Citrus disease and pest causation link analysis method based on knowledge graph reasoning

Abstract

The present application relates to the technical field of knowledge graph, in particular to a citrus disease and pest cause link analysis method based on knowledge graph reasoning, comprising the following steps: mapping environment abnormal features and disease state features into environment mutation event nodes and disease appearance event nodes respectively, constructing graph edges between the environment mutation event nodes and the disease appearance event nodes, extracting time stamps and spatial three-dimensional coordinates corresponding to each event node, and combining orchard wind direction vectors and water flow gradient vectors to generate a citrus spatio-temporal event knowledge graph. The present application can extract the node sequence most consistent with the spatio-temporal propagation rule from multiple candidate association paths by reversely tracing the retained graph edge direction and combining the edge spatio-temporal excitation weight and the benchmark occurrence probability to form a logarithmic likelihood cumulative excitation score, and further generate a citrus disease and pest cause evolution link, thereby improving the pertinence of disease and pest inducing cause identification.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to a method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning. Background Technology

[0002] Knowledge graphs are intelligent information organization technologies that use graph structures to organize and express entities, events, attributes, and relationships. They use nodes to represent objects, events, or concepts, and edges to represent the relationships between objects. They combine attribute information, time information, spatial information, and semantic constraints to build a data network with logical association capabilities.

[0003] Existing knowledge graph technologies primarily use nodes to represent objects, events, or concepts, and edges to represent the relationships between objects. They construct logically connected data networks through attribute information, temporal information, spatial information, and semantic constraints. In practice, they focus more on the expression of entity relationships and static semantic organization, lacking fine-grained constraints on whether events possess a true propagation direction, the role of propagation media, spatial consistency, and temporal validity. This leads to the simultaneous presence of causal relationships, co-occurrence relationships, proximity relationships, and manually labeled relationships in the graph, making it difficult to directly transform relationship types into reliable causal links in complex scenarios. Improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning, comprising the following steps:

[0006] Environmental anomaly features and disease status features are mapped to environmental mutation event nodes and disease manifestation event nodes, respectively. Graph edges are constructed between the environmental mutation event nodes and the disease manifestation event nodes. The timestamps and spatial three-dimensional coordinates corresponding to each event node are extracted. A spatiotemporal event knowledge graph of citrus is generated by combining the orchard wind direction vector and water flow gradient vector. Spatial relative coordinate vectors are obtained based on the spatial three-dimensional coordinates of the event nodes at both ends of the graph edges.

[0007] Based on the spatial relative coordinate vector, the orchard wind direction vector, and the water flow gradient vector, the direction cosine similarity and water flow cosine similarity are calculated. Combined with the preset medium influence weight, the medium bias influence factor is generated. According to the mapping relationship between the row coordinates corresponding to the environmental mutation event node and the column coordinates corresponding to the disease manifestation event node, the medium bias influence factor is filled into the relationship matrix to generate a spatial anisotropic bias matrix.

[0008] The time difference is calculated using the timestamps of the disease manifestation event node and the environmental mutation event node. The time difference is then substituted into the decay kernel function to obtain the time decay weight. The time decay weight is combined with the matrix elements in the spatial anisotropy bias matrix to generate the spatiotemporal excitation weights corresponding to the graph connection directions. The spatiotemporal excitation weights and graph connection directions that are greater than the set trigger limit are filtered and retained.

[0009] Preferably, the method further includes:

[0010] Assign a preset baseline occurrence probability to the disease manifestation event node; trace back along the preserved graph edge direction to extract the edge spatiotemporal excitation weights that constitute the associated path; calculate the log-likelihood cumulative excitation score corresponding to the associated path by combining the edge spatiotemporal excitation weights with the baseline occurrence probability; extract the node sequence associated with the log-likelihood cumulative excitation score with the largest value, and splice them according to the timestamp order to generate the causal evolution link of citrus diseases and pests.

[0011] Preferably, the steps for obtaining the citrus spatiotemporal event knowledge graph are as follows:

[0012] The environmental anomaly features are analyzed one by one according to the time of occurrence, location of occurrence, and type of anomaly. The analyzed environmental anomaly features are mapped to environmental mutation event nodes. The disease status features are analyzed one by one according to the time of disease manifestation, location of disease manifestation, and disease category. The analyzed disease status features are mapped to disease manifestation event nodes. Graph edges are constructed between the environmental mutation event nodes and disease manifestation event nodes according to the chronological correspondence of the events. The timestamps and spatial three-dimensional coordinates corresponding to each event node are extracted. The orchard wind direction vector and water flow gradient vector are retrieved and written into the propagation reference attributes of the corresponding graph edges to generate a citrus spatiotemporal event knowledge graph.

[0013] Preferably, the step of obtaining the spatial relative coordinate vector is as follows:

[0014] Based on the citrus spatiotemporal event knowledge graph, the starting event node and ending event node of each graph connection are read one by one. The starting event node is checked to see if it belongs to the environmental mutation event node, and the ending event node is checked to see if it belongs to the disease manifestation event node. For graph connections that meet the verification results, the spatial three-dimensional coordinates corresponding to the environmental mutation event node are extracted. For graph connections that meet the verification results, the spatial three-dimensional coordinates corresponding to the disease manifestation event node are extracted. The coordinate correspondence is established according to the starting event node and ending event node of the same graph connection to obtain the spatial three-dimensional coordinates corresponding to the event nodes at both ends of the graph connection.

[0015] Based on the spatial three-dimensional coordinates corresponding to the event nodes at both ends of the graph connection, the spatial three-dimensional coordinates of the disease manifestation event node corresponding to each graph connection are read. The horizontal, vertical, and triangular coordinates of the disease manifestation event node are used as minuends. The spatial three-dimensional coordinates of the environmental mutation event node corresponding to the same graph connection are read. The horizontal, vertical, and triangular coordinates of the environmental mutation event node are used as subtractors. directional subtraction of the horizontal coordinate difference, vertical coordinate difference, and triangular coordinate difference are performed respectively. The three coordinate differences are combined in the order of horizontal coordinate difference, vertical coordinate difference, and triangular coordinate difference to obtain the spatial relative coordinate vector.

[0016] Preferably, the step of obtaining the dielectric bias influence factor is as follows:

[0017] From the citrus spatiotemporal event knowledge graph, the orchard wind direction vector and water flow gradient vector corresponding to the edge directions of the graph are read one by one. The sum of the squares of the coordinate components of the spatial relative coordinate vector is calculated, and the square root of the sum of the squares of the coordinate components is taken. If the magnitude of the spatial relative coordinate vector is zero, the direction cosine similarity and water flow cosine similarity are both set to zero. If the magnitude of the spatial relative coordinate vector is non-zero, the magnitudes of the orchard wind direction vector and the water flow gradient vector are calculated respectively. For the orchard wind direction vector with a non-zero magnitude, the corresponding coordinates of the orchard wind direction vector and the spatial relative coordinate vector are multiplied. The cosine similarity is obtained by multiplying the cosine coordinates by the product of the orchard wind direction vector magnitude and the spatial relative coordinate vector magnitude. For the water flow gradient vector with a non-zero magnitude, the cosine similarity of the water flow gradient vector and the spatial relative coordinate vector is calculated by multiplying the cosine coordinates by the product of the cosine coordinates and the spatial relative coordinate vector magnitude. The cosine similarity of the water flow gradient vector is obtained by dividing the cosine similarity of the cosine coordinates by the product of the water flow gradient vector magnitude and the spatial relative coordinate vector magnitude. The direction cosine similarity is assigned to the orchard wind direction vector with a zero magnitude, and the water flow gradient vector with a zero magnitude is assigned to the water flow cosine similarity.

