A machine learning based plague monitoring method

Abstract

The present application relates to the technical field of machine learning, and discloses a plague monitoring method based on machine learning. Host animal density data, medium flea index data, and temperature, humidity and vegetation index data of fixed monitoring points are acquired; the fixed monitoring points are taken as graph nodes, a heterogeneous space adjacency matrix is constructed according to geographical distance and water system flow direction, time sequence connection edges are constructed according to monitoring time stamps, and spatio-temporal graph structure data is generated; the spatio-temporal graph structure data is input into a spatio-temporal graph convolution network, and the host and medium features of adjacent graph nodes are aggregated based on the heterogeneous space adjacency matrix through a spatial graph convolution layer. The monitoring data is reconstructed into spatio-temporal graph structure data conforming to the topology of a physical epidemic source, the host and medium correlation features between adjacent regions are fused through the spatio-temporal graph convolution network, and the output plague occurrence probability distribution graph has continuity in the physical space and reflects the spatial distribution form of the internal risk of the epidemic source.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology and discloses a plague monitoring method based on machine learning. Background Technology

[0002] The conventional implementation of machine learning-based plague surveillance methods in existing technologies involves tabulating and concatenating host animal density data, vector flea index data, and temperature, humidity, and vegetation index data collected from environmental meteorological sensors at various fixed monitoring points. After obtaining the concatenated multidimensional feature vector, conventional solutions typically employ machine learning models such as Long Short-Term Memory networks or random forests, treating data collected from different geographical locations and at different times as independent and identically distributed samples to train the model. In the prediction phase, the model directly outputs the probability of plague occurrence at a specific monitoring point within a future timeframe based on the input feature vector.

[0003] The aforementioned existing technical solutions treat monitoring data from different geographical locations as independent samples when processing plague foci monitoring data. This severs the physical spatial coupling and linkage between rodent migration activities, flea parasitism transfer, and microenvironmental meteorological changes within natural plague foci at different monitoring points. Because no data association method reflecting the physical topology of the foci is introduced, conventional models cannot capture the heterogeneous interaction characteristics of host, vector, and environmental factors in adjacent spatial regions. Consequently, in areas with wide host activity ranges and complex habitats, the output spatial distribution of plague risk appears discrete, failing to provide early warning and location information with physical spatial continuity. Summary of the Invention

[0004] The purpose of this invention is to provide a plague monitoring method based on machine learning, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A plague monitoring method based on machine learning includes: acquiring host animal density data, vector flea index data, and temperature, humidity, and vegetation index data collected by environmental meteorological sensors at fixed monitoring points within the plague source area;

[0007] Using the fixed monitoring points as graph nodes, a heterogeneous spatial adjacency matrix is ​​constructed based on the geographical distance between the fixed monitoring points and the direction of water flow, and temporal connection edges are constructed based on the monitoring timestamps of the fixed monitoring points to generate spatiotemporal graph structure data.

[0008] The spatiotemporal graph structure data is input into a pre-constructed spatiotemporal graph convolutional network. The spatial graph convolutional layer in the spatiotemporal graph convolutional network aggregates the host and medium features of adjacent graph nodes based on the heterogeneous spatial adjacency matrix. The temporal convolutional layer in the spatiotemporal graph convolutional network extracts the seasonal and periodic dynamic features of the host and medium features along the temporal connection edges to obtain the spatiotemporal feature vector.

[0009] The spatiotemporal feature vector is input into a fully connected layer, and the fully connected layer outputs a probability distribution map of plague occurrence at the fixed monitoring point within a future preset time window.

[0010] Preferably, the step of constructing a heterogeneous spatial adjacency matrix based on the geographical distance between the fixed monitoring points and the direction of water flow includes: calculating the Euclidean distance between any two fixed monitoring points and using the reciprocal of the Euclidean distance as the distance weight;

[0011] Extract the water flow direction vector between any two fixed monitoring points. When the water flow direction vector points from the upstream fixed monitoring point to the downstream fixed monitoring point, assign a directional weight; otherwise, assign a zero value.

[0012] Multiply the distance weight by the direction weight to obtain the elements of the initial adjacency matrix;

[0013] The elements of the initial adjacency matrix are subjected to row normalization to generate the heterogeneous spatial adjacency matrix.

[0014] Preferably, the step of constructing a time-series connection edge based on the monitoring timestamps of the fixed monitoring points includes: sorting the monitoring timestamps of the same fixed monitoring point in chronological order, and constructing a directed time-series connection edge from the current timestamp to the next timestamp;

[0015] Calculate the time difference between two adjacent monitoring timestamps, and input the time difference into a sinusoidal periodic function to calculate the time decay coefficient;

[0016] The time decay coefficient is used as the edge weight of the directed temporal connection edge to generate a set of temporal connection edges containing non-uniform time interval weight information.

[0017] Preferably, the step of aggregating the host and media features of adjacent graph nodes based on the heterogeneous spatial adjacency matrix through the spatial graph convolutional layer in the spatiotemporal graph convolutional network includes: concatenating the host animal density data and the media flea index data into a node feature matrix;

[0018] Perform matrix multiplication between the heterogeneous space adjacency matrix and the node feature matrix to obtain the neighbor aggregation feature matrix;

[0019] The neighbor aggregation feature matrix is ​​input into the linear transformation layer of the spatial graph convolutional layer. The learnable weight matrix in the linear transformation layer is used to perform feature mapping on the neighbor aggregation feature matrix, and the spatial aggregation feature is output.

