Rural governance situational awareness methods, devices, electronic equipment and storage media

By constructing a spatiotemporal knowledge graph and a spatiotemporal dynamic perception model, the problems of data silos and lack of multimodal integration in rural governance have been solved, enabling a comprehensive and dynamic perception of the rural governance situation and improving the scientific nature and efficiency of governance decisions.

CN122311595BActive Publication Date: 2026-11-10BEIJING ACADEMY OF AGRICULTURE & FORESTRY SCIENCES +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610169221.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-02-05
Publication Date
2026-11-10
Estimated Expiration
2046-02-05

AI Technical Summary

Technical Problem

The existing rural governance system suffers from data silos, a lack of multimodal integration, and superficial intelligent analysis, which prevents the formation of a comprehensive, real-time, and interconnected governance situation awareness and scientific decision-making capabilities.

Method used

By constructing a spatiotemporal knowledge graph, acquiring multimodal data and performing task-adaptive feature fusion, and combining it with a spatiotemporal dynamic perception model, a comprehensive and dynamic perception of the rural governance situation can be achieved.

Benefits of technology

It has enabled a comprehensive, dynamic, and in-depth understanding of the rural governance situation, improved the scientific nature and efficiency of governance decisions, and enhanced the level of intelligence in rural governance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122311595B_ABST
    Figure CN122311595B_ABST
Patent Text Reader

Abstract

This invention relates to the field of artificial intelligence technology, providing a method, device, electronic device, and storage medium for rural governance situation awareness. The method includes: using a spatiotemporal knowledge graph with each governance grid unit as a node and the target relationships between governance grid units as edges; acquiring multimodal data from each governance grid unit and extracting initial modal features; performing feature fusion on the initial modal features based on the target governance task to obtain fused governance features; determining spatiotemporal graph data based on the fused governance features and the spatiotemporal knowledge graph; and determining the perception result of the rural governance situation based on the spatiotemporal graph data. This invention constructs a spatiotemporal knowledge graph that integrates spatial and functional relationships, structurally associating rural governance elements scattered across different business systems; and combines fused governance features with the spatiotemporal knowledge graph to determine spatiotemporal graph data, effectively assisting limited grassroots governance personnel in coping with complex and ever-changing governance scenarios, thereby improving the level of intelligence in rural governance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to methods, devices, electronic devices and storage media for rural governance situational awareness. Background Technology

[0002] With the in-depth implementation of my country's rural revitalization strategy and the comprehensive advancement of digital village construction, the modernization of rural governance has become an important issue in improving grassroots governance capabilities.

[0003] However, many prominent problems still exist in current rural governance practices that urgently need to be addressed. Traditional rural governance models rely primarily on manual inspections and experience-based judgment. Existing urban models are ill-suited to the characteristics of rural data: dispersed spatial distribution, low monitoring coverage, and a high proportion of unstructured data. This makes it difficult to achieve comprehensive perception of multi-dimensional governance objects such as the rural environment, safety, and infrastructure. Simultaneously, various monitoring data are scattered across different business systems, forming "data silos" and lacking effective integration and correlation analysis mechanisms. This results in severe fragmentation of governance information, failing to provide comprehensive support for decision-making. At the technical level, although some areas have deployed intelligent devices such as video surveillance and IoT sensors, the massive amounts of heterogeneous data generated by these devices have not been deeply mined and intelligently analyzed. Existing systems generally lack the ability to fuse and process multimodal data, failing to effectively integrate different types of data such as video images, sensor values, and text information, making it difficult to construct a complete governance situational awareness picture. Furthermore, the limited number of grassroots governance personnel often struggles to cope with complex and ever-changing governance scenarios, necessitating the use of intelligent technologies to improve governance efficiency and the scientific nature of decision-making. Summary of the Invention

[0004] This invention provides a method, device, electronic device, and storage medium for rural governance situation awareness, which addresses the shortcomings of existing technologies such as fragmented sensing methods, severe data silos, lack of multimodal fusion, and shallow intelligent analysis, resulting in the inability to form a comprehensive, real-time, and interconnected rural governance situation awareness and scientific decision-making capability.

[0005] This invention provides a method for situational awareness in rural governance, comprising the following steps.

[0006] A spatiotemporal knowledge graph is determined; the spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationship between each governance grid unit as an edge; the target relationship is obtained by fusing the spatial relationship and functional relationship between each governance grid unit. Acquire multimodal data for each governance grid unit and extract the initial modal features corresponding to each governance grid unit; Based on the target governance task, the initial modal features corresponding to each governance grid unit are subjected to task-adaptive feature fusion to obtain the fused governance features of each governance grid unit at different collection time points. Based on the aforementioned fusion governance features and the aforementioned spatiotemporal knowledge graph, spatiotemporal graph data is determined; Based on the spatiotemporal map data, the perception results of the rural governance situation are determined.

[0007] According to a rural governance situation awareness method provided by the present invention, determining the awareness result of the rural governance situation based on the spatiotemporal map data includes: The spatiotemporal map data is input into the spatiotemporal dynamic perception model to obtain the perception result output by the spatiotemporal dynamic perception model; The spatiotemporal dynamic perception model includes a spatial graph convolution module, a temporal convolution module, a spatiotemporal cross-attention module, and an output module; The spatial graph convolution module is used to extract the spatial dependency features of each governance grid unit in the spatiotemporal graph data; The temporal convolution module is used to extract the temporal evolution characteristics of the governance state of each governance grid unit in the spatiotemporal graph data; The spatiotemporal cross-attention module is used to fuse the spatial dependence features and the temporal evolution law features to obtain comprehensive spatiotemporal features; The output module is used to generate the perception result based on the integrated spatiotemporal features.

[0008] According to the present invention, a rural governance situation awareness method is provided, wherein based on a target governance task, task-adaptive feature fusion is performed on the initial modal features corresponding to each governance grid unit to obtain fused governance features of each governance grid unit at different collection time points, including: Determine the task query vector corresponding to the target governance task; Based on the task query vector and the initial modal features corresponding to each of the governance grid units, the attention weight of each of the initial modal features is determined; The initial modal features are weighted and fused based on the attention weights to obtain the fused governance features.

[0009] According to a rural governance situation awareness method provided by the present invention, determining the task query vector corresponding to the target governance task includes: From the task vector library, match the task query vector corresponding to the target governance task; The task vector library stores multiple governance task types and corresponding task query vectors for those governance task types.

[0010] According to a rural governance situation awareness method provided by the present invention, the step of determining the awareness result of the rural governance situation based on the spatiotemporal map data further includes: Based on the perceived results of the rural governance situation, governance events are identified; Obtain the original descriptive information corresponding to the governance event, and determine the comprehensive event feature vector based on the perception results of the rural governance situation and the original descriptive information; The comprehensive event feature vector is input into a preset multi-label classification network to obtain the classification probability of the governance event belonging to at least one preset event category, as output by the multi-label classification network. Based on the indicator values ​​of the governance events across multiple preset dimensions and the fuzzy inference rule base, the quantitative priority of the governance events is determined.

[0011] According to the rural governance situation awareness method provided by the present invention, the spatiotemporal knowledge graph is determined based on a spatial graph and a functional association graph; The weight of the first edge of the spatial relationship edge in the spatial graph is calculated based on the geographical distance and attribute similarity between each governance grid unit; The weight of the second edge of the functional association edge in the functional association graph is the functional association strength; the functional association strength is calculated based on the interaction data between each governance grid unit.

[0012] According to a rural governance situation awareness method provided by the present invention, the step of determining the weights of the edges of the spatiotemporal knowledge graph includes: Determine the adjacency matrix; the adjacency matrix is ​​used to characterize the graph structure of the spatiotemporal knowledge graph, and the values ​​of each element of the adjacency matrix are determined based on the weights of the first edge and the second edge. The weights of the edges in the spatiotemporal knowledge graph are obtained by setting the values ​​of elements in the adjacency matrix that are below a preset weight threshold to zero.

[0013] The present invention also provides a rural governance situational awareness device, comprising the following units: The first determining unit is used to determine the spatiotemporal knowledge graph; the spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationship between each governance grid unit as an edge; the target relationship is obtained by fusing the spatial relationship and functional relationship between each governance grid unit. The acquisition unit is used to acquire the multimodal data of each governance grid unit and extract the initial modal features corresponding to each governance grid unit; The fusion unit is used to perform task-adaptive feature fusion on the initial modal features corresponding to each of the governance grid units based on the target governance task, so as to obtain the fused governance features of each of the governance grid units at different collection time points. The second determining unit is used to determine the spatiotemporal graph data based on each of the fusion governance features and the spatiotemporal knowledge graph; The sensing unit is used to determine the sensing results of the rural governance situation based on the spatiotemporal map data.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the rural governance situational awareness method as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the rural governance situational awareness method as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the rural governance situational awareness method as described above.

