A large model-based event risk detection method, device and medium

Abstract

The embodiment of the application discloses a kind of event risk detection methods, equipment and medium based on large model, belong to risk detection technical field, solve the problems that prior art scene poor adaptability, rely on sample and artificial rule, so that the risk detection accuracy and efficiency of complex and diverse event are lower. Including, based on the industry knowledge and industry scene information obtained, event risk detection processing template is constructed;Wherein, event risk detection processing template is used to provide detection processing logic for event risk detection model;Obtain the event basic data corresponding to the event to be detected, retrieve event related knowledge data in the preset vector database based on event basic data;Event basic data and event related knowledge data are filled into event risk detection processing template to obtain input prompt word;Call event risk detection model, based on event risk detection model and input prompt word carry out event risk detection, and output detection conclusion.

Description

Technical Field

[0001] This application relates to the field of risk detection technology, and in particular to an event risk detection method, device and medium based on a large model. Background Technology

[0002] In fields such as public safety and emergency management, rapid and accurate risk detection of incidents is a key basis for effective handling and resource allocation.

[0003] In existing technologies, event risk detection mainly relies on three assessment methods: expert experience assessment, assessment based on multi-dimensional weighted indicators, and assessment based on traditional machine learning models.

[0004] Expert assessments rely heavily on the domain experts' personal knowledge and on-the-spot judgment, but their response speed and accuracy are significantly insufficient when dealing with sudden, novel, or complex events. Secondly, assessments based on multi-dimensional weighted indicators involve pre-setting a series of risk assessment indicators and assigning weights to each, then calculating a risk score through weighted calculations. However, the construction of the indicator system and the setting of weights often depend on subjective experience; when event information is incomplete or indicator calculations are biased, the accuracy of the assessment results drops sharply. Furthermore, while detection methods based on traditional machine learning models have achieved a degree of automation, their effectiveness heavily depends on a large number of high-quality, high-cost labeled samples. In the public safety field, such samples are often scarce and have limited coverage. Moreover, once the model is trained, its assessment logic becomes fixed, making it difficult to adapt to changes in business rules.

[0005] Events are complex and diverse, and are characterized by being unstructured, having various event types, and being highly dynamic and evolving. As a result, existing technologies often suffer from poor adaptability to different scenarios and heavy reliance on samples and manual rules, leading to low accuracy and efficiency in risk detection for such complex and diverse events. Summary of the Invention

[0006] This application provides a method, device, and medium for event risk detection based on a large model, which addresses the following technical problems: existing technologies often suffer from poor scene adaptability and heavy reliance on samples and manual rules, resulting in low accuracy and efficiency in risk detection for complex and diverse events.

[0007] The embodiments of this application adopt the following technical solutions:

[0008] This application provides an event risk detection method based on a large model. It includes: constructing an event risk detection processing template based on acquired industry knowledge and industry scenario information; wherein the event risk detection processing template provides detection processing logic for the event risk detection model; acquiring basic event data corresponding to the event to be detected, and retrieving event-related knowledge data from a pre-set vector database based on the basic event data; filling the basic event data and event-related knowledge data into the event risk detection processing template to obtain input prompt words; calling the event risk detection model, performing event risk detection based on the event risk detection model and the input prompt words, and outputting detection conclusions; wherein the detection conclusions include at least one of the following: event risk level, multi-dimensional detection criteria, and event alarm identifier.

[0009] In one implementation of this application, an event risk detection and processing template is constructed based on acquired industry knowledge and industry scenario information. Specifically, this includes: constructing multiple detection modules based on industry knowledge and industry scenario information; wherein each detection module includes at least one of the following: a task description module, a processing logic module, a constraint and restriction module, a background knowledge module, an output specification module, an event input module, a risk detection case module, a reflection and inspection module, and a risk alarm module; wherein the processing logic module includes at least one of the following dimensions: alarm or clue frequency dimension, hazard level change dimension, change in involved personnel dimension, change in personnel's emotions dimension, change in involved tools dimension, change in handling method dimension, change in involved items or amounts dimension, and change in involved behavior dimension. The event risk detection and processing template is constructed based on these multiple modules and dimensions.