[0020] Preferably, the step of extracting the seasonal and periodic dynamic features of the host and media features through the temporal convolutional layer in the spatiotemporal graph convolutional network along the temporal connection edges includes: arranging the spatial aggregation features into a temporal feature sequence according to the order of the temporal connection edges;

[0021] Multiple dilated causal convolution kernels of different widths are used to perform convolution operations on the temporal feature sequence to extract temporal features at different time scales;

[0022] The time-series features at different time scales are input into a gated linear activation unit, and the seasonal and periodic dynamic features are output through the gated linear activation unit.

[0023] Preferably, the step of inputting the spatiotemporal feature vector into a fully connected layer and outputting a probability distribution map of plague occurrence at the fixed monitoring point within a future preset time window through the fully connected layer includes: flattening the spatiotemporal feature vector into a one-dimensional vector;

[0024] The one-dimensional vector is sequentially input into multiple cascaded fully connected hidden layers, wherein the fully connected hidden layers use rectified linear units as activation functions;

[0025] The output vector of the last fully connected hidden layer is input into the Sigmoid activation function, and the output vector is mapped to a probability matrix with values ​​between 0 and 1.

[0026] The elements in the probability value matrix are rearranged according to the geographical coordinates of the fixed monitoring points to generate the plague occurrence probability distribution map.

[0027] Preferably, after generating the heterogeneous spatial adjacency matrix, the method further includes: acquiring historical seasonal precipitation data of the area where the fixed monitoring point is located;

[0028] The average precipitation increment for each month is calculated based on the historical seasonal precipitation data, and the average precipitation increment is multiplied by a preset water level elevation coefficient to obtain the water system flow adjustment factor.

[0029] The water system flow adjustment factor is superimposed on the directional weight of the corresponding water system flow direction in the heterogeneous spatial adjacency matrix to generate an adaptive spatial adjacency matrix that dynamically changes with the month.

[0030] The adaptive spatial adjacency matrix replaces the heterogeneous spatial adjacency matrix and is input into the spatiotemporal graph convolutional network.

[0031] Preferably, before aggregating the output space features, the step of performing feature mapping on the neighbor aggregating feature matrix through the learnable weight matrix in the linear transformation layer further includes: calculating the cosine similarity between the central node features in the node feature matrix and the features of each neighbor node in the neighbor aggregating feature matrix;

[0032] The cosine similarity is input into the Softmax function to calculate the spatial attention coefficient matrix;

[0033] Perform a Hadamard product operation between the spatial attention coefficient matrix and the neighbor aggregation feature matrix to obtain a weighted neighbor aggregation feature matrix.

[0034] The weighted neighbor aggregated feature matrix is ​​input into the linear transformation layer for feature mapping.

[0035] Preferably, before outputting the seasonal and periodic dynamic features through the gated linear activation unit, the method further includes: performing a fast Fourier transform on the time-series features at different time scales to obtain a frequency domain feature sequence;

[0036] Extract low-frequency components with frequencies within a preset low-frequency range from the frequency domain feature sequence, and extract high-frequency components with frequencies within a preset high-frequency range from the frequency domain feature sequence.

[0037] The periodic trend characteristics are obtained by performing an inverse fast Fourier transform on the low-frequency components, and the residual fluctuation characteristics are obtained by performing an inverse fast Fourier transform on the high-frequency components.

[0038] The periodic trend feature and the residual fluctuation feature are added together and then input into the gated linear activation unit.

[0039] Preferably, after generating the plague occurrence probability distribution map, the method further includes: obtaining the geographic information system vector boundary polygon of the plague source area;

[0040] The vector boundary polygon of the geographic information system is rasterized into a binary mask matrix with the same resolution as the probability value matrix. In the binary mask matrix, the elements inside the epidemic source area are set to specific values, and the elements outside the epidemic source area are set to zero values.

[0041] The probability raster matrix corresponding to the plague occurrence probability distribution map is multiplied element-wise with the binary mask matrix to remove the probability values ​​of the area outside the plague source, and the plague occurrence probability distribution map with boundary alignment is output.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. This invention reconstructs monitoring data into a spatiotemporal graph structure that conforms to the topology of the physical plague foci by using fixed monitoring points as graph nodes, constructing a heterogeneous spatial adjacency matrix based on geographical distance and water flow direction, and constructing temporal connection edges containing non-uniform time interval weight information based on monitoring timestamps. In the operation of the spatiotemporal graph convolutional network, the spatial graph convolutional layer aggregates the host and vector features of adjacent graph nodes based on the above-mentioned heterogeneous spatial adjacency matrix, and the temporal convolutional layer extracts seasonal and periodic dynamic features along the temporal connection edges through expanded causal convolution kernels of different widths, enabling the model to directly integrate the host migration and vector transfer associations between adjacent regions during the feature extraction stage. The fully connected layer rearranges the fused spatiotemporal feature vectors into a probability distribution map according to the geographical coordinates of the fixed monitoring points, so that the output plague occurrence early warning results have physical spatial continuity and reflect the spatial distribution pattern of risk within the plague foci.