[0017] The present invention provides a method, device, electronic device, and storage medium for rural governance situation awareness, which determines a spatiotemporal knowledge graph. The spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationships between each governance grid unit as edges. The target relationships are obtained by fusing the spatial and functional relationships between each governance grid unit. Multimodal data of each governance grid unit is acquired, and initial modal features are extracted. Based on the target governance task, the initial modal features are subjected to task-adaptive feature fusion to obtain fused governance features. Based on the fused governance features and the spatiotemporal knowledge graph, spatiotemporal graph data is determined. Based on the spatiotemporal graph data, the perception result of rural governance situation is determined. This invention constructs a spatiotemporal knowledge graph that integrates spatial and functional relationships, enabling the structured association of rural governance elements scattered across different business systems. This overcomes the problems of fragmented governance information and lack of effective integration and correlation analysis mechanisms caused by data silos in traditional governance. By combining integrated governance characteristics with the spatiotemporal knowledge graph to determine spatiotemporal graph data, it effectively assists limited grassroots governance personnel in coping with complex and ever-changing governance scenarios. This achieves a comprehensive, dynamic, and in-depth perception of the rural governance situation, improves the scientific nature and efficiency of governance decisions, and enhances the level of intelligence in rural governance. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts of the rural governance situation awareness method provided by the present invention.

[0020] Figure 2 This is the second flowchart of the rural governance situation awareness method provided by the present invention.

[0021] Figure 3 This is a schematic diagram of the rural governance situational awareness device provided by the present invention.

[0022] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] This invention provides a method for situational awareness in rural governance, which aims to solve the problems of single monitoring methods, fragmented information, and passive response in traditional rural governance models by constructing a technical framework that integrates multi-source heterogeneous data and spatiotemporal correlation, thereby achieving comprehensive, dynamic, and intelligent awareness of the state of rural governance. Figure 1 This is one of the flowcharts illustrating the rural governance situational awareness method provided by this invention, such as... Figure 1 As shown, the method includes the following: Step 110: Determine the spatiotemporal knowledge graph; the spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationship between each governance grid unit as an edge; the target relationship is obtained by fusing the spatial relationship and functional relationship between each governance grid unit.

[0025] Specifically, firstly, a spatiotemporal knowledge graph is defined. This spatiotemporal knowledge graph is used to reflect a model organized in the form of structured graph data, capable of simultaneously representing the geospatial distribution and functional interactions of rural governance elements.

[0026] Spatiotemporal knowledge graphs abstractly model rural governance areas using nodes and edges. Specifically, each governance grid unit serves as a node, and the target relationships between these grid units form the edges. Here, a governance grid unit refers to a specific area within the rural geographic space. In practice, the granularity of the governance grid units can be flexibly set according to the required precision of governance. For example, a governance grid unit can be an administrative village or township; alternatively, it can be a natural village, a villagers' group, or even a specific functional area such as an industrial park, a water source protection area, or a major traffic intersection. By using these governance grid units as nodes, discrete geographic spaces can be transformed into a networked, computable structure.

[0027] Among them, the target relationship is used to comprehensively reflect the strength of the inherent connections between governance grid units. The target relationship is obtained by integrating the spatial and functional relationships between each governance grid unit. The target is not a single-dimensional relationship, but a composite relationship obtained by integrating spatial and functional relationships.

[0028] Here, spatial relationships are used to reflect the physical spatial correlation between governance grid units; that is, the closer the geographical proximity of the units, the stronger their mutual influence and correlation. For example, spatial relationships can be quantified based on the geographical distance and / or attribute similarity between governance grid units. Even if two governance grid units are not directly adjacent, if their functional attributes are highly similar, such as both being tourist-oriented villages, they may still have a strong spatial relationship. The geographical distance between governance grid units can be Euclidean distance or actual road network distance, and the attribute similarity can be the similarity of characteristics such as population structure, industrial composition, and land use type. This embodiment of the invention does not specifically limit these characteristics.

[0029] Here, functional relationships are used to reflect the social, economic, and other functional interactions between governance grid units that are not geographically proximate. Functional relationships transcend the limitations of physical distance and can capture deeper dynamic connections. For example, functional relationships can be calculated based on interaction data between governance grid units. Interaction data can include population flow data, economic transaction data, event co-occurrence data, social relationship data, etc. Population flow data can be the frequency of personnel exchanges obtained through mobile phone signaling data analysis; economic transaction data can be the flow and volume of goods transactions between villages recorded by e-commerce platforms; event co-occurrence data can be the occurrence of certain types of events simultaneously or sequentially in specific villages in historical governance event records; social relationship data can be the interaction relationships between residents of different villages on social media, etc. This embodiment of the invention does not specifically limit these aspects.

[0030] Step 120: Obtain multimodal data for each governance grid unit and extract the initial modal features corresponding to each governance grid unit.

[0031] Specifically, multimodal data from each governance grid unit can be acquired, and the corresponding initial modal features can be extracted. The multimodal data reflects a dataset that can describe the state of the same governance grid unit from different perspectives and in different forms. This solves the problem of the one-sidedness of information from a single data source.

[0032] In this embodiment, multimodal data is acquired by integrating a data acquisition system that covers the entire scenario. Multimodal data may include video surveillance data. IoT sensor data Drone inspection data Social media data and government data Video surveillance data can be video streams collected by high-definition cameras deployed at village entrances, public places, and around key facilities, used to capture dynamic events and behaviors. IoT sensor data can be time-series numerical data collected by deployed environmental monitoring stations, infrastructure monitoring equipment, etc. Drone inspection data can be remote sensing images collected over a wide area in the air by drones equipped with high-definition or multispectral cameras, used for inspecting farmland, forest land, buildings, etc. Social media data can be text, images, and short video information related to local governance obtained from social media platforms, used to perceive public opinion and public demands. Government data can be data integrated from internal government business systems, such as public complaints, emergency reports, population flow information, etc., and this embodiment of the invention does not specifically limit this.

[0033] Here, multimodal data can be used ={ , , , , }express.

[0034] In one specific embodiment, video surveillance data acquisition utilizes fixed cameras, mobile monitoring equipment, and infrared thermal imaging devices to collect 1080P to 4K high-definition video streams encoded in H.264 / H.265. The system employs a high-frequency acquisition of 25fps for key areas and a standard frequency of 15fps for general areas, covering village entrances, public places, and the vicinity of key facilities. All data includes complete identification information such as timestamps and geographic coordinates. An IoT sensor network enables real-time monitoring of the environment and infrastructure. Environmental monitoring includes air quality, water quality parameters (pH value, dissolved oxygen, heavy metal content, etc.), noise levels, and meteorological data. Infrastructure monitoring covers key indicators of bridge health, road conditions, and water supply systems. The sensor network supports multiple communication protocols, including Long Range Wide Area Network (LoRaWAN), Narrowband Internet of Things (NB-IoT), and Wireless Fidelity (WiFi), with acquisition intervals ranging from 1 minute to 1 hour set according to monitoring needs. The drone inspection system provides wide-area monitoring capabilities from an aerial perspective, acquiring high-resolution remote sensing images containing RGB and multispectral data, and simultaneously recording GPS coordinates, flight altitude, and other trajectory information. Inspections cover farmland, woodlands, water bodies, and buildings, employing a combination of real-time transmission and offline processing to ensure data quality. Social media data collection gathers multimodal information such as text, images, videos, and geographic locations from social media platforms. The system uses keyword filtering to focus on governance-related topics such as environmental pollution and safety accidents, and utilizes natural language processing technology to analyze public sentiment, providing data support for public opinion monitoring.

[0035] Here, the initial modal features are used to reflect numerical vectors extracted from the original multimodal data that can characterize the core information of the original multimodal data.

[0036] The process of extracting initial modal features is usually accomplished using deep learning models designed for different data types.

[0037] To fully explore the deep features of various data types, dedicated feature extraction functions were designed for different modalities. The original multimodal data Mapped to a higher-order feature vector of uniform dimension .

[0038] For example, for video surveillance data, a 3D Convolutional Neural Network (3D-CNN) can be used to extract spatiotemporal features from video clips that can characterize dynamic behaviors such as vehicle movement and crowd gathering. Specifically, for surveillance videos of rural public areas, waterways, and roads, a 3D convolutional neural network can be used to learn spatiotemporal features. Unlike 2D CNNs, which only extract spatial features, the convolutional kernels of a 3D convolutional neural network perform convolution operations in the temporal dimension, effectively capturing the spatiotemporal evolution patterns of dynamic behaviors such as vehicle movement, crowd gathering, and water flow. This process can be described as follows: ; in, Output video surveillance data To characterize the dynamic changes in video content, i.e., spatiotemporal features, This represents a three-dimensional convolutional neural network.

