Citrus huanglongbing early-stage collaborative awareness method based on Internet of Things

Abstract

The invention relates to a Citrus Huanglongbing early collaborative sensing method based on Internet of Things, in particular to the field of Internet of Things, by performing confidence-driven local collaborative filtering at a sensor node, environmental noise and individual difference interference are effectively inhibited, the signal-to-noise ratio of early weak disease features is remarkably improved, and the accuracy of the Citrus Huanglongbing early collaborative sensing method is improved. A dynamic space-time association graph constructed by a convergence gateway accurately depicts space-time association of an abnormal mode, a cloud heterogeneous graph neural network further excavates a deep mode which is robust to global interference and sensitive to local coupling anomaly from the graph, and finally, a collaborative training mechanism based on federated learning is utilized to improve the robustness of the abnormal mode on the premise of protecting data privacy. And the group intelligence of distributed data is gathered, and a lightweight diagnosis model capable of reasoning in real time on the edge side is trained, so that the early-stage, accurate, low-power-consumption and sustainable-evolution intelligent perception of the citrus huanglongbing is realized.

Description

Technical Field

[0001] This invention relates to the field of the Internet of Things (IoT), and more specifically, to an IoT-based method for early collaborative sensing of citrus Huanglongbing (HLB). Background Technology

[0002] Citrus Huanglongbing (HLB), a devastating disease, relies heavily on early identification and intervention for control. Traditional manual inspection methods are inefficient and slow, failing to meet the control needs of large-scale orchards. With the development of IoT technology, particularly low-power wide-area network technologies such as narrowband IoT, it has become possible to deploy wireless monitoring networks composed of multiple types of sensor nodes in large hilly citrus orchards. These networks typically consist of numerous low-cost multispectral, leaf temperature, and microenvironment sensor nodes, aiming to achieve early automatic disease detection through continuous collection of plant physiological and ecological parameters. However, in actual deployment scenarios, the hilly terrain leads to significant microclimate differences, including diurnal temperature variations and irregular temperature patterns. Environmental factors such as light, intermittent rainfall, and local shading are complex and variable. At the same time, due to cost and energy consumption limitations, the accuracy of a single sensor node is limited, and it is difficult to achieve completely uniform network coverage. Furthermore, the inherent physiological differences between individual plants objectively exist. In this complex context, the physiological changes in plants caused by early diseases, such as subtle changes in chlorophyll content or slight abnormalities in leaf temperature, appear extremely weak in the single-point raw data of a single sensor, resulting in a very low signal-to-noise ratio. These potential early disease signals are easily completely drowned out by strong environmental background noise and individual plant differences, making the traditional method of threshold discrimination based on a single sensor very unreliable and difficult to put into practical application.

[0003] Currently, most IoT technology applications for agricultural pest and disease monitoring focus on monitoring and alerting for single or a few independent environmental indicators such as soil moisture and meteorological parameters. Existing solutions have significant limitations in the early identification of diseases involving plant physiological states. On the one hand, some studies use high-precision handheld or airborne spectrometers for point or area scanning, which can obtain high-quality data, but the equipment is expensive, difficult to implement continuous monitoring at fixed points, and lacks deep collaboration with large-scale, low-cost IoT sensing networks. On the other hand, even when using multi-sensor combinations, existing solutions typically treat each sensor node as an independent data source, employing a simple cloud-based data aggregation and post-processing model, such as processing multi-source data... Independent threshold judgment followed by decision-making voting, or simple weighted average fusion, fail to fully utilize the inherent networked correlation characteristics between IoT nodes. Specifically, they fail to effectively coordinate the correlation between the sensing data of spatially adjacent nodes in the temporal and spatial dimensions, nor do they deeply explore the complementarity between data from different modal sensors. As a result, they cannot achieve the synergistic suppression of widespread environmental noise and the synergistic enhancement and extraction of sparse and weak early disease features from a mechanistic perspective. Therefore, how to design a lightweight algorithm framework that can effectively coordinate multi-source heterogeneous sensor nodes to achieve reliable perception and fusion of weak early disease features in complex farmland environments has become the core bottleneck restricting the practical application of this technology. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an Internet of Things-based early collaborative sensing method for citrus Huanglongbing (HLB) to solve the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes the following steps: Step S1: At the sensor node deployed on the citrus plant, in response to the multimodal time-series data generated by the sensor node itself from multiple types of sensors including spectral sensors, infrared temperature sensors and micro-deformation sensors, and the associated sensor data generated by a single type of sensor received from neighboring nodes, a collaborative sensing confidence level is calculated based on a preset sliding time window. This confidence level integrates the consistency of the multimodal time-series data of the sensor node itself and the spatial consistency of the associated sensor data with the neighboring nodes. The filtered data features corresponding to the collaborative sensing confidence level exceeding a preset threshold are then uploaded to the aggregation gateway as preliminary abnormal feature data packets. Step S2: At the aggregation gateway, preliminary anomaly feature data packets uploaded from multiple sensor nodes are received and cached. Based on the timestamp and node location information carried by each preliminary anomaly feature data packet, a weighted dynamic time warping method is used to align the preliminary anomaly feature data packets from different nodes and different sensors. The weight assigned to each data point in the weighted dynamic time warping method comes from the collaborative sensing confidence calculated by the sensor node that generated the data point at the corresponding time. Based on the spatial proximity relationship between nodes and the correlation of the aligned data, a dynamic spatiotemporal association graph containing multimodal node attributes and multi-relationship edge weights composed of aligned preliminary anomaly feature data packets is constructed. Step S3: At the processing unit in the cloud, the dynamic spatiotemporal correlation graph is input into a pre-trained heterogeneous graph neural network model. Through the graph attention mechanism and cross-modal message passing of the heterogeneous graph neural network model, a low-dimensional cooperative anomaly embedding vector is calculated for each node in the dynamic spatiotemporal correlation graph. Step S4: Between the aggregation gateway and the cloud, a lightweight classification model is collaboratively trained using a federated learning framework based on the collaborative anomaly embedding vector and its corresponding pre-labeled historical tags. The trained lightweight classification model is then deployed on the aggregation gateway, enabling the aggregation gateway to perform local inference on the real-time generated collaborative anomaly embedding vector and output the early occurrence probability of citrus Huanglongbing. In a preferred embodiment, step S1, which calculates a collaborative sensing confidence level that integrates the consistency of the sensor node's own multimodal temporal data and the spatial consistency of its associated sensor data with neighboring nodes, specifically includes: First, a local anomaly index reflecting the degree of abnormal coupling between multimodal data of the sensor node itself is calculated. The calculation process is as follows: within the sliding time window, for the multi-type sensors of the sensor node itself, the total number of multi-type sensors is determined; all possible combinations of two different sensors are selected from the multi-type sensors, and the total number of all possible combinations is determined; for each sensor combination, the first data sequence and the second data sequence formed by the two sensors in the combination within the current time window are obtained; the normalized cross-correlation coefficient between the first data sequence and the second data sequence is calculated; the difference between the average value of the first data sequence within the current time window and its long-term historical average value, and the difference between the average value of the second data sequence within the current time window and its long-term historical average value are obtained respectively. The first intermediate value is obtained by subtracting the absolute value of the number 1 from the normalized cross-correlation coefficient; the second intermediate value is obtained by subtracting the difference of the second data sequence from the difference of the first data sequence and taking the sign of the subtraction result; the contribution value corresponding to the sensor combination is obtained by multiplying the signs of the first and second intermediate values; for all possible combinations, the corresponding contribution value is calculated, all contribution values ​​are summed, and the summation result is divided by the total number of all possible combinations to obtain the local anomaly index. Secondly, an average spatial consistency residual reflecting the degree of spatial data difference between the sensor node and its neighboring nodes is calculated. The calculation process is as follows: For each sensor node and its neighboring node, a specified type of associated sensor data shared by both is selected as the local node data and the neighboring node data, respectively; the absolute value of the difference between the local node data and the neighboring node data is calculated; the standard deviation of the specified type of sensor data in the long-term history is obtained; the absolute value is divided by the standard deviation to obtain the normalized difference; the physical distance between the sensor node and the current neighboring node is obtained, and a preset distance attenuation coefficient is obtained; the exponential function value is calculated with the natural constant as the base and the quotient of the negative physical distance and the distance attenuation coefficient as the exponent; the normalized difference is multiplied by the exponential function value to obtain the spatial consistency residual corresponding to the current neighboring node; for all neighboring nodes, the corresponding spatial consistency residuals are calculated, and the arithmetic mean of these spatial consistency residuals is taken as the average spatial consistency residual. Finally, the collaborative perception confidence score is generated through a fusion calculation. This calculation takes the local anomaly index and the average spatial consistency residual as inputs. The local anomaly index is multiplied by the average spatial consistency residual to obtain the first product as the numerator. The first harmonic coefficient, the second harmonic coefficient, and a positive coefficient are obtained. The first harmonic coefficient is multiplied by the local anomaly index to obtain the second product. The second harmonic coefficient is multiplied by the average spatial consistency residual to obtain the third product. The second product, the third product, and the positive coefficient are added to obtain the sum as the denominator. The numerator is divided by the denominator to obtain the collaborative perception confidence score.

