A multi-level community data linkage analysis processing method and system

Abstract

The application provides a multi-level community data linkage analysis processing method and system, and belongs to the technical field of smart communities. The method comprises the following steps: modeling the space-time causal relationship of five-level space units from community individuals to urban areas, generating five-level space-time causal graph data, and constructing a multi-level data correlation framework based on the graph data; implementing federated knowledge distillation according to the multi-level data correlation framework, generating a global federated model through multi-agency local model parameter encryption aggregation; and synchronously establishing a privacy protection mechanism, performing perturbation processing on original data by using differential privacy technology, and generating a federated training data set meeting safety requirements. By constructing five-level space-time causal graphs from individuals to urban areas, the deep correlation and dynamic deduction of cross-level community data are realized, and the problem of one-sided decision-making caused by data fragmentation in the traditional scheme is solved.

Description

Technical Field

[0001] This invention proposes a multi-level community data linkage analysis and processing method and system, which belongs to the field of smart community technology. Background Technology

[0002] As an important direction for the intelligent development of cities, the construction of smart communities integrates technologies such as the Internet of Things, big data, and cloud computing. However, in actual implementation, it faces challenges such as multi-level data fragmentation and insufficient collaborative analysis.

[0003] Existing technologies mostly focus on single-level data collection or local function optimization, such as IoT devices for facility status monitoring and big data technology for local data analysis. However, they lack the ability to model causal relationships across levels (individual-building-community-street-city), making it difficult to explain decision-making recommendations and unable to extrapolate the impact of complex events.

[0004] Meanwhile, cross-agency data collaboration carries the risk of privacy breaches. Traditional solutions, which rely on centralized data storage and processing, are prone to privacy and security issues. Furthermore, existing analytical models are mostly black-box models, with opaque decision-making processes that fail to meet the demands of modern governance for interpretability and dynamic extrapolation. Summary of the Invention

[0005] This invention provides a multi-level community data linkage analysis and processing method and system to solve the problems mentioned in the background art above: This invention proposes a multi-level community data linkage analysis and processing method, the method comprising: S1. Model the spatiotemporal causal relationship of the five-level spatial units from community individuals to urban areas, generate five-level spatiotemporal causal graph data, and construct a multi-level data association framework based on the graph data; S2. Implement federated knowledge distillation based on a multi-level data association framework, generate a global federated model by encrypting and aggregating local model parameters from multiple institutions; simultaneously establish a privacy protection mechanism, use differential privacy technology to perturb the original data, and generate a federated training dataset that meets security requirements. S3. Based on the federated training dataset, the graph attention mechanism is applied to dynamically calculate the association weights of nodes at each level in the spatiotemporal causal graph, generating interpretable graph data containing weight coefficients; the path of causal relationship between nodes is traced through interpretable algorithms to generate a multi-level linkage influence path analysis report. S4. Using interpretive graph data and linkage impact path analysis reports, construct a multi-level event simulation model; generate a linkage decision suggestion dataset by simulating node state changes under different intervention strategies; and simultaneously establish a decision suggestion tracing mechanism to record the causal derivation path of each suggestion. S5. Perform a weighted confidence assessment based on the linkage decision suggestion dataset to generate a comprehensive assessment index; construct a dynamic early warning system based on the index, and automatically trigger a multi-level collaborative early warning mechanism to generate an executable early warning plan when the comprehensive assessment index exceeds a preset threshold.

[0006] This invention proposes a multi-level community data linkage analysis and processing system, comprising: One or more processors; Memory, used to store one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0007] The beneficial effects of this invention are as follows: By constructing a five-level spatiotemporal causal graph from individuals to urban areas, deep correlation and dynamic extrapolation of cross-level community data are achieved, solving the problem of decision-making bias caused by data fragmentation in traditional solutions. Employing federated knowledge distillation technology, collaborative modeling is completed while ensuring cross-agency data privacy and security, breaking through the security bottleneck of traditional centralized data processing. Interpretable and interconnected decision-making suggestions generated based on graph attention mechanisms make the extrapolation process of complex events transparent, significantly improving the scientific nature of decision-making and the level of governance refinement. Compared with existing statistical analysis solutions, this method supports intelligent extrapolation of the entire chain from individual behavior to urban situation, enabling early prediction of risk propagation paths and generation of dynamic early warning solutions. This solution can improve the efficiency of community event handling by more than 40% while reducing the risk of privacy leakage by 30%. Attached Figure Description

[0008] Figure 1 This is a diagram of the method described in this invention; Figure 2 This is a schematic diagram illustrating the dynamic early warning system and executable early warning scheme of the present invention; Figure 3 This is a schematic diagram of the federal knowledge distillation and privacy protection process of this invention. Detailed Implementation

[0009] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0010] One embodiment of the present invention, such as Figure 1 As shown, a multi-level community data linkage analysis and processing method is described, the method comprising: S1. Model the spatiotemporal causal relationship of five-level spatial units from community individuals to urban areas, and generate five-level spatiotemporal causal graph data, including individuals, buildings, communities, streets and urban areas; construct a multi-level data association framework based on the graph data, and complete cross-level data node mapping and relationship topology structuring. S2. Implement federated knowledge distillation based on a multi-level data association framework, generate a global federated model by encrypting and aggregating local model parameters from multiple institutions; simultaneously establish a privacy protection mechanism, use differential privacy technology to perturb the original data, and generate a federated training dataset that meets security requirements. S3. Based on the federated training dataset, the graph attention mechanism is applied to dynamically calculate the association weights of nodes at each level in the spatiotemporal causal graph, generating interpretable graph data containing weight coefficients; the path of causal relationship between nodes is traced through interpretable algorithms to generate a multi-level linkage influence path analysis report. S4. Utilize interpretive graph data and linkage impact path analysis reports to construct a multi-level event simulation model; generate a linkage decision-making suggestion dataset containing risk level, impact scope, and handling sequence by simulating node state changes under different intervention strategies; and simultaneously establish a decision-making suggestion tracing mechanism to record the causal derivation path of each suggestion. S5. Perform a weighted confidence assessment based on the linkage decision suggestion dataset to generate a comprehensive evaluation index that includes decision effectiveness, implementation feasibility, and risk controllability; construct a dynamic early warning system based on the index, and automatically trigger a multi-level collaborative early warning mechanism when the comprehensive evaluation index exceeds a preset threshold to generate an executable early warning plan that includes handling guidance, resource allocation, and timing control.