[0018] Read the preset medium influence weight corresponding to the direction cosine similarity, read the preset medium influence weight corresponding to the water flow cosine similarity, multiply the direction cosine similarity by the preset medium influence weight corresponding to the direction cosine similarity, multiply the water flow cosine similarity by the preset medium influence weight corresponding to the water flow cosine similarity, sum the two product results to obtain a weighted sum result, perform exponential scaling on the weighted sum result, record the exponentially scaled value to the corresponding graph connection direction, and generate the medium bias influence factor.

[0019] Preferably, the step of obtaining the spatial anisotropic bias matrix is ​​as follows:

[0020] The medium bias influence factor is read one by one according to the direction of the graph connection, the environmental mutation event node corresponding to the medium bias influence factor is located, the row coordinates of the environmental mutation event node in the relationship matrix are extracted, the disease manifestation event node corresponding to the medium bias influence factor is located, the column coordinates of the disease manifestation event node in the relationship matrix are extracted, and the medium bias influence factor is filled into the corresponding position in the relationship matrix according to the mapping relationship between the row coordinates of the environmental mutation event node and the column coordinates of the disease manifestation event node. Empty or zero value markers are reserved for the relationship matrix positions where no graph connection direction is formed, and a spatial anisotropic bias matrix is ​​generated.

[0021] Preferably, the steps for obtaining the edge spatiotemporal excitation weights and the graph connection directions are as follows:

[0022] The graph connection directions are read one by one from the citrus spatiotemporal event knowledge graph. The environmental mutation event node at the starting point of the graph connection direction is located and the timestamp of the environmental mutation event node is extracted. The disease manifestation event node at the ending point of the graph connection direction is located and the timestamp of the disease manifestation event node is extracted. The timestamp of the disease manifestation event node is subtracted from the timestamp of the environmental mutation event node to obtain the time difference value. The positive orientation of the time difference value is judged one by one. If the time difference value is less than or equal to zero, the corresponding graph connection direction is blocked. If the time difference value is greater than zero, the corresponding graph connection direction is retained. The retained time difference value is substituted into the preset time decay calculation item to obtain the time decay weight.

[0023] The preserved graph connection directions are read, and the environmental mutation event nodes in the preserved graph connection directions are determined as the starting point and the disease manifestation event nodes in the preserved graph connection directions are determined as the ending point. According to the row position of the environmental mutation event nodes in the spatial anisotropic bias matrix and the column position of the disease manifestation event nodes in the spatial anisotropic bias matrix, the matrix elements in the spatial anisotropic bias matrix are extracted. The matrix elements are multiplied by the time decay weight corresponding to the same graph connection direction to generate the edge spatiotemporal excitation weight corresponding to the graph connection direction.

[0024] Based on the edge spatiotemporal activation weights corresponding to the graph connection directions, all retained graph connection directions within the citrus spatiotemporal event knowledge graph are read one by one. The edge spatiotemporal activation weights corresponding to each graph connection direction are extracted. The edge spatiotemporal activation weights are compared numerically with a set trigger limit. If the edge spatiotemporal activation weight is greater than the set trigger limit, the edge spatiotemporal activation weight and the graph connection direction corresponding to the edge spatiotemporal activation weight are retained. If the edge spatiotemporal activation weight is less than or equal to the set trigger limit, the edge spatiotemporal activation weight and the graph connection direction corresponding to the edge spatiotemporal activation weight are removed. This yields edge spatiotemporal activation weights greater than the set trigger limit and the corresponding graph connection directions.

[0025] Preferably, the step of obtaining the log-likelihood cumulative excitation score is as follows:

[0026] The disease manifestation event nodes are read from the citrus spatiotemporal event knowledge graph. The node identifier, disease manifestation timestamp, and three-dimensional coordinates of the disease manifestation event nodes are checked. A preset baseline occurrence probability is written to the disease manifestation event nodes. Based on the edge spatiotemporal excitation weights that are greater than the set trigger limit and the corresponding graph connection direction, the disease manifestation event nodes are read in reverse along the preserved graph connection direction. The preceding environmental mutation event nodes are located level by level. The edge spatiotemporal excitation weights corresponding to each graph connection direction that constitutes a traceability relationship are extracted. The traceability relationship is organized according to the connection order from the environmental mutation event nodes to the disease manifestation event nodes to obtain the associated path and all the edge spatiotemporal excitation weights within the associated path.

[0027] Based on the associated path and all the edge spatiotemporal excitation weights within the associated path, the edge spatiotemporal excitation weights corresponding to each of the graph connection directions within the associated path are read one by one. A logarithmic transformation is performed on each edge spatiotemporal excitation weight. The baseline occurrence probability corresponding to the disease manifestation event node is read. A logarithmic transformation is performed on the baseline occurrence probability. All the edge spatiotemporal excitation weights after the logarithmic transformation are accumulated in the tracing order of the graph connection directions within the associated path. The accumulated result is combined and summed with the logarithmic value of the baseline occurrence probability to obtain the log-likelihood cumulative excitation score corresponding to the associated path.

[0028] Preferably, the steps for obtaining the causal evolutionary link of citrus diseases and pests are as follows:

[0029] The log-likelihood cumulative excitation score corresponding to each associated path is read one by one. All log-likelihood cumulative excitation scores are sorted according to their numerical values. The log-likelihood cumulative excitation score with the largest value is extracted. The node sequence associated with the log-likelihood cumulative excitation score with the largest value is locked. The timestamps corresponding to each event node in the node sequence are read. The environmental mutation event nodes and disease manifestation event nodes in the node sequence are arranged according to the order of the timestamps. All the arranged event nodes are spliced ​​together in sequence to generate the causal evolution link of citrus diseases and pests.

