Police risk multi-level early warning system based on large model cooperative calculation

Abstract

The application provides a police risk multi-level early warning system based on large model collaborative calculation, comprising: a data module that collects and pre-processes multi-source police data in real time, converts the data into text, vision and space-time initial vectors, stores the vectors into a vector database, and identifies initial risk events through an event detector to generate early warning tasks; a calculation module that analyzes tasks and calls corresponding vectors through a collaborative scheduling hub, processes and generates feature vectors and space-time attributes through a large model cluster of text analysis, vision analysis and space-time prediction adjusted in the police field; and an early warning module that calculates a comprehensive risk value based on a attention mechanism to fuse multi-modal features, and realizes individual-level instant early warning, regional-level situation early warning and trend-level deduction early warning in combination with a early warning threshold and space-time attributes. The system can realize multi-level, self-adaptive and accurate early warning of police risks, and improve the intelligent level of public security risk prevention and control.

Description

Technical Field

[0001] This invention relates to the field of public safety technology, specifically to a multi-level early warning system for police risks based on large-scale collaborative computing. Background Technology

[0002] Traditional police risk early warning systems primarily rely on single data sources and rule engines for risk assessment, resulting in insufficient multimodal data fusion capabilities and limited early warning accuracy. Existing methods typically employ static thresholds or simple models for analysis, which struggle to cope with complex and ever-changing police scenarios, leading to high false alarm and false negative rates. Furthermore, traditional systems lack the ability to model the spatiotemporal evolution of risk events, hindering multi-level, adaptive early warning responses and limiting the improvement of public safety risk prevention and control effectiveness.

[0003] Current technologies have significant shortcomings in real-time collaborative computation of multi-source heterogeneous police data. Text, visual, and spatiotemporal data are often processed independently, lacking fusion methods driven by cross-modal semantic alignment and attention mechanisms, resulting in insufficient extraction of key risk features. At the same time, traditional models struggle to support cross-scale risk extrapolation from the individual level to the trend level, failing to meet the urgent needs of modern policing for intelligent and precise early warning. Summary of the Invention

[0004] In order to solve the technical problems mentioned in the background art, the purpose of this invention is to provide a multi-level early warning system for police risks based on large-scale model collaborative computing.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A multi-level early warning system for police risks based on large-scale model collaborative computation includes:

[0007] Data module: Collects police data in real time, preprocesses the police data, converts the preprocessed police data into initial vectors, and stores the initial vectors in a preset vector database. The initial vectors include text initial vectors, visual initial vectors, and spatiotemporal initial vectors. An event detector is set up to monitor the preprocessed police data in real time and identify initial risk events based on preset risk rules to generate early warning tasks.

[0008] The computation module includes a collaborative scheduling center and a large model cluster. The collaborative scheduling center generates sub-tasks by parsing the early warning task and retrieves and calls the corresponding initial vectors from the vector database according to the sub-tasks. The large model cluster includes a text analysis model, a visual analysis model, and a spatiotemporal prediction model adjusted with data from the police field. The large model cluster processes the initial vectors to generate feature vectors and spatiotemporal attributes.

[0009] The feature vectors include text feature vectors, visual feature vectors, and spatiotemporal feature vectors;

[0010] The text analysis model is built using the Transformer-XL architecture, which includes an input layer, a multi-head attention layer, a semantic enhancement layer, and an output layer.

[0011] The initial text vector is input into the input layer for domain-adaptive embedding, as detailed below.

[0012] Aligning the general semantic space to the policing domain enhances the ability to recognize police terminology and obtains embedding vectors aligned with the policing domain. ;

[0013] The multi-head attention layer calculates the relative positional attention of 12 heads to capture long-distance semantic dependencies and temporal relationships of alarm events, thereby obtaining attention features. The formula is as follows:

[0014]

[0015] in, , and The first The query matrix, key matrix, and value matrix of the header. For the first Head and First Head relative position encoding matrix, For the number of attention heads, The dimension of the key or query vector;

[0016] The semantic enhancement layer integrates prior knowledge of crime types to obtain risk characteristics;

[0017] The output layer fuses attention features and risk features to generate a text feature vector;

[0018] The visual analysis model is constructed using a multi-scale attention convolutional network, specifically including an input layer, a spatial pyramid layer, an attention layer, and an output layer.