[0010] In one implementation of this application, the process involves obtaining basic event data corresponding to the event to be detected, and retrieving event-related knowledge data from a pre-set vector database based on the basic event data. Specifically, this includes: determining the modality type of the event data corresponding to the event to be detected; wherein the modality type includes at least one of field data, image data, voice data, and spatiotemporal trajectory data; matching a corresponding feature extraction strategy based on the modality type to extract risk features; wherein the risk features include at least one of field features, image features, voice features, and trajectory features; inputting the extracted risk features into the pre-set vector database, filtering candidate data of the same modality and cross-modality based on risk feature labels, and determining associated data with a matching degree greater than a preset threshold from the candidate data based on feature vector similarity; determining the event subject dimension, and mapping the field data and associated data corresponding to each risk feature to the corresponding subject dimension sequentially based on a pre-set feature subject mapping relationship; and concatenating the mapped data sequentially based on the subject dimension order to obtain event-related knowledge data.

[0011] In one implementation of this application, a corresponding feature extraction strategy is matched based on the modality type to extract risk features. Specifically, this includes: if the data is field data, semantic parsing and structured information extraction are performed to construct an event logic graph, and field features are extracted based on the event logic graph; wherein, field features include at least one of entity features, logical relationship features, event chain features, and conflict features; if the data is image data, keyframes are identified, and image features are extracted from the keyframes; wherein, image features include at least one of target entity features, scene environment features, and image semantic association features; if the data is speech data, the speech data is converted into text data, and speech features are extracted during the conversion process; wherein, speech features include at least one of speech-to-text conversion features and emotional intent features; if the data is spatiotemporal trajectory data, trajectory analysis is performed based on time series data to construct a trajectory time series graph, and trajectory features are extracted based on the trajectory time series graph; wherein, trajectory features include at least one of trajectory morphology features, trajectory association features, and trajectory anomaly features.

[0012] In one implementation of this application, before filling the event basic data and event-related knowledge data into the event risk detection and processing template, the method further includes: performing semantic analysis on the event basic data to extract event semantic units; wherein, the event semantic unit includes at least one of entity units, relation units, and attribute units; and performing semantic analysis on the event-related knowledge data to extract knowledge semantic units; wherein, the knowledge semantic unit includes at least one of rule units, case units, and concept units; determining the association relationship between the event semantic units and the knowledge semantic units based on semantic similarity, and matching the relevant event semantic units and the relevant knowledge semantic units to preset semantic slots based on the association relationship, so as to concatenate the event basic data and the event-related knowledge data to generate input prompt words.

[0013] In one implementation of this application, an event risk detection model is invoked, and event risk detection is performed based on the event risk detection model and the input prompt words, and a detection conclusion is output. Specifically, this includes: invoking the event risk detection model, inputting the input prompt words into the event risk detection model, and driving the event risk detection model to perform risk analysis and deduction based on the prompt words; outputting an initial detection conclusion through the event risk detection model based on a preset output specification; verifying the initial detection conclusion, and outputting the final detection conclusion after the verification is passed.

[0014] In one implementation of this application, after performing event risk detection based on the event risk detection model and input prompt words, and outputting the detection conclusion, the method further includes: constructing an event dynamic graph based on the multi-dimensional detection criteria in the detection conclusion, simulating the risk through a pre-set time-series graph neural network, and generating a risk evolution path; obtaining the environmental characteristics and external intervention variables corresponding to the event to be detected to construct a simulated environment; inputting the risk evolution path into different simulated environments for deduction to generate a risk evolution scenario set; analyzing each path in the risk evolution scenario set, determining the scenario branch points, and generating a directorized path based on the path probability, the maximum risk level, and the scenario branch points, so as to determine the risk evolution prediction result based on the directorized path.

