A large model-based police case whole-process intelligent processing method and system

Abstract

This invention relates to the field of police information processing technology, and discloses a method and system for intelligent processing of the entire police incident process based on a large model. The method includes: acquiring raw police incident input data to generate structured police incident data; performing entity alignment and association matching between the structured police incident data and preset police data resources to generate multi-dimensional feature data; constructing a police incident association graph based on the multi-dimensional feature data, and calling a large model to perform risk reasoning on the police incident association graph to obtain risk assessment results; generating handling suggestions and collaborative handling objects corresponding to the warning level based on the risk assessment results and preset risk scoring rules; pushing the handling suggestions to the collaborative handling objects, collecting handling feedback data, and updating the police incident association graph, preset risk scoring rules, and large model calling parameters based on the handling feedback data. This invention can improve the efficiency of structured police incident processing, the accuracy of risk assessment, and the collaborative handling capability.

Description

Technical Field

[0001] This invention relates to the field of police information processing technology, and in particular to a method and system for intelligent processing of the entire police incident process based on a large model. Background Technology

[0002] During the police response process, incident information typically originates from multiple data sources, including caller ID, alarm text, images, videos, location information, and existing police data. This data is characterized by inconsistent formats, complex semantic expressions, and high real-time requirements. Therefore, it needs to undergo element extraction, data correlation, risk assessment, and dispatch processing before it can provide effective support for police command.

[0003] While existing police incident handling systems can record, transfer, and store incidents, they still rely heavily on human experience for incident element extraction, risk assessment, and tiered response in practical applications. On one hand, unstructured data such as call recordings and alarm texts are difficult to convert quickly and accurately into structured incident data, leading to incomplete extraction of key information or inconsistent standards. On the other hand, existing systems lack the ability to automatically correlate multi-source police data, including population databases, case databases, vehicle databases, key personnel databases, key location databases, and historical incident databases, making it difficult to promptly identify duplicate incidents, related individuals, related locations, and potential risk chains.

[0004] Furthermore, traditional methods of police incident analysis are typically based on fixed rules or manual judgment, making it difficult to comprehensively reason based on multi-dimensional characteristics such as personnel risk, location sensitivity, event patterns, and spatiotemporal evolution trends. They also struggle to continuously optimize risk scoring rules and analysis models based on feedback from responses. Therefore, it is necessary to provide a fully intelligent police incident processing method and system capable of automated structuring of incidents, multi-source correlation analysis, intelligent risk assessment, tiered early warning and response, and feedback optimization. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for intelligent processing of the entire alarm process based on a large model, aiming to solve the problems of existing technologies.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for intelligent processing of the entire alarm process based on a large model, including: Obtain raw police incident input data, process the raw police incident input data, extract police incident elements, and generate structured police incident data based on the police incident elements; The structured police incident data is aligned and matched with preset police data resources to generate multi-dimensional feature data corresponding to the current police incident. Based on the multidimensional feature data, a crime association graph of the current crime is constructed, and a large model is called to perform risk reasoning on the crime association graph to obtain risk assessment results; Based on the risk assessment results and preset risk scoring rules, the risk score and warning level of the current alarm situation are determined, and corresponding handling suggestions and collaborative handling targets are generated. The proposed handling suggestions are pushed to the collaborative handling objects, and handling feedback data is collected. The alarm correlation graph, the preset risk scoring rules, and / or the calling parameters of the large model are updated based on the handling feedback data.

[0007] In one optional approach, the processing of the original alarm input data includes: the original alarm input data including alarm voice data, alarm text data, image data, video data and / or location data; performing real-time speech recognition on the alarm voice data to obtain alarm transcription text; filtering redundant sentences and semantically completing the alarm transcription text to obtain standardized alarm text; and calling a large language model to perform semantic parsing on the standardized alarm text to extract alarm elements including personnel, events, time, location, items, causes, and results.

[0008] In one optional approach, generating structured police incident data based on the incident elements includes: matching the address information in the incident elements with a standard address database to obtain standardized location information; performing field normalization processing on the personnel, vehicle, item, and event descriptions in the incident elements; and performing deduplication, error correction, and field mapping on the normalized incident elements to generate structured police incident data for incident analysis.

[0009] In one optional approach, the step of aligning and matching the structured police incident data with preset police data resources includes: matching the personnel entities, location entities, vehicle entities, item entities, and event entities in the structured police incident data with data in population databases, case databases, vehicle databases, key personnel databases, key location databases, and historical police incident databases, respectively; determining historical police incidents, associated personnel, associated vehicles, associated locations, and associated events related to the current police incident based on the matching results; and generating the multidimensional feature data based on the historical police incidents, associated personnel, associated vehicles, associated locations, and associated events.

[0010] In one optional approach, the multidimensional feature data includes at least two of the following: personnel risk features, location sensitivity features, event pattern features, and spatiotemporal evolution features; wherein, the personnel risk features are used to characterize the historical alarms, historical cases, key personnel tags, and / or repeated associations of personnel related to the incident; the location sensitivity features are used to characterize the key location attributes, historical incident frequency, and / or surrounding risk status of the incident location; the event pattern features are used to characterize the semantic similarity, type similarity, and / or similarity of handling results between the current incident and historical incidents; and the spatiotemporal evolution features are used to characterize the degree of aggregation, diffusion trend, and / or recurrence trend of the current incident in the time and space dimensions.

[0011] In one optional approach, the step of constructing an alarm association graph based on the multidimensional feature data and using a large model to perform risk reasoning on the alarm association graph includes: constructing the alarm association graph using personnel, locations, events, items, and organizations as graph nodes, and alarm relationships, case relationships, location relationships, vehicle relationships, item relationships, and historical handling relationships as graph edges; retrieving similar alarms matching the current alarm from a historical alarm feature database based on vector similarity; and inputting the alarm association graph, the similar alarms, and the multidimensional feature data into the large model to obtain risk assessment results that characterize the probability of repeated alarms, risk evolution paths, and key hidden danger information.

[0012] In one optional approach, determining the risk score and warning level of the current alarm based on the risk assessment results and preset risk scoring rules includes: calculating the risk score of the current alarm based on at least two of the following: the degree of harm of the alarm, the scope of impact, the frequency of historical occurrence, the risk weight of personnel, the sensitivity weight of the location, the emotional characteristics of the alarm, and the characteristics of the risky items; comparing the risk score with a preset classification threshold to determine the red warning level, yellow warning level, or blue warning level corresponding to the current alarm.