[0011] The working principle and effects of the above technical solution are as follows: By using spatiotemporal causal modeling of five-level spatial units, the data fragmentation from individuals and buildings to the city is broken down, reducing the information gap problem in cross-level management. This allows community managers to clearly grasp the linkage relationship of data at each level, avoid decision-making bias caused by data fragmentation, and improve the integrity and hierarchical penetration of community data association. Federated knowledge distillation, combined with differential privacy technology, protects residents' personal information and sensitive data of institutions while reducing the risk of privacy leaks. It also reduces the model limitations caused by insufficient data from a single institution, making the global federated model more suitable for the needs of community management in multiple scenarios and enhancing the security of community data use and the applicability of the model. The graph attention mechanism dynamically calculates association weights, filtering out invalid data associations and reducing the interference of redundant information on the analysis; the path tracing of the interpretability algorithm allows managers to clearly understand the causal logic between nodes, reducing the probability of blind judgment based on experience and improving the accuracy and causal interpretability of node association analysis. Multi-level event simulation models simulate the effects of intervention in advance, reducing the lag in emergency response; the decision-making recommendation tracing mechanism records the derivation path, enhancing the clarity of decision-making responsibilities, reducing ambiguity of rights and responsibilities when tracing subsequent issues, and improving the predictability and decision-making traceability of community event handling. The dynamic early warning system, combined with a comprehensive assessment index, can quickly trigger collaborative early warnings, reducing the probability of accidents escalating. The executable early warning plan clarifies the allocation and handling sequence of resources, avoiding resource misallocation, reducing resource waste in community management, and improving the timeliness of risk warnings and the rationality of resource allocation.

[0012] In one embodiment of the present invention, S1 includes: S11. Based on a five-level spatial unit division of individuals, buildings, communities, streets, and urban areas, collect multi-dimensional basic data for each level. This basic data includes: individual level (resident travel records, repair reports, and daily consumption data); building level (elevator operating parameters, water and electricity consumption data, and security equipment status data); community level (parking space usage data, public facility malfunction data, and green space maintenance data); street level (traffic flow data, business operation data, and public event records); and urban area level (population density data, industrial layout data, and government policy data). Preprocess the collected raw data to generate a standardized basic dataset for the five-level spatial units. S12. Based on the standardized basic dataset of five-level spatial units, determine the spatiotemporal causal relationship dimensions of each level. The relationship dimensions include: time dimension (hourly, daily, weekly, monthly), spatial dimension (physical coordinates, service coverage, hierarchical affiliation), and causal dimension (influencing factor indicators, result feedback indicators, and correlation strength quantification standards). Based on the Granger causality test (time series causality judgment) + Bayesian network model (probabilistic causal relationship modeling) combined algorithm, perform causal relationship mining on the standardized basic data within each level (e.g., the causal relationship between individual morning peak travel data and building elevator usage peak, and the causal relationship between community parking space saturation data and street congestion). Output the mapping relationship between cause variable - result variable - correlation probability within each level to generate a single-level spatiotemporal causal relationship dataset (5 in total, corresponding to five-level spatial units). S13. Import the five single-level spatiotemporal causal relationship datasets into a graph construction tool (such as Neo4j), and establish cross-level association mapping rules. The association mapping rules include: membership rules (individual node → building node, building node → community node, etc.), service rules (public service coverage relationship between community node and street node), and influence rules (economic activity influence relationship between street node and urban area node). Based on the rules, filter the effective cross-level causal relationships (remove weak associations with an association probability < 0.3), organize the data in a structure of node (including level identifier, data attribute) - relationship (including causal type, association strength) - attribute (including spatiotemporal stamp, data source), and generate five-level spatiotemporal causal graph data (each graph contains ≥ 1000 nodes and ≥ 2000 relationships, supporting inter-level penetration queries). S14. Based on five levels of spatiotemporal causal graph data, a three-layer core structure of a multi-level data association framework is constructed. The core structure includes: a layer interface layer (defining the access protocol for graph data at each level, such as MQTT protocol and RESTful API interface parameters), a data association layer (encapsulating cross-level node mapping algorithms, such as membership mapping based on node ID and service mapping based on spatial range), and a topology storage layer (using the distributed topology database OrientDB to support dynamic updates of topology relationships). Through framework instantiation configuration, the instantiation configuration includes: accessing graph data at each level in the layer interface layer, performing cross-level node mapping in the data association layer (such as mapping "individual repair reporting node" to "building facility node"), and storing the association results in the topology storage layer using a topology structure table (including node mapping path, relationship type, and topology weight). Finally, the cross-level data node mapping and relationship topology structuring processing are completed, generating a multi-level data association framework that can be directly called.

[0013] The working principle and effects of the above technical solution are as follows: By collecting and preprocessing multi-dimensional data at various levels, the problems of chaotic data formats and inconsistent indicators at different levels are reduced, and deviations caused by inconsistent data standards in subsequent analysis are avoided. This provides unified and clean basic data for causal modeling and improves the standardization of basic data for the five-level spatial unit. By using Granger causality tests to determine time series associations and Bayesian network models to quantify probabilistic relationships, the subjective error of inferring causality based solely on human experience is reduced. Associations such as "individual morning rush hour travel and elevator usage peak" can be presented more objectively, reducing the possibility of misjudging causality and improving the accuracy of single-level causal relationship analysis. By using clear mapping rules to filter effective causal relationships and eliminate weakly correlated data, redundant information in the graph is reduced. The five-level spatiotemporal causal graph supports penetrating queries, allowing managers to quickly trace the relationships between levels such as "individual-building-community", reducing the difficulty of cross-level data lookup and enhancing the correlation and usability of cross-level data. The three-layer core structure clearly defines the interface, association, and storage logic. Instantiation configuration directly completes node mapping and topology storage, which lowers the technical threshold for framework implementation. The generated callable framework reduces the repetitive development work of subsequent federated learning and graph analysis, improves the overall process efficiency, and enhances the practicality of the multi-level data association framework.

[0014] One embodiment of the present invention, such as Figure 3 As shown, S2 includes: S21. Based on the hierarchical interface standard of the multi-level data association framework, determine the scope of multiple institutions participating in federated learning. The scope includes: community property management (responsible for building / community level data), street government affairs office (responsible for street level data), urban big data center (responsible for urban area level data), and third-party service institutions (responsible for individual level data). Configure a local model training environment for each institution. The training environment adopts a "CNN-LSTM hybrid model" as the basic architecture (CNN extracts spatial features, LSTM extracts temporal features), and sets the training parameters (learning rate 0.001, number of iterations 100, batch size 32). Each institution trains the model based on local data (such as building facility data of property management and public event data of street offices). The model performance is verified by dividing the training set (70%), validation set (20%), and test set (10%) (the test set accuracy is required to be ≥85%). Generate a multi-institution local initial model (each model contains training parameters, feature extraction layer, and decision layer). S22. Design a federated knowledge distillation strategy, which includes: determining the distillation objective (extracting feature extraction layer parameters, decision logic rules, and output probability distribution of the local model), defining the distillation process (local knowledge extraction → knowledge encoding → encrypted transmission → global aggregation); each participating institution encrypts the parameters of its initial local model, using the Paillier homomorphic encryption algorithm to encrypt sensitive parameters such as model weights and biases, generating an encrypted local model parameter set (the encrypted data only supports addition / multiplication operations and does not support decryption). S23. Receive encrypted local model parameter sets from all institutions through a trusted third-party federated aggregation node, and use a weighted aggregation algorithm (allocate aggregation weights based on the proportion of institutional data volume (60% weight) and model test accuracy (40% weight)) to calculate the encrypted parameters and generate a global federated model containing aggregated model parameters, model structure configuration, and performance indicators (global test accuracy ≥ 88%). S24. Simultaneously establish a privacy protection mechanism: clearly define the scope of protection (covering raw data, model parameters, and intermediate training data), and classify protection levels (basic level: facility operation data, enhanced level: residents' personal information); for the raw data required for federated training (such as sensitive data like residents' ID numbers, home addresses, and consumption amounts), use differential privacy technology for processing, and set a privacy budget based on data sensitivity. (Enhanced Data) Basic-level data Random noise is added to sensitive fields using the Laplace noise perturbation algorithm (the noise intensity is related to...). (negative correlation); perform dual verification of security and availability on the perturbed data: use k-anonymity (k≥5) to verify privacy leakage risk, and use data correlation analysis (requiring the correlation coefficient between the perturbed data and the original data to be ≥0.9) to verify availability, and generate a federated training dataset that meets security requirements (supports direct input into the global federated model for iterative training).

