A smart police multi-service coordination integrated management system

Abstract

The present application relates to the technical field of wisdom police, and discloses a wisdom police multi-service coordination integrated management system, which comprises a multi-modal perception module, an entity extraction module, a space-time situation prediction module and a coordination scheduling module.The multi-modal perception module performs space-time alignment and cross-modal fusion on police event and dynamic police resource data, and then performs structured updating; the entity extraction module extracts police entities to dynamically update a global police knowledge graph, and intercepts a graph topological substructure of the current police event; the space-time situation prediction module aggregates the spatial features of the substructure and the time sequence features of the historical sequence, and outputs a multi-modal prediction result; the coordination scheduling module maps the prediction result and the resource data into a resource demand state space, uses a multi-agent coordination mechanism and a global optimal allocation strategy to adaptively configure police forces and logistical materials, and generates scheduling instructions.The present application realizes full-link coordination of deep mining of police data and intelligent scheduling of resources.

Description

Technical Field

[0001] This invention relates to the field of smart policing technology, and in particular to a smart policing multi-service collaborative integrated management system. Background Technology

[0002] With the rapid development of smart cities, public security policing systems have integrated massive amounts of heterogeneous data, including video surveillance, emergency call texts, IoT sensor data, and police resource status. However, most existing policing management systems remain at the level of simple data aggregation and visualization, exhibiting the following technical shortcomings:

[0003] First, the system lacks deep semantic fusion capabilities across modal data, often processing text and video data in isolation, failing to effectively extract entity-related clues from complex incidents. Second, traditional situational awareness relies primarily on static threshold alarms or single time-series models, making it difficult to provide accurate spatiotemporal evolution warnings for cross-regional emergencies. Finally, the coordinated scheduling of cross-departmental (such as traffic police and patrol police) and logistical supplies heavily depends on human experience and static contingency plans, lacking an adaptive optimization configuration mechanism for dynamically evolving environments. These deficiencies result in slow cross-departmental collaborative responses and a disconnect between logistical scheduling and frontline needs when facing complex emergencies, failing to meet the closed-loop management requirements of modern policing operations. Summary of the Invention

[0004] This invention provides a smart policing multi-service collaborative integrated management system. By constructing a cross-modal dynamic knowledge graph and a spatiotemporal prediction model, combined with a multi-agent collaborative mechanism, it achieves efficient closed-loop management and control of data fusion perception, situational early warning, and adaptive scheduling of police forces and materials in complex policing environments.

[0005] This invention provides a smart policing multi-service collaborative integrated management system, including a multimodal perception module, an entity extraction module, a spatiotemporal situation prediction module, and a collaborative scheduling module. The multimodal perception module, entity extraction module, and spatiotemporal situation prediction module are connected in sequence, and the collaborative scheduling module is connected to the multimodal perception module and the spatiotemporal situation prediction module respectively.

[0006] The multimodal perception module is used to acquire real-time police incident data and dynamic police resource status data from multiple heterogeneous sources in real time, perform spatiotemporal coding alignment and cross-modal feature fusion on the real-time police incident data to generate multimodal police incident feature data, and perform structured updates on the dynamic police resource status data.

[0007] The entity extraction module is used to extract police entities and the relationships between police entities from the multimodal police incident feature data, so as to construct and dynamically update the global police knowledge graph. The node mapped by the currently processed multimodal police incident feature data is taken as the current police incident node, and the graph topology substructure associated with the current police incident node is extracted from the global police knowledge graph and output.

[0008] The spatiotemporal situation prediction module is used to retrieve historical spatiotemporal evolution sequence data that matches the current alarm node in the spatiotemporal dimension from the historical database, and to perform spatial feature aggregation on the adjacent nodes connected to the current alarm node in the graph topology substructure through a preset spatiotemporal feature extraction model to obtain spatial evolution features. It also performs temporal feature extraction on the historical spatiotemporal evolution sequence data to obtain temporal evolution features, and outputs multimodal prediction results based on the spatial evolution features and temporal evolution features.

[0009] The collaborative scheduling module is used to map the multimodal prediction results and the updated dynamic police resource status data together into a resource demand state space. It utilizes a multi-agent collaborative mechanism and, based on a globally optimal allocation strategy, performs adaptive resource allocation of police force, police equipment, and logistical supplies within the resource demand state space. It then generates and outputs cross-business collaborative scheduling instructions to guide cross-police force deployment and material distribution.

[0010] Furthermore, the multimodal perception module includes a heterogeneous data access unit, a spatiotemporal coding alignment unit, a cross-modal feature fusion unit, and a structured data storage unit that are connected in sequence via communication.

[0011] The heterogeneous data access unit is used to acquire real-time multi-source heterogeneous real-time police event data and dynamic police resource status data;

[0012] The spatiotemporal coding alignment unit is used to extract the location information of the real-time alarm event data from different sources and uniformly map it into a geohash code with a preset grid precision to complete spatial alignment; and to extract the timestamp of the real-time alarm event data and uniformly convert it into a standard timestamp, and adopt a sliding time window mechanism with a preset step size to divide asynchronous data in the same geohash code grid and the same sliding time window into the same analysis feature frame to complete time alignment.

