Social governance command and dispatch system and method based on multi-source data fusion and intelligent control

Abstract

The application discloses a social governance command and dispatch system and method based on multi-source data fusion and intelligent control, and particularly relates to the technical field of social governance command and dispatch, and through the collection and cleaning of multi-source heterogeneous data, a dynamic evolution graph is modeled by using a graph neural network based on node embedding, potential abnormalities are captured in time by calculating a node dynamic correlation abnormality coefficient, high-quality event feature vectors are extracted by combining attention mechanism for cross-modal feature fusion, the stability in the data fusion process of different modal data is monitored by a modal feature fusion fluctuation coefficient, a data fusion risk index is output by comprehensively combining the node dynamic correlation abnormality coefficient and the modal feature fusion fluctuation coefficient, dynamic risk monitoring and self-adaptive iterative optimization in the data fusion stage are realized, an event evolution trend is predicted by using a time series modeling method, the social governance state is classified and identified, the social governance problem is modeled as a state transition process, and an optimal resource scheduling strategy is generated based on reinforcement learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of social governance command and dispatch, more specifically, the present application relates to a social governance command and dispatch system and method based on multi-source data fusion and intelligent control. BACKGROUND

[0002] Currently, with the acceleration of urbanization and the increasing complexity of social structure, the traditional social governance command and dispatch system faces many challenges, such as single data source, insufficient information timeliness, reliance on artificial experience in decision-making process, and lag in resource allocation response.

[0003] With the development of technologies such as Internet of Things (IoT), big data analysis, and artificial intelligence (AI), the ability to acquire and process multi-source heterogeneous data has significantly improved. However, existing command and dispatch systems have not achieved deep fusion of multi-source data and intelligent decision-making control, and lack effective monitoring of dynamic correlation analysis processes and cross-modal feature fusion processes for different source data, resulting in data fusion risks in the multi-source data fusion stage and difficulty in effectively responding to social governance needs in complex and dynamic environments.

[0004] Therefore, there is an urgent need for a social governance method that can fuse multi-source data, achieve intelligent control, and dynamically command and dispatch, in order to improve response speed, optimize resource allocation, and enhance the resilience and efficiency of the social governance system. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a social governance command and dispatch system and method based on multi-source data fusion and intelligent control to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] The social governance command and dispatch method based on multi-source data fusion and intelligent control comprises the following steps:

[0008] Step S1, collect multi-source heterogeneous data, including sensor data, video monitoring data, social media information, and geographic information data, and perform data cleaning;

[0009] Step S2, construct a dynamic evolution graph based on node embedding graph neural network for different source data, and analyze the abnormal situation of dynamic correlation between different nodes to calculate the node dynamic correlation abnormal coefficient;

[0010] Step S3, perform cross-modal feature fusion to extract event feature vectors based on attention mechanism, and analyze the fluctuation of different modal feature fusion to calculate the modal feature fusion fluctuation coefficient;

[0011] Step S4: Construct a data fusion risk assessment model based on the node dynamic association anomaly coefficient and modal feature fusion fluctuation coefficient, output the data fusion risk index, and iteratively optimize the data fusion process;

[0012] The data fusion risk assessment model is based on the following formula: sjr=ww1*DAE(t)+ww2*MTZ(t), where sjr is the data fusion risk index, DAE(t) is the node dynamic association anomaly coefficient calculated at time step t, MTZ(t) is the modal feature fusion fluctuation coefficient calculated at time step t, and ww1 and ww2 represent the preset proportional coefficients of the node dynamic association anomaly coefficient and the modal feature fusion fluctuation coefficient, respectively, and both ww1 and ww2 are greater than 0;

[0013] Step S5: Based on the iterative optimization results of the data fusion process, use time series modeling methods to predict the evolution trend of events and classify and identify the state of social governance.

[0014] Step S6: Based on the social governance status identification results, the social governance problem is transformed into a state transition problem of a dynamic system and the optimal resource scheduling decision is generated.

[0015] In a preferred embodiment, in step S2, a graph neural network based on node embeddings constructs a dynamic evolution graph from data from different sources, as follows:

[0016] Each type of data object is defined as a node in a dynamic evolution graph. Node embedding technology is used to map data of different modalities to the same latent space to obtain node embedding vectors. Edge weights are generated based on the similarity between node embedding vectors.

[0017] The edge weights are compared with a preset edge weight threshold. If the edge weight is greater than the threshold, an edge is created between node i and node j, resulting in the graph structure at time step t: GJ t =(V t E t ), where V t Let E be a set of nodes. t This is the set of edges after similarity filtering;

[0018] Establish a dynamic evolution mechanism to continuously evolve the graph structure over time step t:

[0019] Node data vectors are updated over time;

[0020] Update the node embedding vector;

[0021] Update the edge weights between nodes, and update the edge set;

[0022] The dynamic evolution diagram sequence is obtained through the above dynamic evolution mechanism.

