Grain storage process control and insect and mildew early warning ventilation regulation system

Abstract

The application discloses a kind of grain storage process control and insect mould early warning ventilation regulation system, and the application relates to the technical field of intelligent management of grain storage, including perception layer, edge computing layer, cloud digital twin and decision layer and execution control layer, the perception layer adopts the multimodal integrated probe tube array of temperature measurement, humidity, NDIR CO2, MOS VOC array and acoustic emission in one, edge computing layer utilizes independent component analysis algorithm and decouples CO2 composite signal into three independent components of grain itself breathing, mould metabolism and pest breathing, combined with multi-source feature cascade network realizes insect mould accurate classification early warning and three-dimensional space quantitative estimation, cloud layer is based on the anisotropy porosity field of dynamic reconstruction, to build the digital twin model of granary by physical information neural network, fuse 72 hours fine weather forecast and play ventilation dew risk, through multi-objective reinforcement learning optimization ventilation scheme, execution control layer is realized closed-loop control by PLC driven variable frequency fan and electric air valve.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology for grain storage, specifically a grain storage process control and pest and mold early warning ventilation regulation system. Background Technology

[0002] During grain storage, pests and mold (hereinafter referred to as pests and mold) are the two main threats that cause grain loss. Early and accurate warning and precise intervention are the core technical requirements for green grain storage.

[0003] Existing multi-parameter fusion grain condition monitoring and automatic ventilation control systems typically consist of distributed temperature measuring cables, single-point or multi-point CO2 sensors, a grain condition monitoring main control server, and an alarm module with manually set thresholds. The workflow is as follows: the temperature measuring cables periodically collect three-dimensional temperature data of the entire grain pile and upload it to the server. When the CO2 concentration exceeds the preset threshold, the system issues an alarm. The operator reviews the temperature distribution map and CO2 concentration curve, and, based on experience, judges whether there is a risk of insects and mold. After the operator decides to start ventilation, the fan runs at a fixed frequency until the temperature and humidity inside the warehouse reach the target value.

[0004] The above solution has three fundamental flaws:

[0005] First, the early warning timing is severely delayed and the false alarm rate is high. Grain has extremely low thermal conductivity (about 0.1 to 0.2 W / m·K), and the temperature sensors inside the grain pile are slow to respond to local heat sources inside the grain pile. By the time the temperature rise signal is transmitted to the sensor, pests or mold have often developed to the middle or late stages, missing the early intervention window. At the same time, the increase in CO2 concentration in the grain pile is caused by the combined effects of three sources: grain respiration, mold reproduction and metabolism, and feeding by hidden stored grain pests. The existing system alarms based on a single CO2 concentration exceeding the threshold, and cannot distinguish the contribution of the above three sources, resulting in a large number of invalid false alarms, and the risk of missing real dangers increases accordingly.

[0006] Second, it is impossible to achieve spatial localization and quantitative assessment of insect and mold lesions. The existing temperature monitoring system can only sense macroscopic areas of temperature anomalies and cannot accurately provide three-dimensional spatial quantitative information on insect density distribution or latent mold development, resulting in a lack of precise targeting of intervention measures.

[0007] Third, blind ventilation can lead to condensation, which in turn exacerbates the risk of insects and mold. Traditional ventilation control models treat grain piles as homogeneous porous media, ignoring the anisotropic porosity distribution caused by factors such as grain grading and impurity accumulation. This results in severely uneven distribution of ventilation airflow inside the grain pile, creating ventilation dead zones in the core area. The temperature and humidity in these areas cannot be effectively regulated, making them hotspots for insects and mold growth. At the same time, existing systems lack meteorological forecasting capabilities. Blind ventilation when external weather conditions are unfavorable can cause localized condensation inside the grain pile, significantly increasing local water activity and providing suitable breeding conditions for high-risk molds and hidden pests. Summary of the Invention

[0008] The technical problems to be solved by this invention are: how to accurately decouple insect and mold signals from the complex background of grain respiration before the temperature of the grain pile changes imperceptibly, so as to achieve early classification and early warning and three-dimensional spatial quantitative estimation; and how to construct a digital twin model that considers the heterogeneous characteristics of anisotropic porosity of the grain pile, integrates fine weather forecasts to actively avoid the risk of condensation, and achieves precise targeted ventilation control.

[0009] This invention provides a grain storage process control and pest and mold early warning ventilation regulation system, including a sensing layer, an edge computing layer, a cloud digital twin and decision layer, and an execution control layer, forming a complete closed-loop control system of sensing-decoupling early warning-twin optimization-adaptive execution.

