (301) adopts a variable frequency centrifugal fan; the dehumidification unit (302) adopts a rotary dehumidifier; the refrigeration unit (303) adopts a variable frequency compressor refrigeration system; the air conditioning unit (304) comprises a nitrogen generator, a carbon dioxide generator and an ethylene adsorber; and the pest control unit (305) adopts an ultrasonic wave insect expeller in cooperation with a low-toxicity trapping device. ​ ​ 2.The intelligent monitoring and early warning warehouse for agricultural products according to claim 1, characterized in that, ​ 3.The intelligent monitoring and early warning warehouse for agricultural products according to claim 2, characterized in that, ​ 4. The intelligent monitoring and early warning warehouse for agricultural products according to claim 1, characterized in that, ​ 5. The intelligent monitoring and early warning warehouse for agricultural products according to claim 1, characterized in that, The logical steps of the multi-parameter coupled dynamic threshold optimization algorithm include: S1, data preprocessing: removing outliers, filling missing values and standardizing the environmental parameters and physiological state parameters collected by the sensor; S2, agricultural product category feature matching: according to the category and maturity grade of the stored agricultural products, the corresponding basic threshold data set is called from the cloud platform module (5); S3, multi-parameter coupling weight calculation: the influence weight of each parameter on the storage quality of agricultural products is determined by the analytic hierarchy process, a judgment matrix is constructed, and the weight vector is obtained after consistency test; S4, dynamic threshold iterative update: based on the real-time collected standardized parameter vector, the dynamic threshold correction coefficient is calculated combined with the weight vector, and the dynamic threshold vector is generated; S5, threshold validity verification: the deviation rate of dynamic threshold from actual parameters is calculated every hour, if the deviation rate exceeds 5% for 3 times in a row, steps S3~S4 are re-executed to adjust the weight vector and correction coefficient.

6. The intelligent monitoring and early warning warehouse for agricultural products of claim 1, wherein, The logical steps of the LSTM-attention mold risk prediction algorithm include: S1, data set construction: based on the historical data collected by the sensor network module (1), including environmental parameter sequence, physiological state parameter sequence and corresponding mold result label, a training data set is constructed, wherein the time window length of environmental parameter sequence and physiological state parameter sequence is 24 hours, and the sampling interval is 1 hour; S2, feature extraction: time domain feature extraction and frequency domain feature extraction are performed on the parameter sequence to construct a feature matrix; S3, LSTM feature encoding: the feature matrix is input into the LSTM network, the LSTM network contains 3 layers of hidden layers, the number of hidden units in each layer is 128, the activation function is tanh, the activation functions of the forget gate, input gate and output gate are sigmoid, and the hidden state is calculated; S4, Attention mechanism weighting: introducing self-attention mechanism to the hidden state sequence output by LSTM; S5, risk level prediction: the weighted hidden state is input into the full connection layer, the full connection layer contains 2 layers, the output dimension of the first layer is 64, the activation function is ReLU, the output dimension of the second layer is 3 (corresponding to 3 mold risk levels), and the activation function is softmax, the risk level probability distribution is calculated; S6, model iterative optimization: the prediction error is calculated by using cross-entropy loss function, the network parameters are updated by Adam optimizer, and the model is fine-tuned once every 7 days using the historical data of the cloud platform module (5).

7. The intelligent monitoring and early warning warehouse for agricultural products of claim 1, wherein, The logical steps of the distributed collaborative control algorithm include: S1, control target determination: according to the risk level output by the LSTM-attention prediction algorithm and the multi-parameter coupled dynamic threshold, the control target vector is determined, wherein each target parameter is the midpoint of the optimal interval of the dynamic threshold; S2, execution unit state evaluation: the current running state of each execution unit is collected to construct a state matrix, wherein there are 5 execution units (ventilation, dehumidification, refrigeration, gas regulation and pest control), and 3 state dimensions (running state, power and fault state); S3, control priority ranking: based on the deviation of each parameter from the target value and the storage demand of agricultural products, the control priority is determined; S4, cooperative control instruction generation: according to the control priority, the control task is allocated to each execution unit, and the specific logic is as follows: If the temperature deviation is the largest: preferentially start the refrigeration unit (303), and adjust the dehumidification unit (302) according to the humidity parameter (if the humidity is higher than the target value, the dehumidification unit (302) and the refrigeration unit (303) are linked to run; if the humidity is lower than the target value, the refrigeration unit (303) runs alone, and the power of the dehumidification unit is reduced); If the humidity deviation is the largest: preferentially start the dehumidification unit (302), and control the running time of the ventilation unit (301); If the gas concentration deviation is the largest: start the gas regulation unit (304) to adjust the proportion of oxygen and carbon dioxide, and close the ventilation unit (301); The pest control unit (305) runs independently, and starts during the non-high-priority control execution period according to the dust concentration and pest monitoring data; S5, control effect feedback and adjustment: collect parameters every 15 minutes, calculate the deviation of the parameters from the target value, if the deviation is not reduced, adjust the running parameters of the execution unit, and re-calculate the control priority, and iteratively optimize the control instruction. 8.The intelligent monitoring and early warning warehouse for agricultural products of claim 1, wherein, The warning level of the warning module (4) is divided into three levels, corresponding to different warning methods: First level warning (low risk): the parameters of agricultural product storage environment approach the upper / lower limit of dynamic threshold, the mold risk prediction probability is ≤30%, the local sound and light warning unit (401) is triggered, and the remote push unit (402) sends a text reminder to the administrator APP; Second level warning (medium risk): the parameters exceed the dynamic threshold ±5%, the mold risk prediction probability is >30% but ≤60%, the local sound and light warning unit is triggered, the remote push unit (402) sends a text and voice reminder to the administrator APP and WeChat, and calls the administrator; Third level warning (high risk): the parameters exceed the dynamic threshold ±10%, the mold risk prediction probability is >60%, the local sound and light warning unit (401) is triggered, the remote push unit (402) pushes warning information to the administrator, warehouse manager and enterprise management at multiple levels, and links the fire fighting system. 9.The intelligent monitoring and early warning warehouse for agricultural products according to claim 1, characterized in that, The cloud platform module (5) is connected to the user terminal through one of Web or APP, and the cloud platform module (5) comprises: A data storage unit (501) which adopts a distributed database to store raw data collected by sensors, preprocessed data, algorithm model parameters and warning records; A model iteration unit (502) which adopts a federal learning framework to jointly train the LSTM-attention prediction algorithm combined with the anonymized data of multiple warehouses, avoids data privacy leakage, and improves the model generalization ability; A remote monitoring unit (503) which supports Web and APP access, provides real-time parameter viewing, historical data query, warning record export and execution unit control functions; The data visualization unit (504) displays environmental parameter change trend, agricultural product physiological state distribution, early warning level statistics, and supports custom report generation.

10. The intelligent monitoring and early warning warehouse for agricultural products of claim 3, wherein, The fault self-diagnosis module (6) periodically detects the communication state and battery power of the sensor node (101), triggers a sensor fault warning when communication fails for three consecutive times or the battery power is less than 10%, detects the CPU utilization and memory occupancy of the edge computing module (2), automatically closes non-core processes when the CPU utilization exceeds 80% for one hour or the memory occupancy exceeds 90%, detects the running current and fault code of the execution unit, triggers an execution unit fault warning when the current is abnormal or the fault code is not 0, and records fault information.

Images