[0013] In one optional approach, generating handling suggestions and collaborative handling targets corresponding to the warning level includes: retrieving a handling plan fragment matching the current incident based on the warning level, the risk assessment result, and a preset police handling knowledge base; calling a large model to fuse the handling plan fragment and the multi-dimensional feature data of the current incident to generate handling suggestions including at least two of the following: verification points, handling steps, control points, key points of inspection, and precautions; and determining at least one of the following as collaborative handling targets: the command center, the local police station, specialized police units, and mobile police terminals, based on the location information, event type, and risk level of the current incident.

[0014] In one alternative approach, the method is executed collaboratively by multiple intelligent agents, including an element analysis agent, a data association agent, a multi-dimensional judgment agent, a scoring and early warning agent, a handling and feedback agent, and a model enhancement agent. These multiple intelligent agents distribute tasks, transmit status updates, and share results via a message bus, enabling structured processing tasks, data association tasks, risk assessment tasks, early warning push tasks, and feedback optimization tasks to be executed serially and / or in parallel.

[0015] Secondly, this application provides an intelligent emergency response system based on a large model, comprising: The alarm input module is used to acquire raw alarm input data, which includes alarm voice data, alarm text data, image data, video data and / or location data; The structured processing module is used to identify, transform, and semantically parse the original police incident input data, extract police incident elements, and generate structured police incident data based on the police incident elements. The data association module is used to perform entity alignment and association matching between the structured police incident data and preset police data resources to generate multi-dimensional feature data corresponding to the current police incident. The intelligent analysis module is used to construct an alarm correlation graph of the current alarm based on the multi-dimensional feature data, and call a large model to perform risk reasoning on the alarm correlation graph to obtain risk analysis results; The early warning and response module is used to determine the risk score and early warning level of the current alarm based on the risk assessment results and preset risk scoring rules, and to generate response suggestions and collaborative response targets corresponding to the early warning level; The feedback optimization module is used to push the handling suggestions to the collaborative handling objects, collect handling feedback data, and update the alarm association map, the preset risk scoring rules, and / or the calling parameters of the large model based on the handling feedback data.

[0016] Through the above technical solutions, the beneficial effects of this invention are as follows: By processing the original police incident input data, structured police incident data is generated, reducing the reliance on manual input and judgment; by aligning and matching the structured police incident data with preset police data resources, the current police incident is no longer treated as an isolated data point; by constructing a police incident association graph based on multi-dimensional feature data and calling a large model for risk reasoning, the accuracy and comprehensiveness of police incident risk assessment are improved; by determining the risk score and warning level based on the risk assessment results and preset risk scoring rules, and generating corresponding handling suggestions and collaborative handling targets, the classification and dispatch of police incidents can have a unified quantitative basis, improving the pertinence of handling suggestions and the efficiency of multi-department collaborative handling. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, 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 the present invention, and the embodiments in the accompanying drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating an intelligent emergency response method based on a large model, provided as an embodiment of this application; Figure 2 This application provides a schematic diagram of a process for processing and generating structured alarm input data, as an embodiment of the present application. Figure 3 This application provides a schematic diagram of a process for associating and matching structured police incident data with preset police data resources, as an embodiment of the present application. Figure 4 A schematic diagram of a police incident association graph construction and risk reasoning process provided in an embodiment of this application; Figure 5 A schematic diagram of an alarm correlation map provided in one embodiment of this application; Figure 6 A schematic diagram illustrating a risk scoring, graded early warning, and response suggestion generation process provided in an embodiment of this application; Figure 7 This is a schematic diagram of a multi-agent cooperative execution architecture provided in one embodiment of this application; Figure 8 This is a schematic diagram of the structure of an intelligent emergency response system based on a large model, provided as an embodiment of this application.

[0019] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the embodiments of the present invention and the technical features thereof can be combined with each other.

[0021] In this embodiment of the invention, the original police report input data may include alarm voice data, alarm text data, image data, video data, location data, and other data that can reflect the content of the police report. Pre-set police data resources may include population databases, case databases, vehicle databases, key personnel databases, key location databases, historical police report databases, police business databases, equipment databases, police handling knowledge bases, and other data resources set according to actual business scenarios. The large model may be a large language model, a multimodal large model, or a model cluster composed of a speech recognition model, a natural language processing model, an image recognition model, a vector retrieval model, and an inference model. The calling parameters of the large model may include at least one of the following: prompt word template, search scope, context window length, output format constraints, risk threshold, model confidence threshold, temperature parameter, calling priority, and model routing rules.

[0022] Example 1 This embodiment provides a method for intelligent processing of the entire police incident process based on a large model. This method can be executed by a server, a police platform, a command and dispatch platform, or a computing device deployed with intelligent police incident processing programs. The computing device may include a processor, a memory, a communication module, and a data interface. The memory stores program instructions executable by the processor, which, when executed, implement the method described in this embodiment.

[0023] like Figure 1 As shown, the method includes the following steps.

[0024] S101: Obtain the original police incident input data, process the original police incident input data, extract the police incident elements, and generate structured police incident data based on the police incident elements.

[0025] In this step, the original police report input data can come from emergency calls, online emergency platforms, mobile police terminals, video surveillance systems, emergency call order systems, location systems, and other police operational systems. For emergency call voice data, the system can access the real-time conversation between the dispatcher and the caller via a streaming interface and use a speech recognition model to convert the voice data into transcripts in real time. For emergency text data, the system can directly read the text content or perform character correction, sentence segmentation, and semantic standardization. For image and video data, the system can identify people, vehicles, locations, objects, actions, and abnormal behaviors using image recognition or video analysis models. For location data, the system can obtain the location of the emergency terminal, base station location, map coordinates, or the location of the incident recorded by the police system.

[0026] After obtaining processable text or recognition results, the system calls a large language model to perform semantic analysis on the original police report input data and extract police report elements. These elements can include personnel, event, time, location, item, cause, and result. They can also be expanded to include contact information, vehicle information, scene status, hazardous materials, personnel emotions, casualties, amount involved, and number of alarms, depending on actual business needs. These personnel, event, time, location, item, cause, and result elements constitute the "seven W's" elements, used to describe the basic factual relationships of the current police report.

[0027] After extracting the key elements of a police report, the system writes these elements into structured fields according to preset field templates, generating structured police report data. For example, the system writes the caller's name, ID number, contact information, related persons, time of incident, location of incident, type of incident, vehicle involved, items involved, cause of incident, current outcome, the caller's emotional state, and on-site hazards into corresponding fields. This method transforms previously conversational, unstructured, or semi-structured police report information into structured data that is easy to calculate, retrieve, and analyze.

[0028] S102: Perform entity alignment and association matching between the structured police incident data and preset police data resources to generate multi-dimensional feature data corresponding to the current police incident.