[0030] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0031] This invention transforms environmental anomaly features and disease state features into environmental mutation event nodes and disease manifestation event nodes, respectively, and binds these event nodes with timestamps, three-dimensional spatial coordinates, orchard wind direction vectors, and water flow gradient vectors. This imbues the graph structure, which originally only expressed entity associations, with spatiotemporal propagation constraints. By measuring the spatial orientation relationship between environmental mutation event nodes and disease manifestation event nodes using spatial relative coordinate vectors, and combining this with orchard wind direction vectors and water flow gradient vectors to calculate direction cosine similarity and water flow cosine similarity, the invention incorporates the actual media effects of the orchard, such as wind propagation and water flow diffusion, into the association strength calculation. This ensures that graph edges no longer remain static semantic connections but form connections with... The propagation bias relationship of spatial anisotropy differences is studied. By generating edge spatiotemporal excitation weights through time difference, time decay weights, and matrix elements within the spatial anisotropy bias matrix, and eliminating graph connection directions that do not conform to the temporal sequence or are below the set trigger limit, the interference of invalid associations, reverse associations, and accidental co-occurrence associations on cause judgment can be reduced. By tracing back the retained graph connection directions and combining the edge spatiotemporal excitation weights and the baseline occurrence probability to form a log-likelihood cumulative excitation score, the node sequence that best conforms to the spatiotemporal propagation law can be extracted from multiple candidate association paths, thereby generating the causal evolution link of citrus pests and diseases and improving the targeting of pest and disease induction cause identification. Attached Figure Description

[0032] Figure 1 This is a schematic diagram illustrating the influence factor of dielectric bias.

[0033] Figure 2 This is a schematic diagram of the spatiotemporal excitation weights at the edge. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0035] Please see Figure 1-2 This invention provides a technical solution: a method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning, comprising the following steps:

[0036] Environmental anomaly features and disease status features are mapped to environmental mutation event nodes and disease manifestation event nodes, respectively. Graph edges are constructed between environmental mutation event nodes and disease manifestation event nodes. The timestamps and spatial three-dimensional coordinates corresponding to each event node are extracted. A spatiotemporal event knowledge graph of citrus is generated by combining the orchard wind direction vector and water flow gradient vector. Spatial relative coordinate vectors are obtained based on the spatial three-dimensional coordinates of the event nodes at both ends of the graph edges.

[0037] Based on spatial relative coordinate vectors, orchard wind direction vectors, and water flow gradient vectors, the direction cosine similarity and water flow cosine similarity are calculated. Combined with the preset medium influence weights, the medium bias influence factor is generated. According to the mapping relationship between the row coordinates of the environmental mutation event node and the column coordinates of the disease manifestation event node, the medium bias influence factor is filled into the relationship matrix to generate a spatial anisotropic bias matrix.

[0038] The time difference is calculated using the timestamps of disease manifestation event nodes and environmental mutation event nodes. The time difference is then substituted into the decay kernel function to obtain the time decay weight. The time decay weight is combined with the matrix elements in the spatial anisotropy bias matrix to generate the spatiotemporal excitation weights of the edges corresponding to the graph connection directions. Edge spatiotemporal excitation weights and graph connection directions that are greater than the set trigger limit are selected and retained.

[0039] Assign a preset baseline probability of occurrence to disease manifestation event nodes; trace back along the preserved graph edges to extract the spatiotemporal excitation weights that constitute the associated paths; combine the spatiotemporal excitation weights with the baseline probability of occurrence to calculate the log-likelihood cumulative excitation score corresponding to the associated paths; extract the node sequence associated with the largest log-likelihood cumulative excitation score and splice them in chronological order according to the timestamps to generate the causal evolution link of citrus diseases and pests.

[0040] The steps to obtain the citrus spatiotemporal event knowledge graph are as follows:

[0041] The environmental anomaly features are analyzed one by one according to the time of occurrence, location of occurrence, and type of anomaly. The analyzed environmental anomaly features are mapped to environmental mutation event nodes. Similarly, the disease status features are analyzed one by one according to the time of disease manifestation, location of disease manifestation, and disease category. The analyzed disease status features are mapped to disease manifestation event nodes. Graph edges are constructed between environmental mutation event nodes and disease manifestation event nodes according to the chronological correspondence of events. The timestamps and spatial three-dimensional coordinates corresponding to each event node are extracted. Orchard wind direction vectors and water flow gradient vectors are retrieved and written into the propagation reference attributes of the corresponding graph edges to generate a spatiotemporal event knowledge graph for citrus.

[0042] Specifically, based on the collected environmental anomaly and disease status data, each environmental anomaly record is first structured and analyzed to extract the anomaly occurrence timestamp, three-dimensional spatial coordinates (horizontal, vertical, and longitudinal), and anomaly type identifier (e.g., three consecutive days of temperatures above 30 degrees Celsius, or heavy rainfall exceeding 50 mm within 24 hours). This structured information is then encapsulated into an environmental mutation event node. Similarly, each disease status record is analyzed to extract the disease manifestation timestamp, three-dimensional spatial coordinates, and disease category identifier (e.g., citrus Huanglongbing, citrus canker), and mapped to disease manifestation event nodes. All pairs of environmental mutation and disease manifestation event nodes are iterated through, based on a preset causal time window. For example, the disease manifestation timestamp must be later than the environmental mutation timestamp, and the time difference between the two must be between 3 and 21 days. If this condition is met, a directed graph connection is constructed between the two nodes, with the direction pointing from the environmental mutation event node to the disease manifestation event node. Then, the average wind direction and speed data and surface runoff simulation data within the time period (i.e., between the occurrence of the environmental event and the manifestation of the disease event) corresponding to each graph connection are retrieved from the historical database of the meteorological station and soil moisture sensor deployed in the orchard. These data are converted into normalized orchard wind direction vectors and water flow gradient vectors. For example, a level 3 northeast wind is converted into a three-dimensional vector (0.707, 0.707, 0), and the water flow direction along a specific slope is converted into a vector (0.1, -0.98, -0.15). These two vectors are then attached to the corresponding graph connection as propagation reference attributes. Finally, all event nodes, graph connections, and their attributes are integrated to generate a citrus spatiotemporal event knowledge graph containing information on time, space, event type, and propagation medium.

[0043] The steps to obtain spatial relative coordinate vectors are as follows:

[0044] Based on the knowledge graph of spatiotemporal events of citrus, the starting event node and ending event node of each graph connection are read one by one. The starting event node is checked to see if it belongs to the environmental mutation event node and the ending event node belongs to the disease manifestation event node. For graph connections that meet the verification results, the spatial three-dimensional coordinates corresponding to the environmental mutation event node are extracted. For graph connections that meet the verification results, the spatial three-dimensional coordinates corresponding to the disease manifestation event node are extracted. The coordinate correspondence is established according to the starting event node and ending event node of the same graph connection to obtain the spatial three-dimensional coordinates corresponding to the event nodes at both ends of the graph connection.

[0045] Based on the spatial three-dimensional coordinates of the event nodes at both ends of the graph connection, the spatial three-dimensional coordinates of the disease manifestation event node corresponding to each graph connection are read. The horizontal, vertical, and triangular coordinates of the disease manifestation event node are used as minuends. The spatial three-dimensional coordinates of the environmental mutation event node corresponding to the same graph connection are read. The horizontal, vertical, and triangular coordinates of the environmental mutation event node are used as subtractors. directional subtraction of the horizontal, vertical, and triangular coordinate differences is performed respectively. The three coordinate differences are combined in the order of horizontal, vertical, and triangular coordinate differences to obtain the spatial relative coordinate vector.