[0006] In a preferred embodiment, the process of uploading the filtered data features corresponding to the collaborative perception confidence level exceeding a preset threshold as preliminary anomaly feature data packets to the aggregation gateway specifically involves: The sensor node has a preset confidence threshold. When the calculated collaborative sensing confidence is greater than or equal to the confidence threshold, the sensor node determines that the data in the current sliding time window constitutes a potential collaborative anomaly and then triggers the data upload process. The sensor node encapsulates its own multimodal time series data in the current window, the collaborative sensing confidence calculated this time, the data acquisition timestamp, and the node identifier into a preliminary anomaly feature data packet and sends it to the aggregation gateway. If the calculated collaborative sensing confidence level is less than the confidence level threshold, the sensor node discards the multimodal time series data within the current sliding time window and only sends a heartbeat signal indicating that the status is normal to the aggregation gateway.

[0007] In a preferred embodiment, the specific operation of aligning the preliminary anomaly feature data packets from different nodes and different sensors using the weighted dynamic time warping method in step S2 is as follows: At the aggregation gateway, multiple received preliminary anomaly feature data packets are cached; the preliminary anomaly feature data packets corresponding to the two sensor nodes that need to be aligned are extracted from the cache; for each preliminary anomaly feature data packet, the multimodal time-series data contained therein and the sequence of collaborative sensing confidence scores calculated and stored in the preliminary anomaly feature data packet in step S1, arranged in chronological order, are extracted and denoted as the collaborative sensing confidence score sequence; a weighted distance is defined to calculate the distance between the data vectors from the multimodal time-series data corresponding to any two time points in the two preliminary anomaly feature data packets. The calculation process of this weighted distance is as follows: first, calculate the distance between the data vectors from the multimodal time-series data corresponding to any two time points in the two preliminary anomaly feature data packets. The first intermediate value is obtained by dividing the Euclidean distance between the data vectors corresponding to the two time points. The second intermediate value is obtained by dividing the Euclidean distance by the normalization factor. The third intermediate value is obtained by multiplying the collaborative sensing confidence scores from their respective collaborative sensing confidence sequences for each of the two time points. The second intermediate value is then compared with the positive coefficient, and the largest value is taken as the third intermediate value. Finally, the first intermediate value is divided by the third intermediate value to obtain the weighted distance. Based on the calculated weighted distance, an optimal warping path is found between the time points of the two preliminary anomaly feature data packets using a dynamic time warping algorithm. This path minimizes the sum of the overall weighted distances that align the two data packets in time. According to the optimal warping path, the multimodal time series data in the two preliminary anomaly feature data packets are resampled and aligned so that they are mapped to a common warped time axis, and the aligned multimodal time series segments with the same time length are output.

[0008] In a preferred embodiment, the specific operation of constructing a dynamic spatiotemporal correlation graph containing multimodal node attributes and multi-relationship edge weights composed of preliminarily aligned anomaly feature data packets based on the spatial proximity relationship between nodes and the correlation of aligned data is as follows: Each sensor node is abstracted as a vertex in a dynamic spatiotemporal correlation graph. Attributes are generated for each vertex by calculating the statistical features of the aligned multimodal time series segment corresponding to that node on the common regularized time axis. These statistical features include the mean vector of the multimodal time series segment across all sensor type dimensions, and its variance vector across all sensor type dimensions. Then, the weights of two types of connection edges between any two graph vertices are calculated. The first type is the physical proximity edge weight, which is calculated as follows: obtain the geographical distance between the sensor nodes represented by the two graph vertices; obtain a preset distance bandwidth parameter; calculate the exponential function value with the quotient obtained by dividing the square of the negative geographical distance by twice the square of the distance bandwidth parameter as the exponent using the natural constant as the base. This value is the physical proximity edge weight. The second type is data association edge weights, the calculation process of which is as follows: Obtain the aligned multimodal temporal segments of the two graph vertices and their corresponding co-aware confidence sequences; traverse all time point pair combinations of the two segments on the common regular time axis; for each time point pair, calculate the Euclidean distance between the two data vectors in the pair; find the maximum value among all time point pair combinations of Euclidean distances; for each time point pair, calculate a normalized complementary distance by subtracting the ratio of its Euclidean distance to the maximum Euclidean distance; for the same time point pair, use a preset kernel function that calculates weight values ​​based on two co-aware confidences to calculate a weight value according to the two co-aware confidences corresponding to the pair; multiply the normalized complementary distances of all time point pairs by their corresponding weight values, and sum the product results to obtain a weighted distance sum; sum the weight values ​​of all time point pairs to obtain a total weight sum; divide the weighted distance sum by the total weight sum to obtain the data association edge weights. Ultimately, the dynamic spatiotemporal relational graph is composed of a set of all graph vertices, the attributes of each graph vertex, and edges connecting graph vertices that include both physical proximity edge weights and data relational edge weights.

[0009] In a preferred embodiment, step S3, specifically the operation of inputting the dynamic spatiotemporal correlation graph into a pre-trained heterogeneous graph neural network model, is as follows: The dynamic spatiotemporal correlation graph output in step S2 is used as the input of the heterogeneous graph neural network model. The heterogeneous graph neural network model processes each sensor node in the dynamic spatiotemporal correlation graph as a vertex in the graph. Each vertex has an attribute vector, which is derived from the statistical features of the aligned multimodal time-series segments in step S2. The heterogeneous graph neural network model distinguishes the connection relationships between vertices into two predefined types: the first is physical proximity relationship, whose edge weights are calculated using the physical proximity edge weights in step S2; the second is data association relationship, whose edge weights are calculated using the data association edge weights in step S2. During the processing of the heterogeneous graph neural network model, a set of independent, learnable model parameters is maintained for each relationship type. The model parameters include a linear transformation matrix for feature transformation, an attention calculation vector, and a bias vector. The heterogeneous graph neural network model performs progressively deeper calculations on the input dynamic spatiotemporal correlation graph through multiple sequentially connected processing layers. In each layer, the heterogeneous graph neural network model calculates an updated feature representation vector for each vertex in the dynamic spatiotemporal correlation graph.

[0010] In a preferred embodiment, the specific operations of the graph attention mechanism and cross-modal message passing through the heterogeneous graph neural network model are as follows: In each processing layer of the heterogeneous graph neural network model, for each target vertex currently being processed, and each neighboring vertex connected to it through the physical proximity relationship or the data association relationship, an attention weight calculation corresponding to the connection relationship is performed. This attention weight calculation corresponding to the connection relationship takes the feature representation vector of the target vertex in the current layer, the feature representation vector of the neighboring vertex in the current layer, and the edge weight corresponding to the relationship connecting the two as inputs. First, the edge weight corresponding to the relationship is converted into a vector through a learnable mapping function. Then, the feature representation vector of the target vertex and the feature representation vector of the neighboring vertex are processed respectively through a learnable linear transformation matrix corresponding to the relationship type. Finally, the processed target vertex feature representation vector, the neighboring vertex feature representation vector, and the vector converted from the edge weight are concatenated to form a combined vector. The combined vector is multiplied by a learnable attention vector corresponding to the relationship type, and the result is input into a preset nonlinear function with incomplete suppression characteristics to obtain the original attention score representing the importance of the neighboring vertex to the target vertex under this relationship. Subsequently, the original attention scores of all neighboring vertices of the target vertex under the same relationship are normalized. The normalization process uses an exponential function and a summation operation to make the sum of the standardized attention weight coefficients of all neighboring vertices for the same target vertex and the same relationship equal to one, thereby obtaining the standardized attention weight coefficients. Based on the calculated attention weight coefficients, the heterogeneous graph neural network model performs cross-modal message passing, that is, it sums the feature representation vectors of all neighboring vertices of the target vertex under each relation according to the corresponding attention weight coefficients, thereby obtaining the aggregated message vector passed to the target vertex under that relation.

[0011] In a preferred embodiment, the specific operation of calculating a low-dimensional collaborative anomaly embedding vector for each node in the dynamic spatiotemporal correlation graph is as follows: After obtaining the aggregated message vector under each relationship, for each target vertex in the current layer, a gated fusion mechanism is used to update its feature representation vector. This mechanism calculates a gated vector for physical proximity relationships and data association relationships respectively. The calculation process of each gated vector is as follows: the current feature representation vector of the target vertex is processed through a learnable linear transformation matrix corresponding to the relationship, and a corresponding learnable bias vector is added. The result is then input into an S-shaped growth function, which maps each element of the output vector to between zero and one, thereby obtaining a gated vector used to control the contribution of the corresponding relationship message vector to the final feature update. The aggregated message vector of physical proximity is first processed through a learnable linear transformation matrix, and then multiplied element-wise by its gate vector to obtain the weighted physical proximity message; similarly, the aggregated message vector of data association is first processed through a learnable linear transformation matrix, and then multiplied element-wise by its gate vector to obtain the weighted data association message. Then, the feature representation vector of the target vertex itself is processed by a learnable linear transformation matrix, added to the two weighted message vectors mentioned above, and then passed through a linear rectified nonlinear activation function to generate a new feature representation vector of the target vertex in this processing layer. After the above iterative processing with a preset number of layers, the final feature representation vector of each vertex is output as a low-dimensional collaborative anomaly embedding vector.