[0019] The initial visual vector is input into the input layer, and the optimized feature map Z is obtained by reconstructing the initial visual vector.

[0020]

[0021] in, for Convolution kernel, As the initial visual vector, To modify the activation function of the linear unit;

[0022] The spatial pyramid layer performs multi-scale feature fusion on the optimized feature map, capturing behavioral patterns of different granularities to obtain multi-scale fused features.

[0023] The attention layer focuses on regions of abnormal behavior to obtain a behavioral attention map;

[0024] The output layer performs channel weighting and dimension normalization to generate standardized visual feature vectors.

[0025] The spatiotemporal prediction model is constructed using a spatiotemporal graph convolutional network, specifically including an input layer, a propagation layer, an influence layer, and an output layer.

[0026] The spatiotemporal initial vector is input into the input layer to construct a dynamic spatiotemporal graph;

[0027] The propagation layer alternately performs graph convolution and temporal modeling to simulate the risk diffusion process and capture the evolution patterns of events, thereby obtaining a spatiotemporal state matrix. ;

[0028] The influence layer is used to calculate the diffusion radius. Predict the spatial impact range of risk events;

[0029] The output layer performs graph feature pooling and spatiotemporal attribute parsing to generate spatiotemporal feature vectors. and spacetime attribute triples :

[0030]

[0031] in, The core location, latitude and longitude, and t represent the duration of the risk event. This is a graph-level pooling operation;

[0032] The dimensions of the text feature vector, visual feature vector, and spatiotemporal feature vector are unified through a preset first fully connected layer;

[0033] The early warning module calculates a comprehensive risk value based on feature vectors using a weighted fusion algorithm based on an attention mechanism.

[0034] The comprehensive risk value The calculation steps are as follows:

[0035] 1) Generate the query vector, key vector, and value vector of the feature vector through a preset attention perception layer:

[0036]

[0037] in, , and These are the query vector, key vector, and value vector, respectively. For feature vectors, Representing text, visual, and time and space respectively. , , , , and These are trainable parameters;

[0038] 2) Calculate the attention score of the feature vector by scaling the dot product attention. :

[0039]

[0040] in, This is the transpose of the key vector. The dimension of the key vector;

[0041] 3) Based on the attention score For the value vector Weighting is performed to generate modal features. ;

[0042] 4) Overall Risk Value Through the preset second fully connected layer and The activation function is calculated as follows:

[0043]

[0044] in, and These are trainable parameters;

[0045] Based on the warning threshold of the comprehensive risk value and the aforementioned spatiotemporal attributes, a multi-level early warning system is generated, including individual-level real-time early warning, regional-level situational early warning, and trend-level deductive early warning.

[0046] Furthermore, the initial text vector is obtained by using the BERT model to perform context-aware acquisition of the text data;

[0047] The visual initial vector is obtained by extracting multi-layer convolutional features from the image data using a visual encoder based on the ResNet-50 architecture.

[0048] The spatiotemporal initial vector is obtained by extracting time period features and spatial grid encoding from spatiotemporal data, and then mapping it through a multilayer perceptron.

[0049] The initial vectors are organized according to a preset time window and data source identifier, stored in the vector database, and a density-based index is established to support similarity retrieval.

[0050] Furthermore, the event detector identifies the initial risk event based on a set of configurable preset rules;

[0051] When the event detector identifies an initial risk event, it generates the early warning task. The early warning task includes at least a unique event identifier, event type, event occurrence time, core data index associated with the event, and a preliminary description and key features of the event.

[0052] Furthermore, the sub-tasks include text processing requirements, visual processing requirements, and spatiotemporal processing requirements;

[0053] The collaborative scheduling center decomposes specific data processing requirements based on the event type, preliminary description, and core data index of the event association of the early warning task, and classifies these requirements into three sub-tasks: text processing requirements, visual processing requirements, and spatiotemporal processing requirements.