[0046] Specifically, based on the constructed citrus spatiotemporal event knowledge graph, a process of traversing all graph edges in the graph is initiated. For each graph edge being processed, the identifiers and type attributes of its starting and ending event nodes are first read. Then, a node type check is performed. The goal of this check is to confirm whether the starting event node has the type label "environmental mutation event" and whether the ending event node has the type label "disease manifestation event". For example, if the starting node type of a graph edge is "high temperature and drought" and the ending node type is "citrus red spider mite manifestation", then the graph edge passes the check. If the starting or ending types do not match, for example, the starting point is "artificial pruning" or the ending point is "environmental restoration", then the graph confidence is insufficient, and the graph edge is temporarily blocked. Marking them as invalid and skipping subsequent processing, for all graph edges that pass the type check, the spatial three-dimensional coordinates are then extracted from the data structure of their starting event nodes. These coordinates are usually recorded by a positioning system deployed in the orchard (such as a positioning base station based on differential GPS or UWB) when the event occurs. Similarly, the corresponding spatial three-dimensional coordinates are extracted from the ending event nodes of the same graph edge. These two sets of starting and ending spatial three-dimensional coordinates extracted from the same graph edge are paired to establish a one-to-one coordinate correspondence. This correspondence is stored as intermediate data to calculate the subsequent spatial relative positions. Finally, a series of spatial three-dimensional coordinates corresponding one-to-one with the valid graph edges are obtained, including the locations of environmental mutations and disease manifestations at both ends of the event nodes.

[0047] Based on the spatial three-dimensional coordinates of the event nodes at both ends of the graph connection obtained in the previous step, the spatial relative coordinate vector is calculated for each graph connection. Specifically, firstly, the spatial three-dimensional coordinates of the disease manifestation event node are read from the coordinate correspondence associated with the currently processed graph connection, and represented as (x2, y2, z2), where x2 is the horizontal coordinate, y2 is the vertical coordinate, and z2 is the subtrahend. These three coordinate components are then set as minuends. Next, the spatial three-dimensional coordinates of the environmental mutation event node corresponding to the same graph connection are read, represented as (x1, y1, z1), and these three coordinate components are set as subtrahends. Targeted coordinate difference calculations are then performed, specifically calculating the horizontal coordinate difference (x2 - x1), the vertical coordinate difference (y2 - y1), and the vertical coordinate difference (z2 - ... (z1) This calculation process reveals the magnitude and direction of the spatial displacement projected from the point of environmental anomaly to the point of disease manifestation along three basic axes. For example, if the coordinates of the environmental mutation event node are (10.5, 25.2, 3.1) and the coordinates of the disease manifestation event node are (12.8, 24.0, 3.5), then the difference in horizontal coordinates is 2.3 meters, the difference in vertical coordinates is -1.2 meters, and the difference in vertical coordinates is 0.4 meters. This indicates that the disease has moved 2.3 meters horizontally relative to the location of the environmental anomaly, moved 1.2 meters vertically in the opposite direction, and increased in height by 0.4 meters. The three coordinate differences obtained by calculation are combined in a fixed order (horizontal coordinate difference, vertical coordinate difference, vertical coordinate difference) to form a three-dimensional vector, which is the spatial relative coordinate vector.

[0048] The steps for obtaining the dielectric bias influence factor are as follows:

[0049] From the citrus spatiotemporal event knowledge graph, the orchard wind direction vector and water flow gradient vector corresponding to the edge directions are read one by one. The sum of the squares of the coordinate components of the spatial relative coordinate vector is calculated, and the square root of the sum of the squares of the coordinate components is taken. If the magnitude of the spatial relative coordinate vector is zero, the direction cosine similarity and water flow cosine similarity are both set to zero. If the magnitude of the spatial relative coordinate vector is non-zero, the magnitudes of the orchard wind direction vector and the water flow gradient vector are calculated respectively. For orchard wind direction vectors with non-zero magnitudes, the co-position of the orchard wind direction vector and the spatial relative coordinate vector is calculated. The coordinate multiplication accumulation is calculated by dividing the accumulated coordinate multiplication by the product of the orchard wind direction vector magnitude and the spatial relative coordinate vector magnitude to obtain the direction cosine similarity. For the water flow gradient vector with a non-zero magnitude, the corresponding coordinate multiplication accumulation is calculated by dividing the accumulated coordinate multiplication by the corresponding coordinate multiplication by the product of the water flow gradient vector magnitude and the spatial relative coordinate vector magnitude to obtain the water flow cosine similarity. The orchard wind direction vector with a zero magnitude is assigned a direction cosine similarity value of zero, and the water flow gradient vector with a zero magnitude is assigned a water flow cosine similarity value of zero.

[0050] Read the preset medium influence weight corresponding to the direction cosine similarity, read the preset medium influence weight corresponding to the water flow cosine similarity, multiply the direction cosine similarity by the preset medium influence weight corresponding to the direction cosine similarity, multiply the water flow cosine similarity by the preset medium influence weight corresponding to the water flow cosine similarity, sum the two product results to obtain a weighted sum result, perform exponential scaling on the weighted sum result, record the exponentially scaled value to the corresponding graph connection direction, and generate the medium bias influence factor.

[0051] Specifically, based on the generated citrus spatiotemporal event knowledge graph, each graph edge and its associated propagation reference attributes, namely the orchard wind direction vector and the water flow gradient vector, are read one by one. Simultaneously, the spatial relative coordinate vector corresponding to the graph edge is retrieved. First, the magnitude of the spatial relative coordinate vector is calculated, specifically by calculating the sum of the squares of its coordinate components. Then, the square root operation is performed on the result. If the calculated magnitude is zero due to spatial overlap between the two event points, the direction cosine similarity and water flow cosine similarity corresponding to the graph edge are directly assigned to zero. If the magnitude is non-zero, further processing continues, calculating the magnitudes of the orchard wind direction vector and the water flow gradient vector respectively. For orchard wind direction vectors with non-zero magnitudes, the dot product (i.e., the product of corresponding coordinate components) with the spatial relative coordinate vector is calculated. The sum of the two vectors is used to calculate the dot product of the two vector magnitudes. Then, the dot product is divided by the product of the magnitudes of the two vectors. The direction cosine similarity is obtained according to the cosine similarity calculation method. The value ranges from -1 to 1, which represents the consistency between the direction of disease spread and the wind direction. Similarly, for the water flow gradient vector with a non-zero magnitude, the dot product of its magnitude and the spatial relative coordinate vector is calculated and divided by the product of the magnitudes of the two vectors to obtain the water flow cosine similarity. This process quantifies the degree of agreement between the direction of disease spread and the direction of water flow. If the magnitude of the orchard wind direction vector or the water flow gradient vector is found to be zero during this process (e.g., no wind or flat terrain with no runoff), the corresponding direction cosine similarity or water flow cosine similarity is directly assigned to zero to avoid the error of dividing by zero in subsequent calculations. Finally, the similarity calculation of all graph edges is completed.