[0012] In a preferred embodiment, step S4, specifically the operation of collaboratively training a lightweight classification model using a federated learning framework, is as follows:A global lightweight classification model is initialized on the cloud server, and its initial model parameters are distributed to each aggregation gateway. Each aggregation gateway maintains a local dataset consisting of multiple data samples, each containing a co-anomaly embedding vector generated in step S3 and a corresponding historical label pre-annotated by an agronomist. In each round of federated training, the cloud server selects a subset of aggregation gateways to participate in training and sends the model parameters of the current global lightweight classification model to these selected aggregation gateways. Upon receiving the model parameters of the global lightweight classification model, each selected aggregation gateway uses its local dataset to refine the model. The model is trained locally to obtain local model parameter updates. Simultaneously, each participating aggregation gateway calculates a contribution metric reflecting the contribution of its local dataset to the global model update. The specific calculation process for this contribution metric is as follows: First, a data volume normalization factor is calculated, which is equal to the total number of samples in the current gateway's local dataset divided by the maximum value among the total number of samples in the local datasets of all participating aggregation gateways in this round. Second, a gradient difference metric is calculated by: calculating the average gradient vector of the loss function obtained by the current gateway on its local dataset based on the received global model parameters; and calculating the average of the above average gradient vectors of all participating aggregation gateways in this round, denoted as [the average gradient vector]. The global average gradient vector is used as the basis for calculating the Euclidean norm of the difference between the current gateway's average gradient vector and the global average gradient vector, thus obtaining the gradient deviation of the current gateway. The sum of the gradient deviations of all participating aggregation gateways in this round of training is calculated. The gradient deviation of the current gateway is divided by the sum of the gradient deviations and the sum of positive coefficients to obtain the normalized gradient difference ratio of the current gateway. Then, a factor is calculated and subtracted from the normalized gradient difference ratio. Finally, the data volume normalization factor is multiplied by the result of subtracting the normalized gradient difference ratio, and the product is the contribution index of the aggregation gateway. Each aggregation gateway encrypts and uploads the calculated contribution index and local model parameter updates to the cloud server. After the server collects the contribution metrics and local model parameter updates uploaded by all participating gateways, it normalizes the contribution metrics to obtain the aggregate weight of each gateway. Then, the cloud server uses these aggregate weights to perform a weighted average of the received local model parameter updates from each gateway. Next, the model parameters of the global lightweight classification model before this round of training are added to the global aggregate learning rate multiplied by a weighted sum. The weighted sum is equal to the product of the aggregate weight of each gateway and its uploaded local model parameter update. The sum of such products for all participating gateways is then obtained to obtain the new model parameters of the global lightweight classification model after this round of training. These new model parameters are then applied to update the global lightweight classification model.Repeat the above process multiple times until the model performance converges or reaches the preset number of training rounds, resulting in a trained global lightweight classification model.

[0013] In a preferred embodiment, the specific operation of deploying the trained lightweight classification model on the aggregation gateway, enabling the aggregation gateway to perform local inference on the real-time generated collaborative anomaly embedding vectors and output the early occurrence probability of citrus Huanglongbing, is as follows: After federated training is completed, the cloud server distributes the parameters of the final global lightweight classification model to all aggregation gateways. Each aggregation gateway loads the parameters of the global lightweight classification model locally and completes the deployment of the global lightweight classification model. After deployment, the aggregation gateway continuously receives the co-anomaly embedding vector generated in real time in step S3 and inputs it into the locally deployed global lightweight classification model for forward inference calculation. The global lightweight classification model processes the input co-anomaly embedding vector and outputs a value between zero and one, which serves as the probability of early Huanglongbing occurrence in the citrus plant where the corresponding sensor node is located.

[0014] The beneficial effects of this invention are as follows: By performing confidence-driven local collaborative filtering at the sensor nodes, environmental noise and individual difference interference are effectively suppressed, significantly improving the signal-to-noise ratio of early weak disease characteristics. The dynamic spatiotemporal correlation graph constructed by the aggregation gateway accurately depicts the spatiotemporal correlation of abnormal patterns. The heterogeneous graph neural network in the cloud further mines deep patterns from the graph that are robust to global interference and abnormally sensitive to local coupling. Finally, based on the collaborative training mechanism of federated learning, the collective wisdom of distributed data is aggregated under the premise of protecting data privacy, and a lightweight diagnostic model that can be used for real-time reasoning at the edge is trained, thereby realizing early, accurate, low-power and sustainable intelligent perception of citrus Huanglongbing. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1 The above describes an IoT-based collaborative sensing method for early detection of Huanglongbing (HLB) in citrus, which includes the following steps: Step S1: At the sensor node deployed on the citrus plant, in response to the multimodal time-series data collected from its own multi-type sensors including spectral sensors, infrared temperature sensors and micro-deformation sensors, and the associated sensor data received from neighboring nodes from one type of sensor, a collaborative sensing confidence score is calculated based on a preset sliding time window. This score integrates the consistency of the sensor node's own multimodal time-series data and the spatial consistency of the associated sensor data with neighboring nodes. The filtered data features corresponding to the collaborative sensing confidence score exceeding a preset threshold are then uploaded to the aggregation gateway as preliminary abnormal feature data packets. Step S2: At the aggregation gateway, receive and cache preliminary anomaly feature data packets uploaded from multiple sensor nodes. Based on the timestamps and node location information carried by each preliminary anomaly feature data packet, use a weighted dynamic time warping method to align the preliminary anomaly feature data packets from different nodes and different sensors. The weight assigned to each data point in the weighted dynamic time warping method comes from the collaborative sensing confidence calculated by the sensor node that generated the data point at the corresponding time. Based on the spatial proximity relationship between nodes and the correlation of the aligned data, construct a dynamic spatiotemporal association graph containing multimodal node attributes and multi-relationship edge weights composed of aligned preliminary anomaly feature data packets. Step S3: In the cloud or at a processing unit with preset computing capabilities, input the dynamic spatiotemporal correlation graph into a pre-trained heterogeneous graph neural network model. Through the graph attention mechanism and cross-modal message passing of the heterogeneous graph neural network model, calculate a low-dimensional cooperative anomaly embedding vector for each node in the dynamic spatiotemporal correlation graph. Step S4: Between the aggregation gateway and the cloud, a lightweight classification model is collaboratively trained using a federated learning framework based on the collaborative anomaly embedding vector and its corresponding pre-labeled historical tags. The trained lightweight classification model is then deployed on the aggregation gateway, enabling the aggregation gateway to perform local inference on the real-time generated collaborative anomaly embedding vector and output the early occurrence probability of citrus Huanglongbing.