[0029] In this step, the system first identifies entity information in the structured police incident data. This entity information can include personnel entities, location entities, vehicle entities, item entities, event entities, and organization entities. Subsequently, the system matches these entity entities with preset police data resources. For example, it matches personnel entities with the population database, key personnel database, and historical case-related personnel data; it matches vehicle entities with the vehicle database and case-related vehicle database; it matches location entities with the standard address database, key location database, and historical police incident location database; and it matches event entities with event types in the case database, historical police incident database, and police business database.

[0030] During entity alignment, the system can employ one or more of the following methods: exact matching, fuzzy matching, address standardization matching, alias matching, identity identifier matching, and vector similarity matching. For example, for address information, the system first converts colloquial addresses into standard addresses and then matches them with a standard address database; for personnel information, the system can perform comprehensive matching based on name, ID number, mobile phone number, alarm number, historical associations, etc.; for event descriptions, the system can convert event text into event vectors and calculate similarity with event vectors in a historical alarm database.

[0031] Based on the matching results, the system identifies historical incidents, associated personnel, associated vehicles, associated locations, and associated events related to the current incident, and further generates multidimensional feature data. This multidimensional feature data can include personnel risk characteristics, location sensitivity characteristics, event pattern characteristics, and spatiotemporal evolution characteristics. Specifically, personnel risk characteristics can characterize whether related personnel have historical alarms, historical cases, key personnel tags, repeated alarms, or cross-regional associations; location sensitivity characteristics can characterize whether the incident location is a key location, a densely populated location, a historically high-incidence location, or a sensitive area; event pattern characteristics can characterize the semantic similarity, type similarity, handling outcome similarity, and behavioral pattern similarity between the current incident and historical incidents; and spatiotemporal evolution characteristics can characterize the degree of clustering, diffusion trend, recurrence trend, and cross-regional spread trend of the current incident in both time and space dimensions.

[0032] Through this step, the system can automatically associate a single police report with multi-source police data resources, so that the current police report no longer exists as an isolated work order, but forms a multi-dimensional feature data system with personnel background, location background, event background and spatiotemporal background.

[0033] S103: Construct an alarm correlation graph of the current alarm based on the multidimensional feature data, and call the large model to perform risk reasoning on the alarm correlation graph to obtain risk assessment results.

[0034] In this step, the system constructs a crime association graph for the current incident based on multi-dimensional feature data. Specifically, it can use people, locations, events, items, and organizations as graph nodes, and alarm relationships, case relationships, location relationships, vehicle relationships, item relationships, organizational relationships, and historical handling relationships as graph edges to construct the graph structure corresponding to the current incident. For example, a location relationship is formed between the person who reported the incident and the location of the incident; an alarm relationship is formed between the person who reported the incident and historical incidents; a vehicle relationship is formed between the vehicle involved and related personnel; an event pattern similarity relationship is formed between the current event and historical events; and a historical handling relationship is formed between the handling unit and historical handling results.

[0035] After constructing the police incident association graph, the system can retrieve similar police incidents matching the current incident from the historical police incident feature database based on vector similarity. Specifically, the system can encode the structured text, multi-dimensional feature data, and graph summary of the current incident into vectors and calculate the similarity with the incident vectors in the historical police incident feature database. If the similarity exceeds a preset similarity threshold, the corresponding historical incident is identified as a similar incident. Similar incidents can be used to determine whether the current incident is a duplicate incident, multiple reports from the same person, multiple occurrences of the same incident, multiple occurrences in the same location, or consecutive occurrences of the same type of event.

[0036] Subsequently, the system inputs the police incident correlation graph, similar incidents, and multi-dimensional feature data into the large model, and calls the large model to perform risk reasoning. The large model can output risk assessment results based on the relationships between graph nodes, historical similar cases, police handling knowledge, and current incident elements. These risk assessment results may include the probability of repeated incidents, risk evolution paths, key hidden danger information, explanations of risk causes, potential impact scope, key personnel or locations that may be involved, and suggested areas for investigation. Through the semantic understanding and logical reasoning capabilities of the large model, the system can identify complex risk patterns that are difficult to cover using traditional fixed rules.

[0037] S104: Determine the risk score and warning level of the current alarm based on the risk assessment results and preset risk scoring rules, and generate handling suggestions and collaborative handling objects corresponding to the warning level.

[0038] In this step, the system calculates the risk score of the current alarm based on the risk assessment results and preset risk scoring rules. The preset risk scoring rules may include scoring factors such as the severity of the alarm, the scope of impact, historical frequency, personnel risk weight, location sensitivity weight, alarm emotional characteristics, characteristics of risky items, probability of recurring alarms, speed of risk evolution, and on-site danger level. Each scoring factor can be configured with a corresponding weight, and the system calculates the risk score of the current alarm based on the values ​​and weights of each scoring factor.

[0039] As an example, risk scores can be set from 0 to 100. If the current incident involves violent injury, hazardous chemicals, personal injury, or large-scale public safety risks, the weight of the incident severity factor can be increased; if the current incident occurs in a school, hospital, station, business district, key unit, or historically high-incidence location, the weight of location sensitivity can be increased; if the same person, the same location, or the same event occurs multiple times in a short period of time, the scores corresponding to historical frequency and probability of repeated incidents can be increased; if the alarm voice or alarm text contains strong expressions of panic, threat, or distress, the score corresponding to the alarm emotional characteristics can be increased.

[0040] The system compares the risk score with preset grading thresholds to determine the corresponding warning level for the current incident. For example, an incident with a risk score greater than or equal to 80 points can be classified as a red warning level, an incident with a risk score greater than or equal to 60 points but less than 80 points can be classified as a yellow warning level, and an incident with a risk score less than 60 points can be classified as a blue warning level. The above thresholds are only examples and can be adjusted according to the characteristics of local incidents, police resource allocation, and business rules in actual deployment.

[0041] After determining the warning level, the system generates corresponding handling suggestions and collaborative handling targets. Specifically, the system can retrieve handling plan fragments matching the current situation based on the warning level, risk assessment results, and a preset police handling knowledge base; then, it calls a large model to fuse the handling plan fragments and the multi-dimensional feature data of the current situation to generate handling suggestions. The handling suggestions may include key verification points, handling steps, control points, key points of inspection, personnel protection measures, evidence preservation requirements, on-site safety precautions, police force allocation suggestions, and follow-up visit suggestions, etc.

[0042] Meanwhile, the system determines the entities to coordinate with based on the location, type, and risk level of the current incident. These entities can include the command center, local police station, specialized police units, mobile emergency response terminals, grassroots units, and relevant social forces. For example, for high-risk violent incidents, the system can suggest joint action by the command center, local police station, and relevant specialized police units; for emergencies at key locations, the system can simultaneously push notifications to the local police station, location management unit, and mobile emergency response terminals; and for low-risk requests for assistance, the system can push notifications to grassroots units or relevant departments for handling.