[0023] In a preferred embodiment, the acquisition logic of the node dynamic association anomaly coefficient is as follows:

[0024] The edge weight of different time steps is acquired in the process of evolving the graph structure with time step t;

[0025] The association weight change rate of the computing node between the current time step and the last time step is calculated: bh ij (t) = w ij (t+1) - w ij (t), wherein bh ij (t) is the association weight change rate of node i and node j at time step t, w ij (t) is the edge weight of node i and node j at time step t, w ij (t+1) is the edge weight of node i and node j at time step t+1;

[0026] The local fluctuation standard deviation of the association weight change rate is calculated in a preset sliding window τ: Wherein σjb τ is the local fluctuation standard deviation of the association weight change rate in the sliding window τ, bh ij (k) is the association weight change rate of node i and node j at time step k, is the average value of the association weight change rate in the sliding window τ;

[0027] The node dynamic association anomaly factor is calculated: Wherein yc ij (t) is the node dynamic association anomaly factor, is a small positive number to prevent division by zero;

[0028] For each node i, the node dynamic association anomaly factors of all its neighbor nodes are summarized,

[0029] The node dynamic association anomaly coefficient is calculated: Wherein DAE(t) is the node dynamic association anomaly coefficient calculated at time step t, jh(i) is the neighbor set of node i, |jh(i)| is the total number of nodes in the neighbor set of node i, and I is the total number of nodes in the dynamically evolving graph.

[0030] In a preferred embodiment, in step S3, cross-modal feature fusion is performed based on an attention mechanism and an event feature vector is extracted, specifically as follows:

[0031] Feature extraction is performed on data of different sources to obtain modal feature representations;

[0032] To depict the dynamic changes of the contribution degree of different modalities to event features, a cross-modality attention mechanism is introduced to calculate the global importance weight of each modality, and a query vector, a key vector and a value vector are generated for each modality feature vector respectively;

[0033] The cross-modality attention score matrix is as follows: Wherein AZ m,m is the attention weight matrix of modality m to modality m', which represents the attention degree of modality m to modality m', and Softmax() is a normalization operation, CX m is the query vector of modality m, JZ m' T is the transpose of the key vector of modality m', and JL is the feature dimension of modality m;

[0034] The fusion attention weight of each modality is obtained by integrating all inter-modality attentions.

[0035] Based on the attention weight, the modalities are weighted and fused to form a unified event feature vector.

[0036] In a preferred embodiment, the acquisition logic of the modality feature fusion volatility coefficient is as follows:

[0037] The inter-modality synergy value is calculated for modality m and modality m' to quantify the mutual dependence degree of different modality features in time series evolution: Wherein IH(m,m') is the synergy value between modality m and modality m', P(xe m ,xe m ') is the joint probability distribution, P(xe m ) is the marginal probability distribution of modality m, P(xe m ') is the marginal probability distribution of modality m', F m and F m ' represent the feature value set of modality m and modality m' respectively;

[0038] The feature data matrix FF m (t) is fused with the synergy value: Wherein FFfusi(t) is the fused feature data matrix, and M is the total number of modalities;

[0039] Based on autoregression and heteroscedasticity, a dynamic volatility prediction model is constructed to calculate the volatility rate σfusi(t) of the fused feature data matrix, and the specific formula of the model is as follows: Wherein σfusi(t) 2 represents the volatility rate variance of the modality feature fusion at time step t, denotes the predicted value of the modal feature fusion at the past time step u, σfusi(t-v) 2 denotes the volatility variance of the modal feature fusion at the past time step v, c0is a constant term, b u is a weight coefficient of the autoregressive term, d v is a weight coefficient of the heteroscedastic term, p is the order of the autoregressive term, and q is the order of the heteroscedastic term.

[0040] The modal feature fusion volatility coefficient is calculated as: where MTZ(t) is the calculated modal feature fusion volatility coefficient at time step t, σfusi(t) is the volatility of the fused feature data matrix, σfusi m (t) is the volatility of modal m at time step t.

[0041] In a preferred embodiment, the data fusion risk index is compared with a preset data fusion risk index threshold, and the data fusion process is iteratively optimized, as follows:

[0042] If the data fusion risk index is greater than the data fusion risk index threshold, an iteration instruction is generated.

[0043] If the data fusion risk index is less than or equal to the data fusion risk index threshold, no iteration instruction is generated.

[0044] In a preferred embodiment, in step S5, the time series modeling method is used to predict the event evolution trend and classify the social governance state based on the iterative optimization result of the data fusion process, as follows:

[0045] The event feature vector is serialized in chronological order, and the serialized event feature vector is used as input; the long short-term memory network is used to learn the dynamic pattern of the event feature evolution over time, and the evolution trend sequence at future time is predicted; a multi-class label system of social governance state is preset; the prediction probability of each class is calculated by the classifier; and the current social governance state is determined according to the maximum probability principle.

[0046] In a preferred embodiment, in step S6, the social governance problem is converted into a state transition problem of a dynamic system and an optimal resource scheduling decision is generated according to the social governance state recognition result, as follows:

[0047] The social governance process is modeled as a Markov decision process, and the following definitions are made: state set G: composed of social governance recognition results; action set F: including emergency scheduling, human resource allocation, material resource allocation, etc.; reward function R(g, f): depicting the immediate income obtained by performing action f in state g;

[0048] The objective function of maximizing long-term cumulative income is constructed as: where CL * is the optimal policy, gamma is a discount factor that controls the importance of future rewards, mean(·) represents the operation of taking the average, g t is the state at time step t, f t is the action at time step t

[0049] The reinforcement learning method is used to learn the mapping strategy of state to action and output the optimal strategy.

[0050] In a preferred embodiment, the social governance command and dispatch system based on multi-source data fusion and intelligent control comprises a heterogeneous data acquisition module, a dynamic evolution correlation module, a modal feature fusion module, an iterative analysis module, an evolution trend classification module, and a resource scheduling module.

[0051] The heterogeneous data acquisition module is used to acquire multi-source heterogeneous data, including sensor data, video monitoring data, social media information, and geographic information data, and perform data cleaning.

[0052] The dynamic evolution correlation module is used to construct a dynamic evolution graph based on a node embedding graph neural network, analyze the abnormal situation of dynamic correlation between different nodes, and calculate the node dynamic correlation abnormal coefficient.

