A Distributed Detection Method and System for Multiple Grain Quality Indicators Based on Edge Collaboration

Abstract

This invention discloses a distributed detection method and system for multiple indicators of grain quality based on edge collaboration, belonging to the field of grain quality monitoring technology. The method includes: synchronously collecting multispectral response data of the target grain pile and microenvironmental parameters of the node locations; based on the microenvironmental parameters, performing real-time calibration of the multispectral response data and inputting it into a quality feature extraction model to obtain a multidimensional initial indicator set for each node; combining the microenvironmental parameters and spatial topological relationships of each node to analyze the spatial correlation of the mold risk index in the node network and calculate the local anomaly diffusion potential; based on the coupling relationship between the local anomaly diffusion potential and moisture content and protein content, performing spatiotemporal dynamic mold risk field simulation to generate a risk field distribution; and performing spatial data fusion and state optimization evaluation on the multidimensional initial indicator sets of all nodes to generate a global collaborative quality report. This invention effectively improves the accuracy and timeliness of early warning of mold growth in grain piles.

Description

Technical Field

[0001] This invention relates to the field of grain quality monitoring technology, specifically to a distributed detection method and system for multiple grain quality indicators based on edge collaboration. Background Technology

[0002] As the grain storage industry develops towards large-scale, refined, and intelligent operations, real-time, comprehensive, and accurate monitoring of grain quality has become a core requirement for safe grain storage. Distributed detection technologies targeting key indicators such as mold, moisture, and protein content in grain piles are being widely applied. Existing technologies have enabled the deployment of multiple sensor nodes within grain warehouses, capable of collecting data on the spectral response of grain piles and related data on the microenvironment of each node. Furthermore, models can be used to extract some grain quality indicators, providing fundamental hardware and data support for distributed grain condition monitoring.

[0003] However, traditional distributed grain condition monitoring typically only performs simple statistical averaging or threshold comparison of the detection data from multiple sensor nodes. This fails to effectively identify and quantify the dynamic diffusion trend of mold risk within the grain pile space, making it impossible to make early warning and control decisions for the early spread of mold. It is also difficult to adapt to the actual monitoring needs of the dynamic evolution of mold risk with space and time during grain storage, resulting in insufficient accuracy and timeliness of early warning of mold in grain piles. Summary of the Invention

[0004] This invention provides a distributed detection method and system for multiple indicators of grain quality based on edge collaboration, aiming to solve the technical problem of insufficient accuracy and timeliness of existing technologies for early warning of mold growth in grain piles.

[0005] In view of the above problems, the present invention provides a distributed detection method and system for multiple indicators of grain quality based on edge collaboration.

[0006] In a first aspect, the present invention provides a distributed detection method for multiple indicators of grain quality based on edge collaboration, comprising:

[0007] Multiple edge spectral sensing nodes are deployed inside the grain warehouse to simultaneously collect multispectral response data of the target grain pile and microenvironmental parameters at the location of the nodes.

[0008] Based on the microenvironment parameters, the multispectral response data is calibrated in real time and input into the quality feature extraction model to obtain a multidimensional initial index set for each node, wherein the multidimensional initial index set includes mold risk index, moisture content and protein content.

[0009] By combining the microenvironmental parameters and spatial topology of each node, the spatial correlation of the mold risk index in the node network is analyzed, and the local abnormal diffusion potential is calculated.

[0010] Based on the coupling relationship between the local abnormal diffusion potential and the moisture content and protein content, a spatiotemporal dynamic mold risk field simulation is performed to dynamically generate the risk field distribution.

[0011] Based on the risk field distribution, spatial data fusion and state optimization assessment are performed on the multidimensional initial index set of all nodes to generate a global collaborative quality report.

[0012] Secondly, the present invention provides a distributed detection system for multiple indicators of grain quality based on edge collaboration, comprising:

[0013] The multi-source data synchronous acquisition module is used to simultaneously acquire multispectral response data of the target grain pile and microenvironmental parameters of the node location by deploying multiple edge spectral sensing nodes in the grain warehouse.

[0014] The quality feature extraction module is used to perform real-time calibration of the multispectral response data based on the microenvironment parameters, and input the data into the quality feature extraction model to obtain a multidimensional initial index set for each node, wherein the multidimensional initial index set includes mold risk index, moisture content and protein content.

[0015] The abnormal diffusion analysis module is used to analyze the spatial correlation of the mold risk index in the node network by combining the microenvironment parameters and spatial topology of each node, and to calculate the local abnormal diffusion potential.

[0016] The mold risk field simulation module is used to perform spatiotemporal dynamic mold risk field simulation based on the coupling relationship between the local abnormal diffusion potential and the moisture content and protein content, and dynamically generate the risk field distribution.

[0017] The collaborative quality report output module is used to perform spatial data fusion and state optimization assessment on the multidimensional initial indicator set of all nodes based on the risk field distribution, and generate a global collaborative quality report.

[0018] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0019] This invention provides a distributed detection method and system for multiple indicators of grain quality based on edge collaboration. Through collaborative acquisition and real-time calibration of multi-source data, the accuracy of the detection data is ensured. Spatial correlation analysis and local anomaly diffusion potential calculation are used to accurately capture the spatial distribution characteristics of mold risk. Spatiotemporal simulation of the risk field enables prediction of the dynamic evolution trend of mold risk. Finally, a collaborative quality report is generated through multi-indicator spatial data fusion and global evaluation. This invention achieves early, accurate, and three-dimensional early warning of mold risk in grain piles, improving the accuracy, timeliness, and global collaboration of multi-indicator detection of grain quality, and providing reliable technical support for early warning and scientific prevention and control decisions regarding mold in grain piles. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating the distributed detection method for multiple indicators of grain quality based on edge collaboration provided in an embodiment of the present invention;

[0022] Figure 2 A schematic diagram of the structure of a distributed detection system for multiple indicators of grain quality based on edge collaboration provided in an embodiment of the present invention;

[0023] The components represented by each number in the attached diagram are explained below:

[0024] Multi-source data synchronous acquisition module 11, quality feature extraction module 12, anomaly diffusion analysis module 13, mold risk field simulation module 14, and collaborative quality report output module 15. Detailed Implementation

[0025] This invention provides a distributed detection method and system for multiple indicators of grain quality based on edge collaboration, which is used to address the technical problem that existing technologies are insufficient in terms of accuracy and timeliness for early warning of mold growth in grain piles.

[0026] Example 1, as Figure 1 As shown, this invention provides a distributed detection method for multiple indicators of grain quality based on edge collaboration, the method comprising:

[0027] S100: Multiple edge spectral sensing nodes are deployed within the grain silo to simultaneously collect multispectral response data of the target grain pile and microenvironmental parameters at the node's location.

[0028] In this embodiment of the invention, multiple edge spectral sensing nodes deployed within the grain warehouse synchronously collect multispectral response data of the target grain pile and microenvironmental parameters at the node's location. The occurrence and development of mold growth in grain piles are not only related to the quality of the grain itself but also closely related to the microenvironment of the grain pile. Furthermore, the early release of fungal metabolites such as aflatoxin in grain piles is low, and their characteristic absorption signals are easily masked by the conventional spectral signals of the grain matrix. Conventional spectral acquisition methods struggle to accurately capture this type of characteristic information, failing to provide effective data support for the early identification of subsequent mold risk. Simultaneously, if the spectral data from the sensing nodes is not synchronized with the acquisition of microenvironmental parameters, or if the data dimension is singular, subsequent microenvironment-based spectral calibration will lack a spatiotemporal matching basis, directly affecting the accuracy of quality indicator extraction. Therefore, it is necessary to achieve targeted acquisition of multispectral response data of grain piles and real-time acquisition of microenvironment parameters through the distributed deployment of edge spectral sensing nodes, and ensure the synchronization of the two types of data. In particular, the enhanced acquisition of specific bands sensitive to mold metabolites can amplify the characteristic absorption signals of mold metabolites, improve the identifiability of early mold-related characteristics, and lay a precise and comprehensive multi-source data foundation for subsequent whole-process grain quality testing and risk analysis.

[0029] Step S100 in the method provided in this embodiment of the invention includes:

[0030] The multispectral response data includes conventional spectral data in the visible-near-infrared band, as well as specific narrow-band spectral data collected with targeted enhancement for the characteristic absorption bands of fungal metabolites such as aflatoxin. The microenvironmental parameters include temperature, relative humidity, and carbon dioxide concentration monitored in real time at the nodes.

[0031] First, multiple edge spectral sensing nodes are deployed within the grain silo. Edge spectral sensing nodes are integrated detection nodes that combine multispectral acquisition and microenvironment monitoring functions, enabling on-site data acquisition and temporary storage. Based on the actual size of the grain silo, the shape of the grain pile, and the characteristics of the storage environment, a spatial three-dimensional grid deployment method is used to distribute multiple edge spectral sensing nodes within the grain pile. The deployment must cover the core risk areas for mold growth, corner areas, and high-temperature, high-humidity areas prone to mold growth, ensuring that the spatial distribution of the nodes comprehensively reflects the quality status and microenvironmental characteristics of different locations within the grain pile. Simultaneously, all deployed nodes need to be spatially calibrated and numbered, recording the three-dimensional spatial coordinates (x, y, z) of each node to provide a topological basis for subsequent spatial correlation analysis.

[0032] For example, in a flat-bottomed grain silo with a capacity of 10,000 tons, 90 edge spectral sensing nodes, numbered N1 to N90, are deployed in a 5m×5m planar grid plus three vertical layers. The three-dimensional spatial coordinates of node N1 are (5m, 5m, 1m), corresponding to the front left side of the grain pile surface; the three-dimensional spatial coordinates of node N45 are (25m, 15m, 3m), corresponding to the center of the middle layer of the grain pile; and the three-dimensional spatial coordinates of node N90 are (50m, 30m, 5m), corresponding to the rear right side of the bottom layer of the grain pile. After all nodes are deployed, the coordinates and number information are entered into the grain silo control system for calibration.

[0033] Secondly, multispectral response data of the target grain pile is collected simultaneously. This multispectral response data includes conventional spectral data in the visible-near-infrared band, as well as specific narrow-band spectral data acquired with targeted enhancement based on the characteristic absorption bands of fungal metabolites such as aflatoxin. Multispectral response data refers to the spectral data formed by the light absorption and reflection response signals generated by the grain sample under illumination of different wavelengths of light; it is the core data for extracting grain quality indicators. The visible-near-infrared band refers to the spectral band with wavelengths ranging from 400 to 2500 nm. Spectral data in this band can reflect characteristic information of conventional quality indicators such as grain moisture and protein. Specific narrow-band spectral data refers to narrow-range spectral bands selected for the characteristic absorption bands of fungal metabolites such as aflatoxin. The characteristic absorption bands of fungal metabolites are spectral ranges in which substances such as aflatoxin and fungal spores specifically absorb light at specific wavelengths. The spectral signals in these ranges are key features for identifying early mold growth. Enhanced acquisition refers to amplifying the weak characteristic absorption signals of fungal metabolites by improving the spectral acquisition resolution of this narrow band and extending the integration time, thereby increasing the sensitivity of the data to fungal metabolites.

[0034] Specifically, all deployed edge spectral sensing nodes, under the unified command of the grain warehouse control system, synchronously activate the multispectral acquisition module to simultaneously acquire conventional spectral data in the visible-near-infrared band and specific narrow-band spectral data of the grain pile. The conventional spectral data is acquired at a preset conventional resolution, while the specific narrow-band spectral data is selected based on the characteristic absorption bands of aflatoxin and fungal metabolites. The acquisition resolution of this narrow band is increased from the conventional 10nm to 2nm, and the integration time is extended from the conventional 50ms to 200ms to complete the enhanced acquisition. The acquired multispectral response data is uploaded to the node's own edge computing unit in real time and marked with an acquisition timestamp.

[0035] For example, in the aforementioned grain warehouse, after the central control system issues a data acquisition command, all nodes N1 to N90 synchronously acquire multispectral response data at 10:00:00 on January 20, 2026. Among them, node N45 acquires conventional spectral data in the 400-2500nm band with a resolution of 10nm, forming a complete conventional spectral reflectance curve. At the same time, for the two characteristic absorption bands of aflatoxin B1 at 337nm and 1520nm, enhanced acquisition is completed with a resolution of 2nm and an integration time of 200ms, obtaining high-resolution spectral data in two specific narrow bands. This node integrates the two types of spectral data, marks them with a timestamp of 10:00:00 on January 20, 2026, and temporarily stores them in its own edge computing unit.

[0036] In addition, microenvironmental parameters at the node's location are collected synchronously. These parameters include real-time monitoring of temperature, relative humidity, and carbon dioxide concentration at the node. Microenvironmental parameters refer to local environmental parameters at different spatial locations within the grain pile, and are key environmental factors affecting the occurrence and development of mold growth and the accuracy of spectral data. Temperature and relative humidity directly affect the rate of mold reproduction, while carbon dioxide concentration indirectly reflects the intensity of microbial activity in the grain pile. All three parameters interfere with the spectral response signal of the grain and are crucial for subsequent spectral data calibration.