[0010] The perception layer includes a multimodal integrated probe array arranged in a three-dimensional grid inside the grain warehouse and grain pile. Each probe integrates a temperature measurement unit, a humidity measurement unit, a CO2 detection unit, a VOC detection unit, and an acoustic emission detection unit. The collected data is uploaded to the edge computing layer through a wireless communication module.

[0011] The edge computing layer includes an edge computing gateway, which is responsible for performing data preprocessing, independent component analysis (ICA) decoupling of multi-source signals, and primary feature extraction on multimodal data. It decomposes the CO2 composite signal into three independent components: grain respiration, mold metabolism, and pest respiration. It uses the mutual information of each component with temperature gradient features, VOC normalized feature vector, and acoustic emission feature vector to perform source labeling and output the comprehensive feature vector of each probe.

[0012] The cloud-based digital twin and decision-making layer includes a multi-source feature cascaded network module, an anisotropic porosity digital twin modeling module, a condensation risk prediction module, and a multi-objective reinforcement learning ventilation optimization module. The multi-source feature cascaded network module, through a cascaded fusion architecture of a VOC fingerprint recognition sub-network, an acoustic emission pattern recognition sub-network, a CO2 decoupling component processing sub-network, and a fusion decision network, outputs insect and mold classification early warning results and quantitative estimates of pest density and mold incubation period. It also uses Kriging interpolation to generate a three-dimensional distribution cloud map of the entire warehouse. The anisotropic porosity digital twin modeling module reconstructs the three-dimensional anisotropic porosity field online and uses it as physical constraint parameters to construct a digital twin model of the grain warehouse airflow field and heat and humidity coupling transfer based on a Physical Information Neural Network (PINN). The condensation risk prediction module integrates the detailed weather forecast for the next 72 hours, calls the PINN digital twin model to quickly predict candidate ventilation schemes, and evaluates condensation risk and ventilation effect. The multi-objective reinforcement learning ventilation optimization module adopts a near-end strategy optimization algorithm, uses the PINN digital twin model as the simulation environment, and outputs the optimized ventilation scheme.

[0013] The execution control layer includes a PLC controller, a variable frequency fan driver, and electric dampers for zoned air ducts. The PLC controller adjusts the damper opening and fan frequency according to the optimized ventilation scheme and feeds back the execution status to the edge computing layer in real time, forming a closed loop.

[0014] The present invention has the following beneficial effects:

[0015] 1. This invention uses sound-gas-heat multimodal signal fusion and ICA decoupling to advance the early warning time of insects and molds to several days or even weeks in advance of traditional temperature monitoring methods, significantly reducing the false alarm rate of insect and mold classification early warning and eliminating the high false alarm problem of CO2 single parameter monitoring.

[0016] 2. This invention achieves accurate quantitative estimation and visual positioning of the density of hidden pests and the incubation period of mold in the three-dimensional space of grain warehouses through multi-source feature cascade networks and Kriging interpolation, providing precise target coordinates for intervention decisions.

[0017] 3. This invention constructs a digital twin model of a grain silo with anisotropic porosity to achieve precise targeted ventilation compensation for the heterogeneous characteristics of grain piles, eliminating ventilation dead zones that cannot be eliminated under the traditional homogeneous assumption.

[0018] 4. This invention fundamentally solves the problem of local condensation caused by blind ventilation by integrating 72-hour detailed weather forecasts with anti-condensation ventilation path simulation and multi-objective reinforcement learning adaptive optimization, and actively avoids creating favorable conditions for the outbreak of insects and mold. Attached Figure Description

[0019] Figure 1 This is a diagram of the overall four-layer architecture of the system of this invention;

[0020] Figure 2 A schematic diagram of the internal structure of the multimodal integrated probe and its three-dimensional mesh array layout scheme in a grain warehouse;

[0021] Figure 3 Flowchart of the ICA multi-source signal decoupling algorithm for the edge computing layer;

[0022] Figure 4 A schematic diagram of a cloud-based multi-source feature cascaded network structure and a three-dimensional early warning cloud map of grain warehouse pest density.

[0023] Figure 5 Flowchart for optimization control of anti-condensation ventilation based on anisotropic porosity digital twin;

[0024] Figure 6 This is a schematic diagram of the complete information flow and closed-loop control of the system. Detailed Implementation

[0025] The following is in conjunction with the appendix Figure 1-6The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0026] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0027] I. System Overall Architecture

[0028] Reference Figure 1 The grain storage process control and pest and mold early warning ventilation control system provided by the present invention consists of four layers: perception layer, edge computing layer, cloud digital twin and decision layer, and execution control layer, forming a complete perception-decoupling early warning-twin optimization-adaptive execution closed-loop control system.

[0029] II. Sensing Layer: Multimodal Integrated Probe Array

[0030] Reference Figure 2 The core hardware of the sensing layer is a multimodal integrated probe. This probe uses existing temperature measuring cables as a carrier to extend its functions, and highly integrates the following sensing units inside the pressure-resistant and sealed probe to achieve in-situ non-destructive sensing.