[0044] 2. In constructing the heterogeneous spatial adjacency matrix, historical seasonal precipitation data and water system flow direction vectors are combined to calculate the water system flow adjustment factor, generating an adaptive spatial adjacency matrix that dynamically changes with the month, ensuring that the spatial aggregation process matches the hydrological conditions of different months. In the spatial map convolution calculation, cosine similarity between the features of the central node and its neighboring nodes is introduced to construct a spatial attention coefficient matrix. In the temporal convolution calculation, a fast Fourier transform is used to separate frequency domain features and reconstruct periodic trend features and residual fluctuation features, eliminating interference from irrelevant spatial nodes and separating different frequency components in the time series data. At the end of generating the probability distribution map, a binary mask matrix is ​​constructed using the geographic information system vector boundary polygon and multiplied element-wise with the probability raster matrix, eliminating the probability output of areas outside the epidemic source, ensuring that the final spatial warning boundary is consistent with the actual geographic control boundary. Attached Figure Description

[0045] Figure 1 This is a general flowchart of the basic process for plague monitoring in this invention;

[0046] Figure 2 This is a flowchart of the heterogeneous spatial adjacency matrix construction process of the present invention;

[0047] Figure 3 This is the graph for constructing temporal connection edges and calculating time decay in this invention;

[0048] Figure 4 This is a flowchart of the spatiotemporal graph convolutional network feature extraction process of the present invention;

[0049] Figure 5 This is a flowchart of the adaptive spatial adjacency matrix generation process of the present invention;

[0050] Figure 6 This is a flowchart for generating and aligning the probability distribution map with the boundary in this invention. Detailed Implementation

[0051] Please refer to the attached document. Figure 1 This embodiment provides a plague monitoring method based on the fixed monitoring network of natural plague foci on land. The fixed monitoring points deployed in the foci all meet the deployment requirements of the national plague prevention and control monitoring standards. Each fixed monitoring point has standardized monitoring capabilities for host animals and vector fleas, and is equipped with environmental meteorological sensors and satellite remote sensing data receiving terminals, which can realize the periodic collection and standardized storage of monitoring data.

[0052] In this embodiment, host animal density data, vector flea index data, and temperature, humidity, and vegetation index data collected by environmental meteorological sensors at fixed monitoring points within the epidemic focus area are first acquired. Host animal density data is obtained using a standardized overnight trapping method. The monitoring targets are the main host rodents within the epidemic focus area, and the monitoring frequency is once a month. The data unit is animals / hectare, representing the host animal population size per unit area within the corresponding monitoring point. Vector flea index data is the ratio of the total number of fleas captured on the body surface of host animals to the total number of captured host animals during the monitoring period. The monitoring frequency is synchronized with the host animal density data, representing the parasitic abundance of vector fleas within the corresponding monitoring point.

[0053] The environmental meteorological sensors collect hourly temperature and relative humidity data. Monthly average temperature and relative humidity are calculated for each monitoring period, representing the temperature and moisture characteristics within the monitoring point area, respectively. Vegetation index data uses the Normalized Differential Vegetation Index (NDVI), extracted from a 16-day composite product of medium-resolution satellite remote sensing data. The extraction range is a buffer zone with a preset radius around the fixed monitoring point. The average NDVI value within the buffer zone is calculated, representing the vegetation cover and host habitat quality characteristics around the monitoring point.

[0054] Table 1. Comparison of Dimensions and Physical Meanings of Input Features at Fixed Monitoring Points

[0055] Feature Name Feature Dimension Physical meaning Collection cycle Host animal density 1 Number of host rodents per unit area Monthly Total index of vector fleas 1 Average number of fleas on the body surface of a single host animal Monthly Monthly average temperature 1 Arithmetic mean of hourly temperatures during the monitoring period Monthly Monthly average relative humidity 1 Arithmetic mean of hourly relative humidity during the monitoring period Monthly Normalized Difference Vegetation Index 1 Average NDVI within the monitoring point buffer zone 16 days

[0056] In this embodiment, the above five types of features are aligned in the time dimension and resampled in natural months to generate a 5-dimensional node feature vector for each fixed monitoring point in each monitoring month, providing basic data for subsequent graph structure construction.

[0057] Please refer to the attached document. Figure 2In this embodiment, fixed monitoring points are used as graph nodes. A heterogeneous spatial adjacency matrix is ​​constructed based on the geographical distance between the fixed monitoring points and the direction of water flow. Temporal connection edges are then constructed based on the monitoring timestamps of the fixed monitoring points to generate spatiotemporal graph structure data. Specifically, the geographical coordinates of all fixed monitoring points are first obtained, using latitude and longitude coordinates in the WGS84 coordinate system. These coordinates are then converted to Cartesian coordinates. The Euclidean distance between any two fixed monitoring points is calculated, and the corresponding distance weight is calculated using the following formula:

[0058]

[0059] in, For fixed monitoring points With fixed monitoring points Distance weights between them The distance between two fixed monitoring points in Cartesian coordinates is expressed in kilometers. The preset minimum value is set to a value of [value]. This is used to avoid the anomaly of a zero denominator when the distance is zero.

[0060] Extract the water flow direction vector between any two fixed monitoring points. Extract the watershed water network from the digital elevation model (DEM) data of the epidemic source area. Determine the watershed sub-unit to which each fixed monitoring point belongs and its upstream and downstream relationships. When the water flow direction vector flows from the upstream fixed monitoring point... Pointing to downstream fixed monitoring points When the value is zero, a directional weight is assigned; otherwise, a zero value is assigned. The formula for calculating the directional weight is as follows:

[0061]

[0062] in, For fixed monitoring points With fixed monitoring points Directional weights between them The preset directional weighting coefficient, with a value of 0.8, is used to characterize the promoting effect of water flow direction on the propagation of the host and the medium in space.

[0063] Multiplying the distance weight by the direction weight yields the elements of the initial adjacency matrix, and the corresponding calculation formula is as follows:

[0064]

[0065] in, The first in the initial adjacency matrix Line 1 The elements of the column correspond to fixed monitoring points. to fixed monitoring point The initial association weights.