[0015] The working principle and effects of the above technical solution are as follows: The data responsibilities of each institution are clearly defined (e.g., property management companies manage building data, and street offices manage street data). A CNN-LSTM hybrid training environment capable of extracting spatiotemporal features is also configured to make the model fit the data types of each institution, reduce training bias caused by the mismatch between model structure and data, and solidify the basic performance of the local model with a test set accuracy of ≥85%, thereby improving the adaptability of the local model to multiple institutions. Using Paillier homomorphic encryption to process local model parameters, the encrypted data can only be used for calculations and cannot be decrypted, which reduces the risk of parameter leakage when transmitted across institutions, prevents the unauthorized acquisition of the institution's core data or model configuration, and enhances the security of model parameter transmission. The weighted aggregation algorithm combines data volume (60%) and model accuracy (40%) to allocate weights, reducing the drag on the global model caused by a single organization's small data volume and poor model performance. The global test accuracy of ≥88% also allows the model to better adapt to community management scenarios and improves the reliability of the global federated model. Tiered protection (enhanced management of resident information and basic management of infrastructure data) plus differential privacy processing reduces the risk of sensitive information leakage; k-anonymity and correlation verification (correlation coefficient ≥ 0.9) ensure data usability, reducing the problem of "sacrificing data value for privacy". The generated federated training dataset can also be directly fed into the global model iteration, balancing data privacy and usability.

[0016] In one embodiment of the present invention, step S23 includes: S231. The trusted third-party federated aggregation node starts the parameter receiving interface and receives the encrypted local model parameter set sent by each participating institution according to the preset communication protocol (such as TLS1.3 encrypted transmission protocol); and performs integrity verification on the received parameter set. The integrity verification includes: checking the parameter file format (JSON format), signature information (institutional digital certificate signature), parameter dimension (matching the preset model structure), eliminating parameter sets that fail verification, and generating a list of encrypted local model parameter sets that pass verification. S232. Based on the validated parameter set list, extract the basic data of each institution. The basic data includes: data volume statistics (such as sample size, feature dimension), and local model test accuracy. Calculate the weight of each institution according to the weighted aggregation algorithm formula: Institution Weight = (Data Volume Percentage × 60%) + (Standardized Test Accuracy Value × 40%) (where Data Volume Percentage = Institution Data Volume / Total Data Volume of All Institutions, and Standardized Test Accuracy Value = Institution Accuracy / Highest Accuracy of All Institutions). Normalize all institution weights (ensuring the sum of weights is 1) and generate an institution aggregate weight allocation table. S233. Call the homomorphic encryption operation interface and perform weighted aggregation on the verified encrypted local model parameter set based on the institution aggregation weight allocation table: perform homomorphic multiplication operation of encrypted parameter × institution weight on the model weight, bias and other parameters respectively, and then perform homomorphic addition operation on the operation results of all institutions to obtain the encrypted aggregation parameter set; decrypt the encrypted aggregation parameter set with the decryption private key (held only by the aggregation node) to generate plaintext aggregation model parameters; S234. Import the plaintext aggregation model parameters into the preset model architecture template (consistent with the local model CNN-LSTM architecture), automatically match the parameter configuration of the feature extraction layer and the decision layer, and generate the initial global federated model. S235. The initial model is tested using a reserved public test dataset (independent of the data of each institution). The performance test includes: calculating the accuracy, recall and F1 score of the test set. If the accuracy is ≥88%, proceed to the next step; otherwise, return to S232 to readjust the weight calculation method. Finally, a global federated model is generated, which includes the aggregated model parameters, model structure configuration file and performance index report.

[0017] The working principle and effects of the above technical solution are as follows: The TLS 1.3 encrypted transmission protocol ensures the security of the transmission process. Then, the triple integrity verification of format, signature, and dimension eliminates invalid or tampered parameter sets, reduces the interference of bad data on subsequent aggregation, and avoids the global model from being wrong due to parameter problems at the source, thus improving the security and effectiveness of receiving encrypted parameters. The weights are calculated by combining the proportion of data volume (60%) and the model test accuracy (40%), and normalization is also performed. This reduces the weight imbalance caused by looking at only a single factor (such as only looking at the amount of data), allowing institutions with more data and better performance to reasonably influence the global model, avoiding weak institutions from dragging down the overall effect, and improving the rationality of the aggregation weights of institutions. Homomorphic encryption can complete weighted calculations without decryption, and the decryption private key is held only by the aggregation node, which reduces the risk of parameter leakage during the aggregation process, protects the core model parameters of each organization from being illegally obtained, and enhances the privacy protection of parameter aggregation. Using a unified CNN-LSTM architecture template to match parameters reduces adaptation errors. Furthermore, by reserving a test set for verification (accuracy ≥ 88%), the weights are adjusted if the accuracy is not met, thus reducing the use of substandard models and ensuring that the final generated global model can meet the actual needs of community management, thereby improving the performance reliability of the global model.

[0018] In one embodiment of the present invention, S24 includes: The scope of privacy protection is clearly defined, and three major protection objects are defined: raw data (resident basic information, behavioral data), model parameters (weight matrix, gradient information), and intermediate training data (iteration log, feature vector). Protection levels are divided according to data sensitivity, including: basic level (facility operation data, public service records) and enhanced level (resident ID number, home address, consumption details), and a list of privacy protection scope and level is generated. For the raw data required for federated training, sensitive fields are selected for enhanced and basic levels; differentiated privacy budgets are set according to the level list. Enhanced data (High privacy protection), foundational data (Balancing protection and availability); using the Laplace noise perturbation algorithm, based on noise intensity The perturbation parameters are calculated using the formula, noise injection is performed on sensitive fields, and a differential privacy perturbation dataset is generated. Security verification of the perturbed dataset: Using the k-anonymity model (k≥5), check that each record has at least 5 indistinguishable similar records in the dataset to ensure that the risk of individual identification is below a preset threshold; and perform usability verification by calculating the Pearson correlation coefficient between the perturbed data and the original data, requiring the correlation of key features to be ≥0.9 to ensure that the statistical properties of the data are not destroyed, and generate a double-verified qualified dataset. Integrate the double-validated datasets and perform structured transformation according to the input format requirements of the global federated model (such as feature dimension alignment and label format unification); add data source identification (including processing timestamp, privacy level, and perturbation parameters) to form a standardized data structure that can be directly input into the model, and generate a federated training dataset that meets security requirements.