[0020] In this embodiment, it is specifically necessary to explain that in step S1, the process of calculating a collaborative sensing confidence score that integrates the consistency of the multimodal temporal data of the sensor node itself and the spatial consistency of the associated sensor data with neighboring nodes specifically includes: First, a local anomaly index reflecting the degree of abnormal coupling between multimodal data of the sensor node itself is calculated. The calculation process is as follows: within the sliding time window, for the multi-type sensors of the sensor node itself, the total number of multi-type sensors is determined; all possible combinations of two different sensors are selected from the multi-type sensors, and the total number of all possible combinations is determined; for each sensor combination, the first data sequence and the second data sequence formed by the two sensors in the combination within the current time window are obtained; the normalized cross-correlation coefficient between the first data sequence and the second data sequence is calculated, which is used to quantify the linear correlation between the two data sequences in the window; the difference between the average value of the first data sequence within the current time window and its long-term historical average value, and the difference between the average value of the second data sequence within the current time window and its long-term historical average value are obtained respectively. The long-term historical average value refers to the average value of the data of the sensor in the same period of the past many days (e.g., the past 30 days) within the corresponding time period of the sliding time window, which is used to establish a benchmark for the health status; The first intermediate value is obtained by subtracting the absolute value of the normalized cross-correlation coefficient from the first data sequence. The larger this value, the weaker the correlation between the two sensor data within the window, which may indicate an anomaly. The second intermediate value is obtained by subtracting the difference between the first and second data sequences from the difference in the first data sequence and taking the sign of the subtraction result. This sign indicates whether the deviation of the mean values ​​of the two sensor data from the long-term benchmark is consistent. If the sign is positive, it indicates that the two deviations are in the same direction, further enhancing the possibility of anomalous coupling. The contribution value corresponding to the sensor combination is obtained by multiplying the signs of the first and second intermediate values, thus combining the weakening of correlation with the consistency of deviation direction to evaluate the anomalous contribution of the combination. For all possible combinations, the corresponding contribution value is calculated, and all contribution values ​​are summed. The summation result is then divided by the total number of all possible combinations to obtain the local anomaly index. This index comprehensively reflects the average intensity of the anomalous pattern among all sensor pairs of the node itself. Secondly, an average spatial consistency residual reflecting the degree of spatial data difference between the sensor node and its neighboring nodes is calculated. The calculation process is as follows: For each sensor node and its neighboring nodes, shared, specified type of associated sensor data is selected as the local node's data and the neighboring node's data, respectively. The specified type is, for example, infrared temperature sensor data, because it is sensitive to environmental changes and is typically deployed across all nodes for spatial comparison. The absolute value of the difference between the local node's data and the neighboring node's data is calculated. The standard deviation of this specified type of sensor data over a long historical period is obtained to normalize the difference, eliminating the influence of the sensor's own dimensions and inherent fluctuations, making the differences between different nodes comparable. The absolute value is divided by the standard deviation to obtain the normalized difference. The physical distance between the sensor node and its current neighboring nodes is obtained, along with a preset distance. The distance attenuation coefficient is used to adjust the effective range of spatial association in actual farmland. For example, it can be set to 20 meters, indicating that the influence of nodes beyond this distance is significantly attenuated. An exponential function value is calculated with the natural constant as the base and the quotient of the negative physical distance and the distance attenuation coefficient as the exponent. This function value is between 0 and 1; the closer the distance, the closer the value is to 1, and the farther the distance, the smaller the value, reflecting the weight of spatial proximity. The normalized difference is multiplied by the exponential function value to obtain the spatial consistency residual corresponding to the current neighboring nodes. This residual considers both the magnitude of data differences and the weight of spatial distance. For all neighboring nodes, their corresponding spatial consistency residuals are calculated separately. The arithmetic mean of these spatial consistency residuals is then taken as the average spatial consistency residual, which reflects the overall degree of anomaly of the current node relative to its entire local spatial neighborhood. Finally, the collaborative perception confidence score is generated through a fusion calculation. This calculation takes the local anomaly index and the average spatial consistency residual as inputs, multiplies the local anomaly index by the average spatial consistency residual, and uses the first product as the numerator. This design ensures that the numerator is valid only when both are non-zero, requiring the anomaly to simultaneously possess its own multimodal misalignment and spatial locality. A first harmonic coefficient, a second harmonic coefficient, and a positive coefficient are obtained. The first and second harmonic coefficients are used to balance the contribution ratios of the local anomaly index and the average spatial consistency residual in the denominator, and can be optimized through experimental data; for example, they can be set to 0.5 respectively. The first harmonic coefficient is multiplied by the local anomaly index to obtain the second product. The second harmonic coefficient is multiplied by the mean spatial consistency residual to obtain the third product. The second and third products are added to the positive coefficient to obtain the sum, which is used as the denominator. The numerator is divided by the denominator to obtain the co-perception confidence. This calculation method makes the co-perception confidence monotonically increasing with respect to both the local anomaly index and the mean spatial consistency residual, and its value range is between 0 and positive numbers. The higher the value, the greater the possibility of potential co-perception. The process of uploading the filtered data features corresponding to when the collaborative perception confidence level exceeds a preset threshold as a preliminary anomaly feature data packet to the aggregation gateway is as follows: The sensor node internally presets a confidence threshold, which is used to determine whether the collaborative sensing confidence is high enough to trigger an upload. Its value can be determined by analyzing the confidence distribution of healthy samples and early disease samples in historical data. For example, it can be set to 0.6 to balance the false alarm rate and the false negative rate. When the calculated collaborative sensing confidence is greater than or equal to the confidence threshold, the sensor node determines that the data in the current sliding time window constitutes a potential collaborative anomaly and then triggers the data upload process. The sensor node encapsulates its own multimodal time series data in the current window, the collaborative sensing confidence calculated this time, the data acquisition timestamp, and the node identifier into a preliminary anomaly feature data packet and sends it to the aggregation gateway. The preliminary anomaly feature data packet is structured and encapsulated in a predefined binary or JSON format to ensure the integrity and parsability of the data elements. If the calculated collaborative sensing confidence level is less than the confidence level threshold, the sensor node discards the multimodal time series data within the current sliding time window and sends only a heartbeat signal indicating that the status is normal to the aggregation gateway. The heartbeat signal is a fixed-format message with a very small amount of data, containing only the node identifier and status code, which is used to maintain network connection and significantly save wireless communication bandwidth and node power consumption.

[0021] In this embodiment, it is specifically necessary to explain the following steps in step S2: the weighted dynamic time warping method is used to align the preliminary abnormal feature data packets from different nodes and sensors as follows: At the aggregation gateway, multiple received preliminary anomaly feature data packets are cached; the preliminary anomaly feature data packets corresponding to the two sensor nodes that need to be aligned are extracted from the cache; for each preliminary anomaly feature data packet, the multimodal time-series data contained therein and the sequence of collaborative sensing confidence scores calculated and stored in the preliminary anomaly feature data packet in step S1, arranged in chronological order, are extracted and denoted as the collaborative sensing confidence score sequence; a weighted distance is defined to calculate the distance between the data vectors from the multimodal time-series data corresponding to any two time points in the two preliminary anomaly feature data packets. The calculation process of this weighted distance is as follows: First, the Euclidean distance between the data vectors corresponding to the two time points is calculated; second, a normalization factor is obtained, which is the combined standard deviation of the same type of sensor data corresponding to the sensor nodes that generated these two data packets in the long-term history. The long-term history usually refers to data from multiple past production cycles (such as 30-60 days). The standard deviation is used to eliminate the differences in the dimensions and inherent fluctuation range of different sensors, making the distances between different nodes comparable. The Euclidean distance is divided by the normalization factor to obtain the first intermediate value. Next, the co-sensing confidence scores from their respective co-sensing confidence scores for the two time points are obtained, and the two co-sensing confidence scores are multiplied to obtain the second intermediate value. The second intermediate value is compared with the positive coefficient, and the largest value is taken as the third intermediate value. The positive coefficient is a very small constant, such as 1e-8, which is used to prevent division by zero when both co-sensing confidence scores are zero, and to ensure the numerical stability of the calculation. Finally, the first intermediate value is divided by the third intermediate value to obtain the weighted distance. The design of this weighted distance is such that when both time points have high co-sensing confidence scores, the "distance" between them will be significantly reduced, thereby guiding the subsequent dynamic time warping algorithm to prioritize aligning high-confidence outliers on the time axis, strengthening the temporal correlation of the outlier patterns. Based on the calculated weighted distance, an optimal warping path is found between the time points of two preliminary anomaly feature data packets using a dynamic time warping algorithm. This path minimizes the sum of the overall weighted distances that align the two data packets in time. According to the optimal warping path, the multimodal time series data in the two preliminary anomaly feature data packets are resampled and aligned so that they are mapped to a common warped time axis. Aligned multimodal time series segments with the same time length are output. The length of the common warped time axis is equal to the length of the optimal warping path, which may be longer or shorter than the length of the original sequence. Through the alignment operation, the time asynchrony problem caused by local clock drift or small event triggering delays of different nodes is solved, providing time-consistent input for subsequent graph construction. Based on the spatial proximity between nodes and the correlation of aligned data, the specific steps for constructing a dynamic spatiotemporal association graph containing multimodal node attributes and multi-relation edge weights composed of preliminarily aligned anomaly feature data packets are as follows: Each sensor node is abstracted as a vertex in a dynamic spatiotemporal correlation graph. Attributes are generated for each vertex. These attributes are generated by calculating the statistical features of the aligned multimodal time series segment corresponding to the node on a common regular time axis. The statistical features include the mean vector of the multimodal time series segment across all sensor types, and its variance vector across all sensor types. The mean vector characterizes the average level of sensor readings during the abnormal period, while the variance vector characterizes the degree of fluctuation. Together, they constitute the static feature representation of the node. Then, the weights of two types of connection edges between any two graph vertices are calculated: The first type is the physical proximity edge weight. Its calculation process is as follows: obtain the geographical distance between the sensor nodes represented by the two graph vertices; obtain a preset distance bandwidth parameter, which is used to control the decay rate of spatial influence. Its value can be set according to the typical spatial range of disease spread or environmental influence in the actual orchard. For example, setting it to 20 meters indicates that there is a strong spatial correlation between nodes within this distance; calculate the exponential function value with the natural constant as the base and the square of the negative geographical distance divided by twice the square of the distance bandwidth parameter as the exponent. This value is the physical proximity edge weight. This calculation makes the physical proximity edge weight decay exponentially as the distance between nodes increases. The value range is between (0,1], which intuitively reflects the strength of spatial proximity. The second method is data association edge weighting, which is calculated as follows: First, obtain the aligned multimodal temporal segments of the two graph vertices and their corresponding co-sensory confidence sequences. Second, iterate through all time point pairs on the common regular time axis. For each time point pair, calculate the Euclidean distance between the two data vectors in that pair. Third, find the maximum value among all time point pair combinations of Euclidean distances. Fourth, for each time point pair, calculate the normalized complementary distance of that pair by subtracting the ratio of its Euclidean distance to the maximum Euclidean distance. This value is between 0 and 1, where 1 is used when the two data points are completely identical and 0 is used when the difference is maximum, effectively measuring instantaneous data similarity. Fifth, for the same time point pair, use a pre-defined kernel function that calculates weight values ​​based on two co-sensory confidences to calculate a weight value according to the two co-sensory confidences corresponding to that pair. A specific example of a kernel function is: weight value = exp(-((1-confidence A)²+(1-confidence B)²) / (2*γ²)), where γ is the kernel width parameter, for example, it can be 0.3. This function assigns higher weights to high-confidence point pairs, making the calculation of data association edge weights more focused on those abnormal moments judged as more suspicious in step S1; multiply the normalized complementary distances of all time point pairs by their corresponding weight values, and sum the product results to obtain the weighted distance sum; sum the weight values ​​of all time point pairs to obtain the total weight sum; divide the weighted distance sum by the total weight sum to obtain the data association edge weight, which integrates the weighted similarity at all time points. It is a scalar between 0 and 1. The larger the value, the more similar the data patterns of the two nodes are as a whole after considering the confidence weight, which may suggest potential common influencing factors; Ultimately, the dynamic spatiotemporal association graph consists of a set of all graph vertices, the attributes of each graph vertex, and edges connecting graph vertices that include both physical proximity weights and data association weights. This graph structure explicitly encodes the spatial constraints and data associations between nodes, providing a semantically rich structured input for subsequent deep pattern mining based on graph neural networks.