[0066] Row normalization is performed on the elements of the initial adjacency matrix to generate a heterogeneous adjacency matrix. Row normalization is achieved by using the inverse of the degree matrix, and the corresponding calculation formula is as follows:

[0067]

[0068] in, For heterogeneous space adjacency matrix, The initial adjacency matrix The corresponding degree matrix is ​​a diagonal matrix, and the elements on the diagonal are... , To fix the total number of monitoring points, row normalization is used to eliminate feature aggregation bias caused by differences in the number of neighbors of different nodes.

[0069] Please refer to the attached document. Figure 3 The monitoring timestamps of the same fixed monitoring point are sorted according to chronological order, and directed temporal connections are constructed from the current timestamp to the next timestamp. The direction of the directed temporal connections represents the unidirectional flow of time, ensuring that no future data leakage occurs during feature extraction. The time difference between two adjacent monitoring timestamps is calculated, and the time difference is input into a sine periodic function to calculate the time decay coefficient. The corresponding calculation formula is as follows:

[0070]

[0071] in, Adjacent monitoring timestamps and The time decay coefficient between them The time difference between two monitoring timestamps, in days. The preset cycle length is 365 days, corresponding to the annual cycle characteristics of plague host and vector population changes. The preset offset, set to 0.2, ensures that the time decay coefficient remains positive, preventing feature cancellation due to negative weights. The time decay coefficient is used as the edge weight of directed temporal connections, generating a set of temporal connections containing weights for non-uniform time intervals. This set of temporal connections, together with the heterogeneous spatial adjacency matrix, constitutes the topology of the spatiotemporal graph data. Each graph node corresponds to a feature vector of a fixed monitoring point at a given monitoring timestamp. Spatial correlations are represented by the heterogeneous spatial adjacency matrix, while temporal correlations are represented by weighted temporal connections.

[0072] Please refer to the attached document. Figure 4In this embodiment, spatiotemporal graph structure data is input into a pre-constructed spatiotemporal graph convolutional network. The spatial graph convolutional layer in the spatiotemporal graph convolutional network aggregates host and vector features of adjacent graph nodes based on a heterogeneous spatial adjacency matrix. Furthermore, the temporal convolutional layer in the spatiotemporal graph convolutional network extracts the seasonal and periodic dynamic features of the host and vector features along temporal connection edges, resulting in a spatiotemporal feature vector. Specifically, host animal density data and vector flea index data are concatenated into a node feature matrix, the dimension of which is... ,in To fix the number of monitoring points, For each node's feature dimension, in this embodiment... The value is 5, corresponding to the aforementioned 5 types of input features. The neighbor aggregation feature matrix is ​​obtained by performing matrix multiplication on the heterogeneous space adjacency matrix and the node feature matrix. The corresponding calculation formula is:

[0073]

[0074] in, Aggregate the feature matrix for neighbors, and the dimensions are the same as the node feature matrix. Consistent, For a heterogeneous spatial adjacency matrix, matrix multiplication performs weighted aggregation of the features of each node to its neighboring nodes. The aggregation weights are determined by the elements in the heterogeneous spatial adjacency matrix. The neighbor aggregation feature matrix is ​​input into the linear transformation layer of the spatial graph convolutional layer. The learnable weight matrix in the linear transformation layer performs feature mapping on the neighbor aggregation feature matrix, outputting the spatial aggregation features. The corresponding calculation formula is as follows:

[0075]

[0076] in, The output spatial aggregation features, Here is the learnable weight matrix of the linear transformation layer, with dimension . , The output dimension for spatial features is set to 32 in this embodiment. Here are the learnable bias vectors for the linear transformation layer, with dimension . The linear transformation layer is used to map the aggregated neighbor features to a high-dimensional feature space to extract the spatial association patterns between host and media features.

[0077] The spatial aggregation features are arranged in the order of temporal connection edges to form a temporal feature sequence. The length of the temporal feature sequence is equal to the total number of monitored timestamps. The feature dimension of each time step is Multiple dilated causal convolution kernels of varying widths are used to perform convolution operations on the temporal feature sequences to extract temporal features at different time scales. The formula for dilated causal convolution is as follows:

[0078]

[0079] in, For convolution operations at time steps The output value, The first convolution kernel Each weight value The width of the convolution kernel. For the input time series feature sequence at time step eigenvalues, The dilation rate ensures that the causal property of dilated causal convolutions uses only historical data from the current time step and earlier, without introducing information from future time steps, thus avoiding data leakage. Convolution kernels with different widths and dilation rates can capture features at different time scales, including short-term fluctuations at the monthly scale and seasonal variations at the annual scale. Temporal features at different time scales are input into gated linear activation units (LAUs), which output seasonal and periodic dynamic features. The formula for the LAU operation is as follows:

[0080]

[0081] in, The seasonal and periodic dynamic characteristics of the output of the gated linear activation unit. This is the result of splicing together time-series features at different time scales. , For learnable convolutional kernel weights, , For the corresponding learnable bias, It is the Sigmoid activation function. For the Hadamard product operation, the gated linear activation unit controls the information transmission of temporal features through a gating mechanism, preserving the seasonal and periodic dynamic features related to the risk of plague occurrence while filtering out irrelevant noise information. Spatial aggregation features are concatenated with seasonal and periodic dynamic features to obtain a spatiotemporal feature vector. This vector integrates the spatial correlation and temporal dynamic features of the monitoring data, providing a foundation for subsequent risk prediction.