[0053] The modal feature fusion module is used to perform cross-modal feature fusion to extract an event feature vector based on an attention mechanism, analyze the fluctuation of different modal feature fusion, and calculate the modal feature fusion fluctuation coefficient.

[0054] The iterative analysis module is used to construct a data fusion hidden danger evaluation model based on the node dynamic correlation abnormal coefficient and the modal feature fusion fluctuation coefficient, output a data fusion risk index, and iteratively optimize the data fusion process.

[0055] The evolution trend classification module is used to predict the event evolution trend based on the iterative optimization result of the data fusion process using a time series modeling method, and classify and identify the social governance state.

[0056] The resource scheduling module is used to convert the social governance problem into a state transition problem of a dynamic system based on the social governance state identification result, and generate an optimal resource scheduling decision.

[0057] The technical effects and advantages of the present application are as follows:

[0058] 1、The application integrates sensor, video monitoring, social media and geographic information multidimensional information to ensure the timeliness and comprehensiveness of information, uses a graph neural network based on node embedding to model a dynamic evolution graph, and through the calculation of a node dynamic correlation anomaly coefficient, accurately monitors the changes of dynamic correlation between different data nodes, timely captures potential anomalies, simultaneously, combines an attention mechanism for cross-modal feature fusion, extracts a high-quality event feature vector, and through a modal feature fusion fluctuation coefficient, monitors the stability in the fusion process of different modal data, further guarantees the reliability and consistency of the data fusion result, comprehensively combines the node dynamic correlation anomaly coefficient and the modal feature fusion fluctuation coefficient, outputs a data fusion risk index, realizes dynamic risk monitoring and adaptive iterative optimization in the data fusion stage, effectively avoids subsequent decision deviation caused by low fusion quality, uses a time series modeling method to predict the event evolution trend by taking the optimized event feature vector as input, and based on the trend prediction result, classifies and identifies the social governance state, greatly improves the pertinence and timeliness of event response, simultaneously models the social governance problem as a state transition process, and based on reinforcement learning, generates an optimal resource scheduling strategy, realizes the optimal allocation of manpower, materials, communication resources and other multi-factors in a dynamic environment, and guarantees the maximization of resource utilization and the intelligentization of scheduling decisions in the emergency response process. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to facilitate the understanding of those skilled in the art, the present application will be further described below with reference to the accompanying drawings;

[0060] Figure 1 is a flowchart of the method of embodiment 1 of the present application.

[0061] Figure 2 is a flowchart of the system of embodiment 2 of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0063] Embodiment 1: Figure 1 The present application provides a social governance command and scheduling method based on multi-source data fusion and intelligent control, including the following steps:

[0064] Step S1, collect multi-source heterogeneous data, including sensor data, video monitoring data, social media information and geographic information data, and perform data cleaning;

[0065] Step S2: Based on the node embedding graph neural network, a dynamic evolution graph is constructed from data from different sources, and the abnormal situation of dynamic association between different nodes is analyzed and the node dynamic association anomaly coefficient is calculated.

[0066] Step S3: Based on the attention mechanism, perform cross-modal feature fusion to extract event feature vectors, and analyze the fluctuation of different modal feature fusion to calculate the modal feature fusion fluctuation coefficient;

[0067] Step S4: Construct a data fusion risk assessment model based on the node dynamic association anomaly coefficient and modal feature fusion fluctuation coefficient, output the data fusion risk index, and iteratively optimize the data fusion process;

[0068] Step S5: Based on the iterative optimization results of the data fusion process, use time series modeling methods to predict the evolution trend of events and classify and identify the state of social governance.

[0069] Step S6: Based on the social governance status identification results, the social governance problem is transformed into a state transition problem of a dynamic system and the optimal resource scheduling decision is generated;

[0070] In step S1, multi-source heterogeneous data is collected, including sensor data, video surveillance data, social media information and geographic information data, and the data is cleaned.

[0071] Data cleaning steps include:

[0072] Outlier removal: Outliers are removed using a statistical distribution-based method. The outlier criterion is: |x-μ|>σ, where x is a single data point, μ is the sample mean, and σ is the sample standard deviation.

[0073] Missing value imputation: Imputation is performed using nearest neighbor interpolation (KNN) or local mean.

[0074] Standardization process: The Z-Score standardization method is used to unify the dimensions;

[0075] In step S2, a graph neural network based on node embeddings constructs a dynamic evolution graph from data from different sources, as follows:

[0076] Each type of data object is defined as a node in a dynamic evolution graph. Node embedding techniques are used to map data from different modalities to the same latent space to obtain node embedding vectors. Among them, jd i Let be the node embedding vector, σ be the activation function (e.g., ReLU activation function), W be the learnable weight matrix, and jh(i) represent the set of neighbors of node i. Let i be the original data vector of node i;

[0077] generating an edge weight according to the similarity between the node embedding vectors: where w ij (t) is an edge weight of node i and node j at time step t, jd i (t) is a node embedding vector of node i at time step t, jd j (t) is a node embedding vector of node j at time step t;

[0078] comparing the edge weight with a preset edge weight threshold, if the edge weight is greater than the edge weight threshold, an edge is established between node i and node j, and a graph structure at time step t is obtained: GJ t = (V t , E t ), where V t is a node set, and E t is an edge set after similarity screening;

[0079] establishing a dynamic evolution mechanism to continuously evolve the graph structure with time step t:

[0080] updating the node data vector with time

[0081] updating the node embedding vector jd i (t) to jd i (t+1);