[0052] Based on the calculated direction cosine similarity and water flow cosine similarity, the medium bias influence factor is calculated for each graph edge. This requires setting the influence weights for wind propagation and water flow propagation. and These are preset values ​​determined based on experience or prior knowledge bases, according to the specific transmission characteristics of pests and diseases. For example, for citrus anthracnose spores, which are mainly spread by air currents, It can be set to 0.8, and Setting it to 0.2, conversely, for root rot, which is mainly spread through rain splash and surface runoff, It can be set to 0.7. Set to 0.3, and is usually set to Next, the directional cosine similarity obtained in the previous step is multiplied by its corresponding medium influence weight. At the same time, the cosine similarity of the water flow is multiplied by its corresponding medium influence weight. Then, the two products are added together to obtain a weighted sum, which combines the combined directional influence of wind and water on the spread of disease from the point of environmental abrupt change to the point of disease manifestation. To amplify the influence of the consistent direction and map it to a non-negative interval, the weighted sum is subjected to exponential scaling. Specifically, the natural exponent is taken from the weighted sum, i.e., applying... For example, if the direction cosine similarity is 0.9 and the water flow cosine similarity is 0.2, with weights of 0.8 and 0.2 respectively, then the weighted sum is... The result after exponential scaling is This value, after exponential scaling, is used as the final medium bias influence factor and recorded as a new attribute of the current graph edge.

[0053] The steps to obtain the spatial anisotropic bias matrix are as follows:

[0054] Read the medium bias influence factors one by one according to the direction of the graph connection, locate the environmental mutation event nodes corresponding to the medium bias influence factors, extract the row coordinates of the environmental mutation event nodes in the relation matrix, locate the disease manifestation event nodes corresponding to the medium bias influence factors, extract the column coordinates of the disease manifestation event nodes in the relation matrix, and fill the medium bias influence factors into the corresponding positions in the relation matrix according to the mapping relationship between the row coordinates of the environmental mutation event nodes and the column coordinates of the disease manifestation event nodes. For the relation matrix positions where no graph connection direction is formed, retain empty or zero value labels, and generate a spatial anisotropic bias matrix.

[0055] Specifically, based on the media bias influence factors calculated for each graph edge, a spatial anisotropy bias matrix is ​​constructed. The dimension of this matrix is ​​determined by the total number of environmental mutation event nodes (M) and the total number of disease manifestation event nodes (N) in the knowledge graph. An M-row, N-column zero or null matrix is ​​created as the initial entity for the relation matrix. Simultaneously, a mapping from the node's unique identifier to the matrix row and column indices is established for all environmental mutation event nodes and disease manifestation event nodes. For example, environmental mutation event nodes E1 to EM correspond to row indices 0 to M-1, and disease manifestation event nodes D1 to DN correspond to column indices 0 to N-1. Following the direction of the graph edges, the previously calculated media bias influence factors are read one by one. For each read factor, the starting point of its graph edge, i.e., the environmental mutation event node, is first located, and its position in the relation matrix is ​​queried. The row coordinates (row index) in the matrix are used to locate the endpoint of the graph connection, i.e., the disease manifestation event node. Then, its column coordinates (column index) in the relation matrix are queried, and the value of the medium bias influence factor is filled into the position in the relation matrix uniquely determined by this pair of row and column coordinates. For example, if the medium bias influence factor calculated for a graph connection from node E3 to node D5 is 2.14, and the row index of E3 is 2 and the column index of D5 is 4, then 2.14 is filled into the position of the 2nd row and 4th column of the relation matrix. This process is repeated until the medium bias influence factors of all graph connections are filled into the matrix. For combinations that do not form graph connections between environmental mutation event nodes and disease manifestation event nodes, their corresponding positions in the relation matrix will retain the initial zero value or empty mark. Finally, the filled relation matrix is ​​the spatial anisotropic bias matrix.

[0056] The steps for obtaining the edge spatiotemporal excitation weights and the direction of the graph connections are as follows:

[0057] The process involves sequentially reading the edges of the citrus spatiotemporal event knowledge graph, locating the environmental mutation event node at the starting point of each edge, extracting its timestamp, locating the disease manifestation event node at the ending point of each edge, extracting its timestamp, subtracting the timestamp of the environmental mutation event node from the timestamp of the disease manifestation event node to obtain the time difference, and then performing a positive evaluation on each time difference. If the time difference is less than or equal to zero, the corresponding edge of the graph is blocked; if the time difference is greater than zero, the corresponding edge of the graph is retained. The retained time difference is then substituted into a preset time decay calculation item to obtain the time decay weight.

[0058] Read the preserved graph connection directions, determine the environmental mutation event nodes in the preserved graph connection directions as the starting point, and determine the disease manifestation event nodes in the preserved graph connection directions as the ending point. According to the row position of the environmental mutation event nodes in the spatial anisotropy bias matrix and the column position of the disease manifestation event nodes in the spatial anisotropy bias matrix, extract the matrix elements in the spatial anisotropy bias matrix, and multiply the matrix elements by the time decay weight corresponding to the same graph connection direction to generate the edge spatiotemporal excitation weight corresponding to the graph connection direction.

[0059] Based on the spatiotemporal activation weights corresponding to the graph connection directions, all retained graph connection directions within the citrus spatiotemporal event knowledge graph are read one by one. The spatiotemporal activation weights corresponding to each graph connection direction are extracted. The spatiotemporal activation weights are compared with the set trigger limit values ​​one by one. If the spatiotemporal activation weight is greater than the set trigger limit value, the spatiotemporal activation weight and the graph connection direction corresponding to the spatiotemporal activation weight are retained. If the spatiotemporal activation weight is less than or equal to the set trigger limit value, the spatiotemporal activation weight and the graph connection direction corresponding to the spatiotemporal activation weight are removed. The result is the spatiotemporal activation weight and the corresponding graph connection direction that are greater than the set trigger limit value.

[0060] Specifically, based on the citrus spatiotemporal event knowledge graph, all graph edges are iteratively processed to calculate time decay weights. Each graph edge is read sequentially from the knowledge graph. For each edge, the environmental mutation event node at its starting point is located, and the recorded timestamp is extracted from the node's attributes. For example... Simultaneously, locate the disease manifestation event node at the endpoint of the edge connection in the graph, and extract its corresponding timestamp, for example... Calculate the difference between these two timestamps to obtain the time difference in hours or days. In this example, the time difference is approximately 5.23 days. Next, a positive positivity check is performed on the calculated time difference. This check aims to ensure the temporal plausibility of the causal relationship, meaning the effect must occur after the cause. If the time difference... A value less than or equal to zero implies that the disease manifests before or simultaneously with environmental anomalies, which is logically invalid. Therefore, the corresponding graph connection direction is logically blocked, meaning it is excluded from subsequent calculations. If the time difference... If the value is greater than zero, the direction of the edge in the graph is retained, and its time difference is substituted into a preset time decay kernel function for calculation. This kernel function usually uses an exponential decay model, and its form is: ,in It is a positive decay rate constant, and the setting of this constant is based on the knowledge of the specific disease's incubation period. For example, for diseases with a short incubation period, A larger value (e.g., 0.5) results in a faster weight decay due to a larger time difference, and vice versa (e.g., 0.1). The calculated... The value, i.e. the time decay weight, is added to the attribute of the current graph edge.