[0082] Please refer to the attached document. Figure 6 In this embodiment, the spatiotemporal feature vector is input into a fully connected layer, which outputs a probability distribution map of plague occurrence at fixed monitoring points within a preset future time window. Specifically, the spatiotemporal feature vector is flattened into a one-dimensional vector, the length of which is the product of the number of fixed monitoring points, the length of the time series, and the feature dimension. This one-dimensional vector is then sequentially input into multiple cascaded fully connected hidden layers. The fully connected hidden layers use a modified linear unit (MRU) as the activation function. The MRU's operation formula is as follows:

[0083]

[0084] in, The input values ​​for the fully connected hidden layers are used. Corrected linear units are introduced to introduce non-linear transformations, improving the model's ability to fit complex non-linear relationships. There are two fully connected hidden layers: the first has an output dimension of 128, and the second has an output dimension of 64. The output vector of the last fully connected hidden layer is input to the Sigmoid activation function, mapping the output vector to a probability matrix with values ​​between 0 and 1. The Sigmoid activation function is calculated as follows:

[0085]

[0086] in, The output value of the last fully connected hidden layer is represented by the probability value matrix. Each element in the probability value matrix corresponds to the probability of plague occurrence at a fixed monitoring point within a preset time window in the future. In this embodiment, the preset time window is three months from now. The elements in the probability value matrix are rearranged according to the geographical coordinates of the fixed monitoring points to generate a plague occurrence probability distribution map. During the rearrangement process, grid interpolation is performed according to the Cartesian coordinates of the fixed monitoring points to generate a continuous raster probability distribution map. The spatial resolution of the raster is consistent with the digital elevation model data of the plague source area.

[0087] In this embodiment, by constructing a spatiotemporal graph structure data that fits the physical topology of the epidemic source area by using fixed monitoring points as graph nodes, the joint representation of spatial correlation and temporal dynamics of monitoring data is realized. The host and medium features of adjacent nodes are aggregated through spatial graph convolutional layers, and the seasonal and periodic dynamic features of the features are extracted through temporal convolutional layers. Finally, a plague occurrence probability distribution map with spatial continuity is output, which solves the problems of fragmented spatial correlation of monitoring data and discrete risk output in the prior art.

[0088] Please refer to the attached document. Figure 5In a preferred embodiment, after generating the heterogeneous spatial adjacency matrix, the heterogeneous spatial adjacency matrix is ​​further seasonally dynamically adjusted to generate an adaptive spatial adjacency matrix that dynamically changes with the month. First, historical seasonal precipitation data for the area where the fixed monitoring point is located is acquired. This historical seasonal precipitation data comes from monthly precipitation observation data from meteorological stations within the epidemic source area, with a time span of no less than 10 years, and is used to characterize the seasonal variation pattern of precipitation within the epidemic source area. Based on the historical seasonal precipitation data, the average precipitation increment for each month is calculated. The average precipitation increment is the difference between the multi-year average precipitation for the corresponding month and the annual average monthly precipitation, used to characterize the degree of deviation of precipitation in different months from the annual average level. The average precipitation increment is multiplied by a preset water level elevation coefficient to obtain a water system flow adjustment factor. The water level elevation coefficient is determined based on the topographic slope and river gradient within the epidemic source area, and is used to convert the precipitation increment into the change in water system flow. The formula for calculating the water system flow adjustment factor is:

[0089]

[0090] in, For the first Monthly water system flow adjustment factor This is the preset water level elevation coefficient. For the first The average annual precipitation for each month This represents the average monthly precipitation throughout the year.

[0091] The water system flow adjustment factor is superimposed onto the directional weights of the corresponding water system flow direction in the heterogeneous spatial adjacency matrix to generate an adaptive spatial adjacency matrix that dynamically changes with the month. During the superposition process, the directional weights in the initial adjacency matrix are dynamically adjusted. The formula for the adjusted directional weights is as follows:

[0092]

[0093] in, For the first Fixed monitoring points in each month arrive The adjusted directional weights are dynamically updated according to monthly precipitation changes. In months with higher precipitation, the water system flow increases, and the directional weights are correspondingly increased, representing an increased probability of host and medium propagation along the water system. In months with lower precipitation, the directional weights are correspondingly decreased to match the water system propagation capacity during the dry season. Based on the adjusted directional weights, the initial adjacency matrix is ​​recalculated and row normalized to generate an adaptive spatial adjacency matrix. This adaptive spatial adjacency matrix replaces the heterogeneous spatial adjacency matrix and is input into the spatiotemporal graph convolutional network, enabling the spatial feature aggregation process to match the hydrological conditions changing in different months.

[0094] Table 2 Comparison of Average Precipitation and Water System Flow Adjustment Factor in Different Months in Epidemic Foci Areas

[0095] month Multi-year average precipitation (mm) Monthly average precipitation increase (mm) Water system flow adjustment factor Adjusted direction weights 1 2.3 -22.7 -0.1135 0.7092 2 3.5 -21.5 -0.1075 0.7140 3 10.2 -14.8 -0.0740 0.7408 4 25.6 0.6 0.0030 0.8024 5 45.3 20.3 0.1015 0.8812 6 78.5 53.5 0.2675 1.0140 7 125.8 100.8 0.5040 1.2032 8 108.6 83.6 0.4180 1.1344 9 58.2 33.2 0.1660 0.9328 10 23.5 -1.5 -0.0075 0.7940 11 8.7 -16.3 -0.0815 0.7348 12 3.2 -21.8 -0.1090 0.7128

[0096] In this embodiment, the water level elevation coefficient k is set to 0.005, the initial directional weight α is set to 0.8, and the adjusted directional weight is always kept non-negative to avoid feature aggregation anomalies caused by negative weights. The adaptive spatial adjacency matrix for different months can accurately match the seasonal changes in the water system transmission capacity within the epidemic source area, thereby improving the accuracy of spatial feature aggregation.