[0082] updating the edge weight w ij (t) to w ij (t+1), and updating the edge set E t to E t+1 ;

[0083] obtaining a dynamic evolution graph sequence GJ t = {GJ1, GJ2,..., GJ T} through the above dynamic evolution mechanism, where t = {1, 2,..., T}, and T is a positive integer;

[0084] analyzing and calculating a node dynamic association anomaly coefficient for abnormal situations of dynamic association between different nodes, specifically as follows:

[0085] In this invention, the node dynamic association anomaly coefficient is an important indicator used to measure the degree of abnormal changes in the association relationships between different nodes during dynamic evolution. Specifically, as multi-source heterogeneous data (such as sensor data, video surveillance data, social media information, geographic information, etc.) in the social governance environment continues to evolve, the association relationships between different data nodes also change over time. By dynamically tracking and quantifying the changes in association weights between nodes in continuous time steps, the node dynamic association anomaly coefficient can keenly capture abnormal association patterns between node pairs caused by external disturbances, data distortion, abnormal events, or structural mutations within a specific time period. When the node dynamic association anomaly coefficient is large, it indicates that the association relationships between nodes have changed drastically in a short period of time, which may mean potential system anomalies, information conflicts during data fusion, data source failure, or external event interference, reflecting a certain degree of instability and hidden risks in the data fusion process. Conversely, when the node dynamic association anomaly coefficient is small, it indicates that the evolution of the association relationships between nodes is stable, and the trend of change conforms to the normal evolution logic, indicating that the data fusion process is relatively coherent, the data sources are coordinated and consistent, and the system operates stably and reliably. Iterative optimization of the data fusion process based on node dynamic association anomaly coefficients has significant technical value and beneficial effects. Firstly, by continuously monitoring the changing trajectory of node dynamic association anomaly coefficients, hidden anomaly evolution trends during data fusion can be identified in a timely manner, preventing the overall fusion effect from deteriorating due to local data distortion, single-point failures, or cross-modal heterogeneous conflicts. Furthermore, the introduction of node dynamic association anomaly coefficients enables early warning and risk classification of potential problems in the fusion process, improving resource allocation efficiency and reducing system maintenance costs. Through these mechanisms, this invention can achieve intelligent perception and iterative improvement of the health status of the fusion process in multi-source data fusion and social governance command and dispatch systems, effectively enhancing the overall resilience, stability, and intelligence level of the system, and improving the accuracy and reliability of social governance responses in complex environments. It has significant application and promotion value and broad practical application prospects.

[0086] The logic for obtaining the dynamic association anomaly coefficient of nodes is as follows:

[0087] Obtain edge weights at different time steps as the graph structure evolves with time step t.

[0088] Calculate the rate of change of association weights of nodes between the current time step and the previous time step: bh ij (t)=w ij (t+1)-w ij (t), where bh ij (t) represents the rate of change of the association weights between nodes i and j at time step t, w ij (t) represents the edge weights between nodes i and j at time step t, w ij(t+1) is the edge weight of node i and node j at time step t+1;

[0089] The local fluctuation standard deviation of the correlation weight change rate is calculated in the preset sliding window τ: where σjb τ is the local fluctuation standard deviation of the correlation weight change rate in the sliding window τ, bh ij (k) is the correlation weight change rate of node i and node j at time step k, is the average value of the correlation weight change rate in the sliding window τ;

[0090] The node dynamic correlation anomaly factor is calculated: where yc ij (t) is the node dynamic correlation anomaly factor, is a small positive number to prevent division by zero;

[0091] For each node i, the node dynamic correlation anomaly factors of all its neighbor nodes are summarized,

[0092] The node dynamic correlation anomaly coefficient is calculated: where DAE(t) is the node dynamic correlation anomaly coefficient calculated at time step t, jh(i) is the neighbor set of node i, |jh(i)| is the total number of nodes in the neighbor set of node i, and I is the total number of nodes in the dynamic evolving graph.

[0093] In the dynamic evolving graph, each node and edge itself may carry multi-modal features (such as text description, sensor value, image segment, geographic coordinates, etc.). To further optimize the data fusion process, in step S3, cross-modal feature fusion is performed based on the attention mechanism and event feature vectors are extracted, as follows:

[0094] Feature extraction is performed on data from different sources to obtain modal feature representations: where FF m (t) is the feature data matrix collected by modal m at time step t, is a data matrix with JH rows and JL columns, and specifically JH m is the number of data samples of modal m (such as the number of sensor readings, the number of video frames, and the number of social media messages captured), and JL m is the feature dimension of modal m, i.e., how many features are used to describe each data sample (such as converting each message into a 300-dimensional word vector and extracting 1024-dimensional image features from each video frame);

[0095] To depict the dynamic changes of the contribution degree of different modalities to event features, a cross-modal attention mechanism is introduced to calculate the global importance weight of each modality, and a query vector CX m , a key vector JZ m , and a value vector ZX m are generated for each modality feature vector, respectively: CX m = W CX FF m (t), JZ m = W JZ FF m (t), ZX m = W ZX FF m (t), where W CX , W JZ , and W ZX represent the trainable weight matrices of the query vector, the key vector, and the value vector, respectively.

[0096] The cross-modal attention score matrix is: where AZ m,m ' is the attention weight matrix of modality m to modality m', representing the attention degree of modality m to modality m', Softmax() is a normalization operation, CX m is the query vector of modality m, JZ m' T is the transpose of the key vector of modality m'; JL is the feature dimension of modality m.

[0097] The fusion attention weight α m of each modality is obtained by integrating all inter-modal attentions: where M is the total number of modalities, and mean(AZ m,m ') represents the average value of all values in the attention weight matrix.