[0043] S105: Push the handling suggestions to the collaborative handling object, collect handling feedback data, and update the alarm association map, the preset risk scoring rules and / or the calling parameters of the large model according to the handling feedback data.

[0044] In this step, the system combines the handling suggestions, warning level, risk assessment results, and key elements of the current situation into a decision package, which is then pushed to the corresponding collaborative handling targets via a messaging mechanism. This messaging mechanism can include messages from the police platform, mobile police terminals, command and dispatch systems, SMS, voice reminders, or other internal communication methods. After receiving the decision package, the collaborative handling targets can perform verification, dispatching officers, sealing off areas, conducting checks, coordinating actions, or other handling operations based on the handling suggestions.

[0045] During the handling process, the system collects handling feedback data. This feedback data may include handling progress, arrival time, personnel involved, on-site verification results, supplementary alarm information, handling measures, handling results, whether the situation was escalated, whether there was a false alarm, whether there was a duplicate alarm, whether a case was formed, handling time, and manual review opinions. If the system detects delays in handling progress, missing feedback, or escalation of alarm risk, it can automatically trigger a supervision, re-push, or reassessment process.

[0046] After a police incident is closed, the system performs closed-loop optimization based on the feedback data. Specifically, the system can update the incident association graph based on newly confirmed personnel relationships, location relationships, event relationships, and handling results; it can adjust preset risk scoring rules based on actual handling results, such as adjusting the risk weights corresponding to certain types of locations, events, or personnel tags; and it can optimize the calling parameters of the large model based on the feedback data, such as updating prompt word templates, adjusting the search scope, correcting output format constraints, optimizing similar incident retrieval strategies, or adjusting model routing rules. Thus, the system forms a closed-loop mechanism of "incident input—structured processing—intelligent analysis—tiered handling—feedback optimization," enabling the intelligent handling capability of police incidents to continuously improve with the accumulation of practical handling experience.

[0047] Example 2 like Figure 2 As shown in the illustration, this embodiment further explains the processing and structured generation of raw alarm input data. The system first collects raw alarm input data. Raw alarm input data may include alarm voice data, alarm text data, image data, video data, and / or location data. In practical applications, alarm voice data typically includes continuous dialogue between the caller and the dispatcher, which may contain repetitive expressions, interjections, emotional language, dialectal vocabulary, and supplementary information. Alarm text data may come from network alarms, SMS alarms, mobile alarms, or police work orders, and may contain non-standard address expressions, inconsistent titles, and colloquial descriptions of events. Image and video data may come from on-site footage, monitoring equipment, or mobile terminals; location data can be used to help determine the location of the alarm.

[0048] For alarm voice data, the system performs real-time speech recognition using a speech recognition model to obtain the transcribed text. The speech recognition model can perform frame segmentation, noise reduction, acoustic feature extraction, and text transcription of the call audio, continuously outputting the transcription results during the call. To improve the quality of subsequent parsing, the system can perform speaker separation on the transcribed text to distinguish between the dispatcher's speech and the caller's speech; it can also segment the transcribed text based on speech pauses, semantic boundaries, and keyword recognition results.

[0049] After receiving the transcribed text of the emergency call, the system performs redundant sentence filtering and semantic completion to obtain a standardized emergency call text. Redundant sentence filtering can include removing meaningless interjections, repeated confirmation statements, irrelevant chatter, and obvious errors in identification. Semantic completion can include completing omitted subjects, time expressions, location hierarchy, interpersonal relationships, and event outcomes based on context. For example, if the caller states, "He's here again, causing trouble downstairs," the system can combine the interpersonal and location information in the context to complete it as "A related person is causing disturbance downstairs in a certain residential area."

[0050] Subsequently, the system invokes a large language model to perform semantic analysis on the standardized police report text, extracting elements such as personnel, events, time, location, items, causes, and results. Personnel elements can include the person reporting the incident, the person being reported, the victim, the suspect, related persons, and their identity information; event elements can include event types such as disputes, requests for assistance, theft, fights, disturbances, fraud, and traffic accidents; time elements can include the time of the report, the time of the incident, the duration, and the recurrence time; location elements can include standard addresses, coordinates, buildings, floors, room numbers, and surrounding landmarks; item elements can include vehicles, tools, dangerous goods, property, and evidence; cause elements can include the cause of the dispute, the triggering cause, the cause of historical conflicts, and the cause of the report; and result elements can include casualties, property damage, the state of the scene, whether the situation is under control, and whether the incident is still ongoing.

[0051] After extracting the alarm elements, the system matches the address information within these elements against a standard address database to obtain standardized location information. Specifically, the system can break down colloquial addresses into administrative divisions, roads, residential areas, buildings, units, house numbers, and coordinates, and then match these against the standard address database. If multiple candidate addresses exist, the system can combine location data, the alarm number's origin, historical alarm locations, and contextual descriptions to determine the most probable standard address.

[0052] The system also performs field normalization on the personnel, vehicle, item, and event descriptions within the crime report elements. For example, it converts license plate numbers in different formats to a unified format, maps event types with different descriptions to a unified crime report category, unifies personnel titles, aliases, or relational terms to the personnel entity field, and maps item names to preset item categories. The system further performs deduplication, error correction, and field mapping on the normalized crime report elements to generate structured crime report data for crime report analysis. This provides a standardized data foundation for subsequent data association, graph construction, and risk scoring.

[0053] Example 3 like Figure 3As shown in the illustration, this embodiment describes the process of associating and matching structured police incident data with preset police data resources. After acquiring structured police incident data, the system identifies personnel entities, location entities, vehicle entities, item entities, and event entities. For each type of entity, the system can construct a corresponding entity identifier. For example, personnel entities can use name, ID number, contact information, alarm number, facial features, or historical association number as entity identifiers; location entities can use standard address number, latitude and longitude coordinates, or location number as entity identifiers; vehicle entities can use license plate number, vehicle identification code, vehicle model color, or historical trajectory number as entity identifiers; and event entities can use event type, event semantic vector, occurrence time, and occurrence location as composite identifiers.

[0054] The system matches the aforementioned entities against data in the population database, case database, vehicle database, key personnel database, key location database, and historical police report database. Matching methods can include exact field matching, fuzzy matching, rule-based matching, semantic similarity matching, and graph adjacency matching. For example, when matching a personnel entity against the population database, strong identifiers such as ID card number and mobile phone number can be prioritized; if strong identifiers are missing, weak identifiers such as name, age, address, and relationship description can be used for combined matching. When matching a location entity against the key location database, it can determine whether it falls within the key location range using standard address numbers or spatial distance calculations. When matching an event entity against the historical police report database, it can calculate the text vector similarity between the current event description and historical event descriptions, and combine temporal and spatial distances to determine the degree of similarity.