[0097] In this embodiment, a water system flow adjustment factor is calculated using historical seasonal precipitation data, and the directional weights of the heterogeneous spatial adjacency matrix are dynamically adjusted to generate an adaptive spatial adjacency matrix that varies with the month. This allows the spatial feature aggregation process to adapt to the hydrological conditions of different months, further aligning with the actual patterns of host and vector transmission along water systems within the epidemic source area, and improving the accuracy of spatial correlation feature extraction.

[0098] In a preferred embodiment, during the construction of temporal connection edges, the weight representation of non-uniform time intervals is further optimized, and a spatial attention mechanism is introduced during the feature aggregation process of spatial graph convolution to adaptively weight the features of neighboring nodes. When sorting the monitoring timestamps of the same fixed monitoring point according to chronological order, monitoring timestamps with missing data are marked, and valid monitoring timestamps with non-uniform distribution are retained. Directed temporal connection edges are constructed from the current valid timestamp to the next valid timestamp, eliminating the need for interpolation to fill in missing data and avoiding noise introduced by interpolation operations. The time difference between two adjacent valid monitoring timestamps is calculated, with the value ranging from 30 to 90 days, corresponding to monthly to quarterly monitoring intervals. The time difference is input into a sine periodic function to calculate the time decay coefficient. The time decay coefficient decreases as the time difference increases, indicating that the larger the time interval, the smaller the impact of historical data on the current moment.

[0099] Table 3. Comparison of Time Decay Coefficients for Different Time Intervals

[0100] Time interval (days) Time decay coefficient Temporal connection edge weights 30 0.4564 0.4564 60 0.6972 0.6972 90 0.9000 0.9000 120 1.0000 1.0000 180 0.9000 0.9000 270 0.4564 0.4564 365 0.2000 0.2000

[0101] In this embodiment, the period length T is 365 days, the offset β is 0.2, and the time decay coefficient reaches its peak at a time interval of 120 days, which corresponds to the optimal time correlation window for quarterly monitoring. After the time interval exceeds 180 days, the time decay coefficient drops rapidly, indicating that the impact of historical data of more than half a year on the current moment is reduced. The time decay coefficient can accurately characterize the temporal correlation strength of non-uniform time intervals, avoiding the feature bias caused by fixed time intervals.

[0102] Furthermore, before the spatial graph convolutional layer outputs the spatial aggregated features, spatial attention weighting is applied to the neighbor aggregated feature matrix. The cosine similarity is calculated between the features of the center node in the node feature matrix and the features of each neighbor node in the neighbor aggregated feature matrix. Cosine similarity characterizes the feature similarity between the center node and its neighbors; the higher the similarity, the greater the influence of the corresponding neighbor node on the center node. The formula for calculating cosine similarity is:

[0103]

[0104] in, As the central node with neighboring nodes Cosine similarity between them As the central node eigenvectors, For neighboring nodes eigenvectors, and These are the L2 norms of the corresponding eigenvectors.

[0105] The cosine similarity is input into the Softmax function to calculate the spatial attention coefficient matrix. The Softmax function is used to normalize the cosine similarity of all neighboring nodes of the same central node, so that the sum of the attention coefficients of all neighboring nodes is 1. The calculation formula is as follows:

[0106]

[0107] in, As the central node For neighboring nodes Spatial attention coefficient, As the central node The set of all neighboring nodes.

[0108] The spatial attention coefficient matrix and the neighbor aggregation feature matrix are subjected to a Hadamard product operation to obtain a weighted neighbor aggregation feature matrix. The Hadamard product operation is an element-wise multiplication, so that the features of each neighbor node are weighted by the corresponding attention coefficient. The calculation formula is as follows:

[0109]

[0110] in, The weighted neighbor aggregation feature matrix, The spatial attention coefficient matrix, dimension, and neighbor aggregation feature matrix are shown. Consistent.

[0111] The weighted neighbor aggregation feature matrix is ​​input into the linear transformation layer for feature mapping, and the output is spatial aggregation features. Through the spatial attention mechanism, the feature contribution weights of different neighbor nodes to the central node can be adaptively adjusted, the weights of neighbor nodes with high feature similarity to the central node are increased, the weights of irrelevant neighbor nodes are reduced, the interference of irrelevant spatial nodes is eliminated, and the accuracy of spatial feature aggregation is improved.

[0112] In this embodiment, the time decay coefficient of non-uniform time intervals is calculated by a sinusoidal periodic function, and weighted temporal connection edges are constructed to accurately characterize the temporal correlation strength of the monitoring data. At the same time, a spatial attention mechanism based on cosine similarity is introduced into the spatial graph convolution to adaptively weight the features of neighboring nodes, eliminate the interference of irrelevant spatial nodes, and further improve the accuracy and robustness of spatiotemporal feature extraction.

[0113] In a preferred embodiment, during the extraction of temporal features in the temporal convolutional layer, different frequency components of the temporal features are further separated by frequency domain transformation. Simultaneously, after generating the plague occurrence probability distribution map, the probability distribution map is aligned to the boundaries using geographic information system vector boundaries. Specifically, before inputting temporal features at different time scales into the gated linear activation unit, a Fast Fourier Transform (FFT) is performed on the temporal features at different time scales to obtain a frequency domain feature sequence. The FFT converts the temporal features in the time domain into frequency components in the frequency domain. The formula for the Discrete Fast Fourier Transform is:

[0114]

[0115] in, Frequency in the frequency domain feature sequence The corresponding components, For time-domain temporal feature sequences, The length of the time-series feature sequence. It is the imaginary unit.