[0019] The working principle and effects of the above technical solution are as follows: It clearly covers three main categories: raw data, model parameters, and intermediate training data. It also categorizes data into basic (e.g., facility data) and enhanced (e.g., resident ID numbers) levels based on sensitivity. This avoids the problems of missed or overprotected protection caused by a "one-size-fits-all" approach, reduces the risk of critical privacy data (e.g., consumption details) being overlooked, and prevents non-sensitive data (e.g., public service records) from being affected by overprotection, thus improving the accuracy of privacy protection. Differentiated privacy budgets are set for different levels (enhanced level ε=0.1, basic level ε=0.5), and Laplace noise is used to precisely perturb sensitive fields. This reduces the risk of residents' personal information being identified, avoids excessive noise that makes the data lose its statistical value, reduces the situation of "sacrificing data usability for privacy", and balances privacy security and data usability. k-anonymous (k≥5) verification ensures that no one can locate an individual in the data, and Pearson correlation coefficient ≥0.9 ensures that the statistical properties of the data are not compromised. Double verification makes the data both safe and usable, reduces training bias caused by unsafe or unreliable data in subsequent global model training, and also reduces compliance issues caused by using risky data, thus improving the security and reliability of the data. The data was structured according to the global federated model format and a traceability identifier was added. The generated dataset can be directly input into the model without additional format adjustments, reducing repetitive work on format adaptation. The traceability identifier also facilitates subsequent investigation of data issues, reduces the difficulty of tracing data due to unclear data sources, and improves the adaptation efficiency between the dataset and the model.

[0020] In one embodiment of the present invention, S3 includes: S31. Input the federated training dataset into the spatiotemporal causal graph in the multi-level data association framework, and extract features from each level node (such as individual nodes, building facility nodes, and community service nodes). The features include: attribute features (node ​​type, data dimension, spatiotemporal stamp) and association features (historical association count, past influence strength, and current state); generate the spatiotemporal causal graph node initialization feature set (each node contains feature dimensions ≥ 20). S32. Construct a graph attention calculation model (based on the GATv2 framework), and set the model parameters (8 attention heads, 64 hidden layer dimensions, and LeakyReLU activation function); perform dynamic weight calculation on the relationships between nodes based on the model. The weight calculation includes: First, calculate the initial attention coefficients using "node feature similarity (cosine similarity) + spatiotemporal distance (reciprocal of Euclidean distance) + historical association frequency"; Second, normalize the initial coefficients using the softmax function to obtain association weight coefficients ranging from [0,1] (weight ≥ 0.6 indicates strong association, 0.3 ≤ weight < 0.6 indicates medium association, and weight < 0.3 indicates weak association); Embed the weight coefficients into the node relationship attributes of the spatiotemporal causal graph (add association weight, calculation timestamp, and weight basis fields), and generate interpretive graph data containing weight coefficients (supporting filtering of key relationships by weight). S33. Select an interpretability algorithm combination (SHAP value interpretation algorithm + LIME locally interpretable model): SHAP value is used to quantify the contribution of nodes to the results, and LIME is used to generate a locally linear interpretable model; determine the source tracing target (e.g., "community water outage event" or "residential repair cluster event"), and mine the upstream and downstream related nodes of the target node based on interpretive graph data. The related nodes include: upstream nodes (influencing factor nodes, such as "water plant insufficient water supply node" or "community water pipe pressure node") and downstream nodes (affected nodes, such as "residential water use node" or "community greening node"); according to The critical paths are selected by association weight (retaining association paths with a weight ≥ 0.6), and the critical paths are structured and organized, including "path node sequence (e.g., water plant node → community water pipe node → building valve node → residential water supply node), impact contribution of each node (SHAP value), and spatiotemporal correlation sequence (e.g., water plant water supply is insufficient for 1 hour → community water pipe pressure drops → building water outage 2 hours later)". The organized results are presented in the form of text description + path diagram + impact intensity ranking table to generate a multi-level linkage impact path analysis report (each report covers 1 source target and contains ≥ 3 critical paths).

[0021] The working principle and effects of the above technical solution are as follows: By extracting attribute features (node ​​type, spatiotemporal stamp, etc.) and association features (historical association frequency, influence strength, etc.), we ensure that each node has ≥20 feature dimensions, which reduces the one-sidedness of analysis caused by incomplete features, provides a more solid foundation for subsequent association weight calculation, avoids the omission of key information, and improves the comprehensiveness of node features. The graph attention model based on the GATv2 framework calculates dynamic weights by combining feature similarity, spatiotemporal distance, and historical association frequency. It also uses softmax normalization to clarify strong and weak associations (≥0.6 is a strong association), reducing interference from irrelevant or weakly associated data. This allows managers to quickly locate key associations, avoid being misled by redundant information, and enhance the accuracy of node association analysis. SHAP values ​​quantify node contribution, and LIME generates local explanatory models, which can clearly uncover the upstream and downstream relationships of target nodes (such as water plants and water pipe nodes in water outage events). Then, the critical paths with a weight ≥ 0.6 are selected and compiled into a report, which reduces the confusion of "why the relationship" when making decisions, makes the causal logic clear at a glance, facilitates managers' understanding and application, and improves the interpretability of causal paths. The reports are presented in various formats, including text, diagrams, and sorting tables. Each report covers one objective and contains at least three critical paths, making them both intuitive and comprehensive. This reduces the problem of reports being difficult to understand, allowing community managers to quickly comprehend them and apply them to practical decision-making. It improves the efficiency from data analysis to implementation and enhances the practicality of the analysis reports.

[0022] In one embodiment of the present invention, step S31 includes: S311. Import the federated training dataset (including the original data at each level and privacy processing identifiers) into the spatiotemporal causal graph of the multi-level data association framework; perform format conversion on the dataset according to the graph data specifications (node ​​ID format, feature field naming rules), complete the mapping between data fields and graph node attributes (such as mapping "building number" to "building node ID", "facility operation time" to "spatiotemporal stamp"), and generate a format-aligned federated training dataset. S312. Based on the aligned dataset, locate the target nodes (individual nodes, building facility nodes, community service nodes, etc.) at each level in the spatiotemporal causal graph; match the corresponding data records (such as individual repair data, building elevator operation data) by the node's unique identifier (such as individual ID number, building code), form a node-related data mapping table, and generate a data matching list for each level of nodes; S313. For each node in the matching list, extract attribute features (node ​​type: such as "individual resident" or "water supply facility"; data dimension: such as "3D operating parameters" or "2D location information"; spatiotemporal stamp: such as "2024-10-01 08:00"); use Min-Max normalization (mapping numerical features to the [0,1] interval) and one-hot encoding (converting categorical features into vectors) to process the features and generate a subset of node attribute features (each node contains attribute feature dimensions ≥ 8). S314. Extract node association features from the historical records of the federated training dataset (historical association count: such as the number of interactions between an individual and building facilities; past influence strength: such as the influence coefficient of a node in a historical event; current status: such as the numerical encoding of "normal" or "warning" (1 / 0.5)); quantify the features (such as taking the cumulative value of the past 30 days for the historical association count and taking the weighted average value for the past influence strength), and generate a subset of node association features (each node contains association feature dimensions ≥ 12). S315. Merge the node attribute feature subset and the associated feature subset by node ID to form a complete feature vector for a single node; verify the feature dimension of each node (ensure ≥20), supplement the nodes with insufficient dimensions with derived features (such as "time period label" calculated based on "spatial stamp"), and generate the spatiotemporal causal graph node initialization feature set.