[0054] The collaborative scheduling center will assign the retrieved initial vector and corresponding subtask to the appropriate professional models in the large model cluster for processing.

[0055] Furthermore, the warning thresholds include individual thresholds, regional thresholds, trend thresholds, and range thresholds;

[0056] The individual-level real-time early warning is activated when the overall risk value is greater than or equal to the individual threshold and the diffusion radius is less than or equal to the range threshold.

[0057] When the overall risk value is greater than or equal to the regional threshold and less than the individual threshold, or when the diffusion radius is greater than the range threshold, a regional-level situation warning is initiated.

[0058] When the overall risk value is greater than or equal to the trend threshold and less than the regional threshold, and the spatiotemporal attributes indicate that the risk event has the potential to continue or spread, a trend-level inference warning is initiated.

[0059] Compared with the prior art, the advantages of the present invention are as follows:

[0060] 1. This invention uses a cluster of text, visual, and spatiotemporal large models adapted for the police field to perform integrated processing and feature extraction on multi-source heterogeneous data. It also employs a weighted fusion algorithm based on an attention mechanism, which effectively solves the problems of data isolation and insufficient feature extraction in traditional methods, and significantly improves the accuracy of risk identification.

[0061] 2. This invention dynamically triggers early warnings at different granularities, such as individual, regional, and trend levels, based on the comprehensive risk value and its spatiotemporal attributes, thereby achieving full-cycle coverage of police risks from real-time handling to macro-level projection and enhancing the pertinence and foresight of risk response.

[0062] 3. Through structures such as spatiotemporal graph convolutional networks, the present invention can effectively capture the spatiotemporal diffusion patterns and evolution trends of risk events, and support quantitative prediction of attributes such as the scope of risk impact and duration, thereby improving the ability to perceive and respond to risk dynamics in complex policing scenarios. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a schematic diagram of the system workflow of the present invention;

[0065] Figure 2 This is a schematic diagram of the large model cluster of the present invention;

[0066] Figure 3 This is a schematic diagram of the comprehensive risk value generation process of the present invention. Detailed Implementation

[0067] To achieve the above objectives, this invention provides a multi-level early warning system for police risks based on large-scale model collaborative computation. Please refer to [link to relevant documentation]. Figure 1 The system includes:

[0068] Data module: Collects police data in real time, preprocesses the police data, converts the preprocessed police data into initial vectors, and stores the initial vectors in a preset vector database; sets up an event detector to monitor the preprocessed police data in real time, and identifies initial risk events based on preset risk rules to generate early warning tasks;

[0069] The specific steps for real-time collection of police data are as follows:

[0070] The system connects to and accesses multiple police data sources in real time through a distributed data collection agent deployed within the police intranet. These data sources include, but are not limited to:

[0071] Emergency response system: Real-time streaming of 110 or 122 alarm records, alarm description text, alarm caller information, and initial response feedback;

[0072] Video surveillance platform: Real-time video streams, captured images, and structured event snapshots of key areas, such as intelligent analysis of initial screening results for fights, crowd gatherings, etc.

[0073] Mobile policing terminal: Real-time text and image reports, location information, and on-site audio recordings transcribed into text by law enforcement officers at the scene;

[0074] Population, vehicle and case information database: query or subscribe to updated structured data on demand, such as files of key personnel, information on vehicles involved in cases, information on fugitives, and summaries of historical case files;

[0075] IoT sensing devices: such as vehicle data at checkpoints, WiFi probe data (anonymized MAC addresses), and alarm information from sensors on critical infrastructure;

[0076] Internet public opinion monitoring system: integrates online text information that has been preliminarily screened and involves local public security, safety, and sensitive events.

[0077] The data acquisition process emphasizes low latency and high concurrency to ensure that information can be promptly entered into the system processing flow.

[0078] The data collection interface for police affairs follows a preset security protocol and authorization mechanism.

[0079] Preprocessing includes data anonymization, format standardization, and noise removal;

[0080] The specific steps for data anonymization are as follows:

[0081] Text data: For fields in structured text data that explicitly contain personally identifiable information and sensitive personal information, matching and replacement are performed based on predefined rules. For example, for mobile phone numbers, the first 3 and last 4 digits are retained, and the middle is filled with ****.