[0037] Specifically, all edge spectral sensing nodes simultaneously activate their own microenvironment monitoring modules at the same time they acquire multispectral response data. They collect the temperature, relative humidity, and carbon dioxide concentration at the node's location in real time. During the acquisition process, the acquisition frequency and timestamp of the microenvironment parameters are kept completely consistent with the multispectral response data, achieving spatiotemporal synchronization matching between the microenvironment parameters and the multispectral response data. The acquired microenvironment parameters are also temporarily stored in the node's edge computing unit, forming a corresponding data set with the multispectral response data.

[0038] For example, in the aforementioned grain warehouse, while node N45 was collecting multispectral response data at 10:00:00 on January 20, 2026, it simultaneously collected data from the microenvironment monitoring module, which showed that the temperature at its location was 28°C, the relative humidity was 75%, and the carbon dioxide concentration was 0.8%. This node marked the set of microenvironment parameters with the same timestamp as the multispectral response data and integrated them with the spectral data into a set of temporary data, completing the multi-source data acquisition of a single node; the other nodes all completed the synchronous acquisition of microenvironment parameters and multispectral response data at the same time, forming their respective corresponding temporary data sets.

[0039] In this embodiment of the invention, nodes are distributed in a three-dimensional grid to achieve full-area, blind-spot-free detection of the grain pile space, laying a complete node topology foundation for subsequent spatial correlation analysis. Enhanced acquisition of characteristic bands of mold metabolites effectively amplifies the weak characteristic signals of early mold growth, solving the problem that conventional spectral acquisition is difficult to capture early mold growth information. Synchronous acquisition of multispectral data and microenvironmental parameters ensures the spatiotemporal matching of the two types of data, providing a precise basis for subsequent spectral calibration based on microenvironmental parameters. This ensures the accuracy of subsequent detection and analysis results from the data source, and overall achieves comprehensive and accurate acquisition of multi-source detection data of the grain pile.

[0040] S200: Based on the microenvironment parameters, the multispectral response data is calibrated in real time and input into the quality feature extraction model to obtain a multidimensional initial index set for each node, wherein the multidimensional initial index set includes mold risk index, moisture content and protein content.

[0041] In this embodiment of the invention, the multispectral response data is calibrated in real time based on the microenvironment parameters and input into a quality feature extraction model to obtain a multidimensional initial index set for each node. This multidimensional initial index set includes the mold risk index, moisture content, and protein content. During the acquisition process, the multispectral response data is directly interfered with by local microenvironment parameters in the grain pile. Environmental factors alter the spectral absorption and reflection characteristics of wheat grains, causing the original spectral data to deviate from the true spectral characteristics of the grain. Directly using this data for index extraction would result in significant errors. Furthermore, the mold risk index, moisture content, and protein content of grain quality are interrelated feature indicators. Traditional single-indicator modeling methods cannot uncover the intrinsic correlation between indicators, limiting prediction efficiency and accuracy. Moreover, general models are not adapted to the actual environment and mold occurrence patterns of wheat storage in grain warehouses, easily leading to prediction bias. Therefore, it is necessary to calibrate the spectral data in real time based on microenvironment parameters to eliminate environmental interference. Simultaneously, a multi-task deep learning quality feature extraction model adapted to the wheat storage scenario is constructed to achieve synchronous and accurate extraction of multiple indicators, providing accurate single-node initial index data for subsequent spatial analysis and simulation of mold risk.

[0042] The construction process of the quality feature extraction model is as follows:

[0043] First, based on the actual storage environment and mold occurrence patterns of wheat in grain warehouses, real-time calibrated multispectral response data were collected as a sample spectral data set. According to the actual storage environment and mold occurrence patterns of the target grain warehouse and wheat grain warehouses of the same type and in the same region, the environmental gradient and spectral type for sample collection were determined. Within the aforementioned environmental range, raw multispectral response data of wheat were collected through edge spectral sensing nodes within the grain warehouse, simultaneously acquiring corresponding microenvironmental parameters and completing real-time calibration of the spectral data. After integrating calibrated multispectral response data under different environmental conditions and different degrees of mold, and removing outliers, a sample spectral data set was formed, and data dimensions and formats were standardized.

[0044] For example, taking a target wheat silo with a capacity of 10,000 tons and three surrounding wheat silos of the same type as the data collection objects, under an environmental gradient of temperature = 20~35℃, relative humidity = 60~85%, and carbon dioxide concentration = 0.5~1.2%, wheat conventional spectral data of 400~2500nm and specific narrow band enhanced spectral data of 337nm and 1520nm were collected through the edge spectral sensing nodes of each silo. Corresponding microenvironmental parameters were collected simultaneously and calibrated band by band. After integrating 8,000 sets of effective calibrated multispectral data and removing 20 sets of abnormal data, a standardized sample spectral data set of 8,000 sets was obtained, with the data dimension uniformly set to 251 dimensions.

[0045] Secondly, the true values ​​of mold risk index, moisture content, and protein content for each sample in the sample spectral data set were measured and labeled to obtain the corresponding true indicator set for the sample. For each set of calibrated spectral data in the sample spectral data set, corresponding wheat samples were collected; the mold risk index, moisture content, and protein content of the samples were measured using industry standard methods, with the mold risk index determined by plate counting and linearly converted to the 0-1 range; the true values ​​of the three indicators for each sample were labeled and associated with the corresponding calibrated spectral data, and all labeled results were integrated to form the true indicator set for the sample, ensuring that the sample number was completely consistent with that of the sample spectral data set.

[0046] For example, for wheat samples corresponding to 8000 sets of spectral data, the actual indicators were determined sequentially: the colony count of sample group 1000 was 1.2 × 10⁵ CFU / g, which is equivalent to a mold risk index of 0.62; the moisture content was 14.6% measured by the 105℃ constant weight method; and the protein content was 12.9% measured by the Kjeldahl method. After completing the above determinations and labeling for all 8000 samples, the actual values ​​of all three indicators were integrated to obtain a set of actual indicators for 8000 samples that correspond one-to-one with the sample spectral data set. Each set of data contains three actual values: mold risk index, moisture content, and protein content.

[0047] Furthermore, based on multi-task deep learning, a network architecture for an output quality feature extraction model is constructed. The overall model architecture is defined as an input layer - a shared hidden layer - a task-specific output layer. The input layer dimension matches the feature dimension of the sample spectral data. The shared hidden layer is responsible for extracting the general low-level features of the spectral data. The task-specific output layer designs an independent branch for each quality indicator to achieve simultaneous output of multiple indicators. The number of neurons and activation functions of each layer are designed: the input layer has no activation function, the shared hidden layer uses the ReLU activation function to ensure non-linear expression of features, and the task-specific output layer selects an appropriate activation function according to the numerical range of each indicator. The network architecture is built and standardized to ensure that the input and output dimensions match the sample data and target indicators.

[0048] For example, a multi-task deep learning network architecture is built based on the Python+PyTorch framework. The specific design is as follows: Input layer: 251 dimensions, consistent with the standardized dimensions of the sample spectral data, with no activation function, used to receive calibrated multispectral response data; Shared hidden layer: 3 layers in total, extracting common spectral features for all indicators. The first layer has 128 neurons with ReLU activation; the second layer has 64 neurons with ReLU activation; the third layer has 32 neurons with ReLU activation; Task-specific output layer: 3 independent output branches, corresponding to mold risk index, moisture content, and protein content respectively. The mold risk index branch has 1 neuron with Sigmoid activation, mapping the output to the 0~1 range to match the indicator range; the moisture content and protein content branches each have 1 neuron with Linear activation, providing linear output to match the actual numerical range of the indicator. The network architecture of the quality feature extraction model is built according to the above design, forming an end-to-end multi-task deep learning model.

[0049] Based on this, the quality feature extraction model is trained under supervision using a sample spectral data set and a sample real indicator set until the model's prediction and verification errors for each indicator converge, thus completing the model construction. The sample spectral data set and the sample real indicator set are divided into training, validation, and test sets in an 8:1:1 ratio. The training set is used for updating model parameters, the validation set is used for real-time evaluation of prediction errors, and the test set is used for final model performance verification. The model's training optimizer, batch size, and learning rate strategy are determined, and a loss function suitable for multi-indicator prediction is selected, highlighting the core detection requirement of the mold risk index. The model is trained under supervision using the training set. The prediction error of the validation set is calculated after each training epoch. Training is stopped if the preset convergence condition is met, thus completing the construction of the quality feature extraction model.

[0050] For example, the spectral data set of 8000 samples and the real index set of 8000 samples are simultaneously divided in an 8:1:1 ratio to obtain 6400 training sets, 800 validation sets, and 800 test sets. Training configuration: The Adam optimizer is used, with a batch size of 32, an initial learning rate of 0.001, and the learning rate decays exponentially with each training epoch, with a decay coefficient of 0.95. The weighted mean square error (MSE) is selected as the total loss function, with weights set as follows: mold risk index 0.5, moisture content 0.3, and protein content 0.2. The total loss function formula is: risk, moisture, and protein. Convergence criteria: The mean absolute error (MAE) of the validation set decreases by less than 1 × 10⁻⁴ for 15 consecutive epochs, and the independent MAE of each index satisfies: mold risk index ≤ 0.02, moisture content ≤ 0.3%, and protein content ≤ 0.2%. For example, after 68 training rounds, the average MAE of the validation set decreased by 8×10-5 for 15 consecutive rounds. The independent MAE of each indicator was: mold risk index 0.018, moisture content 0.25%, and protein content 0.18%, all of which met the preset convergence conditions. Training was then stopped, and the construction of the quality feature extraction model was completed.

[0051] Finally, based on the microenvironment parameters, the multispectral response data is calibrated in real time and input into the quality feature extraction model to obtain a multidimensional initial index set for each node. This multidimensional initial index set includes the mold risk index, moisture content, and protein content. Real-time calibration of spectral data refers to the process of correcting the original multispectral response data collected by the S100 node band-by-band using microenvironmental parameters such as temperature, relative humidity, and carbon dioxide concentration collected by the nodes, through environmental factor correction logic, to eliminate environmental interference and restore the true spectral characteristics of wheat. The multidimensional initial index set refers to the set of quality indicators, composed of the mold risk index, moisture content, and protein content, predicted by the quality feature extraction model for each edge spectral sensing node. It serves as the foundational single-node data for subsequent spatial correlation analysis of mold risk and diffusion potential calculation.

[0052] Specifically, for each edge spectral sensing node within the grain warehouse, the raw multispectral response data and corresponding microenvironmental parameters synchronously acquired in the S100 are retrieved. Referring to the standard environmental parameters acquired from wheat spectra, and combining them with preset environmental correction coefficients for each spectral band, the raw multispectral data is calibrated band-by-band in real-time based on the differences between the measured microenvironmental parameters and the standard parameters. After calibration, the spectral data is normalized and adjusted to a standardized format matching the input dimensions of the quality feature extraction model. The standardized calibrated spectral data is input into the trained quality feature extraction model, which simultaneously outputs the prediction results for three indicators corresponding to that node: mold risk index, moisture content, and protein content. The prediction results of the three indicators are integrated and archived to form a multidimensional initial indicator set for that node. All edge spectral sensing nodes sequentially perform the above calibration, prediction, and integration operations to ultimately obtain a unique multidimensional initial indicator set for each node.

[0053] For example, taking node N45 as an example, the original multispectral data and measured microenvironmental parameters of this node in S100 are retrieved: temperature 28℃, relative humidity 75%, carbon dioxide concentration 0.8%; referring to the standard environmental parameters for wheat spectral acquisition, and combining the preset environmental correction coefficients for each band, for all spectral bands such as the 400nm conventional spectral band and the 337nm specific narrow band, calibration is performed band by band according to the difference between the measured and standard environmental parameters. Among them, the original reflectance of the 400nm band is 25%, which is corrected to 23.82% after calibration; 33 The original absorbance in the 7nm band was 68%, which was corrected to 70.21% after calibration. After calibration of all bands, the data was normalized and standardized to 251 dimensions. The 251-dimensional standardized spectral data was input into the trained quality feature extraction model, and the predicted results were: mold risk index 0.65, moisture content 14.2%, and protein content 12.5%. Integrating the above three indicators, the multidimensional initial indicator set for node N45 was obtained {mold risk index: 0.65, moisture content: 14.2%, protein content: 12.5%}. The above process was performed on all edge spectral sensing nodes from N1 to N90, and finally, the multidimensional initial indicator sets for 90 nodes were obtained.