[0015] In one implementation of this application, each path in the risk evolution scenario set is analyzed to determine scenario branch points. Based on path probability, maximum risk level, and scenario branch points, a master-directed automated path is generated. Specifically, this includes: identifying nodes that appear in risk evolution paths exceeding a preset proportion and whose corresponding risk vectors are greater than a preset vector threshold as scenario branch points; constructing a path comprehensive detection function based on preset weights, path probabilities, and maximum risk levels to determine the risk value corresponding to each path; using the path corresponding to the highest risk value as the baseline path; and performing path fusion based on the similarity between the baseline path and other paths after the scenario branch points to generate a master-directed automated path.

[0016] This application provides an event risk detection device based on a large model, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: construct an event risk detection processing template based on acquired industry knowledge and industry scenario information; wherein the event risk detection processing template is used to provide detection processing logic for the event risk detection model; acquire basic event data corresponding to the event to be detected, and retrieve event-related knowledge data from a preset vector database based on the basic event data; fill the basic event data and event-related knowledge data into the event risk detection processing template to obtain input prompt words; call the event risk detection model, perform event risk detection based on the event risk detection model and the input prompt words, and output detection conclusions; wherein the detection conclusions include at least one of the following: event risk level, multi-dimensional detection basis, and event alarm identifier.

[0017] This application provides a non-volatile computer storage medium storing computer-executable instructions. These instructions are configured to: construct an event risk detection and processing template based on acquired industry knowledge and industry scenario information; wherein the event risk detection and processing template provides detection and processing logic for the event risk detection model; acquire basic event data corresponding to the event to be detected, and retrieve event-related knowledge data from a preset vector database based on the basic event data; fill the basic event data and event-related knowledge data into the event risk detection and processing template to obtain input prompts; call the event risk detection model, perform event risk detection based on the event risk detection model and the input prompts, and output a detection conclusion; wherein the detection conclusion includes at least one of the following: event risk level, multi-dimensional detection basis, and event alarm identifier.

[0018] The above-mentioned technical solutions adopted in this application embodiment can achieve the following beneficial effects: Firstly, by combining event risk assessment templates, large models, and knowledge bases, this application embodiment automatically understands and analyzes events based on constraints and specifications, improving the accuracy and efficiency of event risk detection. Secondly, by constructing processing templates, this application embodiment improves the consistency and repeatability of the risk detection process. Furthermore, when updating or adapting to new scenarios, only the template content needs to be adjusted, reducing migration complexity. Thirdly, by retrieving event-related knowledge data from a pre-set vector database based on basic event data, dynamic and accurate knowledge enhancement is achieved, improving the contextual relevance and professionalism of risk detection. Finally, this application embodiment performs event risk detection based on models and input prompts, achieving automated and intelligent risk assessment, which not only improves the accuracy of risk detection results but also increases the efficiency of risk detection. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0020] Figure 1 A flowchart of an event risk detection method based on a large model provided for embodiments of this application;

[0021] Figure 2 This is a schematic diagram of the structure of an event risk detection device based on a large model, provided as an embodiment of this application.

[0022] Figure label:

[0023] 200: Event risk detection device based on large model; 201: Processor; 202: Memory. Detailed Implementation

[0024] This application provides an event risk detection method, device, and medium based on a large model.

[0025] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0026] The technical solutions proposed in the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0027] Figure 1 A flowchart of an event risk detection method based on a large model is provided for embodiments of this application, as follows: Figure 1 As shown, the event risk detection method based on a large model includes the following steps:

[0028] Step 101: Based on the acquired industry knowledge and industry scenario information, construct an event risk detection and processing template.

[0029] In one embodiment of this application, an industry business knowledge base is first constructed. Industry knowledge related to event risk assessment, such as rules, requirements, business constraints, judgment criteria, and processing information, is collected, and this knowledge is vectorized using an embedding model and stored in a vector database.