[0013] The cross-modal feature fusion unit is used to perform cross-modal feature fusion on the aligned real-time alarm event data to generate multimodal alarm feature data;

[0014] The structured storage unit is used to perform structured updates on the dynamic police resource status data and store the multimodal police incident feature data and the updated dynamic police resource status data in the database.

[0015] Furthermore, the specific method by which the cross-modal feature fusion unit performs cross-modal feature fusion includes:

[0016] The alarm text in the real-time alarm event data is mapped into a text feature vector matrix using a text encoding network. ;

[0017] The on-site images or video keyframes in the real-time police incident data are mapped into a visual feature vector matrix using a convolutional neural network. and ;

[0018] The formula for calculating the attention weight matrix using the cross-attention mechanism is as follows:

[0019]

[0020] in, The dimension scaling factor is denoted by T, which indicates matrix transpose. The calculated attention fusion features are input into a multilayer perceptron for dimensionality reduction and the multimodal alarm feature data is output.

[0021] Furthermore, the entity extraction module includes a police entity extraction unit, a graph dynamic update unit, a police incident node mapping unit, and a topology substructure extraction unit that are connected in sequence via communication.

[0022] The police entity extraction unit is used to receive the multimodal police situation feature data, perform joint extraction of entities and relations through the decoding network layer, and output structured knowledge triples.

[0023] The graph dynamic update unit is used to receive the knowledge triples, integrate the knowledge triples and update them into the global police knowledge graph through an entity alignment algorithm;

[0024] The alarm node mapping unit is used to instantiate the currently processed multimodal alarm feature data into a central node of the event type, so as to establish it as the current alarm node;

[0025] The topology substructure extraction unit is used to extract all associated nodes and edges with the current alarm node as the core, using a breadth-first search-based neighbor subgraph extraction algorithm. The extracted associated nodes and edges are then encapsulated into a graph topology substructure and output.

[0026] Furthermore, the specific method by which the police entity extraction unit performs joint entity and relation extraction is as follows:

[0027] The multimodal alarm feature data is mapped to a set of candidate entities;

[0028] For any two candidate entities in the candidate entity set and Their corresponding feature vectors are respectively and The probability of a relationship r between the two is calculated using a relationship classifier. The probability calculation formula is as follows:

[0029]

[0030] in, The weight matrix is ​​pre-trained for the relation r. For the bias term, ║ indicates the concatenation operation of feature vectors;

[0031] Select the relation label with the highest probability and output the knowledge triple containing the candidate entity and its associated relations. .

[0032] Furthermore, the specific method by which the graph dynamic update unit executes the entity alignment algorithm is as follows:

[0033] When newly extracted entities When preparing to write to the global policing knowledge graph, the newly extracted entities are calculated. eigenvectors Feature vectors of similar candidate nodes in the global policing knowledge graph The cosine similarity is calculated using the following formula:

[0034]

[0035] When the calculated cosine similarity exceeds a preset similarity threshold, the two are determined to be the same physical entity and a node attribute fusion update operation is performed.

[0036] When the calculated cosine similarity does not exceed the preset similarity threshold, it is considered as the newly extracted entity in the global police knowledge graph. Create a brand new node.

[0037] Furthermore, the spatiotemporal situation prediction module includes a historical sequence matching and retrieval unit, a spatial feature aggregation unit, a temporal feature extraction unit, and a prediction result generation unit, which are connected in sequence via communication.

[0038] The historical sequence matching and retrieval unit is used to receive the spatiotemporal attributes of the graph topology substructure and the current alarm node, and retrieve the historical spatiotemporal evolution sequence data that matches the current alarm node in the spatiotemporal dimension from the historical database based on the spatiotemporal attributes.

[0039] The spatial feature aggregation unit is used to perform spatial feature aggregation calculation on the adjacent nodes connected to the current alarm node in the graph topology substructure through a pre-trained graph attention network to obtain the spatial evolution features;

[0040] The temporal feature extraction unit is used to call a pre-trained convolutional long short-term memory network, use a one-dimensional convolutional layer to perform sliding window convolution calculation on the historical spatiotemporal evolution sequence data, extract local grid mutation features, input the extracted local grid mutation features into the long short-term memory network unit, and extract temporal dependencies through the forget gate, input gate and output gate mechanism inside the long short-term memory network unit, and output the temporal evolution features representing the historical evolution law;

[0041] The prediction result generation unit is used to concatenate the spatial evolution features and the temporal evolution features to obtain a concatenated feature vector, which is then input into a fully connected layer and processed by an activation function to output the evolution probability of an alarm event occurring within a preset time window. The evolution probability is discretized and mapped to the corresponding warning risk level using a preset threshold mapping rule, and a multimodal prediction result containing the evolution probability and the warning risk level is generated.