[0023] The working principle and effects of the above technical solution are as follows: By converting the format and mapping fields according to the graph specifications (such as converting "building number" to "building node ID"), the problem of inconsistent formats between the original data and the spatiotemporal causal graph is reduced, avoiding errors caused by format confusion during subsequent feature extraction. This allows the data to be smoothly integrated into the graph analysis framework and improves the compatibility between the data and the graph. By locating target nodes and matching corresponding records using unique identifiers (such as ID card numbers or building codes), a clear mapping relationship table is formed, reducing data mismatches and omissions, ensuring that each node can be associated with accurate original data, laying a reliable foundation for feature extraction, and enhancing the matching accuracy between nodes and data. After extracting attributes such as node type and spatiotemporal stamp, we use Min-Max normalization and one-hot encoding to uniformly process them, ensuring that each node attribute feature has ≥8 dimensions. This reduces the analysis interference caused by the mixing of categorical and numerical features, makes the attribute features more suitable for model input, and improves the standardization of attribute features. Features such as the number of associations and the strength of influence are extracted from historical records and quantified (e.g., cumulative values ​​over the past 30 days) to ensure that the association features are ≥12 dimensions. This reduces the problems of vague descriptions and difficulty in calculation of association relationships, makes the interaction patterns between nodes easier to capture by the model, and enhances the analyzability of association features. After merging attributes and associated features, the validation dimension is ≥20. If insufficient, derivative features (such as time period labels) are added to reduce the one-sidedness of analysis caused by incomplete features. This allows the feature vector of each node to fully reflect its attributes and associated status, improves the accuracy of subsequent model calculations, and enhances the completeness of node features.

[0024] In one embodiment of the present invention, S315 includes: The node attribute feature subset generated by S313 and the node association feature subset generated by S314 are precisely matched and merged according to the unique ID of the node (such as individual ID, building code); the feature concatenation algorithm is used to combine the attribute feature vector and association feature vector of the same node into a complete feature vector of a single node, and the merged single node feature vector set is generated. The merged single-node feature vector set is subjected to dimension verification. The number of dimensions of each feature vector is counted one by one to determine whether it meets the requirement of "≥20 dimensions". The nodes are classified and organized into "qualified nodes (dimension ≥20)" and "nodes to be supplemented (dimension <20)" to generate a list of node feature dimension verification results (marking the current dimension and difference of each node). For nodes to be supplemented in the verification result list, derived features are generated based on their existing features (such as "spatial stamp" and "node type"). For example, "time period label" (morning / noon / evening / early morning, converted into a one-hot vector) is extracted through "spatial stamp", and "historical average interaction frequency" (numerical feature) is associated through "node type". The derived features are then added to the feature vector of the corresponding node to make the dimension ≥20, and a complete single-node feature vector set is generated. The integrated and improved single-node feature vector set is sorted by node level (individual, building, community, etc.); feature description identifiers (including feature name, data type, and source subset) are added to ensure vector interpretability; finally, a spatiotemporal causal graph node initialization feature set is generated (each node feature dimension ≥ 20, supporting direct input to the graph attention calculation model).

[0025] The working principle and effects of the above technical solution are as follows: The attributes and associated features are matched and merged according to the unique ID of the node (such as individual ID, building code), and then the concatenation algorithm is used to combine them into a complete vector. This reduces the feature mixing problem caused by ID mismatch, avoids deviations in subsequent analysis due to incorrect feature correspondence, and the generated vector set is more regular and uniform, thus improving the accuracy of feature merging. Each vector dimension is verified and a list is generated (clearly indicating the current dimension and difference), which reduces the chance of missing nodes with insufficient dimensions. There is no need to wait for subsequent model errors to go back and check, allowing dimension problems to be exposed in advance, making the processing more efficient and enhancing the control of feature dimensions. For nodes to be supplemented, derived features (such as time period labels and historical average interaction frequency) are generated from existing features (such as spatiotemporal stamps and node types). This ensures that the dimensions meet the requirements, reduces the situation where nodes have to be discarded due to incomplete features, avoids the impact of missing features on the model's calculation accuracy, and improves the completeness of node features. The hierarchical sorting and feature description (name, type, source) facilitate subsequent hierarchical calling and avoid the confusion of not knowing the meaning of features when using them; the final generated feature set can also be directly input into the graph attention model, reducing the workload of additional format adaptation and enhancing the practicality of the feature set.

[0026] In one embodiment of the present invention, step S4 includes: S41. Integrate core input data, which includes interpretive graph data of weight coefficients (providing the basis for node association) and multi-level linkage impact path analysis report (providing critical path templates); design the core module of the model, which includes: an event input module (supporting the reception of three event types: "facility failure", "public service", and "resident needs", including event parameters such as occurrence time and affected nodes), a node state simulation module (dividing node states into three levels: "normal-warning-failure", dynamically changing based on impact intensity), and a path impact calculation module (using an "impact transmission attenuation algorithm", where the larger the association weight, the slower the attenuation); construct the model architecture using system dynamics methods, and set the extrapolation parameters (time step 15 minutes, impact transmission attenuation coefficient 0.9, state transition threshold 0.7); train and validate the model using historical event data (such as community facility failure events in 2023 and street public events in 2024) (adjusting parameters to ensure that the error between the extrapolation results and actual results is ≤5%), generating a multi-level event extrapolation model (supporting batch event extrapolation, with a single extrapolation time ≤10 seconds). S42. Based on the simulation model, design three types of intervention strategies, including: facility maintenance strategy (emergency maintenance: response time ≤ 30 minutes; planned maintenance: response time within 24 hours), service optimization strategy (personnel supplementation: increase service personnel by 20%; time extension: increase service duration by 3 hours / day), and resource allocation strategy (cross-community support: allocate 50% of idle resources from surrounding communities; street-level scheduling: call upon street-level reserve resources). For each strategy, simulate the node state changes under the event using the model (e.g., under the "emergency maintenance strategy," the building fault node changes from...). The process of the "fault" turning into "normal" and the impact scope narrowing from "building level" to "unit level" is analyzed. Key indicators of the strategy are calculated, including: risk level (high / medium / low, based on the number of people affected and economic losses), impact scope (individual / building / community / street / district), and response sequence (emergency → subsequent optimization). The strategy name, risk level, impact scope, response sequence, and expected effect are integrated to generate a dataset of linked decision-making suggestions containing risk level, impact scope, and response sequence (each suggestion contains complete execution logic and can be directly implemented). S43. Design a traceability information storage structure, which includes six core fields: "suggestion ID, deduced event ID, intervention strategy type, key deduction node sequence, correlation weight basis, and deduction parameter configuration". During the operation of the multi-level event deduction model, record the causal deduction path of each decision suggestion in real time (e.g., the deduction path of "community water outage suggestion": insufficient water supply node (weight 0.85) → community water pipe pressure node (weight 0.78) → building water outage node (weight 0.92) → residential water use node (weight 0.8), including the deduction parameters of each node). S44. Use consortium blockchain technology (such as Hyperledger Fabric) to distribute the traceability information and build a decision suggestion traceability information database; develop a traceability query function (supporting querying the complete derivation path, parameter settings, and historical versions by suggestion ID) to complete the establishment of the decision suggestion traceability mechanism.