[0022] In this embodiment, it is specifically necessary to explain the operation of inputting the dynamic spatiotemporal correlation graph into a pre-trained heterogeneous graph neural network model in step S3 as follows: The dynamic spatiotemporal correlation graph output from step S2 is used as input to the heterogeneous graph neural network model. The model treats each sensor node in the graph as a vertex, with each vertex possessing an attribute vector derived from the statistical characteristics of the aligned multimodal time-series segments in step S2. For example, the attribute vector could be a concatenation of the mean and standard deviation vectors of the aligned segments, thus characterizing the node's anomalous features from both central tendency and dispersion perspectives. The model categorizes the connections between vertices into two predefined types: physical proximity (using the edge weights calculated in step S2) and data association (using the edge weights calculated in step S2). These two types of edge weights encode spatial constraints and potential data correlations, respectively, forming the basis for the model's relational reasoning. During the heterogeneous graph neural network model's processing, an independent, learnable set of models is maintained for each relational type. Model parameters include, but are not limited to, linear transformation matrices, attention vectors, and bias vectors used for feature transformation. These parameters are learned from labeled data during model training using backpropagation algorithms and optimizers (such as the Adam optimizer), enabling the model to adaptively extract patterns useful for disease classification. This allows the heterogeneous graph neural network model to learn different information transformation and aggregation patterns for physical proximity relationships and data association relationships. The heterogeneous graph neural network model progressively deepens the calculation of the input dynamic spatiotemporal association graph through multiple sequentially connected processing layers. The number of processing layers is a hyperparameter, usually set according to the size of the dynamic spatiotemporal association graph and the complexity of the task, such as 2 to 3 layers, to achieve information aggregation from local to global. In each layer, the heterogeneous graph neural network model calculates an updated feature representation vector for each vertex in the dynamic spatiotemporal association graph. This feature representation vector integrates the vertex's own previous feature representation vector and the feature representation vectors of its neighboring vertices connected to it through different relationships. The specific operations of graph attention mechanism and cross-modal message passing through heterogeneous graph neural network models are as follows: In each processing layer of the heterogeneous graph neural network model, for each target vertex currently being processed, and each neighboring vertex connected to it through physical proximity or data association, an attention weight calculation corresponding to that connection relationship is performed. This attention weight calculation takes the feature representation vector of the target vertex in the current layer, the feature representation vector of the neighboring vertex in the current layer, and the edge weight corresponding to the relationship connecting the two as input. First, the edge weight corresponding to the relationship is transformed into a vector through a learnable mapping function. The learnable mapping function can be a simple linear layer that maps scalar edge weights to a vector that matches the node feature dimension, so that the edge weight information can be deeply involved in subsequent attention calculations. Next, the feature representation vectors of the target vertex and the neighboring vertices are processed by a learnable linear transformation matrix corresponding to the relationship type. This linear transformation matrix projects the node features into a subspace related to the relationship, preparing for subsequent similarity calculations. Then, the processed target vertex feature representation vector, the neighboring vertex feature representation vector, and the vector transformed from the edge weights are concatenated to form a combined vector. The combined vector is then multiplied by a learnable attention vector corresponding to the relation type. The result is then input into a pre-defined nonlinear function with incomplete inhibition characteristics, such as the LeakyReLU function. This function allows small negative input values ​​to pass through, avoiding the neuron "death" problem that may occur with the ordinary ReLU function, and helps to stabilize the calculation of attention scores, thus obtaining the original attention score representing the importance of the neighboring vertex to the target vertex under this relation. Subsequently, the original attention scores of all neighboring vertices of the target vertex under the same relation are normalized. The normalization process uses an exponential function and a summation operation, i.e., the softmax function. This function transforms the original attention scores into a probability distribution, making the weight coefficients interpretable, and the sum of the weights of all neighbors is 1, realizing the allocation of attention resources. This ensures that the sum of the standardized attention weight coefficients of all neighboring vertices for the same target vertex and the same relation is one, thus obtaining the standardized attention weight coefficients. Based on the calculated attention weight coefficients, the heterogeneous graph neural network model performs cross-modal message passing, that is, it sums the feature representation vectors of all neighboring vertices of the target vertex under each relation according to the corresponding attention weight coefficients, thereby obtaining the aggregated message vector passed to the target vertex under that relation. This step enables the target vertex to selectively and differentially absorb information from different relations and different neighbors, rather than simply averaging, which is the key to the model's accurate perception. The specific operation for calculating a low-dimensional collaborative anomaly embedding vector for each node in the dynamic spatiotemporal correlation graph is as follows: After obtaining the aggregated message vector under each relationship, for each target vertex in the current layer, a gated fusion mechanism is used to update its feature representation vector. This mechanism calculates a gate vector for physical proximity relationships and data association relationships respectively. The calculation process of each gate vector is as follows: the current feature representation vector of the target vertex is processed by a learnable linear transformation matrix corresponding to the relationship, and a corresponding learnable bias vector is added. The result is then input into a sigmoid function, which compresses the input value to the (0,1) interval. Its output value can be regarded as a "valve opening" to control the proportion of corresponding relationship messages passing through. The sigmoid function maps each element of the output vector to between zero and one, thereby obtaining a gate vector used to control the contribution of the corresponding relationship message vector to the final feature update. The aggregated message vector of physical proximity is first processed through a learnable linear transformation matrix, and then multiplied element-wise with its gate vector to obtain a weighted physical proximity message. Similarly, the aggregated message vector of data association is first processed through a learnable linear transformation matrix, and then multiplied element-wise with its gate vector to obtain a weighted data association message. The element-wise multiplication of the gate vector allows the model to have fine-grained control over different feature dimensions. For example, it may rely more on spatial information in some dimensions and more on data pattern information in others. Then, the feature representation vector of the target vertex itself is processed by a learnable linear transformation matrix, added to the two weighted message vectors mentioned above, and then passed through a linear rectified nonlinear activation function, namely the ReLU function, to introduce nonlinearity, enhance the expressive power of the model, and alleviate the possible gradient vanishing problem, generating a new feature representation vector of the target vertex in this processing layer. After the above iterative processing with a preset number of layers, the final feature representation vector of each vertex is output as a low-dimensional cooperative anomaly embedding vector. For example, if the input vertex attribute vector is 32-dimensional, after two layers of processing, a 16-dimensional cooperative anomaly embedding vector can be output, realizing feature compression and purification. The dimension of the cooperative anomaly embedding vector is lower than that of the input vertex attribute vector. Through training of the heterogeneous graph neural network model, it is optimized to be insensitive to widely consistent patterns in dynamic spatiotemporal relational graphs, but has the ability to distinguish anomaly patterns that appear in local vertex subgraphs through relational coupling. Its effect is that: for globally consistent patterns (such as large-area temperature rise), the feature changes of all nodes and their neighbors are similar, making it difficult to distinguish importance in attention calculation. Moreover, in gating fusion, the message patterns passed by different relations tend to converge, making it impossible to form unique activations. However, for locally coupled anomalies, the affected few nodes and their neighbors under specific relations will form feature patterns that are significantly different from other parts of the graph. This unique pattern is focused under the attention mechanism and is enhanced and refined layer by layer through multi-layer gating aggregation, and finally obtains a strong representation in the embedding vector. This makes the embedding vector an ideal feature for subsequent disease diagnosis.