[0039] For time-series numerical data generated by IoT sensors in meteorology, water quality, and soil, a hybrid architecture combining a one-dimensional convolutional neural network (CNN) and a gated recurrent unit (GRU) is adopted. The CNN is responsible for extracting local temporal patterns (such as sudden heavy rainfall), while the GRU captures long-term temporal dependencies (such as persistent drought trends). This process is represented as follows: ; in, For IoT sensor data, Indicates sensor data characteristics, This represents a one-dimensional convolutional neural network. Indicates gated loop unit For social media and government data, a pre-trained model based on the Transformer architecture, BERT (Bidirectional Encoder Representations from Transformers), is used to extract deep semantic features. BERT can understand the language logic in rural contexts and accurately capture key event information such as "garbage accumulation at the east end of the village" and "neighbors arguing over homesteads." The extraction process is as follows: ; in, For input text of social media data or input text of government data, output Contextual semantic embedding vectors for input text of social media data or government data.

[0040] For static images captured by drones during inspections and by villagers, the efficient EfficientNet network is used for visual feature extraction. The EfficientNet network scales the model's depth, width, and resolution using composite coefficients, significantly reducing computational overhead while maintaining feature discriminative power. It is suitable for identifying static scenes such as illegal buildings, farmland diseases, and road damage. The process is as follows: ; in, Represents static image data. Represents image features.

[0041] Furthermore, after obtaining the multimodal data, the multimodal data can be... Convert to high-quality structured initial modal features The initial modal features include spatiotemporal features, sensor data features, contextual semantic embedding vectors of input text from social media or government data, and drone inspection features.

[0042] Specifically, the data quality inspection module first performs integrity checks, consistency verification, and validity assessments on the collected multimodal data to identify missing values, outliers, and duplicate records. Second, a targeted data cleaning process is implemented: blurry and abnormally exposed frames are removed from video surveillance data; outlier detection and smoothing are performed on IoT sensor data; and spam filtering and sensitive word processing are applied to text data from social media and government data. Linear interpolation, cubic spline interpolation, inverse distance weighting, Kriging interpolation, and methods based on the K-Nearest Neighbors (KNN) algorithm and random forest are used to handle different types of missing values. Third, data standardization operations unify the timestamp format (UTC+8), coordinate system (WGS84), and encoding formats, resolutions, sizes, and units for various data types. Finally, feature vectors are extracted from different modalities, including keyframes and optical flow features in videos, edge texture and color features in images, semantic vectors and sentiment tags in text, and trend and periodic features in time-series data, laying the data foundation for subsequent multimodal fusion and intelligent analysis.

[0043] Step 130: Based on the target governance task, perform task-adaptive feature fusion on the initial modal features corresponding to each governance grid unit to obtain the fused governance features of each governance grid unit at different collection time points.

[0044] Specifically, based on the target governance task, task-adaptive feature fusion can be performed on the initial modal features corresponding to each governance grid unit to obtain the fused governance features of each governance grid unit at different collection time points.

[0045] Here, the term "targeted governance task" reflects the specific direction and focus of current rural governance work. For example, flash flood disaster early warning is one targeted governance task, while environmental pollution monitoring is another. Different tasks have different data focuses.

[0046] Task-adaptive feature fusion refers to a fusion process that is not static but dynamically adjusts the contribution of different initial modal features based on the differences in the target governance task. For example, when performing a flash flood disaster early warning task, the system automatically increases the weight of initial modal features from water level sensors, rainfall sensors, and river monitoring videos; while when performing a dispute mediation task, it places greater emphasis on initial textual modal features from government complaint platforms and social media. This adaptive mechanism makes the fused features more targeted and can effectively filter out noise information irrelevant to the current task.

[0047] Step 140: Determine the spatiotemporal graph data based on the fusion governance features and the spatiotemporal knowledge graph.

[0048] Specifically, after obtaining the various integrated governance features, spatiotemporal graph data can be determined based on these features and the spatiotemporal knowledge graph.

[0049] Here, spatiotemporal graph data is used to reflect graph structure data that evolves in both time and space dimensions. Spatiotemporal graph data can be viewed as a time series, with each moment in the time series corresponding to a graph snapshot. In each graph snapshot, the graph's topology is defined by the spatiotemporal knowledge graph, while the attributes of each node on the graph, i.e., the governance grid unit, are the fused governance features calculated at the corresponding time point.

[0050] The process of constructing spatiotemporal graph data involves mapping the generated, time-series-identified integrated governance features one-to-one with time sequence onto the node attributes of the corresponding governance grid units in the spatiotemporal knowledge graph. This forms a dynamically evolving graph structure sequence, thus comprehensively and structurally depicting the dynamic evolution of rural governance in the spatiotemporal dimension.

[0051] Step 150: Based on the spatiotemporal map data, determine the perception results of the rural governance situation.

[0052] Specifically, after obtaining the spatiotemporal map data, the perceived results of the rural governance situation can be determined based on the spatiotemporal map data.

[0053] Here, the perception results are used to reflect the quantitative assessment and qualitative prediction of the current or future state of rural governance. The perception results can be a situation score representing the risk level of each village, with a higher score indicating a greater risk; they can also be an alarm for abnormal events, indicating that an anomaly may have occurred at a certain time and place; or they can be a predicted path for an event in the future, such as the trend of environmental pollution spread and evolution. This embodiment of the invention does not specifically limit these aspects.

[0054] The method provided in this invention determines a spatiotemporal knowledge graph. The spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationships between these units as edges. The target relationships are obtained by fusing spatial and functional relationships between the governance grid units. Multimodal data from each governance grid unit is acquired, and initial modal features are extracted. Based on the target governance task, the initial modal features are subjected to task-adaptive feature fusion to obtain fused governance features. Based on the fused governance features and the spatiotemporal knowledge graph, spatiotemporal graph data is determined. Based on the spatiotemporal graph data, the perceived results of rural governance status are determined. This invention, by constructing a spatiotemporal knowledge graph that integrates spatial and functional relationships, can structurally associate rural governance elements scattered across different business systems, overcoming the problems of fragmented governance information and lack of effective integration and correlation analysis mechanisms caused by data silos in traditional governance. By combining fused governance features with the spatiotemporal knowledge graph to determine spatiotemporal graph data, it effectively assists limited grassroots governance personnel in coping with complex and ever-changing governance scenarios, thereby achieving a comprehensive, dynamic, and in-depth perception of the rural governance status, improving the scientific nature and efficiency of governance decisions, and enhancing the intelligent level of rural governance.

[0055] In related technologies, rural governance events have obvious spatiotemporal correlation characteristics and dynamic evolution patterns, but traditional methods lack in-depth modeling of spatiotemporal dependencies and cannot accurately capture the spread and diffusion trends of events and potential cascading risks.

[0056] Based on the above embodiments, step 150 includes: Step 150-1: Input the spatiotemporal map data into the spatiotemporal dynamic perception model to obtain the perception result output by the spatiotemporal dynamic perception model; The spatiotemporal dynamic perception model includes a spatial graph convolution module, a temporal convolution module, a spatiotemporal cross-attention module, and an output module; The spatial graph convolution module is used to extract the spatial dependency features of each governance grid unit in the spatiotemporal graph data; The temporal convolution module is used to extract the temporal evolution characteristics of the governance state of each governance grid unit in the spatiotemporal graph data; The spatiotemporal cross-attention module is used to fuse the spatial dependence features and the temporal evolution law features to obtain comprehensive spatiotemporal features; The output module is used to generate the perception result based on the integrated spatiotemporal features.

[0057] Specifically, after obtaining the spatiotemporal map data, the spatiotemporal map data can be input into the spatiotemporal dynamic perception model to obtain the perception results output by the spatiotemporal dynamic perception model.

[0058] Among them, the spatiotemporal dynamic perception model is used to reflect a deep learning network model that can simultaneously process and model the spatial dimension dependence and temporal dimension evolution laws in data. The spatiotemporal dynamic perception model aims to automatically learn the inherent evolutionary patterns and diffusion laws of rural governance trends from complex spatiotemporal graph data.

[0059] Here, the spatiotemporal dynamic perception model may include a spatial graph convolution module, a temporal convolution module, a spatiotemporal cross-attention module, and an output module.

[0060] The spatial graph convolution module is used to extract the spatial dependency features of each governance grid unit in the spatiotemporal graph data. These spatial dependency features reflect the mutual influence and state transfer information between the governance grid unit and its neighboring and functionally related units within the same time slice.

[0061] In practical implementation, the spatial graph convolution module can adopt graph convolution operations based on Chebyshev polynomial approximation, for example. Graph convolution operations update the feature representation of a node by aggregating the neighborhood information of the node, thereby effectively capturing local and global spatial structure information.