[0042] Furthermore, the specific method by which the spatial feature aggregation unit calculates the spatial evolution features is as follows:

[0043] For the current alarm node i and its neighboring node j, their corresponding feature vectors are respectively and Calculate the spatial attention coefficient between nodes. :

[0044]

[0045] Where W is the linear transformation weight matrix shared by all nodes. is the weight vector of a single-layer feedforward neural network, and ║ represents the feature concatenation operation;

[0046] The spatial attention coefficients of all neighboring nodes of node i are normalized using the Softmax function to obtain the normalized attention weights. :

[0047]

[0048] in, Let i be the set of neighboring nodes of node i;

[0049] The normalized attention weights are used to perform weighted summation and nonlinear activation on the features of neighboring nodes to output the spatial evolution features. :

[0050]

[0051] in, It is a non-linear activation function.

[0052] Furthermore, the collaborative scheduling module includes a state space mapping unit, a multi-agent collaborative computing unit, and an instruction generation unit, which are sequentially connected in communication.

[0053] The state space mapping unit is used to extract the evolution probability and early warning risk level from the multimodal prediction results, construct a risk distribution matrix representing the expected gap in police force and materials in each spatial grid within a future preset time window; and extract updated dynamic police resource status data to construct a resource distribution matrix representing the real-time coordinates and available quantities of current global police cars, officers, and logistics equipment; and concatenate the risk distribution matrix and the resource distribution matrix to construct the global state of the resource demand state space in the Markov decision process.

[0054] The multi-agent collaborative computing unit is used to perform decision-making and deduction within the resource demand state space using a pre-set multi-agent collaborative mechanism, and to solve and output a joint action strategy.

[0055] The instruction generation unit is used to decode the joint action strategy into specific business execution instructions, generate cross-business collaborative scheduling instructions for guiding cross-police force deployment and material distribution, and output them.

[0056] Furthermore, the specific method by which the multi-agent cooperative computing unit solves for and outputs the joint action strategy is as follows:

[0057] Execution units with different business attributes are instantiated as independent intelligent agents, and each intelligent agent has an action space that includes dispatch path planning and material allocation quantity.

[0058] Construct a global reward function with the joint optimization objectives of minimizing response time and minimizing the difference between resource supply and demand. The calculation formula is as follows:

[0059]

[0060] in, A collection of intelligent agents, Let i be the estimated response time for agent i to reach the specified risk grid. This is a collection of categories of required logistical supplies. To predict the resource demand for material category j in the forecast results, This represents the actual available supply of resources for category j of materials. Penalties for unmet material shortages and These are preset weighting coefficients;

[0061] A multi-agent deep deterministic policy gradient algorithm is used to train and execute the policy network of each agent to maximize the long-term cumulative expected return and solve for the joint action policy that minimizes the global resource idle rate.

[0062] The beneficial effects of this invention are as follows:

[0063] This invention breaks down underlying data barriers by using a multimodal perception module to spatiotemporally align and cross-modal fuse heterogeneous police incident and dynamic police resource status data into a database. It also utilizes an entity extraction module to dynamically construct a global police knowledge graph, achieving deep semantic association of cross-modal clues. Furthermore, a spatiotemporal situation prediction module aggregates graph topology and historical sequence features to output the evolution probability and risk level of police incidents in the spatiotemporal dimension. Finally, a collaborative scheduling module, based on a multi-agent collaborative mechanism and a globally optimal allocation strategy, adaptively configures police force, police equipment, and logistical supplies within the resource demand state space, generating and outputting cross-business collaborative scheduling instructions. This invention achieves end-to-end automated management and control, from comprehensive perception and proactive prediction to cross-business intelligent scheduling, effectively solving the technical problems of delayed response and resource mismatch in traditional police systems. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the structure of the intelligent policing multi-service collaborative integrated management system of the present invention.

[0065] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0066] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0067] like Figure 1 As shown, the present invention provides a smart policing multi-service collaborative integrated management system, including a multimodal perception module, an entity extraction module, a spatiotemporal situation prediction module, and a collaborative scheduling module. The multimodal perception module, entity extraction module, and spatiotemporal situation prediction module are connected in sequence, and the collaborative scheduling module is connected to the multimodal perception module and the spatiotemporal situation prediction module respectively.

[0068] The multimodal perception module is used to acquire real-time police incident data and dynamic police resource status data from multiple heterogeneous sources in real time, perform spatiotemporal coding alignment and cross-modal feature fusion on the real-time police incident data to generate multimodal police incident feature data, and perform structured updates on the dynamic police resource status data.

[0069] The entity extraction module is used to extract police entities and the relationships between police entities from the multimodal police incident feature data, so as to construct and dynamically update the global police knowledge graph. The node mapped by the currently processed multimodal police incident feature data is taken as the current police incident node, and the graph topology substructure associated with the current police incident node is extracted from the global police knowledge graph and output.