[0116] Extract low-frequency components within a preset low-frequency range and high-frequency components within a preset high-frequency range from the frequency domain feature sequence. The frequency range of the low-frequency range is 0 to 1 / 365. The characteristic components corresponding to a period of 1 year or more characterize the annual periodic trends of changes in plague host and vector populations; the frequency range of the high-frequency interval is 1 / 30. To 1 / 10 The corresponding feature components, with periods ranging from 10 to 30 days, characterize the short-term residual fluctuations of host and vector populations. Inverse Fast Fourier Transform (IFFT) is performed on the low-frequency components to obtain periodic trend features, and IFFT is performed on the high-frequency components to obtain residual fluctuation features. IFFT is the inverse operation of IFFT, converting the frequency domain components back to time domain features. The periodic trend features and residual fluctuation features are added together and input into a gated linear activation unit. Through frequency domain separation and recombination, seasonal and periodic dynamic features in the time series can be accurately extracted, filtering out mid-frequency noise components unrelated to the risk of plague occurrence, thus improving the effectiveness of time series feature extraction.

[0117] Table 4. Parameter Configuration Table for Dilated Causal Convolution Kernels

[0118] kernel number kernel width Expansion rate Experiencing the length of the wild (days) Corresponding time scale 1 3 1 3 Monthly short-term fluctuations 2 3 3 7 Quarterly fluctuations 3 3 12 25 Half-year fluctuations 4 3 24 49 Annual seasonal

[0119] In this embodiment, the time step interval of the time series feature sequence is 1 day. Four sets of dilated causal convolution kernels with different parameters can capture time series features at the monthly, quarterly, semi-annual and annual scales, respectively, covering all key time scales of changes in plague host and vector populations. After frequency domain separation and recombination, the time series features at different scales can be further separated into periodic trends and short-term fluctuation features, providing more accurate time series feature inputs for subsequent risk prediction.

[0120] Further, after generating the plague occurrence probability distribution map, boundary alignment processing is performed on the probability distribution map. The geographic information system (GIS) vector boundary polygon of the plague source area is obtained. The vector boundary polygon represents the legal control boundary of the plague source area, using the WGS84 coordinate system, and contains multiple consecutive coordinate points forming a closed polygonal region. The GIS vector boundary polygon is rasterized into a binary mask matrix with the same resolution as the probability value matrix. During rasterization, each raster cell is checked to see if its point is inside the polygon. If the center point of the raster cell is inside the vector boundary polygon, the corresponding binary mask matrix element is set to 1; if the center point of the raster cell is outside the vector boundary polygon, the corresponding binary mask matrix element is set to 0, generating a binary mask matrix. The probability raster matrix corresponding to the plague occurrence probability distribution map is multiplied element-wise by the binary mask matrix to remove probability values ​​from areas outside the plague source area. The calculation formula is:

[0121]

[0122] in, This is the probability raster matrix after boundary alignment. For the initial probability grid matrix, It is a binary mask matrix. This is the Hadama multiplication operation. Through element-wise multiplication, the probability values ​​of the areas outside the plague source are all set to zero, and only the probability values ​​of the areas inside the plague source are retained. The output is a plague occurrence probability distribution map with aligned boundaries, so that the final early warning result is completely consistent with the actual geographical control boundary of the plague source, which facilitates the implementation of subsequent plague prevention and control measures.

[0123] In this embodiment, the temporal features are separated in the frequency domain by fast Fourier transform, and the periodic trend features and residual fluctuation features are extracted and recombined, which improves the accuracy of temporal dynamic feature extraction. At the same time, a binary mask matrix is ​​constructed by constructing the vector boundary of the geographic information system to perform boundary alignment processing on the probability distribution map, eliminating invalid probability outputs of areas outside the epidemic source, so that the final warning result is consistent with the actual control boundary, and improving the practicality of the warning result.

Claims

1. A plague monitoring method based on machine learning, characterized in that, include: Acquire host animal density data, vector flea index data, and temperature, humidity, and vegetation index data collected by environmental meteorological sensors at fixed monitoring points within the epidemic source area; Using the fixed monitoring points as graph nodes, a heterogeneous spatial adjacency matrix is ​​constructed based on the geographical distance between the fixed monitoring points and the direction of water flow, and temporal connection edges are constructed based on the monitoring timestamps of the fixed monitoring points to generate spatiotemporal graph structure data. The spatiotemporal graph structure data is input into a pre-constructed spatiotemporal graph convolutional network. The spatial graph convolutional layer in the spatiotemporal graph convolutional network aggregates the host and medium features of adjacent graph nodes based on the heterogeneous spatial adjacency matrix. The temporal convolutional layer in the spatiotemporal graph convolutional network extracts the seasonal and periodic dynamic features of the host and medium features along the temporal connection edges to obtain the spatiotemporal feature vector. The spatiotemporal feature vector is input into a fully connected layer, and the fully connected layer outputs a probability distribution map of plague occurrence at the fixed monitoring point within a future preset time window.

2. The plague monitoring method based on machine learning according to claim 1, characterized in that, The step of constructing a heterogeneous spatial adjacency matrix based on the geographical distance between the fixed monitoring points and the direction of water flow includes: calculating the Euclidean distance between any two fixed monitoring points and using the reciprocal of the Euclidean distance as the distance weight; Extract the water flow direction vector between any two fixed monitoring points. When the water flow direction vector points from the upstream fixed monitoring point to the downstream fixed monitoring point, assign a directional weight; otherwise, assign a zero value. Multiply the distance weight by the direction weight to obtain the elements of the initial adjacency matrix; The elements of the initial adjacency matrix are subjected to row normalization to generate the heterogeneous spatial adjacency matrix.