[0061] Based on the selected graph edges with positive time differences and their corresponding time decay weights, and combined with the previously generated spatial anisotropy bias matrix, the spatiotemporal excitation weights of each retained graph edge direction are calculated. Specifically, a retained graph edge is read, its starting point (environmental mutation event node) and ending point (disease manifestation event node) are identified, and the node-to-index mapping established during the construction of the spatial anisotropy bias matrix is ​​used to query the row index of the environmental mutation event node and the column index of the disease manifestation event node in the matrix. Through these row and column indices, a specific matrix element in the spatial anisotropy bias matrix can be located, and the value of this element is the previously calculated value. The medium bias influence factor reflects the anisotropic preference of disease propagation in space. Then, from the currently processed graph edge attributes, the time decay weight calculated in the previous step is read. This weight reflects the physical intuition that the longer the time interval from cause to effect, the lower the direct causal correlation. The matrix element extracted from the matrix (i.e., the medium bias influence factor) is multiplied by the corresponding time decay weight. The product of the two values ​​is defined as the edge spatiotemporal excitation weight corresponding to the graph edge direction. This weight comprehensively considers the influence of disease propagation in two dimensions: spatial medium and time delay. The calculated edge spatiotemporal excitation weight is stored as a new attribute on the corresponding graph edge to provide a quantitative basis for subsequent causal path screening.

[0062] Based on the spatiotemporal excitation weights of the edges calculated for each preserved graph connection, a global filtering and pruning operation is performed. First, a trigger limit needs to be set. This limit is determined based on statistical analysis of historical data and expert experience. One specific method is to calculate the average of all generated spatiotemporal excitation weights. and standard deviation Then trigger the limit. Set it to a value that takes into account both the average level and the degree of dispersion, for example, set it to This setting aims to preserve strong associations that are significantly above the average trigger level. For example, if the mean of all values ​​is 0.8 and the standard deviation is 0.2, the trigger limit is set to... Read each graph edge in the citrus spatiotemporal event knowledge graph that was preserved in the previous steps, extract the edge spatiotemporal trigger weight attached to each graph edge, and then compare this weight with the preset trigger limit. By comparing the numerical values, if the spatiotemporal activation weight of a graph connection edge is greater than the trigger limit (e.g., a weight of 1.2 is greater than 1.1), the connection edge is determined to represent a spatiotemporal association with high confidence. The spatiotemporal activation weight edge and its corresponding graph connection direction are retained. If the spatiotemporal activation weight of a graph connection edge is less than or equal to the set trigger limit (e.g., a weight of 0.9 is less than 1.1), the causal association represented by the connection edge is considered weak, and its activation intensity is insufficient to constitute a significant impact. Therefore, the spatiotemporal activation weight edge edge and its corresponding graph connection direction are removed from the effective connection set of the knowledge graph. Through this traversal and comparison, a simplified graph structure that has been filtered and only contains edge spatiotemporal activation weights greater than the set trigger limit edge and their corresponding graph connection directions is finally obtained.

[0063] The steps for obtaining the cumulative log-likelihood score are as follows:

[0064] Within the citrus spatiotemporal event knowledge graph, the analyzed disease manifestation event nodes are read, and the node identifier, disease manifestation timestamp, and three-dimensional spatial coordinates of the disease manifestation event nodes are verified. A preset baseline occurrence probability is written to the disease manifestation event nodes. Based on the edge spatiotemporal excitation weights that are greater than the set trigger limit and the corresponding graph connection directions, the disease manifestation event nodes are read in reverse along the preserved graph connection directions. The preceding environmental mutation event nodes are located level by level, and the edge spatiotemporal excitation weights corresponding to each graph connection direction that constitutes the traceability relationship are extracted. The traceability relationship is organized according to the connection order from the environmental mutation event nodes to the disease manifestation event nodes to obtain the associated paths and all edge spatiotemporal excitation weights within the associated paths.

[0065] Based on the associated paths and all spatiotemporal excitation weights of the edges within the associated paths, the spatiotemporal excitation weights of the edges corresponding to each graph connection direction within the associated paths are read one by one. A logarithmic transformation is performed on each edge spatiotemporal excitation weight. The baseline occurrence probability corresponding to the disease manifestation event node is read. A logarithmic transformation is performed on the baseline occurrence probability. All edge spatiotemporal excitation weights after logarithmic transformation are accumulated according to the tracing order of the graph connection directions within the associated paths. The accumulated result is combined and summed with the logarithmic value of the baseline occurrence probability to obtain the log-likelihood cumulative excitation score corresponding to the associated path.

[0066] Specifically, based on the filtered and retained graph edges and spatiotemporal excitation weights, for a specific disease manifestation event node to be analyzed, a reverse tracing of the causal path is performed. The target disease manifestation event node is read and locked in the citrus spatiotemporal event knowledge graph. Its unique node identifier, disease manifestation timestamp, and spatial three-dimensional coordinates are verified and recorded. A preset baseline occurrence probability is assigned to this disease manifestation event node. This probability represents the theoretical probability of the disease spontaneously occurring without considering any external environmental mutations. Its value is usually set according to the historical average incidence rate of the disease in a specific region and season; for example, it can be set to 0.01. Then, starting from this disease manifestation event node, a reverse tracing is performed according to the direction of the retained graph edges, that is, along the opposite direction of the arrows on the graph edges. The process involves traversing the graph to find all environmental mutation event nodes that directly point to the disease manifestation event node. Each time a reverse path (i.e., a valid graph edge) is found, it is used as the initial segment of the associated path. The spatiotemporal excitation weights of the edges attached to the graph edge are extracted. The process continues to recursively trace back from these newly located preceding environmental mutation event nodes, searching for earlier environmental mutation event nodes level by level, until no preceding node pointing to the current node can be found. All graph edges traversed during the entire tracing process and their corresponding spatiotemporal excitation weights are organized and recorded in the connection order from the earliest environmental mutation event node to the target disease manifestation event node, forming a complete associated path and an ordered set of all edge spatiotemporal excitation weights on the path.

[0067] Based on multiple associated paths obtained through reverse tracing and the total spatiotemporal excitation weights of edges contained within each path, the overall likelihood of each associated path is quantitatively scored. Specifically, each associated path is processed one by one, reading all graph edges contained within the path and extracting the spatiotemporal excitation weights corresponding to these graph edges. A logarithmic transformation is performed on each extracted edge spatiotemporal excitation weight, i.e., its natural logarithmic value is calculated. This transformation converts a multiplicative relationship into an additive relationship, facilitating subsequent cumulative calculations. Simultaneously, the node corresponding to the disease manifestation event ultimately pointed to by the associated path is read. The preset baseline occurrence probability is calculated, and then logarithmically transformed to obtain the logarithmic value of the baseline occurrence probability. All logarithmically transformed edge spatiotemporal excitation weights within the associated path are summed. This sum reflects the cumulative effect intensity of the cascading environmental mutation events that trigger disease manifestation along the path. This sum is then combined with the previously calculated logarithmic value of the baseline occurrence probability to obtain the final log-likelihood cumulative excitation score for that associated path. For example, if a path contains two excitation weights of 2.5 and 3.0, and the baseline occurrence probability is 0.01, then its score is... This calculation process is repeated for all possible associated paths, generating a log-likelihood cumulative excitation score for each path.