[0070] The spatiotemporal situation prediction module is used to retrieve historical spatiotemporal evolution sequence data that matches the current alarm node in the spatiotemporal dimension from the historical database, and to perform spatial feature aggregation on the adjacent nodes connected to the current alarm node in the graph topology substructure through a preset spatiotemporal feature extraction model to obtain spatial evolution features. It also performs temporal feature extraction on the historical spatiotemporal evolution sequence data to obtain temporal evolution features, and outputs multimodal prediction results based on the spatial evolution features and temporal evolution features.

[0071] The collaborative scheduling module is used to map the multimodal prediction results and the updated dynamic police resource status data together into a resource demand state space. It utilizes a multi-agent collaborative mechanism and, based on a globally optimal allocation strategy, performs adaptive resource allocation of police force, police equipment, and logistical supplies within the resource demand state space. It then generates and outputs cross-business collaborative scheduling instructions to guide cross-police force deployment and material distribution.

[0072] (1) Multimodal sensing module

[0073] In a preferred embodiment, the multimodal perception module includes a heterogeneous data access unit, a spatiotemporal coding alignment unit, a cross-modal feature fusion unit, and a structured data storage unit. Specifically,

[0074] ① Heterogeneous data access unit

[0075] The heterogeneous data access unit is used to collect two types of heterogeneous business data in real time and in parallel through a multi-protocol IoT gateway and API interface bus, including real-time police event data and dynamic police resource status data, specifically:

[0076] a. Real-time police incident data includes unstructured alarm texts from the 110 emergency call system of the command center, photos of incidents uploaded by on-site citizens, and real-time streaming video sequences transmitted back by surveillance cameras or drone (UAV) image transmission systems of the Skynet project.

[0077] b. Dynamic police resource status data includes real-time latitude and longitude coordinates of police officers transmitted through vehicle-mounted and individual police Beidou positioning terminals, as well as structured data such as the current inventory quantity and availability status of logistical supplies (such as individual police equipment and medical first aid kits) obtained through RFID system integration.

[0078] ② Spatiotemporal coding alignment unit

[0079] Because the collected multi-source data naturally differ in sampling frequency and coordinate system, the spatiotemporal coding alignment unit performs strict spatial and temporal alignment processing on the real-time alarm event data. Specifically, spatial alignment involves uniformly mapping location information from different sources (such as road intersection names described in text and relative offset coordinates attached to videos) to geohash encoding with a preset grid precision, ensuring absolute uniformity in spatial resolution. Temporal alignment involves extracting the timestamps of each modal data stream, uniformly converting them to standard Unix timestamps, and using a sliding time window mechanism with a preset step size to group asynchronous data within the same spatiotemporal grid and the same time window into the same analysis feature frame.

[0080] ③ Cross-modal feature fusion unit

[0081] To bridge the semantic gap in unstructured data, the cross-modal feature fusion unit receives aligned real-time police event data and performs fusion calculations using a pre-trained two-stream neural network combined with a cross-attention mechanism. The specific calculation process is as follows:

[0082] First, the alarm text is mapped into a text feature vector matrix using a pre-built text encoding network. Simultaneously, a convolutional neural network is used to map keyframes of on-site images or videos into visual feature vector matrices. and Subsequently, in order to enable the text semantics to automatically focus on highly relevant key cue regions in the visual image, the system performs cross-attention calculation, and the formula for calculating the attention weight matrix is ​​as follows:

[0083]

[0084] in, is the dimension scaling factor for the feature vector, and the superscript T denotes matrix transpose. Finally, the calculated attention fusion features are input into a multilayer perceptron (MLP) for dimensionality reduction and normalization, ultimately outputting high-dimensional multimodal alarm feature data.

[0085] ④ Structured data storage unit

[0086] The structured data storage unit performs a dual-track data persistence operation. On one hand, it works in parallel with the cross-modal feature fusion unit to clean and structure the acquired dynamic police resource status data, extracting key fields such as resource ID, current hash coordinates, and available quantity, and updating the resource status mapping table in memory. On the other hand, it is responsible for the final data storage. The system pushes the generated high-dimensional multimodal police incident feature data into the vector database for feature persistence, while simultaneously writing the updated dynamic police resource status data into a relational database. Both are hard-associated through a globally unique police incident business serial number and a spatiotemporal Geohash primary key. This completes the standardized storage of the data, providing a data foundation for entity extraction and knowledge graph construction.

[0087] (2) Entity extraction module

[0088] In a preferred embodiment, the entity extraction module includes a police entity extraction unit, a graph dynamic update unit, a police incident node mapping unit, and a topology substructure extraction unit. Specifically,

[0089] ①Police entity extraction unit

[0090] The police entity extraction unit receives high-dimensional multimodal police situation feature data output by the multimodal perception module. Since the feature data has integrated visual and textual cross-attention, a multilayer perceptron (MLP) combined with a conditional random field (CRF) decoding network layer is used to jointly extract entities and relationships.