[0054] In this embodiment of the invention, a calibrated spectral sample set adapted to real-world scenarios and a corresponding set of real-world indicator annotations are constructed to provide reliable support for training the quality feature extraction model. A multi-task deep learning architecture can uncover the intrinsic correlations between multiple indicators, and by combining a weighted loss function with explicit convergence conditions, the efficiency, robustness, and prediction accuracy of simultaneous extraction of multiple indicators are improved. Through micro-environment parameter-driven spectral data calibration, environmental interference is effectively eliminated, and true spectral features are restored, enabling accurate and synchronous extraction of multi-dimensional initial indicators at each node. This addresses the pain points of traditional detection methods, such as susceptibility to spectral data interference, low indicator extraction accuracy, and difficulty in simultaneous extraction of multiple indicators. It provides reliable single-node foundational data for subsequent mold risk spatial analysis and simulation, aligning with the needs of edge-collaborative distributed detection and improving the accuracy of grain quality detection.

[0055] S300: Combining the microenvironmental parameters and spatial topology of each node, analyze the spatial correlation of the mold risk index in the node network and calculate the local abnormal diffusion potential.

[0056] In this embodiment of the invention, the spatial correlation of the mold risk index in the node network is analyzed by combining the microenvironmental parameters and spatial topology of each node, and the local anomaly diffusion potential is calculated. The spatial distribution of mold risk in grain piles is not random, but exhibits clustering characteristics due to the synergistic influence of local microenvironment and spatial location. Traditional detection methods cannot quantify the spatial autocorrelation and clustering of the mold risk index in the node network, nor the dynamic trend of mold risk spreading from high-risk areas to the surrounding areas, making it difficult to support accurate early warning and control of mold. Therefore, it is necessary to analyze the spatial correlation of mold risk by combining node spatial topology and microenvironmental parameters, identify high-risk clustering areas, and further calculate the local anomaly diffusion potential to provide spatial correlation and diffusion driving basis for subsequent spatiotemporal simulation of the mold risk field.

[0057] Step S300 in the method provided in this embodiment of the invention includes:

[0058] The spatial correlation of the mold risk index in the node network is analyzed by combining the microenvironmental parameters and spatial topology of each node, including:

[0059] Based on the spatial coordinates of all nodes, a spatial weight matrix is ​​constructed using the inverse distance weighting method;

[0060] Using the spatial weight matrix, the global Moran index of the mold risk index is calculated, and the degree of spatial autocorrelation in the overall network is evaluated.

[0061] For each node, the local Moran index is calculated to identify spatially significant clustering regions, which include high-risk clustering regions and low-risk clustering regions.

[0062] The identified spatial clustering areas are correlated with the microenvironmental parameters of the corresponding nodes to establish a mapping relationship between high-risk clustering and high-temperature and high-humidity microenvironmental parameters.

[0063] First, based on the spatial coordinates of all nodes, a spatial weight matrix is ​​constructed using the inverse distance weighting method. The spatial weight matrix represents the spatial mutual influence between different nodes in the grain storage node network; the closer two points are, the greater their mutual influence; the farther apart they are, the smaller or even zero their influence. The inverse distance weighting method is a commonly used method for calculating spatial weights; the weight is inversely proportional to the distance between nodes, i.e., weight = 1 / distance. This method stipulates that a node's weight for itself is 0, and the distance attenuation coefficient α is set to 1.

[0064] Specifically, the three-dimensional spatial coordinates (x, y, z) of all edge spectral sensing nodes are retrieved and organized into a set of node coordinates. For each pair of nodes, the three-dimensional spatial distance is calculated, and the spatial weight between the nodes is calculated based on the inverse distance weighting method, with the weight of each node itself set to 0. The weights of all node pairs are then organized into a matrix to construct the spatial weight matrix. For example, the coordinates of nodes N1 (5m, 5m, 1m) and N45 (25m, 15m, 3m) are retrieved, and the three-dimensional spatial distance is calculated to be approximately 22.5m. The weight between them is calculated using the inverse distance weighting method to be 1 / 22.5 ≈ 0.044, while the weight of N1 itself is set to 0. The above calculation is repeated for all node pairs from N1 to N90, ultimately constructing a 90×90 spatial weight matrix.

[0065] Secondly, using the spatial weight matrix, the global Moran index of the mold risk index is calculated, and the degree of spatial autocorrelation in the overall network is evaluated.

[0066] Specifically, the global Moran's index of the mold risk index is calculated using the spatial weight matrix, and its spatial autocorrelation in the overall network is evaluated, including:

[0067] Calculate the mean and variance of the mold risk index for all nodes;

[0068] Based on the spatial weight matrix, calculate the weighted covariance sum of the deviations of the mold risk index from the mean among all node pairs;

[0069] The global Moran index value is calculated by substituting the mean, variance, and weighted covariance into the global Moran index formula.

[0070] The global Moran index value is compared with the zero distribution generated by the random permutation test to perform a significance test;

[0071] Based on the significance test results and the sign and magnitude of the global Moran index, it is determined whether the mold risk index has a significant spatial clustering distribution pattern, spatial discrete distribution pattern, or random distribution pattern in the grain warehouse node network.

[0072] First, calculate the mean and variance of the mold risk index for all nodes. The mean is the average of the mold risk indices for all nodes, reflecting the overall average level of mold risk in the grain silo. The variance is the average of the squared deviations of each node's mold risk index from the mean, reflecting the dispersion of the risk index. Summarize the mold risk indices from the multidimensional initial index set for all nodes; calculate the average of all risk indices to obtain the mean; calculate the squared deviation of each risk index from the mean, and then average it again to obtain the variance. For example, summarizing the mold risk indices for nodes N1 to N90 yields an overall mean of 0.42 and a variance of 0.08.

[0073] Secondly, based on the spatial weight matrix, the weighted covariance sum of the deviations of the mold risk index from the mean among all node pairs is calculated. The weighted covariance sum refers to the sum of the products of the deviations of the mold risk index from the mean for all non-repeating node pairs in the grain warehouse node network, calculated based on the spatial weight matrix, reflecting the spatial correlation strength of the risk index. All edge spectral sensing nodes within the grain warehouse are paired without omission or repetition; that is, any two different nodes form only one node pair. For example, node A and node B, and node B and node A are considered the same pair, calculated only once, excluding combinations of nodes themselves. For each node pair, the product of the deviations of their mold risk index from the mean is calculated; this product is multiplied by the corresponding node pair weight in the spatial weight matrix; and the weighted covariance sum of the weighted deviation products of all node pairs is obtained.

[0074] For example, all nodes from N1 to N90 are paired without repetition. Nodes N45 and N44 are one such pair. The mold risk index of node N45 is 0.65, and the mold risk index of node N44 is 0.61. The overall mean is 0.42. The product of their deviations is (0.65-0.42)×(0.61-0.42)=0.0437. Multiplying this by the corresponding weights of 0.05 in the spatial weight matrix, we get 0.002185. This operation is repeated for all pairs of nodes from N1 to N90 without repetition, and the sum is obtained. Finally, the weighted covariance sum is 1.26.

[0075] Furthermore, based on the mean, variance, and weighted covariance sum, the global Moran's index value is calculated using the global Moran's index formula. The global Moran's index is a spatial autocorrelation statistic used to determine whether the mold risk index is randomly distributed, clustered, or uniformly discrete across the entire grain storage node network. Combining the calculated mean, variance, and weighted covariance sum, the global Moran's index value is calculated using spatial autocorrelation statistical logic. For example, based on a mean of 0.42, a variance of 0.08, and a weighted covariance sum of 1.26, a global Moran's index value of 0.38 is calculated.

[0076] Subsequently, the global Moran index value is compared with the zero distribution generated by the random permutation test to perform a significance test.

[0077] The comparison of the global Moran index value with the zero distribution generated through a random permutation test, and the significance test, includes:

[0078] Keeping the spatial weight matrix unchanged, the spatial positions of the mold risk index values ​​among all nodes are randomly rearranged multiple times.

[0079] For each new spatial data generated after random permutation, the corresponding global Moran index value is recalculated.

[0080] All global Moran index values ​​obtained from multiple random permutations are summarized to generate the zero distribution, and the mean and standard deviation of the zero distribution are calculated.

[0081] The calculated actual global Moran index value is compared with the mean of the null distribution to calculate the Z score;

[0082] Based on the Z-score and the standard normal distribution table, determine the P-value corresponding to the actual global Moran's index value;

[0083] The P-value is compared with a preset significance level threshold. If the P-value is less than the significance level threshold, it is determined that the mold risk index has a significant spatial autocorrelation pattern in the grain warehouse node network.

[0084] First, keeping the spatial weight matrix unchanged, the spatial positions of the mold risk index values ​​across all nodes are randomly rearranged multiple times. The random permutation test involves generating a reference distribution under the assumption of spatial random distribution by shuffling the spatial positions of the data, used to test the significance of actual spatial autocorrelation. The constructed spatial weight matrix remains unchanged to ensure that the spatial influence relationships between nodes are not altered; the mold risk index of all edge spectral sensing nodes is extracted to form a risk index set; this risk index set is then randomly rearranged multiple times, generating a new node-risk index correspondence after each rearrangement.

[0085] For example, keeping the spatial weight matrix of nodes N1 to N90 unchanged, the mold risk index of all nodes is extracted to form a set, and the set is randomly rearranged 1000 times. After one rearrangement, the risk index of node N1 is 0.41, the risk index of node N45 is 0.58, and the risk index of the remaining nodes is also randomly assigned. The relationship between the spatial coordinates and weights of the nodes does not change.

[0086] Secondly, for each new spatial data generated after random permutation, the corresponding global Moran's index value is recalculated. For the mold risk index data after each random permutation, the mean and variance of the mold risk index of all nodes are recalculated. Then, based on the spatial weight matrix, the weighted covariance sum of the deviations of the mold risk index from the mean among all pairs of non-repeating nodes is calculated. Finally, combining the obtained mean, variance, and weighted covariance sum, the global Moran's index value corresponding to this random permutation is calculated. For example, for the risk index data after one random permutation, the recalculated mean is 0.43, variance is 0.078, and weighted covariance sum is 1.12. Further calculation yields a global Moran's index value of 0.29 for this permutation. Following this process, the global Moran's index calculation is completed for 1000 random permutations, resulting in 1000 index values.

[0087] Furthermore, all global Moran's index values ​​obtained from multiple random permutations are aggregated to generate the zero distribution, and the mean and standard deviation of the zero distribution are calculated. The zero distribution refers to the theoretical distribution of the global Moran's index when the mold risk index is assumed to be spatially random. It is formed by aggregating global Moran's index values ​​obtained from multiple random permutations, and its mean and standard deviation reflect the central tendency and dispersion of the index under a random distribution. A zero distribution dataset is constructed by aggregating all global Moran's index values ​​obtained from all random permutations; the mean of this dataset is calculated to obtain the zero distribution mean, reflecting the average level of the global Moran's index under a random distribution; the standard deviation of this dataset is calculated to obtain the zero distribution standard deviation, reflecting the dispersion of the global Moran's index under a random distribution.

[0088] For example, 1000 global Moran index values ​​obtained from 1000 random permutations are aggregated to construct a zero-distribution dataset; the mean of this zero distribution is calculated to be 0.02 and the standard deviation is 0.09, that is, in the random distribution scenario, the global Moran index averages 0.02 and the numerical fluctuation range is 0.09.

[0089] Then, the calculated actual global Moran's index value is compared with the mean of the zero distribution to calculate the Z-score. The Z-score measures the degree of deviation between the actual global Moran's index value and the mean of the zero distribution; the greater the deviation, the more significant the difference between the actual value and the random distribution hypothesis. The calculated actual global Moran's index value is extracted, and the difference between the actual value and the mean of the zero distribution is calculated. This difference is then divided by the standard deviation of the zero distribution to obtain the Z-score. For example, if the calculated actual global Moran's index value is 0.38, the mean of the zero distribution is 0.02, and the standard deviation is 0.09, the difference between the actual value and the mean is 0.38 - 0.02 = 0.36. Dividing this difference by the standard deviation of 0.09 yields a Z-score of 4.0.

[0090] Then, based on the Z-score and the standard normal distribution table, the P-value corresponding to the actual global Moran's index value is determined. The P-value refers to the probability of the current actual global Moran's index value and more extreme values ​​occurring under the assumption of a spatially random distribution of the mold risk index. The smaller the P-value, the less likely the random distribution assumption is to hold, and the more significant the actual spatial autocorrelation is. The absolute value of the Z-score is calculated, and the corresponding one-sided tail area is obtained from the standard normal distribution table. The one-sided tail area is multiplied by 2 to obtain the P-value for the two-sided test, which is the significance probability corresponding to the actual global Moran's index. For example, the Z-score is 4.0, and its absolute value is 4.0; the one-sided tail area obtained from the standard normal distribution table is 0.00003; multiplying this area by 2, the P-value corresponding to the actual global Moran's index value is calculated to be 0.00006.