[0031] The temperature measurement unit uses a PT100 platinum resistance sensor to measure the local temperature of the grain pile with an accuracy of ±0.1℃ and a sampling interval of 60 seconds.

[0032] The humidity measurement unit uses a capacitive humidity sensor to measure the local relative humidity of the grain pile with an accuracy of ±1.5%RH.

[0033] The CO2 detection unit uses an NDIR nondispersive infrared CO2 sensor with a range of 0 to 5% volume fraction, an accuracy of ±50ppm, and a response time of no more than 30 seconds. Its working principle is as follows: CO2 molecules have characteristic selective absorption of 4.26 μm wavelength infrared light, and the CO2 concentration is calculated by measuring the attenuation of transmitted light intensity.

[0034] The VOC detection unit uses a MOS sensor array, which includes at least 6 cross-sensitive sensors. Each sensor has a differentiated response to characteristic gases of mold metabolism (including 1-octen-3-ol) and specific pheromones of lepidopteran and coleopteran pests. Together, they form an electronic nose fingerprint feature vector. The MOS sensor uses the principle that the resistance changes after different gas molecules are adsorbed on the surface of metal oxides (including SnO2 and ZnO) to achieve gas detection.

[0035] The acoustic emission detection unit uses a piezoelectric broadband acoustic emission sensor with a detection frequency band of 20 kHz to 300 kHz. It has a built-in preamplifier with a gain of 40 dB and a sampling rate of no less than 1 MHz. Grain storage pests (including corn weevils and grain weevils) can generate characteristic acoustic emission signals in this frequency band during feeding and movement, making it feasible as an early detection method for pests.

[0036] The pressure measurement unit uses a MEMS pressure sensor to assist in the calculation of CO2 concentration temperature and pressure compensation.

[0037] The probe uses a LoRa wireless communication module to upload the collected data to the edge computing gateway. It has an IP68 protection rating and can work stably for a long time in the hot and humid environment of the grain pile. In the tall flat warehouse, the probe array is deployed in a three-dimensional grid pattern with a horizontal spacing of no more than 5 m and 3 to 5 measurement nodes in the vertical direction to ensure that at least one probe covers any area within a 2 m diameter of the core area inside the grain warehouse.

[0038] III. Edge Computing Layer: Data Preprocessing and ICA Decoupling

[0039] Reference Figure 3 The edge computing gateway performs the following processing flow on the raw multimodal data uploaded by the probe array.

[0040] Data preprocessing

[0041] The edge computing gateway receives the temperature data uploaded by each probe. relative humidity CO2 concentration VOC sensor array response vector (in For the number of VOC sensor channels, and acoustic emission time-domain waveforms subscript Number the probe tube. ( (Total number of probes), the data from each probe are aligned according to a unified timestamp to form a synchronous observation matrix.

[0042] Acoustic emission time-domain waveforms for each probe Perform a Fast Fourier Transform to extract the frequency domain power spectrum, and simultaneously calculate the effective value (RMS), peak value, kurtosis, and root mean square frequency, among other time-domain statistical characteristics. Based on the pre-calibrated acoustic emission characteristic frequency band of the pest (the feeding characteristic frequency of the grain weevil is concentrated between 25 and 80 kHz), calculate the proportion of bandpass energy in this frequency band. Forming acoustic emission feature vectors .

[0043] Temperature measured using a probe and air pressure Temperature and pressure compensation was applied to the NDIR sensor measurements to obtain the CO2 volume concentration under standard conditions. :

[0044] ;

[0045] in, This is the raw measurement value (volume fraction) from the NDIR sensor. The current air pressure is expressed in Pa. Standard atmospheric pressure, taken as 101325 Pa. The current temperature is in °C. The standard reference temperature is 0℃. This represents the volume fraction of CO2 after temperature and pressure compensation.

[0046] Baseline subtraction and normalization are performed on the response values ​​of each channel of the MOS sensor array to eliminate the influence of individual sensor differences and environmental drift, resulting in a normalized VOC feature vector. .

[0047] ICA Multi-Source Signal Decoupling

[0048] The CO2 concentration signal in the grain pile is a mixed signal from three sources: grain respiration, fungal metabolism, and pest respiration. (Assuming the entire warehouse...) The CO2 concentration vector after temperature and pressure compensation of each probe constitutes the observation vector. The ICA decoupling steps are as follows.

[0049] Centralization: Calculate the historical mean of the observed vectors ( The total number of historical time steps. For the first (the observation vector at each time step), let Eliminate DC bias of the signal.