[0068] The steps to obtain the evolutionary pathways of citrus diseases and pests are as follows:

[0069] The log-likelihood cumulative excitation score corresponding to each associated path is read one by one. All log-likelihood cumulative excitation scores are sorted according to their numerical values. The log-likelihood cumulative excitation score with the largest value is extracted. The node sequence associated with the log-likelihood cumulative excitation score with the largest value is locked. The timestamps corresponding to each event node in the node sequence are read. The environmental mutation event nodes and disease manifestation event nodes in the node sequence are arranged according to the order of the timestamp records. All the arranged event nodes are spliced ​​together in sequence to generate the causal evolution link of citrus diseases and pests.

[0070] Specifically, based on the cumulative log-likelihood excitation score calculated for each associated path, the final causal evolutionary link extraction is performed. All associated paths and their corresponding cumulative log-likelihood excitation scores are summarized, and all scores are sorted in descending order, with the associated path having the highest score placed first. Since a higher cumulative log-likelihood excitation score indicates a higher probability of the causal relationship chain described by that path, the highest cumulative log-likelihood excitation score after sorting is directly extracted, and the node sequence associated with this highest score is identified. This node sequence consists of a series of environmental mutation event nodes and the final disease manifestation event nodes in a sequence. Next, in order to reconstruct the true timeline of events, it is necessary to read the attributes of each event node (including environmental mutation events and disease manifestation events) in the node sequence, extract their respective timestamp information, and rearrange all event nodes in the node sequence according to the order of these timestamps to ensure that the rearranged sequence follows the order from the earliest event to the latest event. Then, all event nodes arranged in the order of timestamps are sequentially spliced ​​together by their identifiers or names in a textual or logical manner to form a coherent event chain. This chain is the final identified and most likely evolutionary link of citrus disease and pest causes.

[0071] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning, characterized in that, Includes the following steps: Environmental anomaly features and disease status features are mapped to environmental mutation event nodes and disease manifestation event nodes, respectively. Graph edges are constructed between the environmental mutation event nodes and the disease manifestation event nodes. The timestamps and spatial three-dimensional coordinates corresponding to each event node are extracted. A spatiotemporal event knowledge graph of citrus is generated by combining the orchard wind direction vector and water flow gradient vector. Based on the spatial three-dimensional coordinates of the event nodes at both ends of the graph connection, obtain the spatial relative coordinate vector; Based on the spatial relative coordinate vector, the orchard wind direction vector, and the water flow gradient vector, the direction cosine similarity and water flow cosine similarity are calculated, and then processed in combination with the preset medium influence weight to generate the medium bias influence factor. Based on the mapping relationship between the row coordinates of the environmental mutation event nodes and the column coordinates of the disease manifestation event nodes, the medium bias influence factor is filled into the relation matrix to generate a spatial anisotropic bias matrix. The time difference is calculated using the timestamps of the disease manifestation event node and the environmental mutation event node, and the time difference is substituted into the decay kernel function to obtain the time decay weight. By combining the time decay weights with the matrix elements in the spatial anisotropy bias matrix, the spatiotemporal excitation weights of the edges corresponding to the edge connection directions of the graph are generated. The edge spatiotemporal excitation weights and the map connection directions that are greater than the set trigger limit are filtered and retained.

2. The method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning according to claim 1, characterized in that, The method further includes: Assign a preset baseline probability of occurrence to the disease manifestation event node; trace back along the preserved graph edge direction to extract the spatiotemporal excitation weights of the edges that constitute the associated path; The log-likelihood cumulative excitation score corresponding to the associated path is obtained by combining the edge spatiotemporal excitation weight with the baseline occurrence probability. Extract the node sequence associated with the log-likelihood cumulative excitation score with the largest value, and splice them together according to the time stamp order to generate the causal evolution link of citrus diseases and pests.

3. The method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning according to claim 1, characterized in that, The steps for obtaining the citrus spatiotemporal event knowledge graph are as follows: The environmental anomaly features are analyzed one by one according to the time of occurrence, location of occurrence, and type of anomaly. The analyzed environmental anomaly features are mapped to environmental mutation event nodes. The disease status features are analyzed one by one according to the time of disease manifestation, location of disease manifestation, and disease category. The analyzed disease status features are mapped to disease manifestation event nodes. Graph edges are constructed between the environmental mutation event nodes and disease manifestation event nodes according to the chronological correspondence of the events. The timestamps and spatial three-dimensional coordinates corresponding to each event node are extracted. The orchard wind direction vector and water flow gradient vector are retrieved and written into the propagation reference attributes of the corresponding graph edges to generate a citrus spatiotemporal event knowledge graph.

4. The method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning according to claim 1, characterized in that, The steps for obtaining the spatial relative coordinate vector are as follows: Based on the citrus spatiotemporal event knowledge graph, the starting event node and ending event node of each graph connection are read one by one. The starting event node is checked to see if it belongs to the environmental mutation event node, and the ending event node is checked to see if it belongs to the disease manifestation event node. For graph connections that meet the verification results, the spatial three-dimensional coordinates corresponding to the environmental mutation event node are extracted. For graph connections that meet the verification results, the spatial three-dimensional coordinates corresponding to the disease manifestation event node are extracted. The coordinate correspondence is established according to the starting event node and ending event node of the same graph connection to obtain the spatial three-dimensional coordinates corresponding to the event nodes at both ends of the graph connection. Based on the spatial three-dimensional coordinates corresponding to the event nodes at both ends of the graph connection, the spatial three-dimensional coordinates of the disease manifestation event node corresponding to each graph connection are read. The horizontal, vertical, and triangular coordinates of the disease manifestation event node are used as minuends. The spatial three-dimensional coordinates of the environmental mutation event node corresponding to the same graph connection are read. The horizontal, vertical, and triangular coordinates of the environmental mutation event node are used as subtractors. directional subtraction of the horizontal coordinate difference, vertical coordinate difference, and triangular coordinate difference are performed respectively. The three coordinate differences are combined in the order of horizontal coordinate difference, vertical coordinate difference, and triangular coordinate difference to obtain the spatial relative coordinate vector.