[0091] Subsequently, the P-value is compared with a preset significance level threshold. If the P-value is less than the significance level threshold, it is determined that the mold risk index exhibits a significant spatial autocorrelation pattern in the grain warehouse node network. The preset significance level threshold is a critical value used to determine significance, typically set to 0.05. If the P-value is less than this threshold, the actual result is statistically significant; if it is greater than or equal to this threshold, the result is not statistically significant. The preset significance level threshold is 0.05. The calculated actual P-value is compared with this threshold. If the P-value is less than the preset significance level threshold, it is determined that the mold risk index exhibits a significant spatial autocorrelation pattern in the grain warehouse node network; if the P-value is greater than or equal to the preset significance level threshold, it is determined that there is no significant spatial autocorrelation pattern. For example, if the actual calculated P-value is 0.00006, since 0.00006 < 0.05, it is determined that the mold risk index exhibits a significant spatial autocorrelation pattern in the N1~N90 node network, and its spatial distribution is not randomly formed.

[0092] Based on this, according to the significance test results and the sign and magnitude of the global Moran's index, it is determined whether the mold risk index exhibits a significant spatial clustering distribution pattern, a spatial discrete distribution pattern, or a random distribution pattern in the grain warehouse node network. A spatial clustering distribution pattern refers to the mold risk index showing a spatial distribution characteristic where high-risk elements are adjacent to each other, and low-risk elements are adjacent to each other, with a positive and significant global Moran's index. A spatial discrete distribution pattern refers to the mold risk index showing a spatial distribution characteristic where high-risk and low-risk elements are adjacent to each other, with a negative and significant global Moran's index. A random distribution pattern refers to the mold risk index showing no obvious spatial correlation, exhibiting a random distribution, and showing no significant spatial autocorrelation.

[0093] Specifically, based on the significance test results, and combined with the sign and magnitude of the global Moran's index, the spatial distribution pattern is comprehensively determined. The determination is categorized by scenario: if the P-value is ≥ the preset significance level threshold, it is determined to be a random distribution pattern; if the P-value is < the preset significance level threshold and the global Moran's index is positive, it is determined to be a spatially clustered distribution pattern; if the P-value is < the preset significance level threshold and the global Moran's index is negative, it is determined to be a spatially discrete distribution pattern. The determination results are recorded as the basis for subsequent local spatial analysis.

[0094] For example, with a preset significance level threshold of 0.05, a P-value of 0.00006 < 0.05, and a global Moran's index of 0.38, the mold risk index is determined to exhibit a spatially clustered distribution pattern in the N1~N90 node network. If we assume that the P-value is 0.06 > 0.05, it is determined to be a random distribution pattern. If the P-value is 0.03 < 0.05 and the global Moran's index is -0.25, it is determined to be a spatially discrete distribution pattern.

[0095] Furthermore, for each node, a local Moran index is calculated to identify spatially significant clustering regions, which include high-risk clustering regions and low-risk clustering regions.

[0096] First, calculate the local Moran's index for each node. The local Moran's index measures the strength of local spatial autocorrelation between a single node and its spatial neighbors. The sign and magnitude of the value reflect whether the node exhibits risk clustering or dispersion characteristics with its neighbors. Retrieve the constructed spatial weight matrix, the mold risk index of all nodes, and the global mean. For each node, combine its own mold risk index, the mold risk indices of its neighbors, and the spatial weights of the node and its neighbors, and calculate the local Moran's index using local spatial autocorrelation statistical logic. Summarize the local Moran's indices of all nodes to form a set of local Moran's indices. For example, taking node N45 as an example, retrieve the risk indices of its neighbors (0.61, 0.63, 0.58, 0.60) and their corresponding spatial weights (0.05, 0.048, 0.045, 0.046). Calculate the local Moran's index for N45 using local spatial autocorrelation statistical logic, resulting in 0.52. Following the same process, calculate the local Moran's indices for all nodes N1 to N90.

[0097] Secondly, a conditional randomization test is performed on each node to assess the significance of the local Moran's index. The conditional randomization test involves fixing the mold risk index value of the target node and randomly rearranging only the risk indices of its neighboring nodes, eliminating interference from the target node's own value and accurately testing the statistical significance of local spatial autocorrelation. The empirical P-value is calculated based on the conditional null distribution generated from multiple random rearrangements, determining the extreme probability of the actual local Moran's index and used to determine whether the local autocorrelation is significant. For each node, the test is performed as follows: the mold risk index value of the node is fixed without any adjustment; all spatially neighboring nodes of the node are extracted, and the mold risk index values ​​of the neighboring nodes are randomly rearranged multiple times; for the risk indices of the neighboring nodes after each rearrangement, combined with the fixed target node value and the spatial weight matrix, the local Moran's index of the target node is recalculated; the local Moran's index values ​​are summarized to generate the local index conditional null distribution for the node; based on the conditional null distribution, the empirical P-value of the actual local Moran's index value in this distribution is calculated.

[0098] For example, taking node N45 as an example, perform a conditional random permutation test: keep the mold risk index value of N45 fixed at 0.65; extract the risk indices of its neighboring nodes, and perform 1000 random rearrangements of these four indices; after each rearrangement, combine the fixed N45 value, the rearranged neighboring node values, and the corresponding spatial weights to recalculate the local Moran index of N45, resulting in 1000 local Moran index values ​​after 1000 rearrangements; summarize these 1000 index values ​​to generate the conditional null distribution of the local index of N45; compare the actual local Moran index of 0.52 with the conditional null distribution to calculate an empirical P-value of 0.012; calculate the empirical P-values ​​of all nodes N1 to N90 using the same process.

[0099] Finally, statistically significant spatial clustering regions are identified, including high-risk clustering regions and low-risk clustering regions. High-risk clustering regions are defined as regions composed of nodes whose empirical P-value is less than a preset threshold, and whose own mold growth risk index is higher than the global average, and whose neighboring nodes' average risk index is also higher than the global average. Low-risk clustering regions are defined as regions composed of nodes whose empirical P-value is less than a preset threshold, and whose own mold growth risk index is lower than the global average, and whose neighboring nodes' average risk index is also lower than the global average.

[0100] Specifically, a significance level threshold of 0.05 is preset. Nodes with empirical P-values ​​< 0.05 are identified as nodes with significant local spatial autocorrelation. For nodes with significant autocorrelation, their own risk index and the mean risk index of their neighboring nodes are calculated. Classification is then performed: nodes with their own risk index > global mean and the mean risk index of their neighboring nodes > global mean are classified as high-risk cluster nodes; nodes with their own risk index < global mean and the mean risk index of their neighboring nodes < global mean are classified as low-risk cluster nodes. Nodes of the same type are then aggregated to form high-risk clusters and low-risk clusters.

[0101] For example, node N45 has an empirical P-value of 0.012 < 0.05, its own risk index of 0.65 > the global average of 0.42, and the average risk of neighboring nodes of 0.605 > the global average of 0.42. Therefore, N45 is classified as a node in a high-risk cluster. Node N1 has an empirical P-value of 0.03 < 0.05, its own risk index of 0.32 < the global average of 0.42, and the average risk of neighboring nodes of 0.31 < the global average of 0.42. Therefore, N1 is classified as a node in a low-risk cluster. By summarizing all nodes of the same type, high-risk clusters and low-risk clusters are finally formed.

[0102] Then, correlation analysis was performed on the identified spatial clustering areas and their corresponding node microenvironmental parameters to establish a mapping relationship between high-risk clustering and high-temperature and high-humidity microenvironmental parameters. Correlation analysis involves exploring the characteristics of microenvironmental parameters in spatially clustered areas, comparing the differences in microenvironmental parameters between non-clustered areas, and uncovering the intrinsic connection between risk clustering and the microenvironment. Mapping relationship refers to clarifying the correspondence between high-risk clustering areas and specific microenvironmental parameters, establishing a correlation between high-risk clustering and high-temperature and high-humidity conditions. Microenvironmental parameters were extracted from all nodes in high-risk clustering areas, all nodes in low-risk clustering areas, and all nodes in non-clustered areas. Average values ​​for temperature, relative humidity, and carbon dioxide concentration were calculated for each of the three types of areas, forming a set of average microenvironmental parameter values ​​for the three types of areas. The differences in average parameter values ​​between high-risk clustering areas and non-clustered areas were compared to identify the microenvironmental characteristics of high-risk clustering areas. Based on the difference analysis results, the correspondence between high-risk clustering and microenvironmental parameters was clarified, and the core of establishing the mapping relationship between high-risk clustering and high-temperature and high-humidity conditions was determined.

[0103] For example, parameters were extracted by category: the high-risk cluster area contained 22 nodes such as N45 and N46, the low-risk cluster area contained 35 nodes such as N1 and N2, and the non-cluster area contained the remaining 33 nodes. Temperature, relative humidity, and carbon dioxide concentration data were extracted for all nodes in the three types of areas respectively. Statistical mean: the average temperature in the high-risk cluster area was 28℃, the average relative humidity was 75%, and the average carbon dioxide concentration was 0.8%; the average temperature in the non-cluster area was 24℃, the average relative humidity was 68%, and the average carbon dioxide concentration was 0.55%; the average temperature in the low-risk cluster area was 23℃, the average relative humidity was 66%, and the average carbon dioxide concentration was 0.52%. Difference comparison: the temperature, relative humidity, and carbon dioxide concentration in the high-risk cluster area were significantly higher than those in the non-cluster area, with the temperature and humidity differences being the most significant. Establishing a mapping relationship: based on the comparison results, a mapping relationship was established that showed a significant correlation between the high-risk cluster area and the high-temperature and high-humidity microenvironment, that is, under high-temperature and high-humidity conditions, the risk of mold growth is more likely to form a clustered distribution.

[0104] The calculation steps for the local anomalous diffusion potential include:

[0105] High-risk cluster nodes identified based on the local Moran index were used as diffusion source nodes;

[0106] For each diffusion source node, spatially adjacent nodes are identified based on the spatial weight matrix;

[0107] Calculate the difference in mold risk index from the diffusion source node to each spatially adjacent node, and use it as the base value of the risk gradient;

[0108] Calculate the absolute values ​​of the temperature difference, relative humidity difference, and carbon dioxide concentration difference between the diffusion source node and its spatially adjacent nodes as monitored in real time;

[0109] The absolute values ​​of the temperature difference, relative humidity difference, and carbon dioxide concentration difference are normalized and weighted summed to generate a comprehensive microenvironment driving factor.

[0110] Multiplying the basic value of the risk gradient with the comprehensive driving factor of the microenvironment yields the local anomalous diffusion potential from the current diffusion source node to the corresponding spatially adjacent node.

[0111] First, high-risk cluster nodes identified based on the local Moran index were used as diffusion source nodes. Diffusion source nodes refer to all nodes within the high-risk cluster area; they are the potential starting points for the spread of mold risk in the grain pile and also the initiating nodes for calculating local abnormal diffusion potential. For example, 22 nodes, including N45 and N46, were identified as diffusion source nodes.

[0112] Secondly, for each diffusion source node, spatially adjacent nodes are identified based on the spatial weight matrix. Spatially adjacent nodes of a diffusion source node refer to all nodes with weights greater than 0 relative to the diffusion source node, determined based on the constructed spatial weight matrix. These are the surrounding nodes that have spatial influence on the diffusion source node and are potential recipient nodes for risk diffusion. For each diffusion source node, the weight row / column corresponding to that node is retrieved from the constructed spatial weight matrix; all nodes with weight values ​​greater than 0 are selected as the spatially adjacent nodes of that diffusion source node; a list of corresponding spatially adjacent nodes is compiled for each diffusion source node. For example, taking diffusion source node N45 as an example, the weight column corresponding to N45 in the spatial weight matrix is ​​retrieved, and nodes with weights > 0 are selected as N44, N46, N25, and N65. These four nodes are the spatially adjacent nodes of N45; the spatially adjacent nodes of diffusion source node N46 include N45, N47, N26, and N66, etc.

[0113] Further, the difference in mold risk index from the source node to each spatially adjacent node is calculated as the basic risk gradient value. The basic risk gradient value refers to the difference in mold risk index between the source node and a single spatially adjacent node, reflecting the potential basic strength of risk diffusion from the source node to that adjacent node. The larger the difference, the more obvious the basic diffusion trend. For each source node, its own mold risk index value is extracted; the mold risk index values ​​of each spatially adjacent node of the source node are extracted sequentially; the basic risk gradient value pointing to that adjacent node is calculated by subtracting the risk index of the adjacent node from the risk index of the source node. For example, the mold risk index of the diffusion source node N45 is 0.65. The basic values ​​of its risk gradient to each adjacent node are calculated in turn: for N44 (0.61): 0.65-0.61=0.04; for N46 (0.63): 0.65-0.63=0.02; for N25 (0.58): 0.65-0.58=0.07; for N65 (0.60): 0.65-0.60=0.05.