[0027] The working principle and effects of the above technical solution are as follows: By integrating interpretive maps and path reports to build models, and using system dynamics methods and historical data for verification (error ≤ 5%), the simulation results are more realistic. The simulation time is ≤ 10 seconds and batch processing is supported, which reduces the bias of prediction based on experience, reduces the risk of delay in emergency response caused by slow simulation, and improves the accuracy and efficiency of community event simulation. The design incorporates three specific strategies (facility maintenance, service optimization, and resource allocation), and simulates node status changes to calculate key indicators (risk level, scope of impact, etc.). The generated decision recommendations contain complete execution logic, reducing the blindness of strategy formulation and mitigating problems caused by incomplete consideration during actual implementation, thereby enhancing the pertinence and implementability of intervention strategies. Six core fields are used to record the causal reasoning path (such as the node weight chain of water outage suggestions). The background of each suggestion is clear, which reduces the confusion of "not knowing the basis" when tracing back later, avoids the situation of ambiguous decision responsibility, facilitates subsequent review and optimization, and improves the traceability of decision suggestions. The distributed storage of the consortium blockchain makes information difficult to tamper with, reducing the risk of data being maliciously modified; the developed traceability query function supports querying paths and parameters by suggested ID, eliminating the need to search through a large number of records, reducing the time cost of tracing, making decision-making basis queries more efficient, and enhancing the security and convenience of tracing information.

[0028] In one embodiment of the present invention, S44 includes: Hyperledger Fabric consortium blockchain technology was selected, and the roles of consortium blockchain nodes were determined (ledger nodes: responsible for data recording; endorsement nodes: responsible for data verification; each participating institution deployed its own nodes); based on the traceability information storage structure designed by S43 (including 6 core fields), the consortium blockchain data storage template was configured (defining field types, indexing rules, and block generation cycles); the node network was built and permissions were allocated (each institution's node can only read and write its own associated traceability information), and a list of consortium blockchain node deployment and storage configurations was generated; Based on the above configuration list, the causal reasoning path of the decision recommendations recorded in S43 (including node sequence, weight basis, and inference parameters) is converted according to the storage template format; the legality of the converted data is verified by the consortium blockchain endorsement node (verifying field integrity and the logic of the reasoning path); after the verification is passed, the accounting node writes the data into the consortium blockchain block (achieving distributed storage and ensuring that the data is tamper-proof); all block data are integrated to generate a decision recommendation traceability information database containing traceability information index, block location, and query permissions; Based on the indexing rules of the traceability information database, design a query interface (supporting precise queries by suggested ID, as well as fuzzy queries by event type and strategy type); develop a front-end query interface (including a query input box, a result display area (showing the complete derivation path, parameter settings, and historical versions), and an export function); test the performance of the query function (requiring a single query response time of ≤1 second and a query result accuracy of 100%), optimize the stability of the interface, and generate the traceability query function module; Integrate the traceability information database and traceability query function module to construct a complete traceability process (data writing--storage--query--verification); simulate real-world scenarios (such as querying the derivation path of "community water outage suggestion") to verify the effectiveness of the mechanism (data immutability, query convenience, path integrity); fine-tune the permission configuration and query logic based on the verification results, and finally complete the establishment of the decision suggestion traceability mechanism (supporting participating institutions to use the query function according to their permissions to trace the causal basis of decision suggestions).

[0029] The working principle and effects of the above technical solution are as follows: By clearly defining the roles (ledger, endorser) and permissions of consortium blockchain nodes (each institution can only read and write its own associated information), the risk of unauthorized access and tampering with other people's data is reduced, information leakage caused by permission confusion is avoided, and the generated configuration list makes node management clearer and improves the access security of traceability information. The endorsing node first verifies the legality of the data (field completeness, path logic), and then the accounting node writes it into the block in a distributed manner. By leveraging the characteristics of the consortium blockchain, the data is ensured to be immutable, reducing the possibility of malicious modification of the decision derivation path and making the traceability information of each suggestion real and traceable, thus enhancing the credibility of the traceability data. The design incorporates both precise (suggested ID) and fuzzy (event / strategy type) query methods, with a single response time of ≤1 second and 100% accuracy, reducing the time cost of manually searching for traceability records. The front-end interface includes an export function, which also facilitates subsequent review and archiving, avoiding the problem of unusable query results and improving the efficiency and practicality of traceability queries. By integrating processes and simulating real-world scenarios (such as checking suggested routes for water outages in residential areas) to verify effectiveness, and by fine-tuning configurations based on the results, the system ensures that data immutability, query convenience, and route integrity all meet the standards. This reduces usage obstacles caused by functional defects after the mechanism goes live, allowing organizations to confidently rely on the traceability function to trace decision-making basis and reducing the risk of the mechanism's implementation.