5. The method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning according to claim 1, characterized in that, The steps for obtaining the dielectric bias influence factor are as follows: From the citrus spatiotemporal event knowledge graph, the orchard wind direction vector and water flow gradient vector corresponding to the edge directions of the graph are read one by one. The sum of the squares of the coordinate components of the spatial relative coordinate vector is calculated, and the square root of the sum of the squares of the coordinate components is taken. If the magnitude of the spatial relative coordinate vector is zero, the direction cosine similarity and water flow cosine similarity are both set to zero. If the magnitude of the spatial relative coordinate vector is non-zero, the magnitudes of the orchard wind direction vector and the water flow gradient vector are calculated respectively. For the orchard wind direction vector with a non-zero magnitude, the corresponding coordinates of the orchard wind direction vector and the spatial relative coordinate vector are multiplied. The cosine similarity is obtained by multiplying the cosine coordinates by the product of the orchard wind direction vector magnitude and the spatial relative coordinate vector magnitude. For the water flow gradient vector with a non-zero magnitude, the cosine similarity of the water flow gradient vector and the spatial relative coordinate vector is calculated by multiplying the cosine coordinates by the product of the cosine coordinates and the spatial relative coordinate vector magnitude. The cosine similarity of the water flow gradient vector is obtained by dividing the cosine similarity of the cosine coordinates by the product of the water flow gradient vector magnitude and the spatial relative coordinate vector magnitude. The direction cosine similarity is assigned to the orchard wind direction vector with a zero magnitude, and the water flow gradient vector with a zero magnitude is assigned to the water flow cosine similarity. Read the preset medium influence weight corresponding to the direction cosine similarity, read the preset medium influence weight corresponding to the water flow cosine similarity, multiply the direction cosine similarity by the preset medium influence weight corresponding to the direction cosine similarity, multiply the water flow cosine similarity by the preset medium influence weight corresponding to the water flow cosine similarity, sum the two product results to obtain a weighted sum result, perform exponential scaling on the weighted sum result, record the exponentially scaled value to the corresponding graph connection direction, and generate the medium bias influence factor.

6. The method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning according to claim 1, characterized in that, The steps for obtaining the spatial anisotropic bias matrix are as follows: The medium bias influence factor is read one by one according to the direction of the graph connection, the environmental mutation event node corresponding to the medium bias influence factor is located, the row coordinates of the environmental mutation event node in the relationship matrix are extracted, the disease manifestation event node corresponding to the medium bias influence factor is located, the column coordinates of the disease manifestation event node in the relationship matrix are extracted, and the medium bias influence factor is filled into the corresponding position in the relationship matrix according to the mapping relationship between the row coordinates of the environmental mutation event node and the column coordinates of the disease manifestation event node. Empty or zero value markers are reserved for the relationship matrix positions where no graph connection direction is formed, and a spatial anisotropic bias matrix is ​​generated.

7. The method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning according to claim 1, characterized in that, The steps for obtaining the edge spatiotemporal excitation weights and the graph connection directions are as follows: The graph connection directions are read one by one from the citrus spatiotemporal event knowledge graph. The environmental mutation event node at the starting point of the graph connection direction is located and the timestamp of the environmental mutation event node is extracted. The disease manifestation event node at the ending point of the graph connection direction is located and the timestamp of the disease manifestation event node is extracted. The timestamp of the disease manifestation event node is subtracted from the timestamp of the environmental mutation event node to obtain the time difference value. The positive orientation of the time difference value is judged one by one. If the time difference value is less than or equal to zero, the corresponding graph connection direction is blocked. If the time difference value is greater than zero, the corresponding graph connection direction is retained. The retained time difference value is substituted into the preset time decay calculation item to obtain the time decay weight. The preserved graph connection directions are read, and the environmental mutation event nodes in the preserved graph connection directions are determined as the starting point and the disease manifestation event nodes in the preserved graph connection directions are determined as the ending point. According to the row position of the environmental mutation event nodes in the spatial anisotropic bias matrix and the column position of the disease manifestation event nodes in the spatial anisotropic bias matrix, the matrix elements in the spatial anisotropic bias matrix are extracted. The matrix elements are multiplied by the time decay weight corresponding to the same graph connection direction to generate the edge spatiotemporal excitation weight corresponding to the graph connection direction. Based on the edge spatiotemporal activation weights corresponding to the graph connection directions, all retained graph connection directions within the citrus spatiotemporal event knowledge graph are read one by one. The edge spatiotemporal activation weights corresponding to each graph connection direction are extracted. The edge spatiotemporal activation weights are compared numerically with a set trigger limit. If the edge spatiotemporal activation weight is greater than the set trigger limit, the edge spatiotemporal activation weight and the graph connection direction corresponding to the edge spatiotemporal activation weight are retained. If the edge spatiotemporal activation weight is less than or equal to the set trigger limit, the edge spatiotemporal activation weight and the graph connection direction corresponding to the edge spatiotemporal activation weight are removed. This yields edge spatiotemporal activation weights greater than the set trigger limit and the corresponding graph connection directions.

8. The method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning according to claim 2, characterized in that, The steps for obtaining the log-likelihood cumulative excitation score are as follows: The disease manifestation event nodes are read from the citrus spatiotemporal event knowledge graph. The node identifier, disease manifestation timestamp, and three-dimensional coordinates of the disease manifestation event nodes are checked. A preset baseline occurrence probability is written to the disease manifestation event nodes. Based on the edge spatiotemporal excitation weights that are greater than the set trigger limit and the corresponding graph connection direction, the disease manifestation event nodes are read in reverse along the preserved graph connection direction. The preceding environmental mutation event nodes are located level by level. The edge spatiotemporal excitation weights corresponding to each graph connection direction that constitutes a traceability relationship are extracted. The traceability relationship is organized according to the connection order from the environmental mutation event nodes to the disease manifestation event nodes to obtain the associated path and all the edge spatiotemporal excitation weights within the associated path. Based on the associated path and all the edge spatiotemporal excitation weights within the associated path, the edge spatiotemporal excitation weights corresponding to each of the graph connection directions within the associated path are read one by one. A logarithmic transformation is performed on each edge spatiotemporal excitation weight. The baseline occurrence probability corresponding to the disease manifestation event node is read. A logarithmic transformation is performed on the baseline occurrence probability. All the edge spatiotemporal excitation weights after the logarithmic transformation are accumulated in the tracing order of the graph connection directions within the associated path. The accumulated result is combined and summed with the logarithmic value of the baseline occurrence probability to obtain the log-likelihood cumulative excitation score corresponding to the associated path.

9. The method for analyzing the causal links of citrus diseases and pests based on knowledge graph reasoning according to claim 2, characterized in that, The steps for obtaining the causal evolutionary link of citrus diseases and pests are as follows: The log-likelihood cumulative excitation score corresponding to each associated path is read one by one. All log-likelihood cumulative excitation scores are sorted according to their numerical values. The log-likelihood cumulative excitation score with the largest value is extracted. The node sequence associated with the log-likelihood cumulative excitation score with the largest value is locked. The timestamps corresponding to each event node in the node sequence are read. The environmental mutation event nodes and disease manifestation event nodes in the node sequence are arranged according to the order of the timestamps. All the arranged event nodes are spliced ​​together in sequence to generate the causal evolution link of citrus diseases and pests.

Images