[0091] Specifically, the decoding network first maps multimodal feature vectors to a set of candidate entities (such as suspects, vehicles involved in the case, crime scene, police resources, etc.). Then, the system predicts the relationships between entities (such as ownership, peers, crime scene location) using a relation classifier. For any two extracted candidate entities... and Their corresponding feature vectors are respectively and The probability formula for calculating the relationship r between the two is:

[0092]

[0093] in, For the weight matrix pre-trained for relation r, For the bias term, ║ indicates the concatenation operation of feature vectors. The system selects the relation label with the highest probability and outputs structured knowledge triples. .

[0094] ②Map Dynamic Update Unit

[0095] After extracting the triples, the graph dynamic update unit integrates them into the global police knowledge graph. To avoid redundant nodes in the graph, the system introduces an entity alignment algorithm. When newly extracted entities... When preparing to write to a graph database (such as Neo4j), the system calculates... eigenvectors Feature vectors of similar candidate nodes in the global graph Cosine similarity:

[0096]

[0097] If and only if the calculated similarity value exceeds the preset similarity threshold When the similarity value does not exceed the preset similarity threshold, the system determines that the two are the same physical entity and performs a fusion and update operation on the node attributes. At that time, the system provides the following in the global police knowledge graph: Create entirely new nodes; this mechanism ensures the accuracy of graph updates and completes the standardized storage of data.

[0098] ③ Police Incident Node Mapping Unit

[0099] To trigger downstream spatiotemporal situation prediction, the incident node mapping unit is responsible for establishing anchor points between the current business flow and the graph structure. The system generates a unique spatiotemporal event identifier for the currently processed multimodal incident feature data and instantiates it as a central node of an event type in the global police knowledge graph, explicitly defining it as the "current incident node" (set as...). This node also carries the precise geohash code and standard timestamp attribute obtained by the multimodal perception module.

[0100] ④ Topological substructure extraction unit

[0101] The global graph is enormous, and directly performing full-graph operations would cause computational bottlenecks and introduce irrelevant noise. Therefore, the topology substructure extraction unit adopts a K-Hop (K-hop) neighbor subgraph extraction algorithm based on breadth-first search (BFS). Taking the current alarm node as an example... With this as the core, the system extracts all associated nodes and edges whose shortest topological distance does not exceed a preset hop count parameter K (preferably, K is 2 or 3). The system then encapsulates this extracted subset into a graph topological substructure. ,in For inclusion and the set of vertices of adjacent nodes, This represents the corresponding set of edges. Finally, the standardized graph topological substructure is... The output is sent to the spatiotemporal situation prediction module, providing a structured graph network foundation for spatial feature aggregation.

[0102] (3) Spatiotemporal situation prediction module

[0103] In a preferred embodiment, the spatiotemporal situation prediction module includes a historical sequence matching and retrieval unit, a spatial feature aggregation unit, a temporal feature extraction unit, and a prediction result generation unit. Specifically,

[0104] ①Historical sequence matching retrieval unit

[0105] The historical sequence matching retrieval unit first receives the graph topology substructure output by the entity extraction module. and current police situation The system determines the spatiotemporal attributes (including geohash encoding and standard timestamp). Using the geohash encoding of the current incident as the center, the system radiates outwards a spatial grid of preset levels; simultaneously, using the standard timestamp as the reference point, it traces back a preset number of time steps T along the historical timeline. Based on the determined spatial grid and time step T, the system performs a retrieval from a relational historical database, retrieving historical spatiotemporal evolution sequence data that strictly matches the current incident in the spatiotemporal dimension. .

[0106] ② Spatial feature aggregation unit (based on graph attention mechanism)

[0107] The spatial feature aggregation unit extracts the diffusion trend of alarms in the spatial topology, targeting the received map topology substructure. The system employs a pre-trained Graph Attention Network (GAT) as the spatial perception layer of the spatiotemporal feature extraction model. Specifically, in the calculation, for the current alarm node i and its neighboring nodes j (… , Let i be the set of neighboring nodes of node i, and its corresponding feature vectors are respectively and The system first calculates the spatial attention coefficients between nodes. The formula is as follows:

[0108]

[0109] Where W is the linear transformation weight matrix shared by all nodes. is the weight vector of a single-layer feedforward neural network with an attention mechanism, and ║ represents the feature concatenation operation.

[0110] Subsequently, the attention coefficients of all neighboring nodes are normalized using the Softmax function:

[0111]

[0112] Finally, the normalized attention weights are used to perform weighted summation and nonlinear activation on the features of adjacent nodes to output the spatial evolution features. :

[0113]

[0114] in, It is a non-linear activation function; this process enables the model to adaptively identify the associated nodes (such as adjacent high-risk grids or key targets of concern) that have the most significant impact on the evolution of the current situation.

[0115] ③ Temporal feature extraction unit (based on CNN-LSTM architecture)

[0116] In parallel with spatial aggregation, the temporal feature extraction unit is responsible for processing the retrieved historical spatiotemporal evolution sequence data. To capture high-frequency burst features and long-term temporal dependencies, the system uses a pre-trained convolutional long short-term memory network (CNN-LSTM) as the temporal awareness layer of the spatiotemporal feature extraction model.