[0082] For entities that require association analysis but need to be anonymous, use an irreversible cryptographic hash function to generate unique and anonymous identifiers to ensure consistent anonymous representation of the same entity across different records; use data masking techniques, such as retaining only the surname Zhang* for names or using generalized terms.

[0083] For unstructured text data, the Named Entity Recognition (NER) model is used to identify sensitive information in the text. The identified entities are then standardized, partially masked, or completely deleted according to their sensitivity level.

[0084] Image or video data: Apply deep learning-based detection models to locate sensitive areas, and then perform pixelation, blurring, or masking on these areas to ensure that the processed image or video frame retains the overall contextual information of the scene while making it impossible to identify individual identities. Detection models include face detection and license plate detection, etc.

[0085] Desensitization principles: Adhere to the "minimum necessity principle" and the "data minimization principle", only desensitize necessary sensitive information, retain the core value of data for risk analysis, and all desensitization rules must comply with the mandatory regulations of the state and public security system on police data security;

[0086] The format is unified by adopting a standardized technology stack to achieve consistent conversion of multi-source data. Timestamps are unified to ISO 8601 format (including time zone), geographic coordinates are normalized to WGS84 coordinate system, text encoding is forcibly converted to UTF-8 and basic word segmentation is performed, and images and videos are unified in resolution and color space. Numerical data is standardized or normalized.

[0087] Noise removal cleans up noise by combining a pre-defined rule base with statistical methods. Missing values ​​are ignored, marked, or filled according to the importance of the field to ensure data reliability. Statistical methods include outlier detection and data smoothing, etc.

[0088] Real-time statistics of metrics at each step of the preprocessing process, and setting thresholds to trigger alarms.

[0089] Initialization vectors include text initialization vectors, visual initialization vectors, and spatiotemporal initialization vectors;

[0090] The specific steps for converting police data into initial vectors are as follows:

[0091] The initial text vectors are generated using a BERT model adjusted from police corpora (such as case descriptions and legal documents). Context-aware embedding vectors are extracted from the text data. Let the text input be... Then the text initial vector Represented as:

[0092]

[0093] in, The dimension of the initial text vector;

[0094] The visual initialization vector employs a visual encoder based on the ResNet-50 architecture to extract multi-layer convolutional features from the image data. Let the image input be... Visual initial vector Represented as:

[0095]

[0096] in, The dimension of the initial visual vector;

[0097] The spatiotemporal initial vector is obtained by extracting temporal periodic features and spatial grid encoding from spatiotemporal data, and then mapping them to a spatiotemporal initial vector through a multilayer perceptron (MLP):

[0098]

[0099] in, The dimension of the spacetime initial vector;

[0100] All initial vectors are organized by time window and data source identifier, stored in a vector database (such as Milvus or Weaviate), and a density-based index (HNSW) is built to support subsequent similarity retrieval.

[0101] Event detectors and alert task generation:

[0102] The event detector performs preliminary identification of risk events based on a set of configurable preset rules. These rules are typically relatively simple, explicit, and computationally inefficient logical judgments, such as:

[0103] Keyword or phrase matching: The alarm text contains specific keywords such as "armed", "fight", and "gathering"; the online public opinion contains specific sensitive place names combined with negative emotional words;

[0104] Pattern matching: Multiple minor alarms of the same type occur in the same area within a certain period of time; specific key personnel appear in the controlled area;

[0105] Threshold alarm: Real-time pedestrian density in a specific area exceeds a safety threshold; alarm status of critical infrastructure sensors persists for more than a set time.

[0106] Simple model inference: Combining preset rules and a lightweight statistical model to calculate the probability of a risk event occurring can be expressed as a function based on rule matching terms and prior probability functions.