[0062] The formula for the first-order graph convolution operation, which uses Chebyshev polynomial approximation to reduce computational complexity and enhance scalability, is as follows: ; in, Indicates the first The node feature matrix of the layer; Indicates the first The node feature matrix of the layer; L =I- For the normalized Laplace matrix, W It is an adjacency matrix. D For degree matrix, The k-th term of the Chebyshev polynomial is used for the approximate graph convolution kernel. This is a trainable parameter matrix that controls the weights of the convolutional kernel. Nonlinear activation functions, such as the Rectified Linear Unit (ReLU), introduce nonlinearity into the model.

[0063] Here, the temporal convolution module is used to extract the temporal evolution characteristics of the governance state of each governance grid unit in the spatiotemporal map data. These temporal evolution characteristics reflect the trend, periodicity, or abrupt change in the governance state of a single governance grid unit over time. The role of the temporal convolution module is to capture the dynamic changes of each node along the time axis. For example, the periodic increase in traffic flow in a certain area during holidays, or the sudden abnormal jump in a sensor reading, can be captured by this module. In specific implementations, the temporal convolution module can employ a gated temporal convolutional network (Gated TCN). Compared to traditional recurrent neural networks, temporal convolutional networks have advantages such as strong parallel computing capabilities and the ability to capture longer temporal dependencies. Furthermore, gating mechanisms, such as the sigmoid gate in Gated-TCN, allow the model to selectively forget or retain historical information, making it more sensitive to sudden events and more stable in terms of long-term trends.

[0064] The temporal convolution module preserves temporal dependencies while filtering noise through gating mechanisms, enhancing the model's sensitivity to sudden events and its stability against long-term trends. Its structure combines causal convolution with gating mechanisms, as shown in the following formula: ; in, This indicates that the temporal convolution module is at time step. t The output characteristics, This represents a one-dimensional convolution operation. The node feature sequence within the time window [tk,t] is the fused governance feature sequence. , For trainable convolutional kernel weights, and These are the hyperbolic tangent and the Sigmoid function, respectively. The latter serves as a gating unit to control the flow of information, and ⊙ represents element-wise multiplication.

[0065] Here, the spatiotemporal cross-attention module is used to fuse spatial dependency features and temporal evolution pattern features to obtain comprehensive spatiotemporal features. These comprehensive spatiotemporal features reflect a more discriminative feature representation after sufficient interaction and fusion of spatial and temporal information. The spatiotemporal cross-attention module solves the information loss problem caused by simply adding or concatenating spatial dependency features and temporal evolution pattern features. It can dynamically determine whether spatial influence or temporal trend is more important in a specific context. In its implementation, the spatiotemporal cross-attention module can employ a multi-head attention mechanism, using spatial dependency features as queries and temporal evolution pattern features as keys and values, and vice versa. By calculating attention scores, weights are dynamically assigned to features from different dimensions, thereby generating more information-rich comprehensive spatiotemporal features.

[0066] In one specific embodiment, the spatiotemporal cross-attention module introduces a spatiotemporal cross-attention mechanism to dynamically adjust the feature contributions of the spatial and temporal dimensions. This mechanism enables the spatiotemporal dynamic perception model to adaptively focus on key spatiotemporal dimensions based on the current governance task, thereby improving the discriminative power of feature representation. The calculation formula is as follows: ; in, and These are spatial dependence characteristics and temporal evolution characteristics, respectively. and Here, d is the trainable projection matrix, and d is the feature dimension used to scale the dot product result. This is a spatiotemporal attention weight matrix, used for weighted fusion of spatial dependence features and temporal evolution features. Here, the output module is used to generate perception results based on comprehensive spatiotemporal features. In specific implementations, the output module can consist of one or more fully connected layers that map the comprehensive spatiotemporal features to the final output space. For example, the output module can output a risk score for each governance grid unit at the next time step, or output a classification probability indicating whether an anomaly has occurred in that governance grid unit.

[0067] The method provided in this invention, by introducing a spatiotemporal dynamic perception model with three core modules—a spatial graph convolution module, a temporal convolution module, and a spatiotemporal cross-attention module—enables joint modeling of complex spatiotemporal dependencies within a unified framework. This spatiotemporal dynamic perception model not only captures spatial diffusion effects and temporal evolutionary patterns separately, but also dynamically fuses these two types of information through an attention mechanism. This results in a more accurate and in-depth understanding of the evolutionary mechanisms of rural governance, significantly improving the accuracy and reliability of the perception results.

[0068] Based on the above embodiments, step 130 includes: Step 1301: Determine the task query vector corresponding to the target governance task; Step 1302: Based on the task query vector and the initial modal features corresponding to each of the governance grid units, determine the attention weight of each of the initial modal features; Step 1303: Based on the attention weights, the initial modal features are weighted and fused to obtain the fused governance features.

[0069] Specifically, first, the task query vector corresponding to the target governance task is determined. This task query vector reflects the semantic representation of the target governance task and is a high-dimensional numerical vector. This numerical vector acts as a query instruction in subsequent attention calculations, guiding the model to focus on the modal information most relevant to the target governance task. For example, the query vectors for flood control and drought relief tasks and public safety patrol tasks point in different directions in the vector space, respectively implying a focus on water conservancy and meteorological data, and on pedestrian flow and security video data.

[0070] Then, based on the task query vector and the initial modal features corresponding to each governance grid unit, the attention weight of each initial modal feature is determined. The attention weight reflects the contribution of each modal data type, such as video, sensor data, or text, to the generation of fused governance features under the target governance task. Its value is typically between 0 and 1, and the sum of the weights of all modalities is 1. It should be understood that a higher attention weight indicates that the information of that modality is more critical in the target governance task.

[0071] Among them, attention weight A task query vector Q is introduced to achieve dynamic adaptive weight allocation, and its calculation formula is as follows: ; in, Let Q be the initial modal features obtained by the dedicated feature extractor for the i-th modality, and let φ be the task query vector, which is the key to the task adaptation of this mechanism. The task query vector is a learnable embedding vector, and each objective governance task corresponds to a specific Q. and The weight matrix is ​​a learnable matrix. The hyperbolic tangent activation function introduces a nonlinear transformation into the system, enhancing the model's ability to fit complex modal-task relationships. These are learnable weight vectors. The normalized exponential function normalizes the attention scores calculated for all n modalities, ensuring that the final output attention weights are consistent. satisfy =1 and This forms a valid probability distribution.

[0072] Finally, the initial modal features are weighted and fused based on attention weights to obtain the fused governance features. Specifically, each initial modal feature vector is multiplied by its corresponding attention weight, and then all weighted feature vectors are summed. In this way, initial modal features highly relevant to the target governance task will dominate the final fused governance features due to their higher weights, while initial modal features with lower relevance will be suppressed. For example, in the early warning of flash floods and geological disasters, the sensor features of rain gauges and water level gauges, as well as the features of river monitoring videos, will receive higher attention weights, while other features such as social media sentiment text will have lower weights.

[0073] Based on the above embodiments, step 1301 includes: Step 1301-1: Match the task query vector corresponding to the target governance task from the task vector library; The task vector library stores multiple governance task types and corresponding task query vectors for those governance task types.

[0074] Specifically, task query vectors corresponding to the target governance task can be matched from the task vector library. The task vector library reflects a pre-built knowledge base that stores various governance tasks and their corresponding task query vectors. Essentially, the task vector library is a key-value pair store, where the key is the type name of the governance task, such as the string "environmental pollution monitoring," and the value is a predefined or learned high-dimensional task query vector representing the semantics of that governance task.

[0075] The matching process involves the system receiving a specific target governance task. For example, if a user selects public safety prevention and control on the interface, the system will use the task name as the key to search in the task vector library and retrieve the corresponding task query vector.

[0076] Here, the task vector library stores multiple governance task types and corresponding task query vectors. Unlike general attention mechanisms, in this embodiment of the invention, Q is designed as a predefined encoding vector for specific rural governance tasks, rather than a randomly initialized learnable parameter. The scenario-based design of the task query vector Q is as follows: This invention predefines six core rural governance tasks, each task corresponding to a d=256-dimensional query vector: 1) Early warning mission for flash floods and geological disasters When encoding the corresponding task query vector, the focus will be on dimensions related to sensor modalities (such as rainfall and water level) and video modalities (such as river monitoring). The first 128 dimensions of the vector are given a large initial weight (0.8), corresponding to sensor modal features and video feature channels. 2) Environmental pollution monitoring tasks Its task query vector will focus on sensor modalities (such as air quality and water quality) and drone image modalities, and the attention weight of the corresponding feature channels can be set to 0.75; 3) Public safety prevention and control tasks Its task query vector will focus on video modality and social media text modality, and the relevant channel attention weights for video modality and social media text modality can be set to 0.85; 4) Mediation of disputes Its task query vector will focus on text modalities (such as government complaints and social media) and government data modalities, and the attention weight of text-related channels can be set to 0.9; 5) Infrastructure inspection tasks Its task query vector will focus on UAV image modalities and sensor modalities (such as bridge structural health monitoring), and the attention weight of the image-related channels can be set to 0.8; 6) Integrated Situational Awareness Task This is a general task whose task query vector will give balanced weights to all modalities. For example, the attention weight of each modal channel is set to 0.5 for global monitoring without specific task orientation.