[0055] Based on the matching results, the system identifies historical incidents, associated individuals, vehicles, locations, and events related to the current incident. If the current caller, the person reported, or the location involved appears multiple times in the historical incident database, the system can mark it as a duplicate association. If the current incident involves key individuals, vehicles, or locations, the system can write the corresponding tags into the multi-dimensional feature data. If the current event is highly similar to historical events in terms of semantics, location, and time, the system can mark it as a similar incident or a candidate for a duplicate incident.

[0056] When generating multidimensional feature data, the system can generate at least personnel risk characteristics, location sensitivity characteristics, event pattern characteristics, and spatiotemporal evolution characteristics. Personnel risk characteristics can include the number of historical alarms, the number of historical cases, key personnel tags, risk level, number of cross-regional associations, and recent activity levels of relevant personnel. Location sensitivity characteristics can include whether the incident location is a key location, the frequency of historical alarms at that location, the number of surrounding key units, the density of people in the surrounding area, and the density of similar recent alarms. Event pattern characteristics can include the current alarm type, the number of similar historical alarms, the handling results of similar alarms, the semantic similarity of the event, and the probability of escalation. Spatiotemporal evolution characteristics can include the alarm growth trend of the same location within a preset time window, the alarm trajectory of the same person in different locations, the diffusion trend of similar alarms in adjacent areas, and the recurrence of the same event in multiple time periods.

[0057] Through the above processing, the system can construct four core feature databases: personnel risk, location sensitivity, event patterns, and spatiotemporal evolution. These four core feature databases can be used for current police situation analysis, and can also be updated based on feedback data after a police situation is closed, thus forming a continuously accumulating digital asset for policing.

[0058] Example 4 like Figure 4 and Figure 5 As shown, this embodiment illustrates the process of constructing a police incident association graph and inferring risk. After acquiring multi-dimensional feature data, the system constructs a police incident association graph for the current incident, using personnel, locations, events, items, and organizations as graph nodes, and alarm relationships, case relationships, location relationships, vehicle relationships, item relationships, and historical handling relationships as graph edges. Graph nodes may include the current alarm caller node, the alarm recipient node, related personnel nodes, the incident location node, historical incident nodes, involved vehicle nodes, involved item nodes, handling unit nodes, and related organization nodes. Graph edges may include the alarm relationship of the alarm caller initiating the current incident, the location relationship between personnel and locations, the vehicle relationship between personnel and vehicles, the similarity relationship between the current incident and historical incidents, and the historical handling relationship between the handling unit and historical incidents.

[0059] When constructing a crime incident correlation graph, the system can configure attribute information for nodes and edges. Node attributes can include personnel identity information, personnel risk tags, location sensitivity level, event type, item type, organization type, historical frequency, etc. Edge attributes can include relationship type, relationship confidence, relationship occurrence time, relationship source, similarity value, historical handling results, etc. By configuring node and edge attributes, the system can express the correlation relationships of current crime incidents at a more granular level.

[0060] The system retrieves similar police reports matching the current report from a historical police report feature database based on vector similarity. Specifically, the system encodes the structured fields, report summary, event description, multidimensional feature data, and graph relationship summary of the current report into a current report vector; calculates the similarity between the current report vector and historical report vectors in the historical report feature database; and retrieves several similar reports based on the similarity ranking. Similarity calculation can use cosine similarity, Euclidean distance, dot product similarity, or other vector similarity algorithms. To improve retrieval accuracy, the system can also use temporal distance, spatial distance, event type consistency, and personnel overlap as recall filtering conditions.

[0061] Upon receiving similar alerts, the system inputs the alert correlation graph, similar alerts, and multi-dimensional feature data into the large model. The information input into the large model can include a current alert summary, alert elements, a list of correlation graph nodes, a list of correlation graph edges, similar alert summaries, similar alert handling results, a list of risk features, and handling knowledge fragments. To ensure the output is easy for the system to process, the system can constrain the large model to output in a preset format using prompt word templates, such as outputting the probability of repeated alerts, risk evolution paths, key hidden danger information, suggested targets for attention, explanations of risk causes, and confidence levels.

[0062] When performing risk reasoning, the large-scale model can combine personnel associations, location associations, event associations, and historical handling relationships in the data map to determine the potential risks of the current incident. For example, when the same person is involved in similar incidents in multiple jurisdictions, and all related incidents are associated with the same vehicle or location, the large-scale model can infer the risk of cross-regional repeated incidents; when multiple similar incidents occur at a key location within a short period of time, and the alarm text contains descriptions of panic, escalation of conflict, or dangerous materials, the large-scale model can infer a risk diffusion trend; when the current incident is highly similar to historical high-risk incidents in terms of event patterns and handling outcomes, the large-scale model can highlight key hidden dangers and possible evolution paths.

[0063] Finally, the system outputs risk assessment results. These results can include not only quantitative findings, such as the probability of repeat incidents and risk scoring recommendations, but also qualitative findings, such as the causes of the risk, the direction of risk evolution, key areas for investigation, and recommended handling strategies. These risk assessment results are used to generate subsequent risk scores, early warning levels, and handling recommendations.

[0064] Example 5 like Figure 6As shown in the illustration, this embodiment explains the process of generating risk scores, tiered early warnings, and response suggestions. The system first obtains the risk assessment results. These results may include the probability of repeated incidents, risk evolution paths, key hidden danger information, key targets of concern, response results of similar incidents, and risk descriptions output by the large model. The system calculates the risk score of the current incident based on at least two of the following: the severity of the incident, the scope of impact, historical frequency of occurrence, personnel risk weight, location sensitivity weight, alarm emotion characteristics, and characteristics of risky items.

[0065] As an optional implementation, the system can use a weighted scoring model to calculate the risk score. The weighted scoring model can be expressed as follows: the risk score equals the sum of the standardized values ​​of each scoring factor and their corresponding weights. Scoring factors may include factors such as the severity of the incident, the scope of impact, historical frequency, personnel risk, location sensitivity, alarm sentiment, and risky items. The values ​​of each scoring factor can be determined based on preset rules, historical statistical results, or the output of a large model. The weights can be pre-configured by business rules or dynamically adjusted based on feedback data from the handling of incidents.