[0030] Specifically, this embodiment first collects multi-source, multi-modal industry knowledge related to event risk assessment, including but not limited to regulations, emergency plans, historical case records, expert experience manuals, and standard operating procedures. Then, the collected raw knowledge is preprocessed and anonymized. A semantic-based sliding window is used to segment the processed data, ensuring that each text fragment carries a relatively complete semantic unit. Next, an embedding model is used to convert each text slice into a high-dimensional vector. All generated vectorized knowledge fragments, along with their corresponding original text content and metadata, are stored in a pre-set vector database.

[0031] In one embodiment of this application, multiple detection modules are constructed based on industry knowledge and industry scenario information. These multiple detection modules include at least one of the following: a task description module, a processing logic module, a constraint and restriction module, a background knowledge module, an output specification module, an event input module, a risk detection case module, a reflection and inspection module, and a risk alarm module. The processing logic module includes at least one of the following dimensions: alarm or clue frequency dimension, hazard level change dimension, change in involved personnel dimension, change in personnel's emotions dimension, change in involved tools dimension, change in handling method dimension, change in involved items or amounts dimension, and change in involved behavior dimension. Based on these multiple modules and dimensions, an event risk detection processing template is constructed. This event risk detection processing template provides detection processing logic for the event risk detection model.

[0032] Specifically, this application also designs constraints, background knowledge, processing logic, and output specifications for event risk assessment, serving as an event risk detection and processing template. The template defines the core processing logic for event risk assessment and can be used for risk assessment of general events. By modifying the content of the template, it can support the analysis and judgment of personalized events. This event risk detection and processing template mainly includes the following parts:

[0033] 1. Task Description: Define the perspective, purpose, and key outputs of the risk assessment. This part is relatively fixed.

[0034] 2. Processing Logic: Define the risk assessment process, logical constraints, etc. This part can be fine-tuned as needed.

[0035] 3. Constraints and Limitations: Define the limiting factors and constraints of the risk assessment. This section is adjustable.

[0036] 4. Background Knowledge: Define the industry knowledge and business background required for risk assessment. This section can be refined and adjusted according to the specific scenario.

[0037] 5. Output Specifications: Define the format and specifications for the risk assessment output results. This part is relatively fixed but can be fine-tuned as needed.

[0038] 6. Event Input: Define the format and limitations of risk assessment input data. This section is adjustable.

[0039] 7. Risk Assessment Example: Defines a general example of the output for an event risk assessment. This section can be fine-tuned as needed.

[0040] 8. Reflection and Inspection: Define reflection strategies and inspection constraints to ensure consistent output. This part is relatively fixed but can be fine-tuned as needed.

[0041] 9. Risk Alerts: Define the alert constraints and return criteria for risk escalation. This section can be fine-tuned as needed.

[0042] Secondly, this application embodiment also designs the content of each part of the risk detection processing template. For example, this application embodiment has designed the content and verified the effect for a single event scenario in police reporting, and formed a fixed processing template. The content in this processing template can be adjusted as needed when applying subdivided scenarios to meet the analysis and evaluation of personalized events. For example, in the processing logic part of police reporting, the dimensions of event risk detection are defined to include at least the following aspects: frequency of alarms or clues, changes in the degree of harm, changes in the people involved, changes in the emotions of the people involved, changes in the tools involved, changes in the handling method, changes in the items or amounts involved, and changes in the behaviors involved. Through the limited analysis of these dimensions, this application embodiment provides a multi-perspective approach to event risk detection, thereby providing a reliable guarantee for the final risk assessment results.

[0043] Step 102: Obtain the basic event data corresponding to the event to be detected, and retrieve event-related knowledge data from the preset vector database based on the basic event data.

[0044] In one embodiment of this application, the modality type of the event data corresponding to the event to be detected is determined; wherein, the modality type includes at least one of field data, image data, voice data, and spatiotemporal trajectory data. Based on the modality type, a corresponding feature extraction strategy is matched to extract risk features; wherein, the risk features include at least one of field features, image features, voice features, and trajectory features. The extracted risk features are input into a preset vector database, and candidate data of the same modality and cross-modality are filtered in the preset vector database by risk feature labels. Based on feature vector similarity, associated data with a matching degree greater than a preset threshold is determined from the candidate data. The event subject dimension is determined, and based on a preset feature subject mapping relationship, the field data and associated data corresponding to each risk feature are sequentially mapped to the corresponding subject dimension. The mapped data are concatenated sequentially according to the subject dimension order to obtain event-related knowledge data.