[0077] The aforementioned task query vectors are preset based on prior knowledge during system initialization and fine-tuned using small-scale labeled data. In practical applications, the system automatically selects the corresponding Q-vector to participate in attention calculation based on the current governance task, thereby achieving task-driven adaptive fusion.

[0078] Correspondingly, integrated governance characteristics The result is obtained by weighting the features of each modality with their corresponding attention weights: ; This formula enables the selective fusion of information.

[0079] Examples of rural governance applications are as follows: Scenario: Early warning of flash floods and geological disasters (target governance task R= Among the sensor data, rain gauge and water level gauge readings rose sharply, and their attention weights... It will increase significantly. Video data: River monitoring video shows rapid water flow and turbid water quality, attention weight. Second. Text / image data: If there is no directly related report, its weight... , It will automatically decrease.

[0080] Integrated governance characteristics Mainly composed of highly weighted sensors and video features, the system can quickly and accurately trigger flood control emergency responses.

[0081] The method provided in this invention injects domain knowledge into the model in a structured manner by constructing and using a task vector library, making task switching and model adaptive adjustment efficient and interpretable. This approach avoids the complex process of retraining the model for each task change, improves the system's response speed and practicality, and ensures the stability and accuracy of the feature fusion process for different governance tasks.

[0082] In related technologies, existing governance systems mostly adopt a passive response mode, only initiating emergency response after a problem occurs. They lack forward-looking risk warning capabilities, often missing the best intervention opportunity and causing unnecessary expansion of losses.

[0083] Based on the above embodiments, step 150 further includes: Step 151: Based on the perception results of the rural governance situation, determine the governance events; Step 152: Obtain the original description information corresponding to the governance event; and determine the comprehensive event feature vector based on the perception results of the rural governance situation and the original description information. Step 153: Input the comprehensive event feature vector into a preset multi-label classification network to obtain the classification probability of the governance event belonging to at least one preset event category output by the multi-label classification network. Step 154: Determine the quantitative priority of the governance event based on the indicator values ​​of the governance event on multiple preset dimensions and the fuzzy inference rule base.

[0084] Specifically, firstly, governance events can be identified based on the perception results of rural governance situation. These governance events reflect discrete anomalies identified from continuous situational awareness results that require human intervention or attention. For example, when the risk score of a governance grid unit continuously exceeds a preset threshold, or when the model identifies a specific anomaly pattern, the system can determine that a governance event has occurred and record its basic information such as time and location.

[0085] Then, the original descriptive information corresponding to the governance event can be obtained. Based on the perceived results of the rural governance situation and the original descriptive information, a comprehensive event feature vector can be determined. Among them, the original descriptive information is used to reflect the original data related to the governance event that is not generated by the model, such as on-site text descriptions and pictures reported by the public through their mobile phones, or raw data fragments from relevant sensors or cameras.

[0086] Here, the comprehensive event feature vector is used to reflect a unified vector representation that comprehensively describes governance events and integrates multi-source information. The process of determining the comprehensive event feature vector involves fusing the structured perception results generated by the model, such as situation scores, spatiotemporal features, and event propagation probabilities, with unstructured raw descriptive information.

[0087] The perception results include a situational score, which can be achieved using a sliding time window mechanism. Spatiotemporal graph data is input into a trained Spatiotemporal-Temporal Graph Convolutional Network (ST-GCN), and the network outputs a situational score for each node at a future time step. This perception mechanism supports multiple time scales from hours to days, adapting to the timeliness requirements of different governance scenarios. The scoring function is as follows: ; in, For spatiotemporal plot data within a time window, Represents a spatiotemporal graph convolutional network. , For the output layer weights and biases, The situation score for the next time step can be used for anomaly detection or risk level classification.

[0088] Furthermore, based on the spatiotemporal graph convolutional features, a system is constructed to build an event propagation model, simulating the diffusion process of governance events (such as environmental pollution and safety accidents) in spatial and functional networks. Random walk and dynamic system theory are employed to determine the event propagation probability. Defined as: ; in, Indicates an event node, The event feature vectors are extracted by ST-GCN. M is a trainable transformation matrix that models the propagation strength between events. For nodes The model is a set of neighbors. It can quantify the cascading risks of events and provide governance departments with a visualized evolution path and intervention suggestions, enabling a shift from passive response to proactive prevention and control. Represents event nodes The feature vector of any of the neighboring nodes.

[0089] This invention aims to construct a complete intelligent response closed loop, from dynamic perception to precise policy implementation. Based on the risk assessment, event location, and evolution prediction output by the aforementioned dynamic perception module, this method proposes a hierarchical, human-machine collaborative intelligent event handling and decision optimization mechanism. This mechanism, through intelligent event classification, automatic generation and deduction of response plans, dynamic resource scheduling, and continuous evaluation and feedback of response effects, achieves rapid response, scientific decision-making, and efficient handling of rural governance events, significantly improving the agility and accuracy of the governance system.

[0090] This step first performs refined intelligent classification and urgency assessment of the events identified by the dynamic perception module, laying the foundation for subsequent differentiated handling. The system constructs an event classification and priority determination model based on the fusion of multi-label classification and rule engine.

[0091] Furthermore, the fusion governance features output by the dynamic perception module can be... Situation score and the probability of event propagation This information, along with structured information such as the original event description text, location, and involved parties, is encoded into a comprehensive event feature vector. .

[0092] Then, the comprehensive event feature vector can be input into a preset multi-label classification network to obtain the classification probability of the governance event belonging to at least one preset event category.

[0093] The multi-label classification network can be a classification network containing fully connected layers and an attention mechanism to classify governance events using multiple labels. The output layer uses the sigmoid function to independently predict the probability of an event belonging to each preset category, such as natural disasters, safety accidents, public health incidents, conflicts, and infrastructure failures. The formula is expressed as: ; in, Represents the classification probability. 1 represents the Sigmoid function. and represents the classification layer parameters, and MLP stands for Multilayer Perceptron.

[0094] This step aims to automatically classify events. A multi-label classification network is a deep learning model that can simultaneously assign multiple category labels to an event; for example, an event can be classified as both environmental pollution and a safety incident. The network's output is the probability value of the event belonging to each preset category, such as natural disasters, public health incidents, or conflicts.

[0095] Furthermore, the quantitative priority of governance events can be determined based on indicator values ​​across multiple preset dimensions and a fuzzy inference rule base. This step aims to quantitatively assess the urgency of events, providing a basis for resource allocation. The determination method here employs fuzzy logic inference. First, indicator values ​​for the event across multiple preset dimensions are extracted from the comprehensive event feature vector, such as urgency (e.g., spread rate), impact range (e.g., affected population), social attention (e.g., public opinion intensity), and handling complexity. Then, these indicator values ​​are input into a preset fuzzy inference rule base, which may contain expert rules such as "IF high urgency AND large impact range THEN highest priority". Finally, through fuzzy inference and defuzzification processes, a quantitative priority score within the range [0, 1] is obtained, such as... ∈[0,1] and the corresponding levels, such as Level I / Extremely Serious, Level II / Serious, Level III / Significant, Level IV / General.

[0096] After classifying and prioritizing events, a preferred embodiment of the present invention further provides an intelligent mechanism for generating and optimizing disposal solutions, specifically including disposal solution element retrieval and assembly, solution deduction and optimization based on digital twins, and disposal effect evaluation and model iteration.

[0097] Here, the step of retrieving and assembling the elements of the response plan aims to quickly generate a preliminary response plan with a solid foundation. The system can pre-build a structured knowledge base of response plan elements. This knowledge base is a collection that reflects and stores various knowledge elements related to event response. It can include publicly available documents, standardized response procedures extracted from historical successful cases, expert experience, and lists of resources required for different response actions, such as manpower, materials, and equipment. In a preferred implementation, the knowledge base can be stored using a graph database, where nodes can represent specific response actions (such as "evacuating the public"), resources (such as "inflatable boats"), etc., while edges can represent the logical order relationship or historical co-occurrence frequency between them.

[0098] When a new, categorized, and graded governance event occurs, the system automatically retrieves relevant response elements from the response plan element knowledge base based on its category, characteristics, and priority, using semantic matching and graph query technologies. For example, for a high-priority flash flood geological disaster event, the system will retrieve clauses on personnel evacuation from the "Flood Control Emergency Plan," command procedures from successful historical cases of similar disasters, and lists of required rescue teams and supplies.

[0099] Subsequently, the system uses a predefined combinatorial logic template to dynamically assemble these retrieved, fragmented disposal elements into a structurally complete preliminary disposal plan draft, denoted as […]. .