[0066] For example, for alarms involving personal injury, weapons, hazardous chemicals, mass gatherings, fires, and explosions, the alarm severity factor is set higher; for alarms that may affect the order of places such as schools, hospitals, stations, commercial centers, and key units, the impact range factor is set higher; for alarms involving the same location, the same person, or the same event occurring multiple times in a short period of time, the historical frequency factor is set higher; for alarms involving key personnel, individuals with historical cases, or individuals with high-risk labels, the personnel risk factor is set higher; for alarms occurring in key locations or historically high-incidence areas, the location sensitivity factor is set higher; for alarms with strong panic, threats, cries for help, or loss of control in the alarm voice or alarm text, the alarm emotion factor is set higher; for alarms involving dangerous goods, controlled instruments, flammable and explosive materials, or other high-risk materials, the risk material factor is set higher.

[0067] The system compares the risk score with preset grading thresholds to determine the warning level corresponding to the current alarm. Warning levels can include red, yellow, and blue alert levels. A red alert corresponds to a high-risk alarm, indicating that rapid coordination between the command center, local units, and specialized police units is required; a yellow alert corresponds to a medium-risk alarm, indicating that local units need to promptly verify the situation and escalate it as needed; and a blue alert corresponds to a general-risk alarm, indicating that it should be handled according to standard procedures with follow-up monitoring. The warning levels can also be further expanded to include more levels based on actual operational needs.

[0068] After determining the warning level, the system retrieves relevant emergency response plan segments matching the current emergency situation based on the warning level, the risk assessment results, and a pre-set police response knowledge base. The pre-set police response knowledge base may include handling procedures for different types of emergencies, historical best practices, command and dispatch rules, departmental responsibility lists, on-site safety precautions, and collaborative rules. Search methods may include keyword search, tag matching, vector search, and graph relationship search. For example, for violent conflict emergencies, the system can retrieve plans related to personnel isolation, on-site control, evidence preservation, casualty assistance, and follow-up questioning; for emergencies involving key locations, the system can retrieve plans related to location lockdown, personnel evacuation, perimeter security, and local coordination.

[0069] The system calls upon a large model to fuse fragments of the emergency response plan and multi-dimensional feature data of the current incident to generate response suggestions. These suggestions can include at least two of the following: verification points, response steps, control points, key areas for investigation, and precautions. For example, verification points could include verifying the identity of the person reporting the incident, verifying the relationships between the individuals involved, verifying whether there is still danger at the scene, and verifying whether similar repeated incidents exist. Response steps could include initial telephone follow-up, dispatching officers to the scene, isolating the scene, questioning personnel, securing evidence, and reporting an escalation of risk. Control points can be determined based on location information and map data. Key areas for investigation can be determined based on the risk characteristics of individuals, vehicles, and items. Precautions could include police safety, crowd evacuation, evidence protection, and record-keeping.

[0070] The system also determines the collaborative response targets based on the location information, event type, and risk level of the current alarm. Collaborative response targets can include at least one of the following: command center, local police station, specialized police units, and mobile response terminals; they can also include grassroots units, venue management, or other related collaborative units. When determining collaborative response targets, the system can select based on territorial jurisdiction, alarm type responsibility rules, risk level linkage rules, and available police force status. For red alert level alarms, the system can generate a cross-departmental collaborative response list; for yellow alert level alarms, the system can prioritize pushing the list to local units and copy it to the command center; for blue alert level alarms, the system can push the list to regular response units or mobile response terminals.

[0071] Thus, the system can transform risk assessment results into actionable solutions, enabling intelligent processing from incident identification to risk scoring, tiered early warning, generation of handling suggestions, and determination of collaborative handling targets.

[0072] Example 6 like Figure 7As shown, this embodiment illustrates the multi-agent collaborative execution mechanism. The method in this embodiment can be executed collaboratively by multiple agents. These agents may include an element analysis agent, a data association agent, a multi-dimensional judgment agent, a scoring and early warning agent, a handling feedback agent, and a model enhancement agent. Each agent can be a functional module running on the same server or a service program distributed across different computing nodes. The agents distribute tasks, report status, and share results through a message bus, enabling structured processing tasks, data association tasks, risk assessment tasks, early warning push tasks, and feedback optimization tasks to be executed serially and / or in parallel.

[0073] The element analysis agent receives raw alarm input data and invokes speech recognition models, large language models, image recognition models, or multimodal models to perform data recognition and element extraction tasks. Specifically, the element analysis agent can be responsible for alarm speech transcription, text cleaning, semantic completion, KJD element extraction, field normalization, and structured alarm data generation.

[0074] The data association agent is used to perform entity alignment and association matching between structured police incident data and pre-set police data resources. Specifically, the data association agent can access population databases, case databases, vehicle databases, key personnel databases, key location databases, historical police incident databases, and other police business databases to identify historical police incidents, associated personnel, associated vehicles, associated locations, and associated events related to the current police incident, and generate multi-dimensional feature data.

[0075] The multidimensional analysis agent is used to construct a crime scene association graph and invoke large models or model clusters for risk reasoning. Specifically, the multidimensional analysis agent can construct people, locations, events, items, and organizations as graph nodes, and construct alarm relationships, case relationships, location relationships, vehicle relationships, item relationships, and historical handling relationships as graph edges; it can also retrieve similar crime scenes based on vector similarity, and input the graph, similar crime scenes, and multidimensional feature data into the large model to obtain risk assessment results.

[0076] The scoring and early warning intelligent agent is used to calculate risk scores and determine early warning levels based on risk assessment results and preset risk scoring rules. The agent can also access a police response knowledge base, retrieve fragments of response plans, and invoke large-scale models to generate response suggestions and collaborative response targets. The early warning levels, response suggestions, and collaborative response targets generated by the scoring and early warning intelligent agent can be sent to the command and dispatch system, local units, or mobile police terminals via a message bus.

[0077] The intelligent agent for handling feedback is used to collect data on the handling process and the handling results. Specifically, the intelligent agent can track the reception status, on-site status, handling progress, supplementary information, and handling results of each collaborative handling entity in real time. If the intelligent agent detects a delay in progress, missing information, or a change in alarm risk, it can trigger supervision, re-push, or reassessment through the message bus.

[0078] The model-enhancing agent is used to perform reverse optimization of the system based on the feedback data from the handling of incidents. Specifically, the model-enhancing agent can update the nodes and edges in the incident association graph based on the actual handling results, adjust the weights or thresholds in the risk scoring rules, update the similar incident retrieval samples, optimize the prompt word templates and large model calling parameters, and feed the optimization results back to the element analysis agent, data association agent, multi-dimensional judgment agent, and scoring and early warning agent.