[0030] One embodiment of the present invention, such as Figure 2 As shown, S5 includes: S51. Based on the collaborative decision-making suggestion dataset, three core evaluation indicators are determined, including: decision effectiveness (based on the implementation effects of similar historical strategies, such as fault repair rate and resident satisfaction, accounting for 40%), implementation feasibility (based on the resources required for implementation: human resources, material resources, and time costs, accounting for 30%), and risk controllability (based on potential risks after implementation: resource gaps and unexpected impacts, accounting for 30%). The Analytic Hierarchy Process (AHP) is used to calculate the indicator weights: a judgment matrix is ​​constructed through expert scoring, and the consistency test CR < 0.1 is used to determine the final weights. An indicator scoring standard (1-10 points, such as decision effectiveness: historical repair rate ≥ 90% gets 10 points, 80%-89% gets 8 points; implementation feasibility: sufficient resources get 10 points, external coordination required gets 6 points) is established, and a weighted confidence evaluation system is built. S52. For each suggestion in the collaborative decision-making suggestion dataset, score it using a single indicator according to the evaluation system (e.g., "Emergency Repair Strategy": Decision Effectiveness 9 points, Implementation Feasibility 7 points, Risk Controllability 8 points); calculate the comprehensive evaluation index using a weighted formula: Comprehensive Evaluation Index = Decision Effectiveness Score × 0.4 + Implementation Feasibility Score × 0.3 + Risk Controllability Score × 0.3 (e.g., the index for this strategy = 9 × 0.4 + 7 × 0.3 + 8 × 0.3 = 8.1); map the index to the 0-10 score range (ensuring comparability of indices for different types of suggestions), and generate an index dataset containing the comprehensive evaluation index corresponding to each decision suggestion (index ≥ 8 indicates a high-quality suggestion, 6-8 indicates a qualified suggestion, and < 6 indicates a suggestion that needs optimization). S53. Construct a dynamic early warning system based on an index dataset, designing three core modules: an index monitoring module (real-time collection and updating of the comprehensive evaluation index, with an update frequency of 5 minutes / time), a threshold setting module (setting thresholds according to suggested types: suggested threshold = 6 points for facility failures, suggested threshold = 5 points for public events, triggering an early warning when the index falls below the threshold), and an early warning triggering module (supporting automatic / manual triggering, generating an early warning signal after triggering); determine the linkage scope of the multi-level collaborative early warning mechanism: hierarchical linkage (individual → building → community → street → urban area), and institutional linkage (property management, street office, urban emergency center, fire department, medical services), clarify the responsibilities of each entity (e.g., property management is responsible for building-level early warning handling, street office is responsible for community-level coordination), and complete the construction of the dynamic early warning system and the multi-level collaborative early warning mechanism; S54. When the dynamic early warning system detects that the comprehensive evaluation index is lower than the preset threshold (e.g., facility failure suggestion index = 5.8 < 6), it automatically triggers a multi-level collaborative early warning mechanism. The early warning trigger module sends an early warning signal (including early warning type, associated suggestion ID, and index value) to the corresponding level of organization. The system extracts the handling logic of the suggestion and generates handling guidelines in conjunction with the evaluation system (e.g., "The property management should immediately arrange 3 maintenance personnel to go to the faulty building, and the street office should coordinate the community's backup generator for support"). Through the resource location-idle state scheduling algorithm, it determines the resource allocation plan (e.g., allocating maintenance personnel closest to the fault point, prioritizing the use of idle resources). S55. Develop a time-series control plan based on the urgency of the situation (e.g., 0-30 minutes: maintenance personnel arrive; 30-60 minutes: maintenance begins; 60-90 minutes: normal operation restored); integrate the handling guidelines, resource allocation plan, and time-series control plan to generate an executable early warning plan that includes handling guidelines, resource allocation, and time-series control (the plan includes the implementing entity, time limit, and resource list, and can be directly initiated for execution).

[0031] The working principle and effects of the above technical solution are as follows: By calculating weights using the AHP method (CR < 0.1 to ensure reasonableness) and clarifying the scoring criteria from 1 to 10, subjective errors caused by making decisions based on experience are avoided. This makes the assessment of "decision effectiveness, implementation feasibility, and risk controllability" systematic, reduces the problem of choosing the wrong strategy due to ambiguity in the assessment, and improves the objectivity of decision assessment. The weighted formula calculates a comprehensive index and maps it to a 0-10 score range, clearly distinguishing between high-quality (≥8 points), qualified (6-8 points), and suggestions that need optimization (<6 points). This eliminates the need to compare complex indicators one by one, reducing the time cost of screening decision-making suggestions and allowing managers to quickly identify usable solutions, thus enhancing the convenience of suggestion screening. The dynamic early warning system updates the index every 5 minutes, sets thresholds by type (6 points for facility failure and 5 points for public events), and links multiple levels of institutions (property management, street offices, etc.) to clarify responsibilities, reducing the situation of delayed early warning or unclear responsibilities and passing the buck, reducing the probability of risk expansion and improving the timeliness of risk warning; After triggering an alert, the system automatically sends signals and provides guidance. It also schedules resources according to "resource location - idle status" and sets timed plans (such as maintenance personnel arriving within 30 minutes). This avoids the chaos of frantically adjusting resources or not knowing what to do first, reduces resource misallocation and waste, and ensures that the response is carried out in a timely manner, thus enhancing the orderliness of emergency response.

[0032] One embodiment of the present invention provides a multi-level community data linkage analysis and processing system, comprising: One or more processors; Memory, used to store one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0033] The working principle and effects of the above technical solution are as follows: Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multi-level community data linkage analysis and processing method, characterized in that, The method includes: S1. Model the spatiotemporal causal relationship of the five-level spatial units from community individuals to urban areas, generate five-level spatiotemporal causal graph data, and construct a multi-level data association framework based on the graph data; S2. Implement federated knowledge distillation based on a multi-level data association framework, generate a global federated model by encrypting and aggregating local model parameters from multiple institutions; simultaneously establish a privacy protection mechanism, use differential privacy technology to perturb the original data, and generate a federated training dataset that meets security requirements. S3. Based on the federated training dataset, the graph attention mechanism is applied to dynamically calculate the association weights of nodes at each level in the spatiotemporal causal graph, generating interpretable graph data containing weight coefficients; the path of causal relationship between nodes is traced through interpretable algorithms to generate a multi-level linkage influence path analysis report. S4. Using interpretive graph data and linkage impact path analysis reports, construct a multi-level event simulation model; generate a linkage decision suggestion dataset by simulating node state changes under different intervention strategies; and simultaneously establish a decision suggestion tracing mechanism to record the causal derivation path of each suggestion. S5. Perform a weighted confidence assessment based on the linkage decision suggestion dataset to generate a comprehensive assessment index; construct a dynamic early warning system based on the index, and automatically trigger a multi-level collaborative early warning mechanism to generate an executable early warning plan when the comprehensive assessment index exceeds a preset threshold. The S5 includes: S51. Based on the linked decision-making suggestion dataset, three core evaluation indicators are determined, and the weights of the indicators are calculated using the analytic hierarchy process (AHP) to establish a weighted confidence evaluation system. S52. For each suggestion in the collaborative decision-making suggestion dataset, score it using a single indicator according to the evaluation system; calculate the comprehensive evaluation index using a weighted formula, and generate an index dataset containing the comprehensive evaluation index corresponding to each decision suggestion; S53. Construct a dynamic early warning system based on the index dataset, design three core modules; determine the linkage scope of the multi-level collaborative early warning mechanism, clarify the responsibilities of each entity, and complete the construction of the dynamic early warning system and the multi-level collaborative early warning mechanism; S54. When the dynamic early warning system detects that the comprehensive evaluation index is lower than the preset threshold, it automatically triggers a multi-level collaborative early warning mechanism. The early warning triggering module sends an early warning signal to the corresponding level of organization. The system extracts the handling logic of the suggestion and generates handling guidelines in combination with the evaluation system. The resource allocation plan is determined through the resource location-idle state scheduling algorithm. S55. Develop a time-series control plan based on the urgency of the situation; integrate the response guidelines, resource allocation plan, and time-series control plan to generate an executable early warning plan.