[0098] The features of each modality are weighted and fused based on the attention weight to form a unified event feature vector SJ(t): where SJ(t) represents the event feature vector, and MeanPool(FF m (t)) represents the average value of the feature of each data sample in modality m.

[0099] The fluctuation of the fusion of different modalities is analyzed to calculate the modality feature fusion fluctuation coefficient, as follows:

[0100] In the present application, the modal feature fusion fluctuation coefficient is used to measure the fluctuation degree of different modalities in the feature fusion process. This coefficient calculates the fluctuation amplitude of different modal features in the whole feature fusion process, thereby evaluating whether there is a large fluctuation or instability in the fusion process. A larger modal feature fusion fluctuation coefficient indicates that there is a large difference or fluctuation in the feature contribution between different modalities in the data fusion process. This situation may cause some modalities to have too much influence in a particular time or situation, thereby affecting the stability and decision accuracy of the system. A larger modal feature fusion fluctuation coefficient means that the influence of some modalities on the final decision in the feature fusion process may be unstable, resulting in a large performance fluctuation of the system when facing new or changing input data, thereby reducing the robustness and reliability of the overall system. On the contrary, a smaller modal feature fusion fluctuation coefficient indicates that the features between modalities tend to be stable in the fusion process, the contribution of each modality is more balanced, and the fused features are less affected by a single modality. At this time, the stability of the system is better, the synergy between different modal features is more sufficient, and the overall feature fusion effect is smoother, avoiding the abnormal fluctuation of a certain modality feature from having too much influence on the overall decision result. A smaller modal feature fusion fluctuation coefficient can ensure the stable operation of the system in different situations, and enhance the accuracy and robustness of the system when processing multi-source heterogeneous data. The demand for iterative optimization of the data fusion process based on the modal feature fusion fluctuation coefficient evaluation has significant beneficial effects. First, the modal feature fusion fluctuation coefficient as a quantitative index can help the system identify potential problems in the fusion of different modalities. By monitoring the modal feature fusion fluctuation coefficient, the system can timely discover the abnormal phenomenon of excessive contribution of a certain modality, and the introduction of the modal feature fusion fluctuation coefficient can also effectively promote the adaptive ability of the system. Through the calculation of the modal feature fusion fluctuation coefficient, the system not only can improve the performance of the current data fusion, but also can have stronger adaptability and robustness when facing future complex situations, thereby providing more accurate and stable support in the social governance command and dispatch system. It can effectively reduce the uncertainty caused by data fluctuation, improve the intelligent level of the system, and ensure the accurate fusion and effective utilization of multi-source data in dynamic environment.

[0101] The acquisition logic of the modal feature fusion fluctuation coefficient is as follows:

[0102] The inter-modal synergy degree value is calculated for modal m and modal m' to quantify the mutual dependence degree of different modal features in time evolution: where IH(m, m') is the synergy degree value between modal m and modal m', P(xe m , xe m ') is the joint probability distribution (indicating the feature value xe mand xe m The probability of simultaneous occurrence), P(xe m The marginal probability distribution of modal m (indicating the probability of the occurrence of feature value xe m The marginal probability distribution of modal m' (indicating the probability of the occurrence of feature value xe m The marginal probability distribution of modal m' (indicating the probability of the occurrence of feature value xe m F m and F m respectively represent the feature value set of modal m and modal m';

[0103] Fuse the feature data matrix FF m (t) with the synergy value: Where FFfusi(t) is the fused feature data matrix, and M is the total number of modes;

[0104] Based on autoregression and heteroscedasticity, a dynamic volatility prediction model is constructed to calculate the volatility σfusi(t) of the fused feature data matrix, and the specific formula of the model is as follows: Where σfusi(t) 2 represents the volatility variance of modal feature fusion at time step t, represents the predicted value of modal feature fusion at past time step u, σfusi(t-v) 2 represents the volatility variance of modal feature fusion at past time step v, c0 is a constant term (used to adjust the base level of the dynamic volatility prediction model, which is set according to actual conditions), b u is the weight coefficient of the autoregressive term (indicating the influence weight of past predicted values on current volatility), d v is the weight coefficient of the heteroscedasticity term (indicating the influence weight of past volatility variance on current volatility), p is the order of the autoregressive term (indicating the number of past predicted values considered by the dynamic volatility prediction model), and q is the order of the heteroscedasticity term (indicating the number of past volatility variances considered by the dynamic volatility prediction model);

[0105] Calculate the modal feature fusion volatility coefficient: Where MTZ(t) is the modal feature fusion volatility coefficient calculated at time step t, σfusi(t) is the volatility of the fused feature data matrix, and σfusi m (t) is the volatility of modal m at time step t;

[0106] In step S4, a data fusion risk assessment model is constructed according to the node dynamic correlation anomaly coefficient and the modal feature fusion volatility coefficient, and a data fusion risk index is output, and the data fusion process is iteratively optimized, as follows:

[0107] According to the node dynamic correlation anomaly coefficient, the modal feature fusion fluctuation coefficient, a data fusion hidden danger evaluation model is constructed, and a data fusion risk index is output. The formula on which the data fusion hidden danger evaluation model is based is as follows: sjr=ww1*DAE(t)+ww2*MTZ(t), wherein sjr is the data fusion risk index, DAE(t) is the node dynamic correlation anomaly coefficient calculated at time step t, MTZ(t) is the modal feature fusion fluctuation coefficient calculated at time step t, ww1 and ww2 respectively represent preset proportion coefficients of the node dynamic correlation anomaly coefficient and the modal feature fusion fluctuation coefficient, and ww1 and ww2 are both greater than 0;

[0108] It should be noted that the above formulas are all dimensionless numerical calculations. Common dimension removal methods include Min-Max normalization and Z-Score standardization, which are not described here. ww1 and ww2 are set according to actual conditions. For example, an expert weighting method is used, that is, experts in the relevant field are invited to determine the preset proportion coefficients of each index through professional opinion surveys and comprehensive evaluations. For example, ww1 and ww2 can be 0.5 and 0.5.