[0107] P( |D)∝P(D∣ )×P( )

[0108] Wherein, P( |D) represents the posterior probability, based on police data for event D. The probability of occurrence, P(D| Let be the likelihood function, representing the event... Given that the event occurs, the conditional probability of observing the current data D is P( ) represents the prior probability, indicating the probability of an event before the data is observed. The initial probability of occurrence;

[0109] When the probability of a risk event exceeds a preset risk threshold, it is determined to be an initial risk event. The risk threshold is set based on the event risk tolerance, historical data and experience, the cost balance between false positives and false negatives, and industry standards and compliance requirements.

[0110] When the event detector identifies an initial risk event, it generates a structured warning task. The warning task includes at least the event's unique identifier, event type, event occurrence time, core data index associated with the event, and a preliminary description and key features of the event.

[0111] The computation module includes a collaborative scheduling center and a large model cluster. The collaborative scheduling center generates sub-tasks by parsing the early warning task and retrieves and calls the corresponding initial vectors from the vector database according to the sub-tasks. The large model cluster includes a text analysis model, a visual analysis model, and a spatiotemporal prediction model adjusted with data from the police field. The large model cluster processes the initial vectors to generate feature vectors and spatiotemporal attributes.

[0112] The subtasks include text processing requirements, visual processing requirements, and spatiotemporal processing requirements;

[0113] Feature vectors include text feature vectors, visual feature vectors, and spatiotemporal feature vectors;

[0114] The collaborative scheduling center first conducts an in-depth analysis of the task. Based on the event type, preliminary description, and core data index of the event association of the early warning task, it decomposes the specific data processing requirements and classifies these requirements into three sub-tasks: text processing requirements, visual processing requirements, and spatiotemporal processing requirements.

[0115] Subsequently, for each subtask, the scheduling center generates a structured data retrieval instruction, which specifies the key retrieval conditions such as the data type, time range, and spatial range of the required initial vector. This instruction is sent to the vector database, and similarity retrieval and retrieval are performed through HNSW to obtain the text initial vector, visual initial vector, and spatiotemporal initial vector most relevant to the current warning task.

[0116] Finally, the scheduling center accurately allocates the retrieved initial vector set and corresponding processing requirements to the appropriate specialized models within the large model cluster for processing. Please refer to [link to relevant documentation]. Figure 2 ;

[0117] The text analysis model is built using the Transformer-XL architecture, which includes an input layer, a multi-head attention layer, a semantic enhancement layer, and an output layer.

[0118] The initial text vector is input into the input layer for domain-adaptive embedding, as follows:

[0119] Aligning the general semantic space to the policing domain enhances the ability to recognize police terminology and obtains embedding vectors aligned with the policing domain. The formula is as follows:

[0120]

[0121] in, and These are the police terminology mapping matrix and the bias term, respectively, and represent the trainable parameters of the model. The initial vector is the text vector, and LayerNorm is the layer normalization operation;

[0122] The multi-head attention layer captures long-distance semantic dependencies and temporal relationships of alarm events by calculating the relative positional attention of 12 heads, and obtains attention features, as shown in the following formula:

[0123]

[0124] in, , and The first The query matrix, key matrix, and value matrix of the header are the trainable parameters of the model. For the first Head and First The relative position encoding matrix of the heads is generated based on the time difference of the alarm. Here, represents the number of attention heads, and represents the preset hyperparameters. Embedded vector The key or the dimension of the query vector;

[0125] Embedded vector The key or query vector is obtained by embedding the vector. Generate by performing a linear transformation;

[0126] The semantic enhancement layer integrates prior knowledge of crime types to enhance the ability to identify high-risk events, reduce false alarm rates, and obtain risk characteristics. :

[0127]

[0128] in, Crime type vectors are obtained through encoding from a police knowledge base. Here, represents the feature fusion matrix, and represents the trainable parameters of the model. The activation function for the Gaussian error linear unit;

[0129] The output layer fuses attention features and risk features to generate a text feature vector with a unified dimension. :

[0130]

[0131] in, Here, is the dimensionality transformation matrix, and are the trainable parameters of the model. This is the hyperbolic tangent activation function, used to limit the range of eigenvalues.

[0132] The visual analysis model is constructed using a multi-scale attention convolutional network, which specifically includes an input layer, a spatial pyramid layer, an attention layer, and an output layer.