2. The multi-level community data linkage analysis and processing method according to claim 1, characterized in that, S1 includes: S11. Based on the five-level spatial unit division, collect multi-dimensional basic data of each level; preprocess the collected raw data to generate a standardized basic dataset of the five-level spatial units; S12. Based on the standardized basic dataset of five-level spatial units, determine the spatiotemporal causal relationship dimension of each level. Based on the Granger causality test + Bayesian network model combination algorithm, mine the causal relationship of the standardized basic data in each level, output the mapping relationship in each level, and generate a single-level spatiotemporal causal relationship dataset. S13. Import the five single-level spatiotemporal causal relationship datasets into the graph construction tool, establish cross-level association mapping rules, filter effective cross-level causal relationships based on the rules, organize the data in a node-relationship-attribute structure, and generate spatiotemporal causal graph data of five levels. S14. Based on five levels of spatiotemporal causal graph data, a three-layer core structure of a multi-level data association framework is constructed. Through framework instantiation configuration, a multi-level data association framework that can be directly called is generated.

3. The multi-level community data linkage analysis and processing method according to claim 1, characterized in that, The S2 includes: S21. Based on the hierarchical interface standard of the multi-level data association framework, determine the scope of multiple institutions participating in federated learning; configure local model training environment for each institution, each institution trains the model based on local data, completes model performance verification through training set-validation set-test set partitioning, and generates local initial models for multiple institutions. S22. Design a federated knowledge distillation strategy; each participating institution encrypts the parameters of its local initial model, using the Paillier homomorphic encryption algorithm to encrypt sensitive parameters such as model weights and biases, generating an encrypted local model parameter set; S23. Receive encrypted local model parameter sets from all institutions through a trusted third-party federated aggregation node, calculate the encrypted parameters using a weighted aggregation algorithm, and generate a global federated model. S24. Establish a privacy protection mechanism simultaneously; for the raw data required for federated training, use differential privacy technology to process it, set a privacy budget according to the data sensitivity, add random noise to sensitive fields through the Laplace noise perturbation algorithm, and perform dual verification of security and availability on the perturbated data to generate a federated training dataset that meets security requirements.

4. The multi-level community data linkage analysis and processing method according to claim 3, characterized in that, S23 includes: S231. The trusted third-party federated aggregation node starts the parameter receiving interface, receives the encrypted local model parameter set sent by each participating institution according to the preset communication protocol, and performs integrity verification on the received parameter set to generate a list of verified encrypted local model parameter sets. S232. Based on the verified parameter set list, extract the basic data of each institution, calculate the weight of each institution according to the weighted aggregation algorithm formula; normalize the weights of all institutions and generate an institution aggregation weight allocation table. S233. Call the homomorphic encryption operation interface, perform weighted aggregation on the verified encrypted local model parameter set based on the institution aggregation weight allocation table, and then perform homomorphic addition on the operation results of all institutions to obtain the encrypted aggregation parameter set; decrypt the encrypted aggregation parameter set with the decryption private key to generate plaintext aggregation model parameters. S234. Import the plaintext aggregation model parameters into the preset model architecture template, automatically match the parameter configurations of the feature extraction layer and the decision layer, and generate the initial global federated model. S235. Use the reserved public test dataset to perform performance testing on the initial model. If the accuracy is ≥88%, proceed to the next step; otherwise, return to S232 to readjust the weight calculation method; finally, generate the global federated model.

5. The multi-level community data linkage analysis and processing method according to claim 1, characterized in that, The S3 includes: S31. Input the federated training dataset into the spatiotemporal causal graph in the multi-level data association framework, extract features from each level node, and generate the spatiotemporal causal graph node initialization feature set. S32. Construct a graph attention calculation model and set model parameters; perform dynamic weight calculation on the relationship between nodes based on the model; embed the weight coefficients into the node relationship attributes of the spatiotemporal causal graph to generate interpretive graph data containing weight coefficients; S33. Select an interpretable algorithm combination, determine the source tracing target, mine the upstream and downstream related nodes of the target node based on the interpretable graph data, select the key path according to the association weight, and organize the key path in a structured manner; organize the organization results in the form of text description + path diagram + impact intensity ranking table to generate a multi-level linkage impact path analysis report.

6. The multi-level community data linkage analysis and processing method according to claim 5, characterized in that, S31 includes: S311. Import the federated training dataset into the spatiotemporal causal graph of the multi-level data association framework; convert the dataset according to the graph data specification to generate a format-aligned federated training dataset. S312. Based on the aligned dataset, locate the target nodes at each level in the spatiotemporal causal graph; match the corresponding data records by the unique identifier of the node to form a mapping table of node-related data, and generate a data matching list for each level of node; S313. For each node in the matching list, extract attribute features; use Min-Max normalization and one-hot encoding to process the features and generate a subset of node attribute features; S314. Extract node association features from the historical records of the federated training dataset; quantize the features to generate a subset of node association features; S315. Merge the node attribute feature subset and the associated feature subset according to the node ID to form a complete feature vector of a single node; verify the feature dimension of each node, supplement the derived features for nodes with insufficient dimensions, and generate the spatiotemporal causal graph node initialization feature set.

7. The multi-level community data linkage analysis and processing method according to claim 6, characterized in that, The S315 includes: The node attribute feature subset generated by S313 and the node association feature subset generated by S314 are precisely matched and merged according to the unique ID of the node; the feature concatenation algorithm is used to combine the attribute feature vector and association feature vector of the same node into a complete feature vector of a single node, generating a merged single node feature vector set. Perform dimension verification on the merged single-node feature vector set and generate a list of node feature dimension verification results. For nodes to be supplemented in the verification result list, derived features are generated based on their existing features. The derived features are then added to the feature vector of the corresponding node to make the dimension ≥20, thus generating a complete single-node feature vector set. The integrated and improved single-node feature vector set is sorted by node level; feature description labels are added; and finally, the spatiotemporal causal graph node initialization feature set is generated.

8. The multi-level community data linkage analysis and processing method according to claim 1, characterized in that, The S4 includes: S41. Integrate core input data; design core modules of the model, construct the model architecture using system dynamics methods, and set inference parameters; train and validate the model using historical event data to generate a multi-level event inference model; S42. Based on the inference model, design three types of intervention strategies. For each strategy, simulate the changes in node state under the event through the model; calculate the key indicators of the strategy, and integrate the strategy name, risk level, scope of impact, treatment sequence and expected effect to generate a dataset of linkage decision suggestions. S43. Design a traceability information storage structure to record the causal derivation path of each decision suggestion in real time during the operation of the multi-level event deduction model; S44. Use consortium blockchain technology to distribute the traceability information and build a traceability information database for decision-making suggestions; develop traceability query function and complete the establishment of the traceability mechanism for decision-making suggestions.

9. A multi-level community data linkage analysis and processing system, comprising: One or more processors; Memory, used to store one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 8.

Images