[0114] Next, the absolute values ​​of the real-time monitored temperature difference, relative humidity difference, and carbon dioxide concentration difference between the diffusion source node and its spatially adjacent nodes are calculated. The absolute values ​​of the microenvironmental parameter differences refer to the absolute values ​​of the measured differences in temperature, relative humidity, and carbon dioxide concentration between the diffusion source node and its adjacent nodes. These values ​​reflect the degree of difference in the microenvironment between the two and are fundamental data for quantifying the driving role of the microenvironment in risk diffusion. For each diffusion source node, its own real-time monitored microenvironmental parameters are extracted; the same type of real-time microenvironmental parameters are then extracted from each spatially adjacent node of that diffusion source node; the absolute values ​​of the temperature difference, relative humidity difference, and carbon dioxide concentration difference between them are calculated respectively, completing the calculation of the microenvironmental differences for a set of node pairs. For example, the measured microenvironmental parameters of diffusion source node N45 are: temperature 28℃, relative humidity 75%, and carbon dioxide concentration 0.8%. Calculate the absolute value of the microenvironmental difference between it and the adjacent node N44: N44 measured parameters: temperature 26℃, relative humidity 72%, carbon dioxide concentration 0.7%; absolute value of temperature difference: |28-26|=2℃; absolute value of relative humidity difference: |75-72|=3%; absolute value of carbon dioxide concentration difference: |0.8-0.7|=0.1%.

[0115] Then, the absolute values ​​of the temperature difference, relative humidity difference, and carbon dioxide concentration difference are normalized and weighted summed to generate a comprehensive microenvironment driving factor. The comprehensive microenvironment driving factor is a comprehensive quantitative factor obtained by normalizing and weighting the absolute values ​​of the differences in temperature, relative humidity, and carbon dioxide concentration. It reflects the comprehensive driving effect of microenvironmental differences on the spread of mold risk; the larger the factor value, the stronger the driving force of the microenvironment on risk spread. For all node pairs, the absolute values ​​of temperature difference, relative humidity difference, and carbon dioxide concentration difference are normalized to the range of 0 to 1 to eliminate dimensional differences in each parameter. Weights are assigned to the three types of normalized differences. The three types of normalized differences for a single node pair are multiplied by their corresponding weights and then summed to obtain the comprehensive microenvironment driving factor for that node pair.

[0116] For example, taking the microenvironmental difference between diffusion source N45 and adjacent node N44 as an example: 2℃, 3%, and 0.1% are normalized to the 0~1 range, respectively, to obtain normalized values ​​of 0.2, 0.3, and 0.1; combined with the mold occurrence pattern, temperature and relative humidity are the core driving factors, with the weights of temperature and relative humidity both set to 0.4, and the weight of carbon dioxide concentration set to 0.2. The weighted sum is: 0.2×0.4+0.3×0.4+0.1×0.2=0.08+0.12+0.02=0.22; 0.22 is the comprehensive driving factor of the microenvironment from N45 to N44.

[0117] Finally, the basic risk gradient value is multiplied by the comprehensive microenvironmental driving factor to obtain the local anomalous diffusion potential from the current diffusion source node to its corresponding spatial neighbor. Local anomalous diffusion potential: refers to the product of the basic risk gradient value and the comprehensive microenvironmental driving factor. It is a comprehensive quantitative indicator of the potential trend and intensity of mold risk diffusion from the diffusion source node to a specific neighbor node. The larger the value, the higher the probability and intensity of risk diffusion in that direction. For each diffusion source node, the basic risk gradient value pointing to a single neighbor node and the corresponding comprehensive microenvironmental driving factor are extracted; the two are multiplied to obtain the local anomalous diffusion potential from the diffusion source node to that neighbor node; the calculation of the local anomalous diffusion potential from the diffusion source node to all neighbor nodes is completed using the above method.

[0118] For example, the baseline risk gradient value from diffusion source node N45 to its neighboring node N44 is 0.04, and the corresponding comprehensive microenvironmental driving factor is 0.22. Multiplying these two values ​​yields the local anomalous diffusion potential: 0.04 × 0.22 = 0.0088. Similarly, the baseline risk gradient value from N45 to N25 is 0.07, and the calculated comprehensive microenvironmental driving factor is 0.35. The local anomalous diffusion potential is 0.07 × 0.35 = 0.0245, indicating that the potential for risk diffusion from N45 to N25 is higher. Similarly, the local anomalous diffusion potentials from all diffusion source nodes to all neighboring nodes are calculated.

[0119] In this embodiment of the invention, a spatial weight matrix is ​​constructed using the inverse distance weighting method. Combined with global and local Moran's index analysis, the global spatial clustering pattern and local high / low risk clustering areas of mold risk are accurately identified, solving the problem that traditional detection methods cannot quantify spatial correlations. The statistical significance of spatial autocorrelation analysis is ensured through conditional random permutation test. Combined with microenvironment parameter correlation analysis, the mapping relationship between high-risk clustering and high temperature and humidity is clarified, providing a basis for the driving mechanism of risk diffusion. The calculation of local abnormal diffusion potential quantifies the potential intensity of mold risk diffusion from high-risk clustering areas to the surrounding areas, providing a core input for subsequent spatiotemporal dynamic simulation of mold risk fields. Overall, the accuracy and scientific nature of spatial analysis of mold risk are improved, providing a reliable spatial correlation basis for early warning and prevention and control decisions.

[0120] S400: Based on the coupling relationship between the local abnormal diffusion potential and the moisture content and protein content, a spatiotemporal dynamic mold risk field simulation is performed to dynamically generate the risk field distribution.

[0121] In this embodiment of the invention, based on the coupling relationship between the local abnormal diffusion potential and the moisture content and protein content, a spatiotemporal dynamic simulation of the mold risk field is performed to dynamically generate the risk field distribution. While the local abnormal diffusion potential, intrinsic quality parameters of each node, and real-time microenvironment parameters have been obtained, traditional detection methods can only achieve static extraction of risk indicators for single nodes. They cannot quantify the dynamic propagation process and global distribution characteristics of mold risk in three-dimensional space, making it difficult to support accurate forward-looking early warning and prevention. Therefore, it is necessary to couple multiple parameters to construct a risk field evolution model, combine it with initial driving forces for spatiotemporal dynamic simulation, and intuitively generate the risk field distribution, providing a global and dynamic risk visualization basis for subsequent prevention and control decisions.

[0122] Step S400 in the method provided in this embodiment of the invention includes:

[0123] Using high-risk cluster nodes as initial high-risk sources, and real-time moisture and protein content of each node as intrinsic condition parameters for the occurrence and development of local mold growth, combined with real-time monitored temperature, relative humidity and carbon dioxide concentration, a risk field evolution model based on diffusion equations is constructed.

[0124] The local anomaly diffusion potential is used as the initial driving force for risk propagation and input into the risk field evolution model;

[0125] Multi-time-step iterative calculations were performed on a three-dimensional spatial grid of the grain warehouse to simulate the dynamic propagation process of mold risk in space until the simulation results reached a preset stable state.

[0126] Output the mold risk value of each grid cell in the three-dimensional spatial grid at the final time step, as the dynamically generated risk field distribution.

[0127] First, using high-risk cluster nodes as initial high-risk sources, and the real-time moisture and protein content of each node as intrinsic condition parameters for the occurrence and development of local mold, a risk field evolution model based on the diffusion equation is constructed in conjunction with real-time monitored temperature, relative humidity, and carbon dioxide concentration. The risk field evolution model, based on the diffusion equation, is a mathematical model used to simulate the dynamic propagation and evolution of mold risk over time in three-dimensional space. The intrinsic condition parameters refer to the moisture and protein content of each node, which are the core intrinsic quality factors affecting the occurrence and development of mold. The diffusion equation is a fundamental equation used to describe the propagation law of risk in space. In this method, it is adapted to the three-dimensional space of the grain warehouse and coupled with multi-parameter optimization for risk evolution simulation.

[0128] Specifically, the model construction foundation is determined as follows: a pre-defined three-dimensional spatial grid of the grain warehouse is used as the simulation carrier, and nodes in the high-risk cluster area are used as the initial high-risk sources, with initial risk values ​​assigned; core parameters are input: the real-time moisture content and protein content of each node are used as intrinsic condition parameters, and the real-time temperature, relative humidity, and carbon dioxide concentration are used as environmental impact parameters, all of which are incorporated into the model coupling system; an evolution model is constructed: based on the diffusion equation, combined with the initial high-risk sources, intrinsic condition parameters, and environmental parameters, the model coefficients are optimized to adapt to the propagation law of mold risk, and the risk field evolution model is completed.

[0129] For example, a three-dimensional spatial grid of the grain warehouse is pre-defined as 1m×1m×1m, covering a total area of ​​50m×30m×6m. High-risk cluster nodes such as N45 and N46 are used as initial high-risk sources, with N45 assigned an initial risk value of 0.65 and N46 assigned an initial risk value of 0.63. Input the intrinsic condition parameters of each node, such as the moisture content of N45 being 14.2% and the protein content being 12.5%, and the real-time environmental parameters, such as the temperature of N45 being 28℃ and the relative humidity being 75%. Based on the diffusion equation, the model coefficients are optimized to construct a risk field evolution model adapted to the grain warehouse.

[0130] Secondly, the local anomalous diffusion potential is used as the initial driving force for risk propagation and input into the risk field evolution model. The initial driving force refers to the initial energy that propagates the mold risk from the initial high-risk source to the surrounding space. Using the local anomalous diffusion potential as the initial driving force quantifies the initial intensity and direction of risk propagation. The local anomalous diffusion potentials of all diffusion source nodes pointing to their adjacent nodes are retrieved; according to the spatial location of the diffusion source nodes, the corresponding local anomalous diffusion potentials are mapped one by one to the corresponding spatial location in the risk field evolution model; the local anomalous diffusion potential is used as the initial driving force for risk propagation, input into the model, and parameter adaptation is completed.

[0131] For example, the local anomalous diffusion potentials of diffusion source node N45 pointing to its neighboring nodes are retrieved. For instance, the diffusion potentials of diffusion source node N45 pointing to N44 are 0.0088, pointing to N25 are 0.0245, pointing to N46 are 0.0044, and pointing to N65 are 0.011. Based on the coordinates (25m, 15m, 3m) of N45 in the three-dimensional spatial grid, the diffusion potentials in the above four directions are mapped to the corresponding spatial positions of the model, serving as the initial driving force for the spread of risk around N45. Similarly, the local anomalous diffusion potentials of other diffusion source nodes such as N46 and N65 are input into the model.

[0132] Furthermore, iterative calculations with multiple time steps are performed on a three-dimensional spatial mesh of the grain silo to simulate the dynamic propagation of mold risk in space until the simulation results reach a preset stable state. The three-dimensional spatial mesh refers to a preset three-dimensional mesh structure covering the entire grain silo, used to carry the simulation results of the risk field; each mesh cell corresponds to a risk value at a spatial location. The time step refers to the time interval of the dynamic simulation of the risk field, used to simulate the gradual propagation of risk over time. The stable state refers to a state where, within multiple consecutive time steps, the change in the mold risk value of each mesh cell in the three-dimensional spatial mesh is less than a preset threshold, and the simulation results no longer fluctuate significantly.

[0133] Specifically, the simulation time step is set to 1 hour / step to adapt to the rate of mold risk propagation; a stability threshold is set: the continuous change range of the risk value of each grid cell is <0.001; based on the risk field evolution model, and using the input initial driving force, intrinsic condition parameters, and environmental parameters as a basis, the risk propagation calculation for the first time step is performed on the three-dimensional spatial grid to obtain the risk field distribution for that time step; following the same logic, iterative calculations are performed for subsequent time steps. After each time step calculation is completed, the risk field distribution of the current time step is compared with that of the previous time step to determine whether a stable state has been reached; if the stability threshold is met for 5 consecutive time steps, the iterative calculation stops; if not, iteration continues until stability is achieved. For example, based on the model, iterative calculations are performed: after the first hour step, the risk value of the grid cells surrounding N45 increases by 0.02 to 0.05 compared to the initial state; after the tenth hour step, the risk spreads to a range of 10m; when the iterative calculation reaches the 30-hour step, the change in the risk value of all grid cells is less than 0.001 within five consecutive time steps, indicating that the simulation has reached a stable state and the iteration is stopped.

[0134] Finally, the mold risk value of each grid cell in the three-dimensional spatial grid at the final time step is output as the dynamically generated risk field distribution. The risk field distribution refers to the global distribution of the mold risk value corresponding to each grid cell in the three-dimensional spatial grid after the simulation reaches a stable state. It can intuitively present the spatial distribution characteristics of mold risk and the range of high-risk areas within the grain silo. The mold risk values ​​of all grid cells in the three-dimensional spatial grid at the last time step are extracted; the risk values ​​are associated with spatial locations one by one according to the three-dimensional spatial coordinates of each grid cell to form global risk field distribution data; and the dynamically generated risk field distribution is output.

[0135] For example, extract the risk value of the grid cell at a step size of 30 hours, where: the risk value of the grid cell where N45 is located is 0.64; the risk value of the grid cells within a 10m radius around it is 0.4~0.6; the risk value of the grid cells in the edge area of ​​the grain warehouse is 0.1~0.3; associate the risk value of all grid cells with coordinates, and output the dynamically generated distribution of the mold risk field of the entire grain warehouse.