3. The plague monitoring method based on machine learning according to claim 1, characterized in that, The step of constructing a time-series connection edge based on the monitoring timestamps of the fixed monitoring points includes: sorting the monitoring timestamps of the same fixed monitoring point in chronological order, and constructing a directed time-series connection edge from the current timestamp to the next timestamp; Calculate the time difference between two adjacent monitoring timestamps, and input the time difference into a sinusoidal periodic function to calculate the time decay coefficient; The time decay coefficient is used as the edge weight of the directed temporal connection edge to generate a set of temporal connection edges containing non-uniform time interval weight information.

4. The plague monitoring method based on machine learning according to claim 1, characterized in that, The step of aggregating host and vector features of adjacent graph nodes based on the heterogeneous spatial adjacency matrix through the spatial graph convolutional layer in the spatiotemporal graph convolutional network includes: concatenating the host animal density data and the vector flea index data into a node feature matrix; Perform matrix multiplication between the heterogeneous space adjacency matrix and the node feature matrix to obtain the neighbor aggregation feature matrix; The neighbor aggregation feature matrix is ​​input into the linear transformation layer of the spatial graph convolutional layer. The learnable weight matrix in the linear transformation layer is used to perform feature mapping on the neighbor aggregation feature matrix, and the spatial aggregation feature is output.

5. The plague monitoring method based on machine learning according to claim 1, characterized in that, The step of extracting the seasonal and periodic dynamic features of the host and media features through the temporal convolutional layer in the spatiotemporal graph convolutional network along the temporal connection edges includes: arranging the spatial aggregation features into a temporal feature sequence according to the order of the temporal connection edges; Multiple dilated causal convolution kernels of different widths are used to perform convolution operations on the temporal feature sequence to extract temporal features at different time scales; The time-series features at different time scales are input into a gated linear activation unit, and the seasonal and periodic dynamic features are output through the gated linear activation unit.

6. The plague monitoring method based on machine learning according to claim 1, characterized in that, The step of inputting the spatiotemporal feature vector into a fully connected layer and outputting a probability distribution map of plague occurrence at the fixed monitoring point within a future preset time window through the fully connected layer includes: flattening the spatiotemporal feature vector into a one-dimensional vector. The one-dimensional vector is sequentially input into multiple cascaded fully connected hidden layers, wherein the fully connected hidden layers use rectified linear units as activation functions; The output vector of the last fully connected hidden layer is input into the Sigmoid activation function, and the output vector is mapped to a probability matrix with values ​​between 0 and 1. The elements in the probability value matrix are rearranged according to the geographical coordinates of the fixed monitoring points to generate the plague occurrence probability distribution map.

7. The plague monitoring method based on machine learning according to claim 2, characterized in that, After generating the heterogeneous spatial adjacency matrix, the method further includes: acquiring historical seasonal precipitation data for the area where the fixed monitoring point is located; The average precipitation increment for each month is calculated based on the historical seasonal precipitation data, and the average precipitation increment is multiplied by a preset water level elevation coefficient to obtain the water system flow adjustment factor. The water system flow adjustment factor is superimposed on the directional weight of the corresponding water system flow direction in the heterogeneous spatial adjacency matrix to generate an adaptive spatial adjacency matrix that dynamically changes with the month. The adaptive spatial adjacency matrix replaces the heterogeneous spatial adjacency matrix and is input into the spatiotemporal graph convolutional network.

8. The plague monitoring method based on machine learning according to claim 4, characterized in that, Before aggregating features in the output space, the step of performing feature mapping on the neighbor aggregation feature matrix through the learnable weight matrix in the linear transformation layer further includes: calculating the cosine similarity between the central node features in the node feature matrix and the features of each neighbor node in the neighbor aggregation feature matrix. The cosine similarity is input into the Softmax function to calculate the spatial attention coefficient matrix; Perform a Hadamard product operation between the spatial attention coefficient matrix and the neighbor aggregation feature matrix to obtain a weighted neighbor aggregation feature matrix. The weighted neighbor aggregated feature matrix is ​​input into the linear transformation layer for feature mapping.

9. The plague monitoring method based on machine learning according to claim 5, characterized in that, Before outputting the seasonal and periodic dynamic features through the gated linear activation unit, the method further includes: performing a fast Fourier transform on the time-series features at different time scales to obtain a frequency domain feature sequence; Extract low-frequency components with frequencies within a preset low-frequency range from the frequency domain feature sequence, and extract high-frequency components with frequencies within a preset high-frequency range from the frequency domain feature sequence. The periodic trend characteristics are obtained by performing an inverse fast Fourier transform on the low-frequency components, and the residual fluctuation characteristics are obtained by performing an inverse fast Fourier transform on the high-frequency components. The periodic trend feature and the residual fluctuation feature are added together and then input into the gated linear activation unit.

10. The plague monitoring method based on machine learning according to claim 6, characterized in that, After generating the plague occurrence probability distribution map, the method further includes: obtaining the geographic information system vector boundary polygon of the plague source area; The vector boundary polygon of the geographic information system is rasterized into a binary mask matrix with the same resolution as the probability value matrix. In the binary mask matrix, the elements inside the epidemic source area are set to specific values, and the elements outside the epidemic source area are set to zero values. The probability raster matrix corresponding to the plague occurrence probability distribution map is multiplied element-wise with the binary mask matrix to remove the probability values ​​of the area outside the plague source, and the plague occurrence probability distribution map with boundary alignment is output.

Images