[0133] The initial visual vector is input into the input layer and then reorganized to enhance key spatial features, suppress background interference, and obtain an optimized feature map Z.

[0134]

[0135] in, for The convolution kernel is a set of trainable parameters for the model. As the initial visual vector, To modify the activation function of the linear unit;

[0136] The spatial pyramid layer performs multi-scale feature fusion on the optimized feature map, capturing behavioral patterns at different granularities, such as individual actions and group behaviors, to obtain multi-scale fused features. :

[0137]

[0138] For the average pooling of the s-th downsampling, This is a cross-scale feature splicing operation. This is the preset downsampling ratio;

[0139] The attention layer focuses on areas of abnormal behavior (such as the center of a fight) to obtain a behavioral attention map. :

[0140]

[0141] in, The mask is for human pose estimation, obtained through a pre-trained model; ⊙ represents element-wise multiplication. As the activation function, generate 0-1 attention weights. for Convolution kernel;

[0142] The output layer performs channel weighting and dimension normalization to generate standardized visual feature vectors. :

[0143]

[0144] in, Weighted operations for the channel dimension, Here, is the dimensionality transformation matrix, and are the trainable parameters of the model. This is a channel-level normalization operation.

[0145] The spatiotemporal prediction model is constructed using a spatiotemporal graph convolutional network, which specifically includes an input layer, a propagation layer, an influence layer, and an output layer.

[0146] The spatiotemporal initial vector is input into the input layer to construct a dynamic spatiotemporal graph. :

[0147]

[0148] in, The system is a set of nodes, with the smallest granularity of risk perception entities as nodes, such as a single security camera, covering a radius of 80 meters. Let it be the set of edges;

[0149] The propagation layer alternately performs graph convolution and temporal modeling to simulate the risk diffusion process and capture the evolution patterns of events, thereby obtaining the spatiotemporal state matrix. ;

[0150] Influence layer calculation diffusion radius The formula for predicting the spatial impact range of a risk event is as follows:

[0151]

[0152] in, The core region node index set is a set of spatial grid nodes derived from the core data index associated with events through spatial mapping. MLP stands for Multilayer Perceptron. Operations to extract the maximum risk characteristics in the core area. Nodes of the spacetime graph eigenvectors, spatiotemporal state matrix The kth row;

[0153] The output layer performs graph feature pooling and spatiotemporal attribute parsing to generate spatiotemporal feature vectors. and spacetime attribute triples :

[0154]

[0155] in, The core location, latitude and longitude, is obtained by weighted averaging of the node coordinates on the spatiotemporal map; t represents the duration of the risk event, and is a future prediction. This is a graph-level pooling operation;

[0156] The dimensions of text feature vectors, visual feature vectors, and spatiotemporal feature vectors are unified through a pre-defined first fully connected layer.

[0157] The early warning module calculates a comprehensive risk value based on feature vectors using a weighted fusion algorithm based on an attention mechanism. Based on the early warning threshold of the comprehensive risk value and the spatiotemporal attributes, it generates a multi-level early warning system that includes individual-level real-time early warning, regional-level situational early warning, and trend-level deductive early warning.

[0158] Overall Risk Value The calculation steps are as follows, please refer to [link / reference]. Figure 3 :

[0159] 1) The feature vectors are passed through a pre-defined attention perception layer to generate query vectors, key vectors, and value vectors for text feature vectors, visual feature vectors, and spatiotemporal feature vectors, respectively:

[0160]

[0161] in, , and These are the query vector, key vector, and value vector, respectively. For feature vectors, Representing text, visual, and time and space respectively. , , , , and These are trainable parameters;

[0162] 2) Calculate the attention score of the feature vector by scaling the dot product attention. :

[0163]

[0164] in, This is the transpose of the key vector, used to calculate the dot product. The dimension of the key vector;

[0165] 3) Based on attention score value vector Weighting is performed to generate weighted modal features. ;

[0166] 4) Overall Risk Value The modal features, after aggregation of all weighted features, are passed through a pre-defined second fully connected layer and... The activation function is calculated as follows:

[0167]

[0168] in, As the activation function, compress the output to Interval.