[0050] Whitening: Calculate the covariance matrix of the centered observation data. ,right Eigenvalue decomposition yields Calculate the whitening matrix Obtain whitening data ,in The eigenvector matrix, Let be an eigenvalue diagonal matrix such that the components are uncorrelated and have a variance of 1.

[0051] FastICA Iterative Solution: The FastICA algorithm is used, with the maximization of negative entropy as the independence criterion, to separate the vectors corresponding to the three CO2 source components. Perform iterative updates, with the following rules:

[0052] ;

[0053] in For nonlinear functions, take , Its derivative, after each update, is sequentially applied to... Perform orthogonalization ( ) and normalization ) operation, when The difference between the result of the previous iteration and the result of the previous iteration is less than the convergence threshold. Stop iterating when the time comes.

[0054] Source annotation: Constructing a separation matrix Calculate independent components Three decoupling components were obtained. , , By utilizing the mutual information between each component and temperature gradient characteristics (related to grain respiration), VOC fingerprint characteristics (related to mold metabolism), and acoustic emission characteristics (related to pests), the three components are labeled as grain respiration CO2 components. CO2 content from mold metabolism and the amount of CO2 respiration of pests .

[0055] Integrated feature vector output: Decouples the three CO2 components of each probe. VOC normalized feature vector Acoustic emission eigenvectors Temperature gradient between adjacent probes and humidity gradient Together they form a comprehensive feature vector It is then uploaded to the cloud processing module.

[0056] IV. Cloud Layer: Multi-source Feature Cascade Network Module for Insecticidal and Mold Classification and Quantification

[0057] Reference Figure 4 The multi-source feature cascaded network adopts a cascaded fusion architecture, consisting of three feature extraction sub-networks and a fusion decision network.

[0058] VOC fingerprint recognition subnetwork (subnetwork A): Input normalized VOC sensor array response vector After passing through a 3-layer fully connected network (with hidden layer dimensions of 64, 32, and 16 respectively), the output VOC feature embedding vector is obtained. .

[0059] Acoustic emission pattern recognition subnetwork (subnetwork B): Input acoustic emission feature vector After passing through a one-dimensional convolutional layer (used to capture local patterns in the frequency domain) and two fully connected layers, the output acoustic emission feature embedding vector is obtained. .

[0060] CO2 Decoupling Component Processing Subnetwork (Subnetwork C): Inputs the three CO2 components after ICA decoupling and their temporal change rates (using nearly 30 minutes of time-series data), extracts temporal features through a Long Short-Term Memory (LSTM, 32 hidden layer dimensions), and outputs a CO2 dynamic feature embedding vector. .

[0061] Fusion Decision Network: The embedding vectors of the three sub-networks are concatenated to obtain the fusion vector. (in The algorithm first concatenates vectors, then adds two fully connected layers (32 and 16 dimensions respectively), and finally outputs a classification branch and a regression branch in parallel. The classification branch uses Softmax output, and the categories include four types: normal, pest warning, mold warning, and combined pest and mold warning, outputting the probability of each category. The regression branch uses linear output, outputting the estimated pest density at the current probe location. (Unit: heads / kg) and estimated mold incubation period (Unit: days).

[0062] During training, the classification cross-entropy loss is used. With regression mean square error loss Joint loss function of weighted combination:

[0063] ;

[0064] in , These are the weighting coefficients. , One-hot encoding for the true category label. To predict probabilities, , This represents the actual pest density. The value represents the actual incubation period of mold growth.

[0065] The Adam optimizer was used with an initial learning rate of 0.001 and a batch size of 32. Training was stopped when the loss on the validation set decreased by less than 0.1% for 20 consecutive training epochs.

[0066] In a controlled experimental chamber, common stored grain pests such as grain weevils and corn weevils, as well as common stored grain molds such as Aspergillus flavus and Penicillium, were inoculated at different densities. Multimodal signals from multiple probes were collected under different temperature and humidity conditions. The density of pests and the degree of mold were recorded simultaneously by manual microscopic examination. A labeled training dataset was constructed, and a special calibration dataset was constructed for different grain varieties. For new grain varieties or new grain storage environments, a transfer learning strategy was adopted to freeze the weights of the bottom feature extraction layers of subnetworks A, B, and C, and only the fusion decision network was fine-tuned to quickly adapt to new scenarios with a small amount of newly calibrated data.

[0067] 3D spatial early warning map generation: The classification and early warning results and quantitative estimation results of all probes in the whole warehouse are mapped to the 3D coordinate system of the probes. Kriging interpolation is used to perform spatial interpolation on non-probe locations to generate a 3D distribution cloud map of pest density and a 3D distribution cloud map of mold incubation period for the whole warehouse, so as to achieve precise visualization and location of lesions.