[0100] Here, the steps of scheme simulation and optimization based on digital twins aim to simulate, verify, and iteratively optimize the draft preliminary response plan to ensure its feasibility and effectiveness in actual implementation.

[0101] The system can generate the previous step The current real-time situational data (including spatiotemporal knowledge graphs) is imported into a rural digital twin platform. The rural digital twin platform can be a high-fidelity virtual simulation environment that maps to the physical countryside in real time. It can accurately simulate the physical world of the real countryside (such as topography, building distribution, and road traffic flow), social systems (such as population spatial distribution and population behavior patterns), and the status and distribution of governance resources.

[0102] In a digital twin environment, the system will automatically execute... In this process, agent-based model technology can be used to simulate the behavior and interaction of different stakeholders (such as commanders, rescue teams, affected villagers, and social vehicles), thereby dynamically predicting the development of the event under the intervention of the contingency plan and the actual implementation effect of the plan.

[0103] The system evaluates the merits of contingency plans by comparing and analyzing the key performance indicators (KPIs) of different plans (or different parameters of the same plan)—such as estimated completion time, estimated total resource consumption, and the degree of minimization of expected negative impacts. More preferably, reinforcement learning algorithms (such as Q-Learning or policy gradient methods) can be used, with the comprehensive score of KPIs as a reward signal, to automatically iteratively optimize the adjustable parameters in the plan (such as resource deployment locations, evacuation route planning, and response timing). Finally, the system outputs one or more optimal recommended response plans after thorough simulation and optimization, denoted as... It is accompanied by an evaluation report on its expected effects.

[0104] To enable continuous learning and self-evolution of governance capabilities, embodiments of the present invention also include a closed-loop feedback mechanism that evaluates the entire disposal process and uses the results for model iteration.

[0105] After the incident is handled, the system can automatically evaluate the effectiveness based on a multi-dimensional effectiveness evaluation index system. This index system can cover: timeliness (such as the interval from warning to response, and the total time for handling), effectiveness (such as the degree of incident control, the reduction rate of personnel and property losses, and the results of public satisfaction surveys), economic efficiency (such as the actual resource costs invested), and social impact (such as the stability of public opinion during and after the handling, and the restoration of social order). The system automatically collects relevant data and calculates scores for each index.

[0106] Subsequently, reinforcement learning of the model based on evaluation feedback is conducted. Each complete event handling process (from perception, analysis, decision-making to the end of handling) can be abstracted into a "state-action-reward" sequence. Here, "state" refers to the initial situational characteristics when the event occurs; "action" refers to the handling decision ultimately recommended by the system or executed after manual adjustment; and "reward" is the total score calculated based on the above-mentioned effect evaluation index system.

[0107] These sequences are stored in an experience replay pool. The system periodically, for example offline, utilizes this accumulated historical case data to train and fine-tune the AI ​​(Artificial Intelligence) model involved in the method of this invention using deep reinforcement learning algorithms, such as the Deep Deterministic Policy Gradient (DDPG) algorithm. Specifically, this may include: optimizing the risk assessment parameters in the spatiotemporal dynamic perception model to improve perception accuracy; and optimizing the response plan generation model (i.e., the policy network responsible for retrieving and assembling contingency plans) to enable it to generate more effective response decisions.

[0108] Furthermore, to support the efficient and reliable implementation of the aforementioned complete method from perception to processing, this invention also provides a preferred system architecture, namely a hierarchical "cloud-edge-device" collaborative system architecture: The edge-side perception layer, serving as the data source, consists of various smart terminals distributed throughout the countryside. These include high-definition video surveillance cameras, IoT sensor nodes (environment, infrastructure, etc.), drones, and mobile terminals used for information reporting, such as applications on the mobile phones of grid workers or villagers. The edge-side perception layer is responsible for collecting and initially encoding the raw data. It preferably uses lightweight IoT protocols such as Message Queuing Telemetry Transport (MQTT) to report data and supports features like resuming data transmission after network outages and data compression to adapt to the potentially unstable network environment in rural areas.

[0109] Edge computing layer: Deployed in physical locations close to the data source, such as townships or key villages, it is responsible for low-latency, lightweight real-time processing of edge data. Its tasks include: real-time target detection and behavior recognition of video streams, initial anomaly detection of sensor data, and local aggregation and compression of multi-source data. Lightweight models can be deployed on edge nodes, such as YOLOv5s for video analysis and Long Short-Term Memory Auto-Encoder (LSTM-AE) for sensor anomaly detection, to achieve simple early warning and response with high timeliness requirements. An incremental synchronization mechanism is preferred between edge nodes and the cloud, typically uploading only anomaly data and gradient information required for model updates, significantly saving uplink bandwidth.

[0110] The cloud-based intelligent layer, acting as the "brain" of the entire system, is typically deployed in county-level, city-level, or higher-level data centers. This layer carries the core AI algorithm models described in this invention, including multimodal fusion models, spatiotemporal graph convolutional dynamic perception models, event classification and handling scheme generation models, and resource scheduling optimization models. It is responsible for the centralized storage and deep analysis of massive historical and real-time data, complex model training, processing of complex events across regions, and global resource coordination and command decisions. The cloud platform can be deployed using containerization technologies such as Kubernetes to support elastic scaling, canary updates, and A / B testing of the model; model training can employ distributed training frameworks such as Horovod to support parallel processing of multiple graphics processing units (GPUs) and accelerate model iteration.

[0111] Application Interaction Layer: Serving as the interface for human-computer interaction, this layer provides personalized service interfaces for users in different roles, such as commanders, frontline grid workers, and ordinary villagers, through various forms including web portals, mobile applications, and command center large screens. Its functions include: panoramic situation visualization, precise push notifications of intelligent early warning information, issuance and execution feedback of response tasks, and convenient channels for reporting public sentiment.

[0112] The method provided in this invention elevates rural governance from simple situational awareness to intelligent decision-making by adding intelligent event classification and priority determination steps. It can automatically and quickly classify and quantify alarm events, helping governance personnel rapidly identify key and urgent events from massive amounts of information. This enables the scientific and efficient allocation of governance resources, significantly improving the efficiency and accuracy of emergency response.

[0113] Based on the above embodiments, the spatiotemporal knowledge graph is determined based on a spatial graph and a functional association graph; The weight of the first edge of the spatial relationship edge in the spatial graph is calculated based on the geographical distance and attribute similarity between each governance grid unit; The weight of the second edge of the functional association edge in the functional association graph is the functional association strength; the functional association strength is calculated based on the interaction data between each of the governance grid units. The steps for determining the edge weights of the spatiotemporal knowledge graph include: Step 210, determine the adjacency matrix; the adjacency matrix is ​​used to characterize the graph structure of the spatiotemporal knowledge graph, and the values ​​of each element of the adjacency matrix are determined based on the weight of the first edge and the weight of the second edge; Step 220: Set the values ​​of elements in the adjacency matrix that are below a preset weight threshold to zero to obtain the weights of the edges in the spatiotemporal knowledge graph.

[0114] Specifically, the spatiotemporal knowledge graph is determined based on a spatial graph and a functional relationship graph. This two-layer graph structure design can more comprehensively depict the complex relationships between various governance grid units.

[0115] The spatial graph is primarily used to characterize the geographical proximity and attribute similarity among governance grid units. The weight of the first edge of the spatial relationship edge in the spatial graph is calculated based on the geographical distance and attribute similarity among the governance grid units. The spatial graph uses specific functional areas such as administrative villages, natural villages, villager groups, or industrial parks as governance grid units, and abstracts each governance grid unit as a graph node. A rural governance area is defined as N governance grid unit nodes, forming a node set. ={ , ,..., The weight calculation for edges connecting nodes comprehensively considers both the geographical distance decay law and the similarity of node attributes. The formula for the weight of the first edge is as follows: ; in, Indicates the weight of the first edge. This represents the Euclidean distance or actual road network distance between each governance grid unit, while As a distance decay parameter, its value directly regulates the rate of decay of geographical influence. The larger the value, the more gradual the decay effect. (Exponential term) As a distance decay function, following the first law of geography, the qualitative principle of "the closer the distance, the greater the influence" is transformed into a quantitative value between 0 and 1. Meanwhile, attribute similarity... By calculating the cosine similarity or Pearson correlation coefficient between node feature vectors (such as population structure, economic indicators, and land use composition), the functional relationships between nodes are effectively captured. This ensures that even if two governance grid units are physically far apart, a strong connection can still be established in the spatial graph as long as their functional attributes are highly compatible. (Node feature vectors) ∈ It includes M=12 dimensions: permanent resident population, floating population, proportion of elderly population, agricultural output value, industrial output value, service industry output value, cultivated land area, construction land area, forest land area, water area, road density, and public facility density.