[0136] In this embodiment of the invention, a risk field evolution model adapted to the grain warehouse scenario is constructed by coupling local abnormal diffusion potential, intrinsic quality parameters, and environmental parameters, realizing spatiotemporal dynamic simulation of mold risk. Based on a preset three-dimensional spatial grid and multi-time step iterative calculation until the simulation stabilizes, the accuracy and reliability of the risk field distribution are guaranteed. The finally dynamically generated risk field distribution intuitively presents the whole-domain propagation process and spatial distribution characteristics of mold risk in the grain warehouse, solving the pain point that traditional detection cannot dynamically quantify risk propagation and whole-domain distribution. It provides an intuitive and reliable whole-domain risk basis for subsequent accurate early warning of high-risk areas and the formulation of targeted prevention and control measures, which meets the dynamic needs of edge collaborative distributed detection.

[0137] S500: Based on the risk field distribution, perform spatial data fusion and state optimization assessment on the multidimensional initial index set of all nodes to generate a global collaborative quality report.

[0138] In this embodiment of the invention, based on the risk field distribution, spatial data fusion and optimal state assessment are performed on the multidimensional initial indicator sets of all nodes to generate a global collaborative quality report. While the spatiotemporal dynamic simulation of the mold risk field has been completed and the risk field distribution obtained, the multidimensional initial indicator sets of each node are still discrete single-point data, unable to form a continuous spatial distribution covering the entire grain depot area. Simultaneously, traditional grain condition monitoring can only output isolated indicator values, lacking optimal assessment of the overall quality status and visualized, actionable control suggestions, making it difficult to support global collaborative management of grain depots. Therefore, it is necessary to use the risk field distribution as a spatial constraint and achieve spatial data fusion of multidimensional indicators through co-kriging spatial interpolation to complete the optimal assessment of the overall quality status, ultimately generating a global collaborative quality report containing a visualized distribution map and control suggestions, providing a comprehensive, accurate, and operable basis for prevention and control decisions.

[0139] Step S500 in the method provided in this embodiment of the invention includes:

[0140] Using the risk field distribution as a spatial constraint, co-kriging spatial interpolation is performed on the multidimensional initial index set to generate a globally unified spatial distribution of mold risk, moisture, and protein.

[0141] Based on the spatial distribution of mold risk, high-risk areas with risk levels exceeding a preset mold threshold are identified, and their corresponding spatial locations are marked.

[0142] Based on the spatial distribution of moisture and protein, and combined with the mapping relationship of high temperature and high humidity microenvironment parameters, the quality status and potential mold induction conditions of each region are evaluated.

[0143] By integrating the spatial distribution of mold risk, spatial distribution of moisture, spatial distribution of protein, location of high-risk areas, and quality assessment results, a global collaborative quality report is generated, which includes a mold risk heatmap, a key quality distribution map, and comprehensive control recommendations.

[0144] First, using the risk field distribution as a spatial constraint, co-kriging spatial interpolation is performed on the multidimensional initial index set to generate a globally unified spatial distribution of mold risk, moisture, and protein.

[0145] Specifically, using the risk field distribution as a spatial constraint, co-kriging spatial interpolation is performed on the multidimensional initial index set to generate a globally unified spatial distribution of mold risk, moisture content, and protein content, including:

[0146] The risk field distribution is used as the main variable field characterizing the spatial structure of mold risk, and the mold risk index at all nodes is extracted from the multidimensional initial index set as the observed value of the main variable.

[0147] The moisture content and protein content at all nodes were extracted from the multidimensional initial index set and used as the first covariate observation value and the second covariate observation value, respectively.

[0148] Based on the main variable field, the experimental variability function of the mold risk index was calculated and fitted.

[0149] Based on the observed values ​​of the first and second covariates, the experimental variation functions of the moisture content and the protein content were calculated and fitted respectively.

[0150] Based on the spatial correspondence between the observed values ​​of the main variables and the observed values ​​of the covariates, the cross-experimental variability functions of the mold risk index and moisture content, and the mold risk index and protein content were calculated and fitted.

[0151] Based on the experimental variance functions of the mold risk index, the moisture content, the protein content, and the cross-variance function obtained from the fitting, a set of co-kriging equations is constructed.

[0152] Solving the co-kriging equations yields the optimal linear unbiased estimates of the mold risk index, moisture content, and protein content at any unobserved location in the three-dimensional spatial grid of the grain warehouse, and generates globally continuous spatial distributions of the mold risk, moisture, and protein, respectively.

[0153] First, the risk field distribution is used as the main variable field characterizing the spatial structure of mold risk. Mold risk indices at all nodes are extracted from the multidimensional initial index set as observed main variables. The main variable field refers to the risk field distribution output by S400, which serves as a reference field characterizing the spatial structure of mold risk and constrains the spatial structure characteristics of the interpolation process. The observed main variables are the mold risk indices at all nodes extracted from the multidimensional initial index set, which are the main variables for interpolation. The risk field distribution generated by S400 is retrieved and determined as the main variable field characterizing the spatial structure of mold risk. The mold risk indices at all nodes are extracted from the multidimensional initial index set and arranged into a one-dimensional array as the set of observed main variables. For example, using the risk field distribution as the main variable field, the mold risk indices at 90 nodes, such as N1 and N2, are extracted as observed main variables.

[0154] Secondly, the moisture content and protein content at all nodes are extracted from the multidimensional initial index set and used as the first and second covariate observations, respectively. The covariate observations refer to the moisture content and protein content of all nodes extracted from the multidimensional initial index set, which have a spatial correlation with the main variable, thus helping to improve interpolation accuracy. The moisture content and protein content of all nodes extracted from the multidimensional initial index set are arranged into one-dimensional arrays, serving as the sets of the first and second covariate observations. For example, the moisture content of nodes N1, N2, etc., is extracted as the first covariate observation; the protein content of nodes N1, N2, etc., is extracted as the second covariate observation.

[0155] Further, based on the main variable field, the experimental variability function of the mold risk index is calculated and fitted. The experimental variability function is used to quantify the degree of variation of spatial variables at different spatial distances, reflecting the spatial correlation of variables. Distance grouping refers to dividing all node pairs into several groups according to actual spatial distances, with each group corresponding to a distance interval, used to statistically analyze the variable variation characteristics at different distances. All nodes are paired to form non-repeating node pairs, and distance groups are divided according to actual three-dimensional spatial distances; the average difference within each group is calculated to obtain the experimental variability function value: for each distance group, the square of the difference in the mold risk index of each pair of nodes within the group is calculated sequentially, and all squared values ​​are summed and divided by twice the number of node pairs in the group to obtain the experimental variability function value corresponding to that distance group; the distance-experimental variability function values ​​of each distance group are plotted as a scatter plot, and fitted using classical variability function models such as spherical functions / exponential functions to obtain the experimental variability function of the mold risk index.

[0156] For example, the node pairs are grouped into 0-1m, 1-2m…10-11m groups, where the 0-1m group contains 12 node pairs such as N45-N46, and the 1-2m group contains 18 node pairs such as N45-N44. Taking the 0-1m group as an example, the squared difference of the mold risk index for each node pair is calculated, and the sum is 0.0056. Dividing by 12×2=24, the experimental variability function value for this group is approximately 0.00023. The function values ​​for each distance group are plotted as a scatter plot, and the experimental variability function of the mold risk index is fitted as a spherical function model: γ(h)=0.08×(1.5×h / 5-0.5×(h / 5)). 3 ) (h≤5m), γ(h)=0.08 (h>5m).

[0157] Furthermore, based on the observed values ​​of the first and second covariates, the experimental variability functions of the moisture content and the protein content were calculated and fitted, respectively. The experimental variability functions of the moisture content and protein content quantify the degree of variation of moisture content and protein content at different spatial distances, respectively. The fitting method is consistent with the mold risk index, reflecting the spatial correlation of the two types of indicators. The following operations were performed on the moisture content and protein content: following the distance grouping rules described above, all node pairs were divided into the same distance groups according to their three-dimensional spatial distance; for each distance group, the square of the difference in moisture content or protein content between each pair of nodes within the group was calculated, summed, and divided by twice the number of node pairs in that group to obtain the experimental variability function value for the corresponding distance group; the distance-function values ​​were plotted as a scatter plot, and the experimental variability functions of the moisture content and protein content were fitted to obtain these values.

[0158] For example, taking moisture content as an example: the distance group is completely consistent with the distance group of the mold risk index; the square of the moisture difference between N45 and N46 in the 0-1m group is (14.2-14.0). 2 =0.04, the sum of all squares in the group is 0.52, divided by 24, the function value of the group is ≈0.0217; the experimental variation function of water content is fitted as an exponential function model: γ(h)=0.12×(1-e^(-h / 3)); similarly, the experimental variation function of protein content is fitted as γ(h)=0.05×(1-e^(-h / 4)).

[0159] Furthermore, based on the spatial correspondence between the observed values ​​of the main variables and the observed values ​​of the covariates, cross-experimental variograms (CEVs) for the mold risk index and moisture content, and for the mold risk index and protein content, are calculated and fitted. These CEVs include two types: mold risk index-moisture content and mold risk index-protein content. They are used to quantify the degree of spatial covariance between the main variables and covariates, reflect their spatial correlation characteristics, and improve the accuracy of co-kriging interpolation. The cross-experimental variability functions (CEVs) for the mold risk index-moisture content and mold risk index-protein content are calculated as follows: Based on the three-dimensional spatial coordinates of all nodes, they are grouped according to a preset distance interval; within each distance group, node pairs with both nodes having observed values ​​for mold risk index and moisture / protein content are selected; for each valid node pair, the difference in mold risk index and the difference in moisture content or protein content are calculated, and then their product is calculated; the sum of the products of the differences of all node pairs in the group is divided by twice the number of node pairs in that group to obtain the CEV value corresponding to that distance group; the distance-CEV values ​​are plotted as a scatter plot, and the two types of CEVs are fitted to obtain the CEVs.

[0160] For example, taking the cross function of mold risk index-moisture content as an example: the grouping rules of 0-1m, 1-2m, etc. are followed; all 12 node pairs in the 0-1m group have complete observation values ​​and are all retained; the mold risk difference of N45-N46 = 0.65-0.63 = 0.02, the moisture difference = 14.2-14.0 = 0.2, the product = 0.02×0.2 = 0.004; the sum of all products in the group is 0.038, divided by 24, the cross function value of the group is ≈0.00158; the experimental variability function of mold risk-moisture content is obtained by fitting γ(h) = 0.04×(1-e^(-h / 3.5)); similarly, the cross function of mold risk-protein content is obtained by fitting γ(h) = 0.02×(1-e^(-h / 4.2)).

[0161] Based on this, a co-kriging equation system is constructed using the experimental variograms of the mold risk index, moisture content, protein content, and cross-variograms obtained from the fitting. The co-kriging equation system is a linear system of equations constructed based on the experimental variograms of the main variables, covariates, and cross-variograms, and is a mathematical model for achieving optimal linear unbiased estimation. All fitting results are integrated: the experimental variograms of the mold risk index, moisture content, and protein content, as well as the cross-variograms of mold risk-moisture and mold risk-protein are summarized; based on the principle of optimal linear unbiased estimation using spatial interpolation, these functions are substituted into the basic equation framework of co-kriging to construct a simultaneous equation system containing the main variables and covariates, i.e., the co-kriging equation system.

[0162] For example, by integrating the experimental variograms of mold risk, moisture, and protein, and the two types of cross functions, and substituting them into the co-kriging equation framework, a system of equations is constructed for a three-dimensional grid of grain storage. For any unobserved grid cell x0, its estimated mold risk value ŷ(x0) = Σλ i ×y(x i )+Σμ j ×w1(x j )+Σν k ×w2(x k ), where λ i μ represents the principal variable weighting coefficient of the i-th edge sensing node for the unobserved grid cell x0. j νj represents the weighting coefficient of the first collaborative variable of the j-th edge sensing node with respect to the unobserved grid cell x0, and νk represents the weighting coefficient of the second collaborative variable of the k-th edge sensing node with respect to the unobserved grid cell x0. j ) represents the j-th edge sensing node in its spatial coordinates x. j Measured moisture content at w2(x) k ), representing the k-th edge sensing node in its spatial coordinates x. k The measured protein content at the location. Combined with the variogram constraint, a system of linear equations including weighting coefficients is formed, which is the co-kriging equation system.

[0163] Finally, the co-kriging equations are solved to obtain the optimal linear unbiased estimates of the mold risk index, moisture content, and protein content at any unobserved location in the three-dimensional spatial grid of the grain warehouse, and globally continuous spatial distributions of the mold risk, moisture, and protein are generated respectively. The optimal linear unbiased estimate refers to the index estimate obtained by solving the co-kriging equations at the unobserved location, which satisfies the characteristics that the estimated mean equals the true mean and the estimated variance is minimized.