[0109] As can be seen from the above calculation expression, the greater the node dynamic correlation anomaly coefficient and the greater the modal feature fusion fluctuation coefficient, the greater the data fusion risk index, indicating that the data fusion process in the system has a high risk and instability. Conversely, the smaller the node dynamic correlation anomaly coefficient and the smaller the modal feature fusion fluctuation coefficient, the smaller the data fusion risk index, indicating that the hidden danger in the data fusion process is smaller and the result of data fusion is more reliable.

[0110] The data fusion risk index is compared with a preset data fusion risk index threshold, and the data fusion process is iteratively optimized as follows:

[0111] If the data fusion risk index is greater than the data fusion risk index threshold, it indicates that there is abnormal fluctuation or unstable correlation in the current data fusion process, and the fusion quality of the system is insufficient to support the accurate identification and decision-making generation of subsequent social governance events. It is necessary to timely adjust and optimize the fusion strategy and generate an iteration instruction.

[0112] It should be noted that when the iteration instruction is generated, the running logic of steps S2, S3 and S4 needs to be re-executed, that is, the dynamic evolution graph is reconstructed -> the cross-modal feature fusion is extracted -> the event feature vector is constructed -> it is judged whether an iteration instruction needs to be generated.

[0113] If the data fusion risk index is less than or equal to the data fusion risk index threshold value, it indicates that the dynamic association relationship between nodes in the current data fusion process is relatively stable, the fluctuation degree of different modal feature fusion is within an acceptable range, the overall data fusion quality is good, the consistency and coordination between data sources and the fusion effect meet the expected standard, and there is no need to generate iteration instructions, which shows that after the cleaning, modeling and fusion processing of multi-source heterogeneous data, the event feature vector formed has high representation accuracy and robustness, and can provide reliable data support for subsequent social governance state identification, event evolution trend prediction and resource scheduling decision-making;

[0114] In step S5, the time series modeling method is used to predict the event evolution trend and classify and identify the social governance state based on the iteration optimization result of the data fusion process, as follows:

[0115] The event feature vector is serialized in time sequence, and the serialized event feature vector {SJ(t-n),...,SJ(t)} is taken as input;

[0116] The long short-term memory network (such as LSTM) is used to learn the dynamic pattern of event feature evolution over time, and the evolution trend sequence of future time is predicted: {y t+1 ,y t+2 ,...,y t+r}, wherein y t+r is the prediction index of the t+rth time step, and the prediction index includes but is not limited to event size, risk level, evolution speed and propagation range;

[0117] A preset multi-class label system of social governance state is TX={normal, warning, emergency};

[0118] The prediction probability PG c (t) of each class is calculated by a classifier (such as Softmax classifier, MLP classifier), wherein c={1,2,3}, PG1(t) is the prediction probability of the normal class, PG2(t) is the prediction probability of the warning class, and PG3(t) is the prediction probability of the emergency class;

[0119] The current social governance state is determined according to the maximum probability principle: wherein State(t) is the social governance state corresponding to the time step t;

[0120] In step S6, the social governance problem is converted into a state transition problem of a dynamic system according to the social governance state identification result, and an optimal resource scheduling decision is generated, as follows:

[0121] Model the social governance process as a Markov decision process, define: state set G: composed of social governance identification results; action set F: including emergency dispatch, human resource allocation, material resource allocation, etc.; reward function R(g, f): depicting the immediate income obtained by executing action f under state g;

[0122] Construct a target function to maximize long-term cumulative income: Where CL * is the optimal strategy, γ is the discount factor (γ ∈ (0, 1)), which controls the importance of future rewards, mean(·) represents the average value operation, g t is the state at time step t, f t is the action at time step t

[0123] Adopt reinforcement learning methods (such as Deep Q-Learning or Proximal Policy Optimization (PPO)) to learn the mapping strategy from state to action and output the optimal strategy CL * ;

[0124] The present application collects and cleans multi-source heterogeneous data, comprehensively integrates sensor, video monitoring, social media and geographic information multi-dimensional information to ensure the timeliness and comprehensiveness of the information, models the dynamic evolution graph based on node embedding graph neural network, and realizes the accurate monitoring of the change of dynamic correlation between different data nodes through the calculation of node dynamic correlation anomaly coefficient, timely captures potential anomalies, at the same time, combines the attention mechanism to perform cross-modal feature fusion, extracts high-quality event feature vectors, and monitors the stability in the fusion process of different modal data through modal feature fusion volatility coefficient, further guarantees the reliability and consistency of the data fusion result, integrates the node dynamic correlation anomaly coefficient and the modal feature fusion volatility coefficient, outputs the data fusion risk index, realizes the dynamic risk monitoring and adaptive iterative optimization in the data fusion stage, effectively avoids the subsequent decision deviation caused by low fusion quality, inputs the optimized event feature vector, adopts time series modeling method to predict the event evolution trend, and classifies the social governance state based on the trend prediction result, greatly improves the pertinence and timeliness of event response, at the same time, models the social governance problem as a state transition process, and generates an optimal resource scheduling strategy based on reinforcement learning, realizes the optimal allocation of human, material, communication resources and other multi-factors in a dynamic environment, and guarantees the maximization of resource utilization and the intelligentization of scheduling decision in the emergency response process.