[0116] Here, the functional association graph is used to characterize functional connections that transcend geographical limitations. The weight of the second edge of each functional association is defined as the functional association strength. The functional association strength is calculated based on the interaction data between the governance grid units.

[0117] Multiple functional dimensions can be defined, such as population flow, economic transactions, and event co-occurrence. Based on the interaction data in these dimensions, such as mobile signaling data, e-commerce transaction records, and historical event reports, the functional correlation strength between two governance grid units can be calculated.

[0118] Here, the nodes in the functional association graph are still governance grid cells, but the weights of the second side are calculated based on the functional interaction strength, as shown in the following formula: ; in, Indicates the weight of the second side. and All of these represent nodes in the functional association diagram. Represents a node In terms of functionality The intensity of interaction on, Represents a node In terms of functionality The intensity of interaction, such as population inflow, frequency of events, or volume of economic transactions. This invention defines... =5 functional dimensions: population mobility intensity (based on mobile phone signaling data), economic transaction frequency (based on e-commerce and financial data), event co-occurrence frequency (based on historical governance events), social connection strength (based on social media interaction), and administrative collaboration frequency (based on government system records). Sub-section The interaction co-occurrence strength of two nodes across all functional dimensions is calculated, similar to the vector dot product. The denominator is a normalization factor, using cosine similarity to ensure that the weight values ​​are in the range of [-1, 1]. In practical applications, absolute values ​​are often used or mapped to the interval [0, 1].

[0119] Here, the steps for determining the edge weights of the spatiotemporal knowledge graph are as follows: First, determine the adjacency matrix. The adjacency matrix is ​​used to represent the graph structure of the spatiotemporal knowledge graph. The values ​​of each element in the adjacency matrix are determined based on the weights of the first and second edges, as shown in the following formula: ; in, The weights of the edges in the spatiotemporal knowledge graph are represented. ∈[0,1] is the balance factor, with a default value of 0.6, and the adjacency matrix A∈ As input to a spatiotemporal graph convolutional network, Indicates the weight of the first edge. This indicates the weight of the second side.

[0120] Then, elements in the adjacency matrix with weights below a preset weight threshold are set to zero to obtain the edge weights of the spatiotemporal knowledge graph. For example, for node pairs with edge weights below the preset weight threshold τ=0.1, their connection weights are reset to 0 to construct a sparse graph structure and improve computational efficiency. This step aims to sparsify the graph by removing edges with weak connections. This effectively filters noise and unimportant connections, reduces the complexity of subsequent graph computations, and makes the structure of the spatiotemporal knowledge graph clearer, highlighting key relationships more effectively.

[0121] The method provided in this invention constructs a two-layer graph structure consisting of a spatial graph and a functional relationship graph, and performs weighted fusion and sparsification processing on it. This enables the creation of a high-quality spatiotemporal knowledge graph that reflects both geographical proximity and reveals deep-seated functional connections. Compared to a single-dimensional relationship graph, this graph can more realistically and comprehensively depict the complex network structure of the rural socio-economic system, improving the accuracy and reliability of rural governance situational awareness.

[0122] Based on any of the above embodiments Figure 2 This is the second flowchart of the rural governance situation awareness method provided by the present invention, as shown below. Figure 2As shown, this method is implemented based on a cloud-edge-device collaborative architecture. First, at the edge perception layer responsible for raw data acquisition, the system performs a multi-source heterogeneous data acquisition step, gathering data from different channels, specifically including video surveillance data, IoT sensor data, drone inspection data, social media data, and government data. Next, this raw data is transmitted to the edge computing layer responsible for low-latency real-time processing for data preprocessing, such as data cleaning, format standardization, and initial anomaly detection. The preprocessed data is then sent to the cloud intelligence layer, which acts as the system's "brain." The cloud intelligence layer first performs a multimodal fusion step based on an attention mechanism. In this step, the system extracts initial video features, sensor features, text features, and image features from different modalities of data and dynamically fuses these features according to the governance task. Then, the cloud intelligence layer continues with a dynamic perception step based on spatiotemporal graph convolution, using an advanced graph neural network model to perform deep analysis of the fused spatiotemporal data to understand and predict the evolution of the situation. Based on dynamic perception, the cloud intelligence layer further performs intelligent event handling and decision optimization steps, automatically generating and deducing optimized response plans. Finally, all perception results, early warning information, and decision-making suggestions are presented to the user through the application interaction layer in the form of a visual interface, and the user's instructions are received, thus forming a complete closed loop.

[0123] The rural governance situational awareness device provided by the present invention is described below. The rural governance situational awareness device described below can be referred to in correspondence with the rural governance situational awareness method described above.

[0124] Based on any of the above embodiments, the present invention provides a rural governance situational awareness device. Figure 3 This is a schematic diagram of the structure of the rural governance situational awareness device provided by the present invention, as shown below. Figure 3 As shown, the device includes: The first determining unit 310 is used to determine a spatiotemporal knowledge graph; the spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationship between each governance grid unit as an edge; the target relationship is obtained by fusing the spatial relationship and functional relationship between each governance grid unit. The acquisition unit 320 is used to acquire the multimodal data of each governance grid unit and extract the initial modal features corresponding to each governance grid unit; The fusion unit 330 is used to perform task-adaptive feature fusion on the initial modal features corresponding to each of the governance grid units based on the target governance task, so as to obtain the fused governance features of each of the governance grid units at different collection time points. The second determining unit 340 is used to determine the spatiotemporal graph data based on each of the fusion governance features and the spatiotemporal knowledge graph; The sensing unit 350 is used to determine the sensing results of the rural governance situation based on the spatiotemporal map data.

[0125] The apparatus provided in this invention determines a spatiotemporal knowledge graph. The spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationships between these units as edges. The target relationships are obtained by fusing spatial and functional relationships between the governance grid units. Multimodal data from each governance grid unit is acquired, and initial modal features are extracted. Based on the target governance task, the initial modal features are subjected to task-adaptive feature fusion to obtain fused governance features. Based on the fused governance features and the spatiotemporal knowledge graph, spatiotemporal graph data is determined. Based on the spatiotemporal graph data, the perceived results of rural governance status are determined. This invention, by constructing a spatiotemporal knowledge graph that integrates spatial and functional relationships, can structurally associate rural governance elements scattered across different business systems, overcoming the problems of fragmented governance information and lack of effective integration and correlation analysis mechanisms caused by data silos in traditional governance. By combining fused governance features with the spatiotemporal knowledge graph to determine spatiotemporal graph data, it effectively assists limited grassroots governance personnel in coping with complex and ever-changing governance scenarios, thereby achieving comprehensive, dynamic, and in-depth perception of the rural governance status, improving the scientific nature and efficiency of governance decisions, and enhancing the intelligent level of rural governance.

[0126] Based on any of the above embodiments, the sensing unit 350 is specifically used for: The spatiotemporal map data is input into the spatiotemporal dynamic perception model to obtain the perception result output by the spatiotemporal dynamic perception model; The spatiotemporal dynamic perception model includes a spatial graph convolution module, a temporal convolution module, a spatiotemporal cross-attention module, and an output module; The spatial graph convolution module is used to extract the spatial dependency features of each governance grid unit in the spatiotemporal graph data; The temporal convolution module is used to extract the temporal evolution characteristics of the governance state of each governance grid unit in the spatiotemporal graph data; The spatiotemporal cross-attention module is used to fuse the spatial dependence features and the temporal evolution law features to obtain comprehensive spatiotemporal features; The output module is used to generate the perception result based on the integrated spatiotemporal features.

[0127] Based on any of the above embodiments, the fusion unit 330 specifically includes: The query vector unit is used to determine the task query vector corresponding to the target governance task; An attention weight unit is determined, which is used to determine the attention weight of each initial modality feature based on the task query vector and the initial modality features corresponding to each governance grid unit; The weighted fusion unit is used to perform weighted fusion of each of the initial modal features based on the attention weights to obtain the fused governance features.

[0128] Based on any of the above embodiments, the query vector determination unit is specifically used for: From the task vector library, match the task query vector corresponding to the target governance task; The task vector library stores multiple governance task types and corresponding task query vectors for those governance task types.

[0129] Based on any of the above embodiments, a classification prediction unit is further included, wherein the classification prediction unit is specifically used for: Based on the perceived results of the rural governance situation, governance events are identified; Obtain the original descriptive information corresponding to the governance event, and determine the comprehensive event feature vector based on the perception results of the rural governance situation and the original descriptive information; The comprehensive event feature vector is input into a preset multi-label classification network to obtain the classification probability of the governance event belonging to at least one preset event category, as output by the multi-label classification network. Based on the indicator values ​​of the governance events across multiple preset dimensions and the fuzzy inference rule base, the quantitative priority of the governance events is determined.