[0164] Specifically, numerical methods such as Gaussian elimination or conjugate gradient method are used to solve the co-kriging equations to obtain the weight coefficients corresponding to each unobserved grid cell. Based on the weight coefficients and the observed values ​​at the nodes, the optimal linear unbiased estimates of the mold risk index, moisture content, and protein content at all unobserved locations in the three-dimensional spatial grid of the grain warehouse are calculated. The optimal linear unbiased estimates of all three-dimensional grid cells are summarized, including the observed values ​​at the node locations and the estimated values ​​at the unobserved locations. The estimated value of each grid cell is associated with its three-dimensional coordinates to generate spatial distributions of mold risk, moisture content, and protein content, respectively. All three types of distributions are globally unified and continuous spatial data.

[0165] For example, taking the unobserved cell with coordinates (25.5m, 15.5m, 3m) in the three-dimensional grid of the grain warehouse as an example, the weight coefficients λ1=0.6, μ1=0.15, ν1=0.05 are obtained by solving the system of equations. Combined with the node observation values, the estimated value of mold risk is ≈0.64; the estimated value of moisture content is ≈14.1%; and the estimated value of protein content is ≈12.4%. The estimated values ​​of all 1m×1m×1m grid cells in a 50m×30m×6m grain warehouse are summarized as follows: Spatial distribution of mold risk: 0.6~0.65 for the surrounding grid cells in the core mold risk zone N45, 0.4~0.6 for the diffusion zone, and 0.1~0.3 for the edge zone; Spatial distribution of moisture: 14.0%~14.2% in the core mold risk zone, 13.5%~14.0% in the diffusion zone, and 12.5%~13.5% in the edge zone; Spatial distribution of protein: small fluctuations across the entire region, 12.3%~12.5% ​​in the core mold risk zone, and 11.8%~12.2% in the edge zone; All three types of data are correlated by coordinates to form a globally continuous spatial distribution.

[0166] Secondly, based on the spatial distribution of mold risk, high-risk areas with risk levels exceeding a preset mold threshold are identified, and their corresponding spatial locations are marked. A high-risk mold area refers to a continuous region composed of grid cells in the spatial distribution of mold risk whose risk level exceeds the preset mold threshold. The preset mold threshold is set based on grain depot safety management standards; exceeding the preset mold threshold is considered high-risk. The preset mold risk threshold is 0.5. All grid cells in the spatial distribution of mold risk are traversed, and grid cells with risk values ​​> 0.5 are selected. The selected grid cells are then integrated into a continuous region according to three-dimensional coordinates, and their spatial range is marked. For example, after traversing the grid cells, grid cells with risk values ​​> 0.5 are concentrated in the range of x: 20-30m, y: 10-20m, and z: 2-4m.

[0167] Furthermore, based on the spatial distribution of moisture and protein, and combined with the mapping relationship of high temperature and high humidity microenvironment parameters, the quality status and potential mold induction conditions of each region are assessed. Quality status assessment combines the spatial distribution of moisture and protein to evaluate whether the grain quality of each region meets safety standards. Potential mold induction conditions refer to analyzing whether each region possesses the inherent quality conditions and external environmental conditions for mold growth, based on the mapping relationship between moisture / protein distribution and high temperature and high humidity microenvironment. Quality Status Assessment: A preset safety threshold of 13.5% for moisture content and a reference threshold of 12.0% for protein content are established. The spatial distribution of moisture and protein is assessed according to the following rules: Moisture > 13.5% and Protein > 12.0%: Area prone to mold growth; Moisture ≤ 13.5% and Protein ≤ 12.0%: Safe quality area; Other combinations: Quality warning area. Potential Mold Inducing Condition Analysis: Combining the mapping relationship between high-risk clustering and high temperature and humidity, the moisture / protein distribution in each area is correlated with real-time temperature and humidity distribution, analyzed according to the following rules: Moisture > 13.5% + Temperature > 25℃ + Relative Humidity > 70%: High inducing condition area; Meeting only 1-2 conditions: Medium inducing condition area; Not meeting any of the conditions: Low inducing condition area.

[0168] For example, regarding quality conditions: High-risk areas with moisture content of 14.0%~14.2% > 13.5% and protein content of 12.3%~12.5% ​​> 12.0% are classified as areas prone to mold growth; peripheral areas with moisture content of 12.5%~13.5% and protein content of 11.8%~12.0% are classified as quality-safe areas; and grain storage corner areas with moisture content of 13.5%~14.0% and protein content of 11.9%~12.1% are classified as quality warning areas. Inducing conditions: High-risk areas with temperatures of 28℃ > 25℃ and relative humidity of 75% > 70%, combined with high moisture content, are classified as high-inducing-condition areas; quality warning areas with temperatures of 24-25℃ and relative humidity of 68-70% are classified as medium-inducing-condition areas; and quality-safe areas with temperatures of 22-24℃ and relative humidity of 65-68% are classified as low-inducing-condition areas.

[0169] Finally, by integrating the spatial distribution of mold risk, moisture content, protein content, high-risk area locations, and quality assessment results, a comprehensive collaborative quality report is generated, including a mold risk heatmap, a key quality distribution map, and integrated control recommendations. This comprehensive report integrates the spatial distribution of multi-dimensional indicators across the entire region, high-risk areas, and quality assessment results. It includes visual charts and actionable control recommendations, serving as the basis for grain depot management decisions. Visualization involves converting the spatial distribution of mold risk into a mold risk heatmap, using red / orange / yellow / green to indicate high / medium / low / no risk respectively, and converting the spatial distributions of moisture and protein into key quality distribution maps, marking numerical ranges and safety thresholds. Differentiated control recommendations are formulated for different risk levels, quality conditions, and inducing conditions in different regions, focusing on high-risk areas. The mold risk heatmap, key quality distribution map, and integrated control recommendations are integrated into a complete comprehensive collaborative quality report.

[0170] For example, the content of the global collaborative quality report is as follows: Visual charts: Mold risk heat map: The core mold risk area is dark red, the diffusion area is orange, and the edge area is green; Key quality distribution map: The moisture distribution map is marked with the 13.5% safety threshold line, the core mold risk area is dark yellow, and the protein distribution map is light yellow throughout, with only the core mold risk area slightly exceeding the reference threshold; Comprehensive control recommendations: High-risk / high-inducing condition area: Turn on the local ventilation system in this area, lower the temperature to below 25℃ and the relative humidity to below 70%, and take samples to test the mold risk index daily for 7 consecutive days; Quality warning area / medium-inducing condition area: Monitor temperature and humidity every 6 hours, adjust the ventilation frequency to once every 12 hours, and take samples to test once every 3 days; Quality safe area / low-inducing condition area: Maintain the existing ventilation frequency, once every 24 hours, and conduct inspections once a week, without additional intervention.

[0171] In this embodiment of the invention, risk field distribution is used as a spatial constraint, and co-Kriging spatial interpolation is used to fuse discrete multidimensional initial indicators into a continuous spatial distribution across the entire domain, solving the problem that single-point data cannot reflect the overall quality characteristics. Through quality status and inducing conditions assessment, the optimal state assessment of grain conditions across the entire domain is completed, clarifying the mold risk level and inducing conditions in different regions. The final generated global collaborative quality report integrates a visualized distribution map and targeted control suggestions, realizing an upgrade from single-point detection to global collaborative control, providing accurate and implementable decision-making basis for mold prevention and control in grain depots, and improving the comprehensiveness and scientific nature of grain quality control.

[0172] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects:

[0173] This invention provides a distributed detection method and system for multiple indicators of grain quality based on edge collaboration. A basic sensing network is constructed through node deployment and data acquisition. Accurate synchronous acquisition is achieved through model training and multi-dimensional indicator extraction. Risk space correlation analysis and diffusion potential calculation quantify the correlation and diffusion trend of risk space. The entire mold risk propagation process is dynamically presented through risk field simulation. Finally, data fusion and report generation provide a visualized basis for comprehensive control and regulation schemes. This invention achieves a technological upgrade from single-point static detection to comprehensive dynamic control, from isolated indicator extraction to multi-parameter collaborative analysis, and from qualitative risk judgment to quantitative simulation and precise control. It effectively improves the accuracy, comprehensiveness, and timeliness of grain quality mold risk detection, providing strong technical support for the scientific prevention and control of mold in grain depots and ensuring grain quality and safety.

[0174] Example 2, as Figure 2 As shown, this invention provides a distributed detection system for multiple indicators of grain quality based on edge collaboration, the system comprising:

[0175] The multi-source data synchronous acquisition module 11 is used to simultaneously acquire multispectral response data of the target grain pile and microenvironmental parameters of the node location from multiple edge spectral sensing nodes deployed in the grain warehouse.

[0176] The quality feature extraction module 12 is used to perform real-time calibration of the multispectral response data based on the microenvironment parameters, and input the data into the quality feature extraction model to obtain a multidimensional initial index set for each node, wherein the multidimensional initial index set includes mold risk index, moisture content and protein content.

[0177] The abnormal diffusion analysis module 13 is used to analyze the spatial correlation of the mold risk index in the node network by combining the microenvironment parameters and spatial topology of each node, and to calculate the local abnormal diffusion potential.

[0178] The mold risk field simulation module 14 is used to perform spatiotemporal dynamic mold risk field simulation based on the coupling relationship between the local abnormal diffusion potential and the moisture content and protein content, and dynamically generate the risk field distribution.

[0179] The collaborative quality report output module 15 is used to perform spatial data fusion and state optimization assessment on the multidimensional initial index set of all nodes based on the risk field distribution, and generate a global collaborative quality report.

[0180] In one embodiment, the multi-source data synchronization acquisition module 11 is further configured to:

[0181] The multispectral response data includes conventional spectral data in the visible-near-infrared band, as well as specific narrow-band spectral data collected with targeted enhancement for the characteristic absorption bands of fungal metabolites such as aflatoxin. The microenvironmental parameters include temperature, relative humidity, and carbon dioxide concentration monitored in real time at the nodes.

[0182] In one embodiment, the abnormal diffusion analysis module 13 is further configured to:

[0183] The spatial correlation of the mold risk index in the node network is analyzed by combining the microenvironmental parameters and spatial topology of each node, including:

[0184] Based on the spatial coordinates of all nodes, a spatial weight matrix is ​​constructed using the inverse distance weighting method;

[0185] Using the spatial weight matrix, the global Moran index of the mold risk index is calculated, and the degree of spatial autocorrelation in the overall network is evaluated.

[0186] For each node, the local Moran index is calculated to identify spatially significant clustering regions, which include high-risk clustering regions and low-risk clustering regions.

[0187] The identified spatial clustering areas are correlated with the microenvironmental parameters of the corresponding nodes to establish a mapping relationship between high-risk clustering and high-temperature and high-humidity microenvironmental parameters.

[0188] Specifically, the global Moran's index of the mold risk index is calculated using the spatial weight matrix, and its spatial autocorrelation in the overall network is evaluated, including:

[0189] Calculate the mean and variance of the mold risk index for all nodes;

[0190] Based on the spatial weight matrix, calculate the weighted covariance sum of the deviations of the mold risk index from the mean among all node pairs;

[0191] The global Moran index value is calculated by substituting the mean, variance, and weighted covariance into the global Moran index formula.

[0192] The global Moran index value is compared with the zero distribution generated by the random permutation test to perform a significance test;

[0193] Based on the significance test results and the sign and magnitude of the global Moran index, it is determined whether the mold risk index has a significant spatial clustering distribution pattern, spatial discrete distribution pattern, or random distribution pattern in the grain warehouse node network.

[0194] The comparison of the global Moran index value with the zero distribution generated through a random permutation test, and the significance test, includes:

[0195] Keeping the spatial weight matrix unchanged, the spatial positions of the mold risk index values ​​among all nodes are randomly rearranged multiple times.

[0196] For each new spatial data generated after random permutation, the corresponding global Moran index value is recalculated.

[0197] All global Moran index values ​​obtained from multiple random permutations are summarized to generate the zero distribution, and the mean and standard deviation of the zero distribution are calculated.

[0198] The calculated actual global Moran index value is compared with the mean of the null distribution to calculate the Z score;

[0199] Based on the Z-score and the standard normal distribution table, determine the P-value corresponding to the actual global Moran's index value;

[0200] The P-value is compared with a preset significance level threshold. If the P-value is less than the significance level threshold, it is determined that the mold risk index has a significant spatial autocorrelation pattern in the grain warehouse node network.

[0201] The calculation steps for the local anomalous diffusion potential include:

[0202] High-risk cluster nodes identified based on the local Moran index were used as diffusion source nodes;

[0203] For each diffusion source node, spatially adjacent nodes are identified based on the spatial weight matrix;

[0204] Calculate the difference in mold risk index from the diffusion source node to each spatially adjacent node, and use it as the base value of the risk gradient;

[0205] Calculate the absolute values ​​of the temperature difference, relative humidity difference, and carbon dioxide concentration difference between the diffusion source node and its spatially adjacent nodes as monitored in real time;

[0206] The absolute values ​​of the temperature difference, relative humidity difference, and carbon dioxide concentration difference are normalized and weighted summed to generate a comprehensive microenvironment driving factor.