[0125] Embodiment 2: This embodiment is an introduction to the social governance command and dispatch system based on multi-source data fusion and intelligent control, as shown in Figure 2As shown, it includes a heterogeneous data collection module, a dynamic evolution correlation module, a modal feature fusion module, an iterative analysis module, an evolution trend classification module, and a resource scheduling module.

[0126] The heterogeneous data collection module is used to collect multi-source heterogeneous data, including sensor data, video monitoring data, social media information and geographic information data, and perform data cleaning.

[0127] The dynamic evolution correlation module is used to construct a dynamic evolution graph based on node-embedded graph neural networks from different source data, and analyze the abnormal situation of dynamic correlation between different nodes to calculate the node dynamic correlation abnormal coefficient.

[0128] The modal feature fusion module is used to perform cross-modal feature fusion to extract event feature vectors based on an attention mechanism, and analyze the fluctuation of different modal feature fusion to calculate the modal feature fusion fluctuation coefficient.

[0129] The iterative analysis module is used to construct a data fusion hidden danger evaluation model according to the node dynamic correlation abnormal coefficient and the modal feature fusion fluctuation coefficient, output a data fusion risk index, and iteratively optimize the data fusion process.

[0130] The evolution trend classification module is used to predict the event evolution trend based on the iterative optimization results of the data fusion process using a time series modeling method, and classify and identify the social governance state.

[0131] The resource scheduling module is used to convert social governance problems into state transition problems of a dynamic system based on the social governance state identification results and generate optimal resource scheduling decisions.

[0132] The above formulas are all dimensionless values, and the formulas are obtained by software simulation of a large amount of data to obtain a formula for the current real situation. The preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0133] The above-described embodiments can be implemented in part or in whole through software, hardware, firmware or any combination thereof. When implemented in software, the above-described embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When loaded and executed by a computer, the computer instructions or computer programs cause the computer to perform the processes or functions described in the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center through a wired or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. containing a set of one or more available media. The available media can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0134] It should be understood that, in various embodiments of the present application, the size of the sequence number of each process described above does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0135] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and method described above can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0136] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented by other ways.

[0137] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A social governance command and dispatch method based on multi-source data fusion and intelligent control, characterized in that: Comprise the following steps: Step S1, collect multi-source heterogeneous data, including sensor data, video monitoring data, social media information and geographic information data, and perform data cleaning; Step S2, based on node embedding graph neural network, different source data is constructed into dynamic evolution graph, and the abnormal situation of dynamic association between different nodes is analyzed and calculated node dynamic association abnormal coefficient; Step S3, based on attention mechanism, cross-modal feature fusion is carried out to extract event feature vector, and the fluctuation of different modal feature fusion is analyzed to calculate modal feature fusion fluctuation coefficient; Step S4, according to node dynamic association abnormal coefficient, modal feature fusion fluctuation coefficient, data fusion hidden danger evaluation model is constructed, data fusion risk index is output, data fusion process is iteratively optimized; The formula on which the data fusion hidden danger assessment model is based is as follows: , wherein is a data fusion risk index, is a node dynamic correlation anomaly coefficient calculated at a time step t, is a modal feature fusion fluctuation coefficient calculated at the time step t, respectively represent preset proportion coefficients of the node dynamic correlation anomaly coefficient and the modal feature fusion fluctuation coefficient, and are both greater than 0. Step S5, based on the iterative optimization result of data fusion process, the time series modeling method is used to predict the event evolution trend, and the social governance state is classified and identified; Step S6, according to the social governance state identification result, the social governance problem is transformed into the state transition problem of dynamic system and the optimal resource scheduling decision is generated; The logic of obtaining node dynamic association abnormal coefficient is as follows: Obtain the edge weight of different time steps in the process of changing the graph structure with time step t; the rate of change of the association weight of the computing node between the current time step and the previous time step: wherein is the rate of change of the association weight of nodes i and j at time step t, is the edge weight of nodes i and j at time step t, is the edge weight of nodes i and j at time step t+1; In the preset sliding window Calculate the local fluctuation standard deviation of the rate of change of associated weights: ,in For the rate of change of associated weights in the sliding window Local fluctuation standard deviation, Let k be the rate of change of the association weights between nodes i and j at time step k. For the rate of change of associated weights in the sliding window The average value in; a node dynamic correlation anomaly factor: wherein is the node dynamic correlation anomaly factor, is a small positive number to prevent division by zero; For each node summarize the node dynamic correlation anomaly factor of all its neighbor nodes, a node dynamic correlation anomaly coefficient is calculated: wherein is the node dynamic correlation anomaly coefficient calculated at time step t, is the neighbor set of node i, is the total number of nodes in the neighbor set of node i, is the total number of nodes in the dynamic evolving graph; The logic of obtaining modal feature fusion fluctuation coefficient is as follows: For modality and modality The synergy value between modalities is calculated to quantify the mutual dependence of different modal features in the time evolution: where is the synergy value between modalities and modalities , is the joint probability distribution, is the marginal probability distribution of modalities , is the marginal probability distribution of modalities , and represent the feature value set of modalities and modalities , respectively. The feature data matrix is fused with the coordination value to obtain a fused feature data matrix wherein is the fused feature data matrix, is the total number of modes​ Based on autoregression and heteroscedasticity, a dynamic volatility prediction model is constructed, and the volatility of the fused feature data matrix is calculated The specific formula of the model is as follows: Wherein represents the volatility variance of the modal feature fusion at the time step t, represents the predicted value of the modal feature fusion at the past time step , represents the volatility variance of the modal feature fusion at the past time step , is a constant term, is a weight coefficient of the autoregressive term, is a weight coefficient of the heteroscedasticity term, is the order of the autoregressive term, is the order of the heteroscedasticity term; a computed modal feature fusion volatility coefficient: wherein is a computed modal feature fusion volatility coefficient at time step t, is a volatility of the fused feature data matrix, is a volatility of the modal m at time step t. 2.The social governance command and dispatch method based on multi-source data fusion and intelligent control according to claim 1, characterized in that: In step S2, based on node embedding graph neural network, different source data is constructed into dynamic evolution graph, as follows: each type of data object is defined as a node in a dynamic evolution graph, node embedding technology is used to map data of different modalities to the same latent space to obtain node embedding vectors, and edge weights are generated according to the similarity between the node embedding vectors; the edge weight is compared with a preset edge weight threshold, and if the edge weight is greater than the edge weight threshold, an edge is established between the node i and the node j, and a graph structure at the time step t is obtained: wherein is a node set, is an edge set screened according to the similarity; a dynamic evolution mechanism is established, and the graph structure is continuously evolved with the change of the time step t: the node data vector is updated over time; the node embedding vector is updated; the edge weight between the nodes is updated, and the edge set is updated; After the above dynamic evolution mechanism, the dynamic evolution graph sequence is obtained. 3.The social governance command and dispatch method based on multi-source data fusion and intelligent control according to claim 1, characterized in that: In step S3, based on attention mechanism, cross-modal feature fusion is carried out to extract event feature vector, as follows: Feature extraction is performed on data from different sources to obtain modal feature representation; In order to depict the dynamic change of the contribution of different modalities to event features, cross-modal attention mechanism is introduced to calculate the global importance weight of each modality, and query vector, key vector and value vector are generated for each modality feature vector; The cross-modal attention score matrix is: wherein is the attention weight matrix of the modality to the modality , indicating the attention degree of the modality to the modality , is a normalization operation, is the query vector of the modality m, is the transpose of the key vector of the modality ; is the feature dimension of the modality m; Integrate all inter-modal attention to get fusion attention weight of each modality; Based on the attention weight, the modal features are weighted and fused to form a unified event feature vector.