[0023] In this embodiment, the specific operation of collaboratively training a lightweight classification model using the federated learning framework in step S4 is as follows: A global lightweight classification model is initialized on the cloud server, and its initial model parameters are distributed to each aggregation gateway. Each aggregation gateway maintains a dataset locally, consisting of multiple data samples. Each data sample contains a co-anomaly embedding vector generated in step S3, and a corresponding historical label pre-annotated by agronomists. The historical label identifies the health status of the citrus plant corresponding to the sample; for example, a value of zero represents healthy, and a value of one represents the presence of early Huanglongbing symptoms. The historical labels are usually annotated by agronomic experts after verification through field surveys or high-precision detection equipment, constituting the supervision signal for model training. In each round of federated training... In this process, the cloud server selects a subset of all aggregation gateways to participate in training, sending the model parameters of the current global lightweight classification model to these selected gateways. The selection of aggregation gateways can employ random sampling or strategies based on network status, power consumption, etc., to balance training efficiency and resource consumption. After receiving the model parameters of the global lightweight classification model, each selected aggregation gateway uses its local dataset to train the model locally, obtaining local model parameter updates. Local training can employ optimization algorithms such as stochastic gradient descent, performing a certain number of iterations to fit the model on local data. Simultaneously, each participating aggregation gateway calculates a contribution from its local dataset to the global model update. The contribution metric, designed to go beyond simple data volume weighting, measures the "uniqueness" and "information value" of the gateway's data by evaluating the difference between the local data gradient and the global gradient, thus enabling more intelligent allocation of aggregation weights. The specific calculation process is as follows: First, a data volume normalization factor is calculated, which equals the total number of samples in the current gateway's local dataset divided by the maximum value among the total number of samples in the local datasets of all participating aggregation gateways in this round of training. This ensures the basic contribution of data scale. Second, a gradient difference metric is calculated by calculating the loss obtained by the current gateway on its local dataset based on the received global model parameters. The average gradient vector of the function indicates the direction in which the model parameters should be optimized under the current gateway data distribution. The average of the above average gradient vectors of all participating aggregation gateways in this round of training is calculated and denoted as the global average gradient vector, which represents the overall optimization direction of the global data distribution in this round of training. The Euclidean norm of the difference between the current gateway's average gradient vector and the global average gradient vector is calculated to obtain the gradient deviation of the current gateway. This value quantifies the degree of difference between the current gateway's local data distribution and the global average distribution; the smaller the difference, the more consistent its data pattern is with the mainstream, and it may contain more general information. The sum of the gradient deviations of each of the participating aggregation gateways in this round of training is calculated.The normalized gradient difference ratio of the current gateway is obtained by dividing the gradient deviation of the current gateway by the sum of the total gradient deviations and the positive coefficient. The positive coefficient is a very small positive number, such as 1e-8, to prevent division by zero and ensure numerical stability. Then, the result of subtracting the normalized gradient difference ratio is calculated. This value is larger for gateways with smaller gradient deviations, indicating a higher contribution and encouraging the model to absorb more consensus-based knowledge. Finally, the data volume normalization factor is multiplied by the result of subtracting the normalized gradient difference ratio. The product is the contribution index of the aggregation gateway, which combines the data volume and data... The evaluation considers two dimensions: quality (and global consistency). Each aggregation gateway encrypts and uploads its calculated contribution metrics and local model parameter updates to the cloud server. The cloud server collects all contribution metrics and local model parameter updates uploaded by participating gateways and normalizes the contribution metrics to obtain the aggregate weight for each gateway. The normalization process involves dividing the contribution metric uploaded by each gateway by the sum of the contribution metrics uploaded by all participating gateways in this round. This operation is essentially a variant of the Softmax function, converting the contribution into a probability distribution so that the sum of all weights equals 1. Then... The cloud server uses these aggregate weights to perform a weighted average of the received local model parameter updates from each gateway. The weighted average calculation is as follows: First, a preset global aggregate learning rate is obtained. This learning rate controls the step size of the model parameter updates; for example, it can be set to 0.01 to smooth the update process and prevent model oscillation caused by excessively large updates in a single round. Then, the model parameters of the global lightweight classification model before this round of training are added to the global aggregate learning rate and multiplied by a weighted sum. This weighted sum equals the product of each gateway's aggregate weight and its uploaded local model parameter updates, and then this product is applied to all participating gateways. The results are summed to obtain the new model parameters of the global lightweight classification model after this round of training. These new model parameters are then applied to update the global lightweight classification model. This weighted aggregation mechanism makes gateways with high contributions have a greater impact on the global model, effectively guiding the model to learn from data distributions containing high-quality, high-consensus information. The above process is repeated for multiple rounds until the model performance converges or reaches a preset number of training rounds, resulting in a trained global lightweight classification model. The criteria for model convergence can be that the accuracy on the validation set no longer significantly improves after several consecutive rounds, or that a preset number of rounds is reached, such as 100 rounds. The trained lightweight classification model is deployed on the aggregation gateway, enabling the gateway to perform local inference on the real-time generated collaborative anomaly embedding vectors and output the early occurrence probability of citrus Huanglongbing. The specific operation is as follows: After federated training is complete, the cloud server distributes the parameters of the final global lightweight classification model to all aggregation gateways. Each aggregation gateway loads the parameters of this global lightweight classification model locally, completing its deployment. The lightweight classification model has a simple structure, such as containing only 1-2 fully connected layers, ensuring fast inference speed and low memory consumption on resource-constrained aggregation gateways. After deployment, the aggregation gateway continuously receives the co-anomaly embedding vectors generated in real time in step S3 and inputs them into the locally deployed global lightweight classification model for forward inference calculation. The global lightweight classification model processes the input co-anomaly embedding vectors and outputs a value between zero and one, which serves as the probability of early Huanglongbing (HLB) occurrence in the citrus plant corresponding to the sensor node of the embedding vector. This probability value can be used to set different thresholds to trigger different levels of early warning; for example, a probability greater than 0.7 triggers a high-risk warning, and a probability between 0.3 and 0.7 triggers a warning of concern. Simultaneously, an incremental learning loop for continuous model optimization is formed, which is crucial for the long-term effective operation of the system. Each aggregation gateway adds newly generated samples with subsequent manual verification results to its local dataset. The manual verification results are on-site checks of the model's warnings, and the verification results (healthy or diseased) are used as the true labels for the new samples. In subsequent federated learning training cycles, the aggregation gateway uses the local dataset containing the new samples for training. By calculating aggregation weights based on contribution metrics and performing weighted averaging on parameter updates, the knowledge of the new samples is integrated into the updated global lightweight classification model in subsequent training cycles. This allows the global lightweight classification model to continuously adapt to potential changes in disease patterns over time, such as changes in symptoms caused by pathogen mutations or changes in climate conditions. In addition, to ensure data privacy, the aggregation gateway can perform differential privacy perturbation or homomorphic encryption on the local model parameter updates before uploading them. Differential privacy perturbation protects individual sample information by adding calibrated noise, while homomorphic encryption allows the cloud to perform aggregation calculations in encrypted form. Both methods effectively prevent the derivation of original sensitive data from parameter updates.

[0024] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0026] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0029] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0030] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An early-stage collaborative perception method for citrus Huanglongbing based on Internet of Things, characterized in that, Specifically comprising the following steps: Step S1, at the sensor node deployed on the citrus plant, in response to the collected own multi-modal time series data generated by multiple types of sensors including spectrum sensors, infrared temperature sensors and micro deformation sensors and the associated sensor data generated by one type of sensor received from the adjacent node, based on the preset sliding time window, a collaborative perception confidence is calculated which integrates the consistency of the multi-modal time series data of the sensor node itself and the spatial consistency of the associated sensor data of the adjacent node, and when the collaborative perception confidence exceeds the preset threshold, the screened data features corresponding to the collaborative perception confidence are uploaded to the convergence gateway as preliminary abnormal feature data packets; Step S2, at the convergence gateway, receiving and caching the preliminary abnormal feature data packets uploaded from multiple sensor nodes, based on the time stamp and node location information carried by each preliminary abnormal feature data packet, using a weighted dynamic time warping method to align the preliminary abnormal feature data packets from different nodes and different sensors, wherein the weight given to each data point in the weighted dynamic time warping method is derived from the collaborative perception confidence calculated by the sensor node generating the data point at the corresponding time, and according to the spatial proximity relationship between nodes and the correlation of aligned data, a dynamic spatio-temporal association graph is constructed which contains multi-modal node attributes and multi-relation edge weights composed of aligned preliminary abnormal feature data packets; Step S3, at the processing unit in the cloud, inputting the dynamic spatio-temporal association graph into a pre-trained heterogeneous graph neural network model, and calculating a low-dimensional collaborative anomaly embedding vector for each node in the dynamic spatio-temporal association graph through the graph attention mechanism and cross-modal message passing of the heterogeneous graph neural network model; Step S4, between the convergence gateway and the cloud, based on the collaborative anomaly embedding vector and its corresponding pre-labeled historical label, a lightweight classification model is collaboratively trained using a federated learning framework, and the trained lightweight classification model is deployed on the convergence gateway, so that the convergence gateway can perform local inference on the real-time generated collaborative anomaly embedding vector and output the early occurrence probability of citrus Huanglongbing.