[0207] Multiplying the basic value of the risk gradient with the comprehensive driving factor of the microenvironment yields the local anomalous diffusion potential from the current diffusion source node to the corresponding spatially adjacent node.

[0208] In one embodiment, the mold risk field simulation module 14 is also used for:

[0209] Using high-risk cluster nodes as initial high-risk sources, and real-time moisture and protein content of each node as intrinsic condition parameters for the occurrence and development of local mold growth, combined with real-time monitored temperature, relative humidity and carbon dioxide concentration, a risk field evolution model based on diffusion equations is constructed.

[0210] The local anomaly diffusion potential is used as the initial driving force for risk propagation and input into the risk field evolution model;

[0211] Multi-time-step iterative calculations were performed on a three-dimensional spatial grid of the grain warehouse to simulate the dynamic propagation process of mold risk in space until the simulation results reached a preset stable state.

[0212] Output the mold risk value of each grid cell in the three-dimensional spatial grid at the final time step, as the dynamically generated risk field distribution.

[0213] In one embodiment, the collaborative quality report output module 15 is further configured to:

[0214] Using the risk field distribution as a spatial constraint, co-kriging spatial interpolation is performed on the multidimensional initial index set to generate a globally unified spatial distribution of mold risk, moisture, and protein.

[0215] Based on the spatial distribution of mold risk, high-risk areas with risk levels exceeding a preset mold threshold are identified, and their corresponding spatial locations are marked.

[0216] Based on the spatial distribution of moisture and protein, and combined with the mapping relationship of high temperature and high humidity microenvironment parameters, the quality status and potential mold induction conditions of each region are evaluated.

[0217] By integrating the spatial distribution of mold risk, spatial distribution of moisture, spatial distribution of protein, location of high-risk areas, and quality assessment results, a global collaborative quality report is generated, which includes a mold risk heatmap, a key quality distribution map, and comprehensive control recommendations.

[0218] Specifically, using the risk field distribution as a spatial constraint, co-kriging spatial interpolation is performed on the multidimensional initial index set to generate a globally unified spatial distribution of mold risk, moisture content, and protein content, including:

[0219] The risk field distribution is used as the main variable field characterizing the spatial structure of mold risk, and the mold risk index at all nodes is extracted from the multidimensional initial index set as the observed value of the main variable.

[0220] The moisture content and protein content at all nodes were extracted from the multidimensional initial index set and used as the first covariate observation value and the second covariate observation value, respectively.

[0221] Based on the main variable field, the experimental variability function of the mold risk index was calculated and fitted.

[0222] Based on the observed values ​​of the first and second covariates, the experimental variation functions of the moisture content and the protein content were calculated and fitted respectively.

[0223] Based on the spatial correspondence between the observed values ​​of the main variables and the observed values ​​of the covariates, the cross-experimental variability functions of the mold risk index and moisture content, and the mold risk index and protein content were calculated and fitted.

[0224] Based on the experimental variance functions of the mold risk index, the moisture content, the protein content, and the cross-variance function obtained from the fitting, a set of co-kriging equations is constructed.

[0225] Solving the co-kriging equations yields the optimal linear unbiased estimates of the mold risk index, moisture content, and protein content at any unobserved location in the three-dimensional spatial grid of the grain warehouse, and generates globally continuous spatial distributions of the mold risk, moisture, and protein, respectively.

[0226] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A distributed detection method for multiple indicators of grain quality based on edge collaboration, characterized in that, The method includes: Multiple edge spectral sensing nodes are deployed inside the grain warehouse to simultaneously collect multispectral response data of the target grain pile and microenvironmental parameters at the location of the nodes. Based on the microenvironment parameters, the multispectral response data is calibrated in real time and input into the quality feature extraction model to obtain a multidimensional initial index set for each node, wherein the multidimensional initial index set includes mold risk index, moisture content and protein content. By combining the microenvironmental parameters and spatial topology of each node, the spatial correlation of the mold risk index in the node network is analyzed, and the local abnormal diffusion potential is calculated. Based on the coupling relationship between the local abnormal diffusion potential and the moisture content and protein content, a spatiotemporal dynamic mold risk field simulation is performed to dynamically generate the risk field distribution. Based on the risk field distribution, spatial data fusion and state optimization assessment are performed on the multidimensional initial index set of all nodes to generate a global collaborative quality report.

2. The distributed detection method for multiple indicators of grain quality based on edge collaboration according to claim 1, characterized in that, The multispectral response data includes conventional spectral data in the visible-near-infrared band, as well as specific narrow-band spectral data collected with targeted enhancement for the characteristic absorption bands of fungal metabolites such as aflatoxin. The microenvironmental parameters include temperature, relative humidity, and carbon dioxide concentration monitored in real time at the nodes.

3. The distributed detection method for multiple indicators of grain quality based on edge collaboration according to claim 1, characterized in that, Based on the microenvironmental parameters and spatial topology of each node, the spatial correlation of the mold risk index in the node network is analyzed, including: Based on the spatial coordinates of all nodes, a spatial weight matrix is ​​constructed using the inverse distance weighting method; Using the spatial weight matrix, the global Moran index of the mold risk index is calculated, and the degree of spatial autocorrelation in the overall network is evaluated. For each node, the local Moran index is calculated to identify spatially significant clustering regions, which include high-risk clustering regions and low-risk clustering regions. The identified spatial clustering areas are correlated with the microenvironmental parameters of the corresponding nodes to establish a mapping relationship between high-risk clustering and high-temperature and high-humidity microenvironmental parameters.

4. The distributed detection method for multiple indicators of grain quality based on edge collaboration according to claim 3, characterized in that, Using the spatial weight matrix, the global Moran index of the mold risk index is calculated, and its spatial autocorrelation in the overall network is evaluated, including: Calculate the mean and variance of the mold risk index for all nodes; Based on the spatial weight matrix, calculate the weighted covariance sum of the deviations of the mold risk index from the mean among all node pairs; The global Moran index value is calculated by substituting the mean, variance, and weighted covariance into the global Moran index formula. The global Moran index value is compared with the zero distribution generated by the random permutation test to perform a significance test; Based on the significance test results and the sign and magnitude of the global Moran index, it is determined whether the mold risk index has a significant spatial clustering distribution pattern, spatial discrete distribution pattern, or random distribution pattern in the grain warehouse node network.

5. The distributed detection method for multiple indicators of grain quality based on edge collaboration according to claim 4, characterized in that, The global Moran index value is compared with the zero distribution generated by the random permutation test to perform a significance test, including: Keeping the spatial weight matrix unchanged, the spatial positions of the mold risk index values ​​among all nodes are randomly rearranged multiple times. For each new spatial data generated after random permutation, the corresponding global Moran index value is recalculated. All global Moran index values ​​obtained from multiple random permutations are summarized to generate the zero distribution, and the mean and standard deviation of the zero distribution are calculated. The calculated actual global Moran index value is compared with the mean of the null distribution to calculate the Z score; Based on the Z-score and the standard normal distribution table, determine the P-value corresponding to the actual global Moran's index value; The P-value is compared with a preset significance level threshold. If the P-value is less than the significance level threshold, it is determined that the mold risk index has a significant spatial autocorrelation pattern in the grain warehouse node network.

6. The distributed detection method for multiple indicators of grain quality based on edge collaboration according to claim 1, characterized in that, The calculation steps for the local anomalous diffusion potential include: High-risk cluster nodes identified based on the local Moran index were used as diffusion source nodes; For each diffusion source node, spatially adjacent nodes are identified based on the spatial weight matrix; Calculate the difference in mold risk index from the diffusion source node to each spatially adjacent node, and use it as the base value of the risk gradient; Calculate the absolute values ​​of the temperature difference, relative humidity difference, and carbon dioxide concentration difference between the diffusion source node and its spatially adjacent nodes as monitored in real time; The absolute values ​​of the temperature difference, relative humidity difference, and carbon dioxide concentration difference are normalized and weighted summed to generate a comprehensive microenvironment driving factor. Multiplying the basic value of the risk gradient with the comprehensive driving factor of the microenvironment yields the local anomalous diffusion potential from the current diffusion source node to the corresponding spatially adjacent node.

7. The distributed detection method for multiple indicators of grain quality based on edge collaboration according to claim 1, characterized in that, Based on the coupling relationship between the local anomalous diffusion potential and the moisture content and protein content, a spatiotemporal dynamic simulation of the mold risk field is performed to dynamically generate the risk field distribution, including: Using high-risk cluster nodes as initial high-risk sources, and real-time moisture and protein content of each node as intrinsic condition parameters for the occurrence and development of local mold growth, combined with real-time monitored temperature, relative humidity and carbon dioxide concentration, a risk field evolution model based on diffusion equations is constructed. The local anomaly diffusion potential is used as the initial driving force for risk propagation and input into the risk field evolution model; Multi-time-step iterative calculations were performed on a three-dimensional spatial grid of the grain warehouse to simulate the dynamic propagation process of mold risk in space until the simulation results reached a preset stable state. Output the mold risk value of each grid cell in the three-dimensional spatial grid at the final time step, as the dynamically generated risk field distribution.

8. The distributed detection method for multiple indicators of grain quality based on edge collaboration according to claim 1, characterized in that, Based on the aforementioned risk field distribution, spatial data fusion and state optimization assessment are performed on the multidimensional initial indicator set of all nodes to generate a global collaborative quality report, including: Using the risk field distribution as a spatial constraint, co-kriging spatial interpolation is performed on the multidimensional initial index set to generate a globally unified spatial distribution of mold risk, moisture, and protein. Based on the spatial distribution of mold risk, high-risk areas with risk levels exceeding a preset mold threshold are identified, and their corresponding spatial locations are marked. Based on the spatial distribution of moisture and protein, and combined with the mapping relationship of high temperature and high humidity microenvironment parameters, the quality status and potential mold induction conditions of each region are evaluated. By integrating the spatial distribution of mold risk, spatial distribution of moisture, spatial distribution of protein, location of high-risk areas, and quality assessment results, a global collaborative quality report is generated, which includes a mold risk heatmap, a key quality distribution map, and comprehensive control recommendations.

9. The distributed detection method for multiple indicators of grain quality based on edge collaboration according to claim 8, characterized in that, Using the risk field distribution as a spatial constraint, co-kriging spatial interpolation is performed on the multidimensional initial index set to generate a globally unified spatial distribution of mold risk, moisture content, and protein content, including: The risk field distribution is used as the main variable field characterizing the spatial structure of mold risk, and the mold risk index at all nodes is extracted from the multidimensional initial index set as the observed value of the main variable. The moisture content and protein content at all nodes were extracted from the multidimensional initial index set and used as the first covariate observation value and the second covariate observation value, respectively. Based on the main variable field, the experimental variability function of the mold risk index was calculated and fitted. Based on the observed values ​​of the first and second covariates, the experimental variation functions of the moisture content and the protein content were calculated and fitted respectively. Based on the spatial correspondence between the observed values ​​of the main variables and the observed values ​​of the covariates, the cross-experimental variability functions of the mold risk index and moisture content, and the mold risk index and protein content were calculated and fitted. Based on the experimental variance functions of the mold risk index, the moisture content, the protein content, and the cross-variance function obtained from the fitting, a set of co-kriging equations is constructed. Solving the co-kriging equations yields the optimal linear unbiased estimates of the mold risk index, moisture content, and protein content at any unobserved location in the three-dimensional spatial grid of the grain warehouse, and generates globally continuous spatial distributions of the mold risk, moisture, and protein, respectively.

10. A distributed detection system for multiple indicators of grain quality based on edge collaboration, characterized in that, For implementing the distributed detection method for multiple indicators of grain quality based on edge collaboration as described in any one of claims 1-9, the system comprises: The multi-source data synchronous acquisition module is used to simultaneously acquire multispectral response data of the target grain pile and microenvironmental parameters of the node location by deploying multiple edge spectral sensing nodes in the grain warehouse. The quality feature extraction module is used to perform real-time calibration of the multispectral response data based on the microenvironment parameters, and input the data into the quality feature extraction model to obtain a multidimensional initial index set for each node, wherein the multidimensional initial index set includes mold risk index, moisture content and protein content. The abnormal diffusion analysis module is used to analyze the spatial correlation of the mold risk index in the node network by combining the microenvironment parameters and spatial topology of each node, and to calculate the local abnormal diffusion potential. The mold risk field simulation module is used to perform spatiotemporal dynamic mold risk field simulation based on the coupling relationship between the local abnormal diffusion potential and the moisture content and protein content, and dynamically generate the risk field distribution. The collaborative quality report output module is used to perform spatial data fusion and state optimization assessment on the multidimensional initial indicator set of all nodes based on the risk field distribution, and generate a global collaborative quality report.

Images