4. The social governance command and dispatch method based on multi-source data fusion and intelligent control according to claim 1, characterized in that: Compare the data fusion risk index with the preset data fusion risk index threshold to iteratively optimize the data fusion process, as follows: If the data fusion risk index is greater than the data fusion risk index threshold, iteration instruction is generated; If the data fusion risk index is less than or equal to the data fusion risk index threshold, iteration instruction is not needed. 5.The social governance command and dispatch method based on multi-source data fusion and intelligent control according to claim 3, characterized in that: In step S5, based on the iterative optimization result of data fusion process, the time series modeling method is used to predict the event evolution trend, and the social governance state is classified and identified, as follows: The event feature vector is sequenced in time sequence, and the sequenced event feature vector is taken as input; long short-term memory network is used to learn the dynamic pattern of event feature evolution over time, and the evolution trend sequence of future time is predicted; preset multi-class label system of social governance state; the prediction probability of each class is calculated through classifier; According to the maximum probability principle, the current social governance state is determined. 6.The social governance command and dispatch method based on multi-source data fusion and intelligent control according to claim 5, characterized in that: In step S6, the social governance problem is transformed into a state transition problem of a dynamic system according to the social governance state recognition result, and an optimal resource scheduling decision is generated, as follows: Model the social governance process as a Markov decision process, define: state set : composed of social governance identification results; action set : including emergency dispatch, human resource allocation, material resource allocation; reward function : depicts the immediate income obtained by performing action f in state g; The objective function is constructed to maximize the long-term cumulative reward: where is the optimal policy, is the discount factor, controlling the importance of future rewards, denotes the averaging operation, is the state at time step t, is the action at time step t; The reinforcement learning method is used to learn the mapping strategy from state to action and output the optimal strategy.

7. A social governance command and dispatch system based on multi-source data fusion and intelligent control, used to implement the social governance command and dispatch method based on multi-source data fusion and intelligent control in any one of claims 1-6, characterized in that: It includes a heterogeneous data acquisition module, a dynamic evolution correlation module, a modal feature fusion module, an iterative analysis module, an evolution trend classification module, and a resource scheduling module. The heterogeneous data acquisition module is used to collect multi-source heterogeneous data, including sensor data, video monitoring data, social media information, and geographic information data, and perform data cleaning. The dynamic evolution correlation module is used to construct a dynamic evolution graph based on node embedding graph neural network, and analyze the abnormal situation of dynamic correlation between different nodes to calculate the node dynamic correlation abnormal coefficient. The modal feature fusion module is used to perform cross-modal feature fusion based on the attention mechanism to extract event feature vectors, and analyze the fluctuation of different modal feature fusion to calculate the modal feature fusion fluctuation coefficient. The iterative analysis module is used to construct a data fusion hidden danger evaluation model based on the node dynamic correlation abnormal coefficient and the modal feature fusion fluctuation coefficient, output a data fusion risk index, and iteratively optimize the data fusion process. The evolution trend classification module is used to predict the event evolution trend based on the iterative optimization result of the data fusion process using the time series modeling method, and classify and identify the social governance state. The resource scheduling module is used to transform the social governance problem into a state transition problem of a dynamic system according to the social governance state recognition result, and generate an optimal resource scheduling decision.

Images