2. The early-stage collaborative perception method for citrus Huanglongbing based on the Internet of Things according to claim 1, characterized in that: In the step S1, the process of calculating a collaborative perception confidence which integrates the consistency of the multi-modal time series data of the sensor node itself and the spatial consistency of the associated sensor data of the adjacent node specifically comprises: Firstly, a local anomaly index reflecting the abnormal coupling degree between the multi-modal data of the sensor node itself is calculated, and the calculation process is as follows: in a sliding time window, for the multi-type sensors of the sensor node itself, the total number of the multi-type sensors is determined; all possible combinations of two different sensors are selected from the multi-type sensors, and the total number of all possible combinations is determined; for each sensor combination, the first data sequence and the second data sequence formed by the two sensors in the current time window are obtained; the normalized cross-correlation coefficient between the first data sequence and the second data sequence is calculated; the difference between the average value of the first data sequence in the current time window and the long-term historical average value thereof, and the difference between the average value of the second data sequence in the current time window and the long-term historical average value thereof are obtained respectively; The difference between the number one and the absolute value of the normalized cross-correlation coefficient is obtained as a first intermediate value; the difference between the difference value of the first data sequence and the difference value of the second data sequence is obtained, and the sign of the subtraction result is obtained as the sign of the second intermediate value; the first intermediate value is multiplied by the sign of the second intermediate value to obtain the contribution value corresponding to the sensor combination; for all possible combinations, the corresponding contribution values are calculated respectively, and all the contribution values are summed up, and then the sum result is divided by the total number of all possible combinations to obtain the local anomaly index; Secondly, an average spatial consistency residual reflecting the spatial data difference degree between the sensor node and each neighboring node is calculated, and the calculation process is as follows: for the sensor node and each neighboring node, the associated sensor data of a specified type shared by both parties is selected as the node data and the neighboring node data respectively; the absolute value of the difference between the node data and the neighboring node data is calculated; the standard deviation value of the specified type sensor data in the long-term history is obtained; the absolute value is divided by the standard deviation value to obtain the normalized difference value; the physical distance between the sensor node and the current neighboring node is obtained, and a preset distance attenuation coefficient is obtained; the exponential function value with the natural constant as the base number and the negative quotient of the physical distance and the distance attenuation coefficient as the index is calculated; the normalized difference value is multiplied by the exponential function value to obtain the spatial consistency residual corresponding to the current neighboring node; for all neighboring nodes, the corresponding spatial consistency residuals are calculated respectively, and then the spatial consistency residuals are taken as the arithmetic mean value, as the average spatial consistency residual; Finally, the collaborative perception confidence is generated through a fusion calculation, which takes the local anomaly index and the average spatial consistency residual as inputs, multiplies the local anomaly index and the average spatial consistency residual to obtain the first product as the numerator term; the first harmonic coefficient, the second harmonic coefficient and a positive coefficient are obtained; the first harmonic coefficient is multiplied by the local anomaly index to obtain the second product; the second harmonic coefficient is multiplied by the average spatial consistency residual to obtain the third product; the second product, the third product and the positive coefficient are added to obtain the sum result as the denominator term; the numerator term is divided by the denominator term to obtain the collaborative perception confidence. 3.The method of claim 2, wherein the method comprises: The screened data features corresponding to the cooperative perception confidence exceeding the preset threshold are uploaded to the convergence gateway as a preliminary abnormal feature data package, and the process is specifically as follows: A confidence threshold is preset in the sensor node; when the calculated cooperative perception confidence is greater than or equal to the confidence threshold, the sensor node determines that the data in the current sliding time window constitutes a potential cooperative anomaly, and then triggers the data uploading process; The sensor node encapsulates the multi-modal time series data in the current window, the calculated cooperative perception confidence, the data acquisition timestamp and the node identifier into a preliminary abnormal feature data package, and sends it to the convergence gateway; If the calculated cooperative perception confidence is less than the confidence threshold, the sensor node discards the multi-modal time series data in the current sliding time window and only sends a heartbeat signal indicating normal status to the convergence gateway.

4. The early-stage collaborative perception method for citrus Huanglongbing based on the Internet of Things according to claim 3, characterized in that: In step S2, the weighted dynamic time warping method is used to align the preliminary abnormal feature data packages from different nodes and different sensors, and the specific operation is as follows: At the convergence gateway, the received multiple preliminary abnormal feature data packages are buffered; the preliminary abnormal feature data packages corresponding to two sensor nodes to be aligned are extracted from the buffer; For each preliminary abnormal feature data package, the multi-modal time series data contained therein and the sequence of cooperative perception confidences calculated in step S1 and stored in the preliminary abnormal feature data package are extracted in time sequence, denoted as a cooperative perception confidence sequence; a weighted distance is defined to calculate the distance between any two data vectors from the multi-modal time series data corresponding to two time points in the two preliminary abnormal feature data packages, and the calculation process of the weighted distance is as follows: first, the Euclidean distance between the data vectors corresponding to the two time points is calculated; second, a normalization factor is obtained, which is the combined standard deviation of the same type of sensor data corresponding to the sensor nodes producing the two data packages in the long-term history; the Euclidean distance is divided by the normalization factor to obtain a first intermediate value; third, the cooperative perception confidences corresponding to the two time points from their respective cooperative perception confidence sequences are obtained, and the two cooperative perception confidences are multiplied to obtain a second intermediate value; the second intermediate value is compared with the positive coefficient, and the maximum value of the two values is taken as a third intermediate value; finally, the first intermediate value is divided by the third intermediate value to obtain the weighted distance; Based on the calculated weighted distance, a dynamic time warping algorithm is used to find an optimal warping path between the time points of the two preliminary abnormal feature data packages, which minimizes the total weighted distance of the two data packages aligned in time; according to the optimal warping path, the multi-modal time series data in the two preliminary abnormal feature data packages are resampled and aligned, so that they are mapped to a common warping time axis, and the aligned multi-modal time series segments with the same time length are output.

5. The early-stage collaborative perception method for citrus Huanglongbing based on the Internet of Things according to claim 4, characterized in that: The specific operation of constructing a dynamic space-time correlation graph including multi-modal node attributes and multi-relation edge weights composed of the aligned preliminary anomaly feature data packets according to the spatial proximity relationship between nodes and the correlation of the aligned data is as follows: Each sensor node is abstracted as a graph vertex in the dynamic space-time correlation graph, and an attribute is generated for each graph vertex, which is generated by calculating the statistical characteristics of the aligned multi-modal time series segment corresponding to the node on the public normalized time axis. The statistical characteristics include the average value vector of the multi-modal time series segment in all sensor type dimensions and the variance vector of the multi-modal time series segment in all sensor type dimensions. Then, the connection edge weights between any two graph vertices are calculated: The first one is the physical proximity edge weight, and the calculation process is as follows: the geographical distance between the sensor nodes represented by the two graph vertices is obtained; a preset distance bandwidth parameter is obtained; the exponential function value with the natural constant as the base and the square of the negative geographical distance divided by the square of twice the distance bandwidth parameter as the exponent is calculated, and the value is the physical proximity edge weight; The second one is the data correlation edge weight, and the calculation process is as follows: the aligned multi-modal time series segment and the corresponding collaborative perception confidence sequence of each of the two graph vertices are obtained; all time point pair combinations on the public normalized time axis are traversed, and for each time point pair, the Euclidean distance between the two data vectors in the pair is calculated; the maximum value of the Euclidean distances of all time point pair combinations is found; for each time point pair, the normalized complementary distance of the pair is calculated by subtracting the ratio of the Euclidean distance to the maximum Euclidean distance; for the same time point pair, a preset kernel function based on the two collaborative perception confidence calculation weight values is used to calculate a weight value according to the two collaborative perception confidence corresponding to the pair; The normalized complementary distances of all time point pairs are multiplied by the weight values corresponding thereto, and the sum of the product results is obtained to obtain the weighted distance sum; the sum of the weight values of all time point pairs is obtained to obtain the total weight sum; the data correlation edge weight is obtained by dividing the weighted distance sum by the total weight sum; Finally, the dynamic space-time correlation graph is composed of a set of all graph vertices, the attributes of each graph vertex, and the edges connecting the graph vertices and containing the two types of edge weights, i.e., the physical proximity edge weight and the data correlation edge weight.