[0068] V. Cloud Layer: Anti-condensation ventilation optimization module based on anisotropic porosity digital twin

[0069] Reference Figure 5 This module includes four sub-modules: online reconstruction of anisotropic porosity field, PINN digital twin modeling, ventilation path simulation and condensation risk assessment, and multi-objective reinforcement learning ventilation adaptive optimization.

[0070] Online reconstruction of anisotropic porosity field

[0071] When grain is stored in a warehouse, due to the grading and movement of the grains, impurities and fine particles accumulate towards the core of the warehouse, forming a dense area with extremely low porosity (approximately 0.25 to 0.30), while the porosity of the area around the warehouse is higher (approximately 0.40 to 0.45). This invention reconstructs a three-dimensional anisotropic porosity field online based on the following information. ( (in three-dimensional space coordinate vectors)

[0072] Records of grain entry methods (historical records of grain entry at the center and multiple entry points), grain variety and impurity content, historical data on storage sedimentation (using probe depth benchmarks to monitor grain surface sedimentation), inversion calibration of CO2 static diffusion rate in each area (by tracking the concentration response of each probe after injecting a known amount of CO2, and using Tikhonov regularization to invert local porosity), and the porosity field is triggered to update each time new grain is entered into the warehouse, and also triggered to update when significant grain surface sedimentation is detected.

[0073] PINN Physical Information Neural Network Digital Twin Modeling

[0074] With anisotropic porosity field Using physical constraint parameters, a three-dimensional airflow field and heat-humidity coupling digital twin model of a grain silo is constructed based on PINN. The PINN network uses three-dimensional spatial coordinates. and time As input, output the airflow velocity vector at that spatiotemporal point. ,temperature and relative humidity .

[0075] PINN's total loss function consists of the following three parts:

[0076] ;

[0077] in, This refers to the data fitting loss, which is the mean square error between the network output and the measured temperature and humidity data of each probe. The residual loss in the physical equations includes the residuals of the Darcy-Brinkman extended equations (airflow control equations) for porous media with anisotropic porosity fields as parameters, and the residuals of the energy equations (heat conduction-convection coupling) for porous media. Taking the continuity equation for porous media as an example, its form in anisotropic porosity fields is: The equation requires that the divergence residual calculated from the PINN output at the placement point approaches zero, where For local porosity, The airflow velocity vector To account for boundary condition losses, the wind turbine inlet velocity boundary condition and the silo wall boundary condition are incorporated as soft constraints into the loss function. , The weights for each loss are dynamically adjusted during training using an adaptive weighting strategy based on gradient magnitude. After offline pre-training in the cloud, PINN is fine-tuned online for the porosity field of a specific grain warehouse. The fine-tuning can converge in just a few minutes, meeting real-time requirements.

[0078] Ventilation path rehearsal and condensation risk assessment

[0079] When the system triggers an insect and mold warning or determines that cooling and precipitation are needed, the ventilation pre-operation process is initiated.

[0080] Hourly temperatures for the next 72 hours at the location of the grain depot are obtained through a meteorological data interface. relative humidity and dew point temperature Forecast data.

[0081] Based on the opening degree of the air valves in each area of ​​the grain warehouse (discretized into five levels: 0%, 25%, 50%, 75%, and 100%) and the frequency of the variable frequency fan (discretized into four levels: 25Hz, 35Hz, 45Hz, and 50Hz), data is generated through combined sampling. A set of candidate ventilation schemes, each scheme containing the opening vector of all air ducts and valves in the entire warehouse. and the frequency vector of each wind turbine .

[0082] For each candidate ventilation scheme, the boundary conditions corresponding to the damper opening and fan frequency are input into the PINN digital twin model to quickly infer the three-dimensional airflow field of the entire chamber over the next few hours. Temperature field and humidity field The evolution process can be completed in less than 30 seconds per inference, which is much faster than the hours required for traditional computational fluid dynamics (CFD) calculations.

[0083] For each candidate scheme, the condensation risk at each spatial point in the entire warehouse is calculated step by step. A condensation risk is determined to exist when a certain spatiotemporal point meets the following conditions:

[0084] The temperature at this point in space after the introduction of external airflow Lower than the dew point temperature of the external airflow at that moment ,Right now Statistics on the percentage of total open interest risk (Ratio of condensation risk spatial points to total spatial points) to generate a condensation risk cloud map.

[0085] Calculate the comprehensive ventilation effectiveness index for each candidate scheme, including: the average temperature reduction in the target high-risk area. Average rainfall across the entire warehouse Percentage of ventilation dead zones (Percentage of areas with local airflow velocities below 0.01 m / s) and energy consumption estimates (Calculated based on the wind turbine power curve).

[0086] Multi-objective reinforcement learning for adaptive optimization of ventilation

[0087] This invention employs a multi-objective reinforcement learning algorithm to adaptively adjust ventilation strategies online, achieving multi-objective balance optimization.