[0117] First, a sliding window convolution is performed on the sequence data using a one-dimensional convolutional layer (1D-CNN) to extract local grid change features within a short period (e.g., a dense increase in alarms on a street within 15 minutes). Then, the output sequence of the convolutional layer is fed into a Long Short-Term Memory (LSTM) network unit. The LSTM filters out noise information from the historical sequence through its internal forget gate, input gate, and output gate mechanisms, while retaining hidden states that are instructive for future trends. Finally, it outputs the temporal evolution features representing the historical evolution pattern. .

[0118] ④ Prediction Result Generation Unit

[0119] The prediction result generation unit transforms spatial and temporal features into the final business scheduling basis, and the system uses the obtained spatial evolution features. With the aforementioned temporal evolution characteristics Perform feature splicing:

[0120]

[0121] concatenated feature vectors The input is fed into a fully connected layer, and after passing through a Sigmoid activation function, the output is the probability of the spatial grid evolving or spreading into an alarm within a preset time window. Finally, the system discretizes this continuous probability value using a preset threshold mapping rule; that is, when... When the value is within the first preset threshold range (e.g., [0.8, 1.0]), it is mapped and output as a "red" warning risk level; when... When the risk level falls within the second preset threshold range, it is mapped to an "orange" warning risk level; and so on. The final output includes the evolution probability and the determined warning risk level, which is then sent to the collaborative scheduling module for resource allocation simulation.

[0122] (4) Cooperative scheduling module

[0123] In a preferred embodiment, the collaborative scheduling module includes a state space mapping unit, a multi-agent collaborative computing unit, and an instruction generation unit. Specifically,

[0124] ①State-space mapping unit

[0125] The state space mapping unit receives the multimodal prediction results (including the evolution probability and early warning risk level of each spatial grid) output by the spatiotemporal situation prediction module, and the dynamic police resource state data output and updated by the multimodal perception module. These two types of data are jointly mapped into the resource demand state space within a Markov decision process (MDP). At any time step t, the global state of the system is It is composed of a risk distribution matrix and a resource distribution matrix. The risk distribution matrix represents the expected shortage of police force and supplies in each grid within a future preset time window; the resource distribution matrix represents the real-time latitude and longitude coordinates and available quantity of police cars, police officers, and logistical equipment in the current global inventory.

[0126] ② Multi-agent collaborative computing unit

[0127] The multi-agent collaborative computing unit utilizes a pre-built multi-agent collaborative mechanism (Multi-Agent Reinforcement Learning, MARL) to adaptively allocate resources within the resource demand state space. The system instantiates execution units with different business attributes (such as traffic police squadrons, patrol police brigades, and logistics equipment centers) as independent agents. The system is defined to contain N agents, set as follows: Each agent i has its own independent action space. Its actions It includes specific deployment route planning and the quantity of materials allocated.

[0128] To achieve the globally optimal allocation strategy, a unified global reward function was constructed by the multi-agent collaborative computing unit. Based on the technical characteristics of the entire chain of police logistics management, the reward function is designed with "minimizing response time" and "minimizing the difference between resource supply and demand" as joint optimization objectives.

[0129]

[0130] in, The estimated response time for agent i to reach the specified risk grid; This is a collection of categories of required logistical supplies. This refers to the resource demand for material category j in the prediction results. This represents the actual supply of material category j that can be delivered after the action configuration. Indicates penalties for unmet material shortages; and These are the preset weighting coefficients.

[0131] The system employs a multi-agent deep deterministic policy gradient algorithm (MADDPG) to perform centralized training and distributed execution of the policy networks of each agent, thereby maximizing the long-term cumulative expected return. ( As a discount factor, (For the policy network parameters), solve for the joint action policy that minimizes the global resource idle rate and has the fastest response. .

[0132] ③ Instruction generation unit

[0133] The instruction generation unit generates the joint action strategy output by the multi-agent collaborative computing unit. The system decodes the commands into specific business execution instructions. Based on the action strategy, it parses the corresponding police dispatch routes, node takeover permissions, and logistical supply lists, and encapsulates these into cross-business collaborative dispatch instructions according to a preset communication protocol. These instructions include, but are not limited to, remote traffic control instructions sent to traffic police terminals, rapid assault and intervention routes sent to patrol police terminals, and precise material delivery (such as riot control equipment and first-aid kits) allocation instructions sent to UAV (unmanned aerial vehicle) hangars and logistical warehouses. Finally, the instruction generation unit sends the cross-business collaborative dispatch instructions to the corresponding police execution terminals through a unified communication gateway. Simultaneously, the system stores the triggering conditions, instruction content, and received feedback status of this dispatch as business flow data in a relational database, forming a complete business dispatch closed loop.