6. The early-stage collaborative perception method for citrus Huanglongbing based on the Internet of Things according to claim 5, characterized in that: In step S3, the specific operation of inputting the dynamic space-time correlation graph into a pre-trained heterogeneous graph neural network model is as follows: The dynamic space-time correlation graph output in step S2 is input into the heterogeneous graph neural network model, and the heterogeneous graph neural network model processes each sensor node in the dynamic space-time correlation graph as a vertex in the graph, wherein each vertex has an attribute vector derived from the statistical characteristics of the aligned multi-modal time series segment in step S2. The heterogeneous graph neural network model distinguishes the connection relationship between vertices into two predefined types, the first is the physical proximity relationship, and the edge weight value adopts the physical proximity edge weight calculated in step S2, and the second is the data correlation relationship, and the edge weight value adopts the data correlation edge weight calculated in step S2; in the processing process of the heterogeneous graph neural network model, a set of independent and learnable model parameters are maintained for each type of relationship, including linear transformation matrix for feature transformation, attention calculation vector and bias vector, and the heterogeneous graph neural network model performs step-by-step deepening calculation on the input dynamic spatio-temporal association graph through multiple sequentially connected processing layers, and in each layer, the heterogeneous graph neural network model calculates an updated feature representation vector for each vertex in the dynamic spatio-temporal association graph.

7. The early-stage collaborative perception method for citrus Huanglongbing based on the Internet of Things according to claim 6, characterized in that: The specific operation of the graph attention mechanism and cross-modal message passing through the heterogeneous graph neural network model is as follows: In each processing layer of the heterogeneous graph neural network model, for each target vertex being processed as a target and each neighbor vertex connected thereto through the physical proximity relationship or the data correlation relationship, an attention weight calculation corresponding to the connection relationship is performed; the attention weight calculation corresponding to the connection relationship takes the feature representation vector of the target vertex in the current layer, the feature representation vector of the neighbor vertex in the current layer, and the edge weight value corresponding to the relationship connecting the two as inputs; first, the edge weight value corresponding to the relationship is converted into a vector through a learnable mapping function; then, the feature representation vector of the target vertex and the feature representation vector of the neighbor vertex are processed through a learnable linear transformation matrix corresponding to the relationship type respectively; then, the processed target vertex feature representation vector, neighbor vertex feature representation vector and vector converted from the edge weight value are spliced to form a combined vector; The combined vector is dot multiplied with a learnable attention vector corresponding to the relationship type, and the operation result is input into a preset nonlinear function with non-complete inhibition characteristics to obtain an original attention score representing the importance of the neighbor vertex to the target vertex under this relationship; Subsequently, the original attention scores of all neighbor vertices of the target vertex under the same relationship are normalized, and the normalization processing adopts an exponential function and summation operation, so that the sum of the normalized attention weight coefficients of all neighbor vertices of the same target vertex and the same relationship is one, thereby obtaining the normalized attention weight coefficients; Based on the calculated attention weight coefficients, the heterogeneous graph neural network model performs cross-modal message passing, that is, the feature representation vectors of all neighbor vertices of the target vertex under each relationship are weighted and summed according to the corresponding attention weight coefficients, thereby obtaining the aggregated message vector transmitted to the target vertex under this relationship. 8.The method of claim 7, wherein the method comprises: The specific operation of calculating a low-dimensional collaborative anomaly embedding vector for each node in the dynamic spatio-temporal association graph is as follows: After obtaining the aggregated message vectors for each relationship, for each target vertex in the current layer, a gating fusion mechanism is adopted to update its feature representation vector The mechanism calculates a gating vector for each of the physical proximity relationship and the data association relationship, and the calculation process of each gating vector is as follows: the current feature representation vector of the target vertex is processed through a learnable linear transformation matrix corresponding to the relationship, and a corresponding learnable bias vector is added, and then the result is input into an S-shaped growth function, which maps each element of the output vector to between zero and one, thereby obtaining the gating vector for controlling the contribution degree of the corresponding relationship message vector to the final feature update The aggregated message vector of the physical proximity relationship is first processed through a learnable linear transformation matrix, and then multiplied element by element with its gating vector to obtain the weighted physical proximity message; similarly, the aggregated message vector of the data association relationship is first processed through a learnable linear transformation matrix, and then multiplied element by element with its gating vector to obtain the weighted data association message Then, the feature representation vector of the target vertex itself is processed through a learnable linear transformation matrix, and then added to the above two weighted message vectors, and a linear rectifier nonlinear activation function is used to generate the new feature representation vector of the target vertex in the processing layer After the above iteration processing of a preset number of layers, the final feature representation vector of each vertex is output as a low-dimensional collaborative anomaly embedding vector. In step S4, the specific operation of collaboratively training a lightweight classification model using a federated learning framework is as follows: 9.The method of claim 8, wherein the method comprises: A global lightweight classification model is initialized on the cloud server, and the initial model parameters of the global lightweight classification model are distributed to each aggregation gateway; each aggregation gateway locally maintains a data set composed of multiple data samples, wherein each data sample includes a collaborative anomaly embedding vector generated by step S3 and a historical label corresponding to the vector pre-labeled by an agronomist; in each round of federated training, the cloud server selects a part of the aggregation gateways for training from all the aggregation gateways, and sends the model parameters of the current global lightweight classification model to the selected aggregation gateways; after receiving the model parameters of the global lightweight classification model, each selected aggregation gateway performs local training on the model using its local data set to obtain local model parameter updates; at the same time, each aggregation gateway participating in the training calculates a contribution index reflecting the contribution size of its local data set to the global model update; the specific calculation process of the contribution index is as follows: first, a data quantity normalization factor is calculated, which is equal to the total number of samples in the local data set of the current gateway divided by the maximum value in the total number of samples in the local data sets of all aggregation gateways participating in the training in the current round Secondly, a gradient difference measure item is calculated, and the process is as follows: the average gradient vector of the loss function based on the received global model parameters is calculated on the local data set of the current gateway ​ The average of the average gradient vectors of all the aggregation gateways participating in the training in the round is calculated, denoted as a global average gradient vector; the Euclidean norm of the difference between the average gradient vector of the current gateway and the global average gradient vector is calculated to obtain the gradient deviation degree of the current gateway; the sum of the gradient deviation degrees of all the aggregation gateways participating in the training in the round is calculated; the gradient deviation degree of the current gateway is divided by the sum of the gradient deviation degrees and a positive coefficient to obtain the normalized gradient difference ratio of the current gateway; then, one minus the normalized gradient difference ratio is calculated; finally, the data quantity normalization factor is multiplied by the result of one minus the normalized gradient difference ratio, and the product is the contribution degree index of the aggregation gateway; each aggregation gateway uploads the calculated contribution degree index and local model parameter update to the cloud server; after the cloud server collects the contribution degree indexes and local model parameter updates uploaded by all the participating gateways, the contribution degree indexes are normalized to obtain the aggregation weights of each gateway; Then, the cloud server uses the aggregation weights to perform weighted averaging on the received local model parameter updates of the gateways; then, the model parameters of the global lightweight classification model before the current training are added to the global aggregation learning rate multiplied by a weighted sum, and the weighted sum is equal to the aggregation weight of each gateway multiplied by the local model parameter update uploaded by the gateway, and then the sum of the products of all the participating gateways is summed to obtain the new model parameters of the global lightweight classification model after the current training, and the new model parameters are applied to update the global lightweight classification model; the above process is repeated for multiple rounds until the model performance converges or the preset training rounds are reached, and the trained global lightweight classification model is obtained.

10. The early-stage collaborative perception method for citrus Huanglongbing based on the Internet of Things according to claim 9, characterized in that: The trained lightweight classification model is deployed on the aggregation gateway, so that the aggregation gateway can perform local inference on the real-time generated collaborative anomaly embedding vector to output the specific operation of the early occurrence probability of citrus Huanglongbing disease: After the federal training is completed, the cloud server downloads the parameters of the final global lightweight classification model to all the aggregation gateways; each aggregation gateway loads the parameters of the global lightweight classification model locally to complete the deployment of the global lightweight classification model; after the deployment is completed, the aggregation gateway continuously receives the collaborative anomaly embedding vector generated in real time from step S3 and inputs it into the locally deployed global lightweight classification model for forward inference calculation; the global lightweight classification model processes the input collaborative anomaly embedding vector and outputs a value between zero and one as the early occurrence probability of Huanglongbing disease of the citrus plant where the sensor node corresponding to the embedding vector is located.

Images