[0088] state space Definition: State Includes the current comprehensive feature vector of the entire warehouse (temperature and humidity of each probe, decoupled CO2 component, and probability of insect and mold warning), external weather forecast summary (statistics of temperature and humidity forecast for the next 72 hours), and current valve opening vectors. and wind turbine frequency vector .

[0089] Action space Definition: Action To adjust the opening degree of the dampers and the frequency of the fans in each area, a continuous action space design is adopted.

[0090] reward function Definition: Taking multiple objectives into account, the reward function is defined as follows:

[0091] ;

[0092] in, For high-risk areas in the first The average temperature drop over time, For the full position in the first Average precipitation range per hour For the first The percentage of risk volume associated with the full position closing at the end of the trading day. For the first Percentage of dead ventilation area volume during the day For the first Estimated energy consumption of step fan , To promote a positive reward weighting coefficient for both cooling and precipitation effects, To penalize the weighting coefficient of condensation risk, To penalize the weighting factor of ventilation dead zones, The weighting coefficients are used to penalize energy consumption. Each weight can be adjusted by the warehouse keeper on the human-machine interface according to the warehouse management priority.

[0093] The Proximal Policy Optimization (PPO) algorithm is used as the backbone of the reinforcement learning algorithm. Both the policy network and the value network adopt a 4-layer fully connected network with a hidden layer dimension of 256. The PINN digital twin model serves as the simulation environment for reinforcement learning, supporting the policy network to conduct a large number of virtual trial and error training in the digital twin without the need for exploration in the real grain warehouse. As feedback data from the real warehouse accumulates, the policy is continuously updated online with actual data.

[0094] After optimization, the system sorts the ventilation schemes according to the expected value of comprehensive rewards, selects the top-ranked scheme, and sends it to the PLC controller through cloud-edge collaboration to achieve closed-loop control.

[0095] VI. Execution Control Layer: PLC Closed-Loop Control

[0096] Reference Figure 6 After receiving the ventilation plan from the cloud, the PLC controller will adjust the target opening vector accordingly. Drive the electric dampers of each zone's air duct to the corresponding opening degree, with an opening degree adjustment accuracy better than 2%, according to the target frequency vector. The speed of each axial fan is adjusted by a frequency converter driver with a frequency adjustment resolution of 0.1Hz. The PLC monitors the fan current and the feedback signal of the air valve in real time. If a fault occurs (including air valve jamming or fan overload), it immediately reports to the edge gateway and switches to the safety redundancy control mode (frequency reduction protection operation or shutdown of the faulty branch). The execution status is fed back to the edge computing gateway every 60 seconds to update the boundary conditions of the digital twin model and form a complete closed-loop control loop.

[0097] VII. Alternative Implementation Methods

[0098] In the VOC detection unit, an optical sensor based on differential absorption spectroscopy (DOAS) technology can be used to replace the MOS-type sensor array. DOAS technology analyzes the characteristic absorption spectra of gas molecules in the ultraviolet-visible band to simultaneously and quantitatively detect multiple target VOC components in the warehouse. It does not require sensor baseline drift correction, has better selectivity and long-term stability, and is suitable for large-scale grain storage scenarios with higher accuracy requirements. The remaining signal processing, digital twin, and ventilation optimization modules are exactly the same as the above implementation method, only the method of obtaining VOC feature vectors is different.

[0099] In edge computing scenarios with limited computing resources, a lightweight method combining integrated particle filtering and extended Kalman filtering can be used to replace PINN digital twins for online estimation of airflow and thermal humidity fields in grain silos. This method uses a simplified finite element model as the state transition equation and measured data from each probe as the observation value. It updates the anisotropic porosity field and temperature and humidity field in real time through recursive Bayesian estimation. The computational load is much smaller than PINN, and it can run entirely locally on the edge computing gateway, making it suitable for promotion and application in small and medium-sized grain silos with limited resources.