[0079] In one optional implementation, the message bus can maintain the task identifier, task status, input data, output results, and anomaly information for each alarm task. After the element analysis agent completes the structured processing task, it can publish the structured alarm data to the message bus; the data association agent subscribes to the message and performs association matching tasks; the multidimensional analysis agent performs graph construction and risk reasoning tasks after receiving multidimensional feature data; the scoring and early warning agent performs scoring, early warning, and handling suggestion generation tasks after receiving risk assessment results; the handling feedback agent continuously collects feedback data after handling suggestions are pushed; and the model enhancement agent performs model and rule optimization tasks after the alarm is closed. Through this message-driven approach, even if some data sources are delayed, the system can first trigger a preliminary assessment based on the obtained features, and then update the assessment after subsequent data supplementation, thereby improving the system's response speed and robustness.

[0080] Example 7 like Figure 8 As shown in the figure, this embodiment provides an intelligent emergency response system based on a large model. The system includes an emergency response input module, a structured processing module, a data association module, an intelligent analysis module, an early warning and response module, and a feedback optimization module.

[0081] The alarm input module is used to acquire raw alarm input data. This raw alarm input data includes alarm voice data, alarm text data, image data, video data, and / or location data. The alarm input module can connect to alarm receiving systems, network alarm systems, mobile police terminals, video surveillance systems, and location systems to receive and verify the raw alarm input data. Verification may include data format verification, data source verification, timestamp verification, integrity verification, and duplicate submission verification.

[0082] The structured processing module is used to identify, transform, and semantically analyze the original police report input data, extract police report elements, and generate structured police report data based on these elements. The module can call a speech recognition model to perform real-time speech recognition on the police report voice data to obtain transcribed text; it can filter redundant sentences and complete the semantics of the transcribed text to obtain standardized police report text; it can call a large language model to perform semantic analysis on the standardized police report text, extracting police report elements including personnel, events, time, location, items, causes, and results; it can also match address information to a standard address database, normalize the fields of personnel, vehicle, item, and event descriptions, and perform deduplication, error correction, and field mapping on the normalized police report elements to generate structured police report data.

[0083] The data association module is used to perform entity alignment and association matching between the structured police incident data and preset police data resources, generating multi-dimensional feature data corresponding to the current incident. The data association module can connect to population databases, case databases, vehicle databases, key personnel databases, key location databases, and historical incident databases to identify personnel entities, location entities, vehicle entities, item entities, and event entities in the current incident, and match them with data in each of these data resources. Based on the matching results, the data association module determines historical incidents, associated personnel, associated vehicles, associated locations, and associated events, and generates multi-dimensional feature data. The multi-dimensional feature data may include at least two of the following: personnel risk characteristics, location sensitivity characteristics, event pattern characteristics, and spatiotemporal evolution characteristics.

[0084] The intelligent assessment module is used to construct an alarm association graph of the current alarm situation based on the multi-dimensional feature data, and calls a large model to perform risk reasoning on the alarm association graph to obtain risk assessment results. The intelligent assessment module can use personnel, locations, events, items, and organizations as graph nodes, and alarm relationships, case relationships, location relationships, vehicle relationships, item relationships, and historical handling relationships as graph edges to construct the alarm association graph. The intelligent assessment module can also retrieve similar alarm situations from the historical alarm feature database based on vector similarity, and input the alarm association graph, similar alarm situations, and multi-dimensional feature data into the large model to output risk assessment results such as the probability of repeated alarm situations, risk evolution paths, and key hidden danger information.

[0085] The early warning and response module determines the risk score and early warning level of the current emergency based on the risk assessment results and preset risk scoring rules, and generates response suggestions and collaborative response targets corresponding to the early warning level. The module can calculate the risk score based on the severity of the emergency, its scope of impact, historical frequency, personnel risk weight, location sensitivity weight, alarm emotion characteristics, and risk item characteristics, and compare the risk score with preset grading thresholds to determine a red, yellow, or blue early warning level. The module can also retrieve response plan fragments based on the early warning level, risk assessment results, and a preset police response knowledge base, and use a large model to fuse the response plan fragments and multi-dimensional feature data of the current emergency to generate response suggestions including at least two of the following: verification points, response steps, control points, key points of inspection, and precautions. It then determines collaborative response targets based on the location information, event type, and risk level of the current emergency.

[0086] The feedback optimization module is used to push the proposed handling suggestions to the collaborative handling objects, collect handling feedback data, and update the alarm association graph, the preset risk scoring rules, and / or the calling parameters of the large model based on the handling feedback data. The feedback optimization module can receive handling progress, on-site verification results, supplementary alarm information, handling measures, and handling results uploaded by the collaborative handling objects. The feedback optimization module can update the nodes, edges, and attribute information in the alarm association graph based on the handling feedback data; adjust the weights or thresholds of the risk scoring rules based on the actual handling results; and optimize the calling parameters, prompt word templates, and retrieval strategies of the large model based on manual review opinions and handling results.

[0087] In a specific application scenario, when the alarm receiving system receives an alarm call, the alarm input module acquires the alarm voice data and alarm number in real time. The structured processing module converts the call voice into transcribed text and extracts alarm elements such as the caller, time of incident, location of incident, type of incident, persons involved, vehicles involved, and on-site results. The data association module matches the caller, persons involved, vehicles, and locations with population databases, vehicle databases, key personnel databases, key location databases, and historical alarm databases, respectively. It finds that multiple similar alarms have occurred at the same location within the past week, and the associated persons have historical alarm records. The intelligent analysis module constructs an alarm association graph based on the above multi-dimensional feature data, retrieves similar historical alarms, and uses a large model to infer a risk assessment result with a high probability of repeated alarms and an escalating risk trend. The early warning and handling module calculates the risk score based on the risk assessment result and risk scoring rules, determines the current alarm to be at the yellow warning level, generates verification points, on-site handling steps, and collaborative handling targets, and pushes handling suggestions to the local police station and mobile policing terminals. After the incident is resolved, the feedback optimization module collects the on-site verification results and the resolution results, updates the location's sensitive characteristics and historical incident correlations, and optimizes the risk scoring rules for subsequent similar incidents based on the actual results.

[0088] The system described in this embodiment enables automatic alarm entry, data governance and structuring, automatic alarm feature extraction, intelligent alarm analysis, tiered early warning and collaborative response, alarm experience summarization and feedback, and data storage management. The model cluster within the system serves as the underlying intelligent capability, uniformly supporting speech-to-text, natural language processing, image recognition, multimodal understanding, and logical reasoning. The data storage layer of the system can include a relational database, a distributed file system, and a vector database. The relational database stores structured alarm data and response records; the distributed file system stores unstructured data such as audio, video, and images; and the vector database stores alarm feature vectors and historical alarm vectors to support rapid semantic retrieval and similar alarm matching.