[0134] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0135] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A smart policing multi-service collaborative integrated management system, characterized in that, It includes a multimodal perception module, an entity extraction module, a spatiotemporal situation prediction module, and a collaborative scheduling module. The multimodal perception module, entity extraction module, and spatiotemporal situation prediction module are connected in sequence, and the collaborative scheduling module is connected to the multimodal perception module and the spatiotemporal situation prediction module respectively. The multimodal perception module is used to acquire real-time police incident data and dynamic police resource status data from multiple heterogeneous sources in real time, perform spatiotemporal coding alignment and cross-modal feature fusion on the real-time police incident data to generate multimodal police incident feature data, and perform structured updates on the dynamic police resource status data. The entity extraction module is used to extract police entities and the relationships between police entities from the multimodal police incident feature data, so as to construct and dynamically update the global police knowledge graph. The node mapped by the currently processed multimodal police incident feature data is taken as the current police incident node, and the graph topology substructure associated with the current police incident node is extracted from the global police knowledge graph and output. The spatiotemporal situation prediction module is used to retrieve historical spatiotemporal evolution sequence data that matches the current alarm node in the spatiotemporal dimension from the historical database, and to perform spatial feature aggregation on the adjacent nodes connected to the current alarm node in the graph topology substructure through a preset spatiotemporal feature extraction model to obtain spatial evolution features. It also performs temporal feature extraction on the historical spatiotemporal evolution sequence data to obtain temporal evolution features, and outputs multimodal prediction results based on the spatial evolution features and temporal evolution features. The collaborative scheduling module is used to map the multimodal prediction results and the updated dynamic police resource status data together into a resource demand state space. It utilizes a multi-agent collaborative mechanism and, based on a globally optimal allocation strategy, performs adaptive resource allocation of police force, police equipment, and logistical supplies within the resource demand state space. It then generates and outputs cross-business collaborative scheduling instructions to guide cross-police force deployment and material distribution.

2. The intelligent policing multi-service collaborative integrated management system according to claim 1, characterized in that, The multimodal perception module includes a heterogeneous data access unit, a spatiotemporal coding alignment unit, a cross-modal feature fusion unit, and a structured data storage unit that are connected in sequence via communication. The heterogeneous data access unit is used to acquire real-time multi-source heterogeneous real-time police event data and dynamic police resource status data; The spatiotemporal coding alignment unit is used to extract the location information of the real-time alarm event data from different sources and uniformly map it into a geohash code with a preset grid precision to complete spatial alignment. In addition, the timestamps of the real-time alarm event data are extracted and uniformly converted into standard timestamps. A sliding time window mechanism with a preset step size is used to divide asynchronous data in the same geohash coding grid and the same sliding time window into the same analysis feature frame to complete time alignment. The cross-modal feature fusion unit is used to perform cross-modal feature fusion on the aligned real-time alarm event data to generate multimodal alarm feature data; The structured storage unit is used to update the dynamic police resource status data in a structured manner, and to store the multimodal police incident feature data and the updated dynamic police resource status data in the database.

3. The intelligent policing multi-service collaborative integrated management system according to claim 2, characterized in that, The specific methods by which the cross-modal feature fusion unit performs cross-modal feature fusion include: The alarm text in the real-time alarm event data is mapped into a text feature vector matrix using a text encoding network. ; The on-site images or video keyframes in the real-time police incident data are mapped into a visual feature vector matrix using a convolutional neural network. and ; The formula for calculating the attention weight matrix using the cross-attention mechanism is as follows: in, The dimension scaling factor is denoted by T, which indicates matrix transpose. The calculated attention fusion features are input into a multilayer perceptron for dimensionality reduction and the multimodal alarm feature data is output.

4. The intelligent policing multi-service collaborative integrated management system according to claim 1, characterized in that, The entity extraction module includes a police entity extraction unit, a graph dynamic update unit, a police node mapping unit, and a topology substructure extraction unit that are connected in sequence via communication. The police entity extraction unit is used to receive the multimodal police situation feature data, perform joint extraction of entities and relations through the decoding network layer, and output structured knowledge triples. The graph dynamic update unit is used to receive the knowledge triples, integrate the knowledge triples and update them into the global police knowledge graph through an entity alignment algorithm; The alarm node mapping unit is used to instantiate the currently processed multimodal alarm feature data into a central node of the event type, so as to establish it as the current alarm node; The topology substructure extraction unit is used to extract all associated nodes and edges with the current alarm node as the core, using a breadth-first search-based neighbor subgraph extraction algorithm. The extracted associated nodes and edges are then encapsulated into a graph topology substructure and output.

5. The intelligent policing multi-service collaborative integrated management system according to claim 4, characterized in that, The specific method by which the police entity extraction unit performs joint entity and relation extraction is as follows: The multimodal alarm feature data is mapped to a set of candidate entities; For any two candidate entities in the candidate entity set and Their corresponding feature vectors are respectively and The probability of a relationship r between the two is calculated using a relationship classifier. The probability calculation formula is as follows: in, The weight matrix is ​​pre-trained for the relation r. For the bias term, ║ indicates the concatenation operation of feature vectors; Select the relation label with the highest probability and output the knowledge triple containing the candidate entity and its associated relations. .