[0169] After obtaining the comprehensive risk value, it is combined with the spatiotemporal attribute triplet. Based on preset warning thresholds and spatiotemporal attributes, the warning thresholds are divided into individual thresholds, regional thresholds, trend thresholds, and range thresholds, generating the following multi-level warnings;

[0170] Individual-level real-time alert: When the overall risk value is greater than or equal to the individual threshold and the event space range is less than or equal to the range threshold, an individual-level real-time alert is activated.

[0171] In this embodiment, the warning content includes the target's identity and real-time location, and is directly pushed to front-line law enforcement personnel through channels such as police terminals and mobile applications;

[0172] Regional-level situation warning: When the comprehensive risk value is greater than or equal to the regional threshold and less than the individual threshold, or when the event space range is greater than the range threshold, a regional-level situation warning is activated.

[0173] In this embodiment, the regional situation warning not only includes the geographical range of the risk area, but also integrates real-time situation information such as population density, traffic conditions, and police deployment, and pushes it to the command center or local police station for resource scheduling and regional control.

[0174] Trend-level simulation early warning: When the comprehensive risk value is greater than or equal to the trend threshold and less than the regional threshold, and the spatiotemporal attributes show that the risk event has the potential to continue or spread;

[0175] In this embodiment, the trend-level prediction and early warning does not focus on immediate response, but rather predicts the risk evolution path based on historical patterns and current characteristics, generates periodic reports, and assists strategic decision-making departments in formulating long-term prevention and control strategies.

[0176] The warning threshold is set based on historical data verification, industry standards and compliance requirements, expert experience and scenario adaptation, and system goal orientation.

[0177] In summary, this invention effectively solves the problems of isolated multi-source heterogeneous data and insufficient feature extraction in traditional methods by using a dedicated large model cluster adjusted from police data to process the corresponding data separately, and innovatively using a weighted fusion algorithm based on an attention mechanism to calculate the comprehensive risk value. This significantly improves the accuracy of risk identification and reduces the false alarm and false negative rates.