[0100] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A grain storage process control and insect / mold early warning ventilation regulation system, characterized in that, It includes the perception layer, edge computing layer, cloud digital twin and decision-making layer, and execution control layer; The sensing layer includes a multimodal integrated probe array arranged in a three-dimensional grid inside the grain warehouse and grain pile. Each multimodal integrated probe integrates a temperature measurement unit, a humidity measurement unit, a CO2 detection unit, a VOC detection unit, and an acoustic emission detection unit, and uploads the collected data to the edge computing layer through a wireless communication module. The edge computing layer includes an edge computing gateway. The edge computing gateway performs data preprocessing, independent component analysis, multi-source signal decoupling, and primary feature extraction on the received multimodal data. It decomposes the CO2 composite signal into three independent components: the CO2 component from grain respiration, the CO2 component from mold metabolism, and the CO2 component from pest respiration. It then uses the mutual information between each independent component and temperature gradient features, VOC normalized feature vector, and acoustic emission feature vector to label the source and output the comprehensive feature vector of each probe. The cloud-based digital twin and decision-making layer includes a multi-source feature cascaded network module, an anisotropic porosity digital twin modeling module, a condensation risk prediction module, and a multi-objective reinforcement learning ventilation optimization module. The multi-source feature cascaded network module receives comprehensive feature vectors and outputs the three-dimensional spatial insect and mold classification early warning results of the grain warehouse, as well as quantitative estimates of pest density and mold incubation period. The anisotropic porosity digital twin modeling module uses the online reconstructed three-dimensional anisotropic porosity field as physical constraint parameters and constructs a digital twin model of the grain warehouse airflow field and heat and humidity coupling based on a physical information neural network. The condensation risk prediction module integrates the detailed weather forecast for the next 72 hours, calls the digital twin model to perform a prediction of candidate ventilation schemes, and evaluates the condensation risk and ventilation effect of each candidate ventilation scheme. The multi-objective reinforcement learning ventilation optimization module adopts a near-end strategy optimization algorithm, uses the digital twin model as the simulation environment, and outputs the optimized ventilation scheme. The execution control layer includes a PLC controller, a variable frequency fan driver, and electric dampers for the zoned air ducts. The PLC controller adjusts the damper opening and fan frequency according to the optimized ventilation scheme and feeds back the execution status to the edge computing layer in real time to update the boundary conditions of the digital twin model, forming a closed-loop control.

2. The system according to claim 1, characterized in that, In the multimodal integrated probe, the temperature measurement unit uses a PT100 platinum resistance sensor with an accuracy of ±0.1℃ and a sampling interval of 60 seconds; the humidity measurement unit uses a capacitive humidity sensor with an accuracy of ±1.5%RH; the CO2 detection unit uses an NDIR nondispersive infrared CO2 sensor with a range of 0 to 5% volume fraction and an accuracy of ±50 ppm; the VOC detection unit uses a sensor array containing at least 6 cross-sensitive MOS sensors, each sensor having differentiated response characteristics to mold metabolic characteristic gases and pest-specific pheromones; the acoustic emission detection unit uses a piezoelectric broadband acoustic emission sensor with a detection frequency band of 20 kHz to 300 kHz, a built-in preamplifier with a gain of 40dB, and a sampling rate of not less than 1 MHz; the multimodal integrated probe also integrates a MEMS barometric pressure sensor for assisting in CO2 concentration temperature and pressure compensation calculation; the protection level of the multimodal integrated probe is IP68; the transverse spacing of the probe array within the grain silo does not exceed 5 m, and 3 to 5 measurement nodes are set in the longitudinal direction.

3. The system according to claim 2, characterized in that, When performing data preprocessing, the edge computing gateway uses the temperature and pressure measured by the same probe to perform temperature and pressure compensation on the CO2 measurement value of the NDIR sensor. The compensated standard state CO2 volume concentration is arranged according to time steps to form an observation vector. When performing independent component analysis and multi-source signal decoupling, the observation vector is sequentially centered, whitened by principal component analysis, and solved by FastICA iteration. FastICA iteration uses the maximization of negative entropy as the independence criterion. Through iterative updates, orthogonalization, and normalization operations, it converges to the separation matrix, separating the CO2 observation vector into three statistically independent components. Using the mutual information between each independent component and the temperature gradient feature, the VOC normalized feature vector, and the acoustic emission feature vector, the three independent components are labeled as the CO2 component of grain respiration, the CO2 component of mold metabolism, and the CO2 component of pest respiration.

4. The system according to claim 1, characterized in that, The multi-source feature cascaded network module adopts a cascaded fusion architecture, including a VOC fingerprint recognition subnetwork, an acoustic emission pattern recognition subnetwork, a CO2 decoupling component processing subnetwork, and a fusion decision network. The VOC fingerprint recognition subnetwork takes a normalized VOC sensor array response vector as input, passes it through a 3-layer fully connected network, and outputs a VOC feature embedding vector. The acoustic emission pattern recognition subnetwork takes an acoustic emission feature vector as input, passes it through a one-dimensional convolutional layer and a fully connected layer, and outputs an acoustic emission feature embedding vector. The CO2 decoupling component processing subnetwork takes the three decoupled CO2 components after independent component analysis and their temporal change rates as input, extracts temporal features through a long short-term memory network, and outputs... The CO2 dynamic feature embedding vector is used in the fusion decision network. The feature embedding vectors of the three sub-networks are concatenated and then output in parallel through a fully connected layer to form classification and regression branches. The classification branch uses Softmax output and includes four categories: normal, pest warning, mold warning, and combined pest and mold warning. The regression branch uses linear output and outputs the estimated pest density and mold incubation period at the current probe location. The classification warning results and quantitative estimation results of all probes in the warehouse are mapped to the probe's three-dimensional coordinate system. Kriging interpolation is used to perform spatial interpolation on non-probe locations to generate a three-dimensional distribution cloud map of pest density and a three-dimensional distribution cloud map of mold incubation period for the entire warehouse.