6. The intelligent policing multi-service collaborative integrated management system according to claim 4, characterized in that, The specific method by which the graph dynamic update unit executes the entity alignment algorithm is as follows: When newly extracted entities When preparing to write to the global policing knowledge graph, the newly extracted entities are calculated. eigenvectors Feature vectors of similar candidate nodes in the global policing knowledge graph The cosine similarity is calculated using the following formula: When the calculated cosine similarity exceeds a preset similarity threshold, the two are determined to be the same physical entity and a node attribute fusion update operation is performed. When the calculated cosine similarity does not exceed the preset similarity threshold, it is considered as the newly extracted entity in the global police knowledge graph. Create a brand new node.

7. The intelligent policing multi-service collaborative integrated management system according to claim 1, characterized in that, The spatiotemporal situation prediction module includes a historical sequence matching and retrieval unit, a spatial feature aggregation unit, a temporal feature extraction unit, and a prediction result generation unit, which are connected in sequence via communication. The historical sequence matching and retrieval unit is used to receive the spatiotemporal attributes of the graph topology substructure and the current alarm node, and retrieve the historical spatiotemporal evolution sequence data that matches the current alarm node in the spatiotemporal dimension from the historical database based on the spatiotemporal attributes. The spatial feature aggregation unit is used to perform spatial feature aggregation calculation on the adjacent nodes connected to the current alarm node in the graph topology substructure through a pre-trained graph attention network to obtain the spatial evolution features; The temporal feature extraction unit is used to call a pre-trained convolutional long short-term memory network, use a one-dimensional convolutional layer to perform sliding window convolution calculation on the historical spatiotemporal evolution sequence data, extract local grid mutation features, input the extracted local grid mutation features into the long short-term memory network unit, and extract temporal dependencies through the forget gate, input gate and output gate mechanism inside the long short-term memory network unit, and output the temporal evolution features representing the historical evolution law; The prediction result generation unit is used to concatenate the spatial evolution features and the temporal evolution features to obtain a concatenated feature vector, which is then input into a fully connected layer and processed by an activation function to output the evolution probability of an alarm event occurring within a preset time window. The evolution probability is discretized and mapped to the corresponding warning risk level using a preset threshold mapping rule, and a multimodal prediction result containing the evolution probability and the warning risk level is generated.

8. The intelligent policing multi-service collaborative integrated management system according to claim 7, characterized in that, The specific method by which the spatial feature aggregation unit calculates the spatial evolution features is as follows: For the current alarm node i and its neighboring node j, their corresponding feature vectors are respectively and Calculate the spatial attention coefficient between nodes. : Where W is the linear transformation weight matrix shared by all nodes. is the weight vector of a single-layer feedforward neural network, and ║ represents the feature concatenation operation; The spatial attention coefficients of all neighboring nodes of node i are normalized using the Softmax function to obtain the normalized attention weights. : in, Let i be the set of neighboring nodes of node i; The normalized attention weights are used to perform weighted summation and nonlinear activation on the features of neighboring nodes to output the spatial evolution features. : in, It is a non-linear activation function.

9. The intelligent policing multi-service collaborative integrated management system according to claim 1, characterized in that, The collaborative scheduling module includes a state space mapping unit, a multi-agent collaborative computing unit, and an instruction generation unit, which are sequentially connected by communication. The state space mapping unit is used to extract the evolution probability and early warning risk level from the multimodal prediction results, construct a risk distribution matrix representing the expected gap in police force and materials in each spatial grid within a future preset time window; and extract updated dynamic police resource status data to construct a resource distribution matrix representing the real-time coordinates and available quantities of current global police cars, officers, and logistics equipment; and concatenate the risk distribution matrix and the resource distribution matrix to construct the global state of the resource demand state space in the Markov decision process. The multi-agent collaborative computing unit is used to perform decision-making and deduction within the resource demand state space using a pre-set multi-agent collaborative mechanism, and to solve and output a joint action strategy. The instruction generation unit is used to decode the joint action strategy into specific business execution instructions, generate cross-business collaborative scheduling instructions for guiding cross-police force deployment and material distribution, and output them.

10. The intelligent policing multi-service collaborative integrated management system according to claim 9, characterized in that, The specific method by which the multi-agent collaborative computing unit solves for and outputs the joint action strategy is as follows: Execution units with different business attributes are instantiated as independent intelligent agents, and each intelligent agent has an action space that includes dispatch path planning and material allocation quantity. Construct a global reward function with the joint optimization objectives of minimizing response time and minimizing the difference between resource supply and demand. The calculation formula is as follows: in, A collection of intelligent agents, Let i be the estimated response time for agent i to reach the specified risk grid. This is a collection of categories of required logistical supplies. To predict the resource demand for material category j in the forecast results, This represents the actual available supply of resources for category j of materials. Penalties for unmet material shortages and These are preset weighting coefficients; A multi-agent deep deterministic policy gradient algorithm is used to train and execute the policy network of each agent to maximize the long-term cumulative expected return and solve for the joint action policy that minimizes the global resource idle rate.

Images