[0130] Based on any of the above embodiments, the spatiotemporal knowledge graph is determined based on a spatial graph and a functional association graph; The weight of the first edge of the spatial relationship edge in the spatial graph is calculated based on the geographical distance and attribute similarity between each governance grid unit; The weight of the second edge of the functional association edge in the functional association graph is the functional association strength; the functional association strength is calculated based on the interaction data between each governance grid unit.

[0131] Based on any of the above embodiments, it further includes an edge weight determination unit, wherein the edge weight determination unit is specifically used for: Determine the adjacency matrix; the adjacency matrix is ​​used to characterize the graph structure of the spatiotemporal knowledge graph, and the values ​​of each element of the adjacency matrix are determined based on the weights of the first edge and the second edge. The weights of the edges in the spatiotemporal knowledge graph are obtained by setting the values ​​of elements in the adjacency matrix that are below a preset weight threshold to zero.

[0132] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a rural governance situation awareness method, which includes: determining a spatiotemporal knowledge graph; the spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationship between each governance grid unit as an edge; the target relationship is obtained by fusing the spatial and functional relationships between each governance grid unit; acquiring multimodal data of each governance grid unit and extracting the initial modal features corresponding to each governance grid unit; performing task-adaptive feature fusion on the initial modal features corresponding to each governance grid unit based on the target governance task to obtain the fused governance features of each governance grid unit at different collection time points; determining spatiotemporal graph data based on each fused governance feature and the spatiotemporal knowledge graph; and determining the perception result of the rural governance situation based on the spatiotemporal graph data.

[0133] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0134] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the rural governance situation awareness method provided by the above methods. The method includes: determining a spatiotemporal knowledge graph; the spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationship between each governance grid unit as an edge; the target relationship is obtained by fusing the spatial and functional relationships between each governance grid unit; acquiring multimodal data of each governance grid unit and extracting the initial modal features corresponding to each governance grid unit; performing task-adaptive feature fusion on the initial modal features corresponding to each governance grid unit based on the target governance task to obtain the fused governance features of each governance grid unit at different collection time points; determining spatiotemporal graph data based on each fused governance feature and the spatiotemporal knowledge graph; and determining the perception result of the rural governance situation based on the spatiotemporal graph data.

[0135] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the rural governance situation awareness method provided by the above methods. The method includes: determining a spatiotemporal knowledge graph; the spatiotemporal knowledge graph having each governance grid unit as a node and the target relationship between each governance grid unit as an edge; the target relationship being obtained by fusing the spatial and functional relationships between each governance grid unit; acquiring multimodal data of each governance grid unit and extracting the initial modal features corresponding to each governance grid unit; performing task-adaptive feature fusion on the initial modal features corresponding to each governance grid unit based on the target governance task to obtain the fused governance features of each governance grid unit at different collection time points; determining spatiotemporal graph data based on the fused governance features and the spatiotemporal knowledge graph; and determining the perception result of the rural governance situation based on the spatiotemporal graph data.

[0136] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0137] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for situational awareness in rural governance, characterized in that, include: Define the spatiotemporal knowledge graph; The spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationships between each governance grid unit as edges; The target relationship is obtained by integrating the spatial and functional relationships between the governance grid units. Acquire multimodal data for each governance grid unit and extract the initial modal features corresponding to each governance grid unit; Based on the target governance task, the initial modal features corresponding to each governance grid unit are subjected to task-adaptive feature fusion to obtain the fused governance features of each governance grid unit at different collection time points. Based on the aforementioned fusion governance features and the aforementioned spatiotemporal knowledge graph, spatiotemporal graph data is determined; Based on the spatiotemporal map data, the perception results of the rural governance situation are determined; The perception results of determining the rural governance situation based on the spatiotemporal map data include: The spatiotemporal map data is input into the spatiotemporal dynamic perception model to obtain the perception result output by the spatiotemporal dynamic perception model; The spatiotemporal dynamic perception model includes a spatial graph convolution module, a temporal convolution module, a spatiotemporal cross-attention module, and an output module; The spatial graph convolution module is used to extract the spatial dependency features of each governance grid unit in the spatiotemporal graph data; The temporal convolution module is used to extract the temporal evolution characteristics of the governance state of each governance grid unit in the spatiotemporal graph data; The spatiotemporal cross-attention module is used to fuse the spatial dependence features and the temporal evolution law features to obtain comprehensive spatiotemporal features; The output module is used to generate the perception result based on the integrated spatiotemporal features; The spatiotemporal knowledge graph is determined based on a spatial graph and a functional association graph; The weight of the first edge of the spatial relationship edge in the spatial graph is calculated based on the geographical distance and attribute similarity between each governance grid unit; The weight of the second edge of the functional association edge in the functional association graph is the functional association strength; the functional association strength is calculated based on the interaction data between each governance grid unit.

2. The rural governance situational awareness method according to claim 1, characterized in that, The step involves performing task-adaptive feature fusion on the initial modal features corresponding to each governance grid unit based on the target governance task, to obtain the fused governance features of each governance grid unit at different acquisition time points, including: Determine the task query vector corresponding to the target governance task; Based on the task query vector and the initial modal features corresponding to each of the governance grid units, the attention weight of each of the initial modal features is determined; The initial modal features are weighted and fused based on the attention weights to obtain the fused governance features.

3. The rural governance situational awareness method according to claim 2, characterized in that, The determination of the task query vector corresponding to the target governance task includes: From the task vector library, match the task query vector corresponding to the target governance task; The task vector library stores multiple governance task types and corresponding task query vectors for those governance task types.

4. The rural governance situational awareness method according to any one of claims 1 to 3, characterized in that, The process of determining the perceived situation of rural governance based on the spatiotemporal map data further includes: Based on the perceived results of the rural governance situation, governance events are identified; Obtain the original descriptive information corresponding to the governance event, and determine the comprehensive event feature vector based on the perception results of the rural governance situation and the original descriptive information; The comprehensive event feature vector is input into a preset multi-label classification network to obtain the classification probability of the governance event belonging to at least one preset event category, as output by the multi-label classification network. Based on the indicator values ​​of the governance events across multiple preset dimensions and the fuzzy inference rule base, the quantitative priority of the governance events is determined.

5. The rural governance situational awareness method according to claim 1, characterized in that, The steps for determining the edge weights of the spatiotemporal knowledge graph include: Determine the adjacency matrix; the adjacency matrix is ​​used to characterize the graph structure of the spatiotemporal knowledge graph, and the values ​​of each element of the adjacency matrix are determined based on the weights of the first edge and the second edge. The weights of the edges in the spatiotemporal knowledge graph are obtained by setting the values ​​of elements in the adjacency matrix that are below a preset weight threshold to zero.

6. A rural governance situational awareness device, characterized in that, include: The first determining unit is used to determine the spatiotemporal knowledge graph; the spatiotemporal knowledge graph uses each governance grid unit as a node and the target relationship between each governance grid unit as an edge; The target relationship is obtained by integrating the spatial and functional relationships between the governance grid units. The acquisition unit is used to acquire the multimodal data of each governance grid unit and extract the initial modal features corresponding to each governance grid unit; The fusion unit is used to perform task-adaptive feature fusion on the initial modal features corresponding to each of the governance grid units based on the target governance task, so as to obtain the fused governance features of each of the governance grid units at different collection time points. The second determining unit is used to determine the spatiotemporal graph data based on each of the fusion governance features and the spatiotemporal knowledge graph; The sensing unit is used to determine the sensing results of the rural governance situation based on the spatiotemporal map data; The sensing unit is specifically used for: The spatiotemporal map data is input into the spatiotemporal dynamic perception model to obtain the perception result output by the spatiotemporal dynamic perception model; The spatiotemporal dynamic perception model includes a spatial graph convolution module, a temporal convolution module, a spatiotemporal cross-attention module, and an output module; The spatial graph convolution module is used to extract the spatial dependency features of each governance grid unit in the spatiotemporal graph data; The temporal convolution module is used to extract the temporal evolution characteristics of the governance state of each governance grid unit in the spatiotemporal graph data; The spatiotemporal cross-attention module is used to fuse the spatial dependence features and the temporal evolution law features to obtain comprehensive spatiotemporal features; The output module is used to generate the perception result based on the integrated spatiotemporal features; The spatiotemporal knowledge graph is determined based on a spatial graph and a functional association graph; The weight of the first edge of the spatial relationship edge in the spatial graph is calculated based on the geographical distance and attribute similarity between each governance grid unit; The weight of the second edge of the functional association edge in the functional association graph is the functional association strength; the functional association strength is calculated based on the interaction data between each governance grid unit.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the rural governance situational awareness method as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the rural governance situational awareness method as described in any one of claims 1 to 5.

Citation Information

Patent Citations

  • Spatial relationship knowledge graph data model representation method and system

    CN111125294A

  • Grassroots social governance intelligent decision support system and method based on multi-modal fusion

    CN120746326A