5. The system according to claim 4, characterized in that, The multi-source feature cascaded network module adopts a transfer learning strategy for new grain varieties, freezing the weights of the bottom feature extraction layers of the VOC fingerprint recognition subnetwork, acoustic emission pattern recognition subnetwork, and CO2 decoupling component processing subnetwork, and only fine-tuning the fusion decision network to quickly adapt to new grain varieties with a small amount of newly calibrated data.

6. The system according to claim 1, characterized in that, The information used by the anisotropic porosity digital twin modeling module to reconstruct the three-dimensional anisotropic porosity site online includes: records of grain storage methods, archives of grain varieties and impurity content, historical data of storage sedimentation, and the static CO2 diffusion rate of each region calibrated by tracking the concentration response of each probe after injecting a known amount of CO2 into the grain silo and using the Tikhonov regularization inversion method. The anisotropic porosity field is updated each time new grain is added to the warehouse, and also when significant grain surface settlement is detected.

7. The system according to claim 6, characterized in that, The physical information neural network takes three-dimensional spatial coordinates and time as input and outputs the airflow velocity vector, temperature and relative humidity at the corresponding spatiotemporal point. The loss function of the physical information neural network includes three parts: data fitting loss, physical equation residual loss, and boundary condition loss. The physical equation residual loss includes the residuals of the Darcy-Brinkman extended equation of porous media with anisotropic porosity field as parameter and the residuals of the energy equation of porous media. The constraint that the divergence of the product of porosity and airflow velocity vector in the continuity equation of porous media is zero is included in the placement point loss. The weight coefficients of each loss component are dynamically adjusted during training using an adaptive weighting strategy based on gradient magnitude. After being pre-trained offline in the cloud, the physical information neural network is fine-tuned online for the porosity field of a specific grain warehouse.

8. The system according to claim 1, characterized in that, The dew risk prediction module obtains hourly temperature, relative humidity and dew point temperature forecast data for the next 72 hours at the location of the grain depot through a meteorological data interface. A set of candidate ventilation schemes is generated based on the discretized combination of damper opening and fan frequency. The damper opening is discretized into five levels: 0%, 25%, 50%, 75%, and 100%, and the fan frequency is discretized into four levels: 25 Hz, 35 Hz, 45 Hz, and 50 Hz. The boundary conditions corresponding to each candidate solution are input into the physical information neural network digital twin model for rapid reasoning, with a single reasoning time of no more than 30 seconds. The simulation results are evaluated hourly to determine whether each spatial point meets the condensation risk criterion. The condensation risk criterion is that the temperature of the spatial point is lower than the dew point temperature of the external airflow at that moment. The proportion of the total volume at risk of condensation and the proportion of the volume in the ventilation dead zone were statistically analyzed. The threshold for determining the ventilation dead zone was a local airflow velocity of less than 0.01 m / s.

9. The system according to claim 8, characterized in that, The reward function of the multi-objective reinforcement learning ventilation optimization module uses the average temperature drop in the high-risk area and the average precipitation in the whole warehouse as positive reward items, and the proportion of the risk volume of condensation in the whole warehouse, the proportion of the volume of ventilation dead corners, and the estimated energy consumption of the fan as penalty items. The items are weighted and summed using adjustable weight coefficients. Both the policy network and the value network adopt a 4-layer fully connected network with a hidden layer dimension of 256. The physical information neural network digital twin model serves as a simulation environment to support the virtual trial and error training of the policy network, and the policy is updated online after the accumulation of feedback data from the real warehouse. After the optimization is completed, the system sorts the ventilation schemes according to the expected value of the comprehensive reward, selects the top-ranked scheme, and sends it to the PLC controller through cloud-edge collaboration.

10. The system according to claim 1, characterized in that, The PLC controller drives the electric air valves of each zone's air duct according to the target opening vector in the optimized ventilation scheme, with an opening adjustment accuracy better than 2%. The speed of each fan is adjusted by a frequency converter driver according to the target frequency vector, with a frequency adjustment resolution of 0.1 Hz; Real-time monitoring of fan current and damper feedback signals; switching to safe redundancy control mode when a fault occurs. The execution status is fed back to the edge computing gateway every 60 seconds to update the boundary conditions of the digital twin model.

Images