[0089] Those skilled in the art will understand that the modules in the above system can be implemented by software, hardware, or a combination of software and hardware. These modules can be centrally deployed on the same server or distributed across multiple servers or computing nodes. Modules can communicate with each other through interface calls, message queues, data buses, or service gateways. Any system that achieves the functions defined in the claims of this invention should fall within the protection scope of this invention.

[0090] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0091] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for intelligent processing of the entire alarm process based on a large model, characterized in that, include: Obtain raw police incident input data, process the raw police incident input data, extract police incident elements, and generate structured police incident data based on the police incident elements; The structured police incident data is aligned and matched with preset police data resources to generate multi-dimensional feature data corresponding to the current police incident. Based on the multidimensional feature data, a crime association graph of the current crime is constructed, and a large model is called to perform risk reasoning on the crime association graph to obtain risk assessment results; Based on the risk assessment results and preset risk scoring rules, the risk score and warning level of the current alarm situation are determined, and corresponding handling suggestions and collaborative handling targets are generated. The proposed handling suggestions are pushed to the collaborative handling objects, and handling feedback data is collected. The alarm correlation graph, the preset risk scoring rules, and / or the calling parameters of the large model are updated based on the handling feedback data.

2. The method according to claim 1, characterized in that, The processing of the original alarm input data includes: The original alarm input data includes alarm voice data, alarm text data, image data, video data and / or location data; Real-time speech recognition is performed on the alarm voice data to obtain the alarm transcribed text; Redundant sentences are filtered and semantics are completed on the transcribed alarm text to obtain standardized alarm text. The standardized police report text is semantically parsed using a large language model to extract police report elements including personnel, events, time, location, items, causes, and results.

3. The method according to claim 1, characterized in that, The process of generating structured police incident data based on the aforementioned incident elements includes: The address information in the aforementioned alarm elements is matched against a standard address database to obtain standardized location information; The fields of personnel, vehicles, items, and event descriptions in the aforementioned police incident elements are normalized. The normalized police incident elements are deduplicated, corrected for errors, and mapped to fields to generate structured police incident data for police incident analysis.

4. The method according to claim 1, characterized in that, The step of aligning and matching the structured police incident data with preset police data resources includes: The personnel entities, location entities, vehicle entities, item entities, and event entities in the structured police incident data are matched with data in the population database, case database, vehicle database, key personnel database, key location database, and historical police incident database, respectively. Based on the matching results, identify historical police incidents, associated personnel, associated vehicles, associated locations, and associated events that are related to the current police incident; The multidimensional feature data is generated based on the historical police reports, associated personnel, associated vehicles, associated locations, and associated events.

5. The method according to claim 1, characterized in that, The multidimensional feature data includes at least two of the following: personnel risk characteristics, location sensitivity characteristics, event pattern characteristics, and spatiotemporal evolution characteristics. The personnel risk characteristics are used to characterize the historical alarms, historical cases, key personnel tags and / or repeated associations of personnel related to police incidents; The location-sensitive features are used to characterize the key location attributes, historical police incident frequency, and / or surrounding risk status of the crime scene; The event pattern features are used to characterize the semantic similarity, type similarity, and / or similarity of handling results between the current incident and historical incidents; The spatiotemporal evolution features are used to characterize the degree of aggregation, diffusion trend and / or recurrence trend of current incidents in the time and space dimensions.

6. The method according to claim 1, characterized in that, The process of constructing a crime association graph based on the multidimensional feature data and then using a large model to perform risk reasoning on the crime association graph includes: Using people, locations, events, items, and organizations as graph nodes, and alarm relationships, case relationships, location relationships, vehicle relationships, item relationships, and historical handling relationships as graph edges, the aforementioned police incident association graph is constructed; Based on vector similarity, retrieve similar police reports that match the current police report from the historical police report feature database; The alarm correlation map, the similar alarms, and the multidimensional feature data are input into the large model to obtain risk assessment results that characterize the probability of repeated alarms, risk evolution paths, and key hidden danger information.

7. The method according to claim 1, characterized in that, The process of determining the risk score and warning level of the current alert based on the risk assessment results and preset risk scoring rules includes: The risk score of the current incident is calculated based on at least two of the following factors: the severity of the incident, the scope of impact, the frequency of historical occurrences, the risk weight of personnel, the sensitivity weight of the location, the emotional characteristics of the alarm, and the characteristics of the risky items. The risk score is compared with a preset grading threshold to determine the red, yellow, or blue alert level corresponding to the current alert.

8. The method according to claim 1, characterized in that, The generation of handling suggestions and collaborative handling targets corresponding to the warning level includes: Based on the warning level, the risk assessment results, and the preset police response knowledge base, retrieve a response plan fragment that matches the current police situation; The large model is invoked to fuse the multi-dimensional feature data of the aforementioned emergency response plan fragments and the current emergency situation to generate at least two of the following response suggestions: key points for verification, response steps, control points, key points for inspection, and precautions. Based on the location information, event type, and risk level of the current police situation, at least one of the following should be identified as the coordinating entity: the command center, the local police station, specialized police units, and mobile emergency response terminals.

9. The method according to claim 1, characterized in that, The method is executed collaboratively by multiple intelligent agents, including an element analysis agent, a data association agent, a multi-dimensional judgment agent, a scoring and early warning agent, a handling and feedback agent, and a model enhancement agent. The multiple intelligent agents distribute tasks, transmit status, and share results through a message bus, enabling structured processing tasks, data association tasks, risk assessment tasks, early warning push tasks, and feedback optimization tasks to be executed serially and / or in parallel.

10. A large-scale model-based intelligent emergency response system for the entire process, characterized in that: include: The alarm input module is used to acquire raw alarm input data, which includes alarm voice data, alarm text data, image data, video data and / or location data; The structured processing module is used to identify, transform, and semantically parse the original police incident input data, extract police incident elements, and generate structured police incident data based on the police incident elements. The data association module is used to perform entity alignment and association matching between the structured police incident data and preset police data resources to generate multi-dimensional feature data corresponding to the current police incident. The intelligent analysis module is used to construct an alarm correlation graph of the current alarm based on the multi-dimensional feature data, and call a large model to perform risk reasoning on the alarm correlation graph to obtain risk analysis results; The early warning and response module is used to determine the risk score and early warning level of the current alarm based on the risk assessment results and preset risk scoring rules, and to generate response suggestions and collaborative response targets corresponding to the early warning level; The feedback optimization module is used to push the handling suggestions to the collaborative handling objects, collect handling feedback data, and update the alarm association map, the preset risk scoring rules, and / or the calling parameters of the large model based on the handling feedback data.

Images