[0178] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-level early warning system for police risks based on large-scale model collaborative computation, characterized in that: include: Data module: collects police data in real time, preprocesses the police data, converts the preprocessed police data into initial vectors, and stores the initial vectors in a preset vector database. The initial vectors include text initial vectors, visual initial vectors, and spatiotemporal initial vectors. An event detector is set up to monitor the pre-processed police data in real time and identify initial risk events based on preset risk rules to generate early warning tasks; Computation module: Sets up a collaborative scheduling hub and a large model cluster; The collaborative scheduling center generates sub-tasks by parsing the early warning task, and retrieves and calls the corresponding initial vector from the vector database according to the sub-task; the large model cluster includes a text analysis model, a visual analysis model and a spatiotemporal prediction model adjusted by police domain data, and the large model cluster processes the initial vector to generate feature vectors and spatiotemporal attributes; The feature vectors include text feature vectors, visual feature vectors, and spatiotemporal feature vectors; The text analysis model is built using the Transformer-XL architecture, which includes an input layer, a multi-head attention layer, a semantic enhancement layer, and an output layer. The initial text vector is input into the input layer for domain-adaptive embedding, as detailed below. Aligning the general semantic space to the policing domain enhances the ability to recognize police terminology and obtains embedding vectors aligned with the policing domain. ; The multi-head attention layer calculates the relative positional attention of 12 heads to capture long-distance semantic dependencies and temporal relationships of alarm events, thereby obtaining attention features. The formula is as follows: in, , and The first The query matrix, key matrix, and value matrix of the header. For the first Head and First Head relative position encoding matrix, For the number of attention heads, The dimension of the key or query vector; The semantic enhancement layer integrates prior knowledge of crime types to obtain risk characteristics; The output layer fuses attention features and risk features to generate a text feature vector; The visual analysis model is constructed using a multi-scale attention convolutional network, specifically including an input layer, a spatial pyramid layer, an attention layer, and an output layer. The initial visual vector is input into the input layer, and the optimized feature map Z is obtained by reconstructing the initial visual vector. in, for Convolution kernel, As the initial visual vector, To modify the activation function of the linear unit; The spatial pyramid layer performs multi-scale feature fusion on the optimized feature map, capturing behavioral patterns of different granularities to obtain multi-scale fused features. The attention layer focuses on regions of abnormal behavior to obtain a behavioral attention map; The output layer performs channel weighting and dimension normalization to generate standardized visual feature vectors. The spatiotemporal prediction model is constructed using a spatiotemporal graph convolutional network, specifically including an input layer, a propagation layer, an influence layer, and an output layer. The spatiotemporal initial vector is input into the input layer to construct a dynamic spatiotemporal graph; The propagation layer alternately performs graph convolution and temporal modeling to simulate the risk diffusion process and capture the evolution patterns of events, thereby obtaining a spatiotemporal state matrix. ; The influence layer is used to calculate the diffusion radius. Predict the spatial impact range of risk events; The output layer performs graph feature pooling and spatiotemporal attribute parsing to generate spatiotemporal feature vectors. and spacetime attribute triples : in, The core location, latitude and longitude, and t represent the duration of the risk event. This is a graph-level pooling operation; The dimensions of the text feature vector, visual feature vector, and spatiotemporal feature vector are unified through a preset first fully connected layer; The early warning module calculates a comprehensive risk value based on feature vectors using a weighted fusion algorithm based on an attention mechanism. The comprehensive risk value The calculation steps are as follows: 1) Generate the query vector, key vector, and value vector of the feature vector through a preset attention perception layer: in, , and These are the query vector, key vector, and value vector, respectively. For feature vectors, Representing text, visual, and time and space respectively. , , , , and These are trainable parameters; 2) Calculate the attention score of the feature vector by scaling the dot product attention. : in, This is the transpose of the key vector. The dimension of the key vector; 3) Based on the attention score For the value vector Weighting is performed to generate modal features. ; 4) Overall Risk Value Through the preset second fully connected layer and The activation function is calculated as follows: in, and These are trainable parameters; Based on the warning threshold of the comprehensive risk value and the aforementioned spatiotemporal attributes, a multi-level early warning system is generated, including individual-level real-time early warning, regional-level situational early warning, and trend-level deductive early warning.

2. The system according to claim 1, characterized in that, The initial text vector is obtained by using the BERT model to perform context-aware acquisition of the text data. The visual initial vector is obtained by extracting multi-layer convolutional features from the image data using a visual encoder based on the ResNet-50 architecture. The spatiotemporal initial vector is obtained by extracting time period features and spatial grid encoding from spatiotemporal data, and then mapping it through a multilayer perceptron. The initial vectors are organized according to a preset time window and data source identifier, stored in the vector database, and a density-based index is established to support similarity retrieval.

3. The system according to claim 2, characterized in that, The event detector identifies the initial risk event based on a set of configurable preset rules; When the event detector identifies an initial risk event, it generates the early warning task. The early warning task includes at least a unique event identifier, event type, event occurrence time, core data index associated with the event, and a preliminary description and key features of the event.

4. The system according to claim 3, characterized in that, The sub-tasks include text processing requirements, visual processing requirements, and spatiotemporal processing requirements. The collaborative scheduling center decomposes specific data processing requirements based on the event type, preliminary description, and core data index of the event association of the early warning task, and classifies these requirements into three sub-tasks: text processing requirements, visual processing requirements, and spatiotemporal processing requirements. The collaborative scheduling center will assign the retrieved initial vector and corresponding subtask to the appropriate professional models in the large model cluster for processing.

5. The system according to claim 4, characterized in that, The warning thresholds include individual thresholds, regional thresholds, trend thresholds, and range thresholds; The individual-level real-time early warning is activated when the overall risk value is greater than or equal to the individual threshold and the diffusion radius is less than or equal to the range threshold. When the overall risk value is greater than or equal to the regional threshold and less than the individual threshold, or when the diffusion radius is greater than the range threshold, a regional-level situation warning is initiated. When the overall risk value is greater than or equal to the trend threshold and less than the regional threshold, and the spatiotemporal attributes indicate that the risk event has the potential to continue or spread, a trend-level inference warning is initiated.

Images