[0045] Specifically, the basic data of the event is read, including the basic information of the event (time, place, people, cause, process, result, etc.), the timeline backtracking information of the event, the related information of the event (other events, files, manual notes, etc. related to the event), and the processing information of the event (such as historical analysis information, historical handling records, etc.).

[0046] Furthermore, the metadata and content structure of the data are analyzed to determine its modality type, such as structured fields, images, speech, spatiotemporal trajectories, etc. Based on the identified modality type, a corresponding feature extraction algorithm is matched from a pre-defined strategy library. The feature extraction process is as follows: If it is field data, semantic parsing and structured information extraction are performed to construct an event logic graph, and field features are extracted based on the event logic graph; wherein, field features include at least one of entity features, logical relationship features, event chain features, and conflict features. If it is image data, image keyframes are identified, and image features are extracted from the image keyframes; wherein, image features include at least one of target entity features, scene environment features, and image semantic association features. If it is speech data, the speech data is converted into text data, and speech features are extracted during the conversion process; wherein, speech features include at least one of speech-to-text conversion features and emotional intent features. If the data is spatiotemporal trajectory data, trajectory analysis is performed based on time series to construct a trajectory time series graph, and trajectory features are extracted based on the trajectory time series graph; among them, trajectory features include at least one of trajectory morphology features, trajectory correlation features, and trajectory anomaly features.

[0047] Specifically, for field data, the process begins with named entity recognition and relation extraction using natural language processing (NLP) techniques. This identifies core entities such as people and locations within the text, along with their interactions, forming basic triples consisting of subject, predicate, and object. Building upon this, event extraction techniques identify key behaviors like arguments, driving, and alarm calls as event triggers, binding them with relevant time, location, and participants. This constructs a preliminary event logic graph with entities as nodes, relationships as edges, and events as hyperedges. Subsequently, graph algorithms are used to mine features from this event logic graph: entity features are quantified by calculating betweenness centrality and other metrics to identify key people or locations; logical relationship features are extracted by analyzing the complexity and patterns of relationship paths to determine the potential strength of associations between entities; event chain features are constructed and quantified by tracing the sequence and causal relationships of events based on timestamps to determine the dynamic evolution path of risks; finally, information from different sources or paths in the graph is compared to detect logical inconsistencies and contradictions, revealing conflict features as key indicators for assessing information reliability and risk concealment. Ultimately, all obtained features are merged into a field feature vector.

[0048] Specifically, for image data, by calculating the differences between video frames, the keyframes that best represent the core content of the event are selected from the video stream or image sequence. The keyframes are then analyzed using a pre-trained deep neural network model. For example, target entity features are obtained by extracting identifiable people, vehicles, objects and their attributes; scene environment features are determined by extracting the overall atmosphere in the image; and visual content is converted into natural language descriptions through an image description generation model, thereby elevating pixel information into semantic features that can be used for risk reasoning.

[0049] If the data is speech, a speech recognition engine is used to convert the speech data into text data that can be analyzed in order to obtain speech-to-text conversion features. In this process, acoustic parameters such as fundamental frequency, energy, spectrum, and speech rate, as well as the speaker's emotional state, are extracted to determine the emotional intent features.

[0050] If the data is spatiotemporal trajectory data, a trajectory time series graph is constructed based on the time series. In this graph, each node represents a geographical location coordinate and a timestamp, and edges represent movement paths. Based on this, multi-dimensional trajectory features are extracted by calculating statistics such as speed, acceleration, direction changes, and activity hotspots to determine the movement characteristics of individuals or groups. By analyzing the spatiotemporal patterns of intersection, parallelism, and following of different trajectories, trajectory association features are identified. By comparing this trajectory time series graph with a normal pattern database, suspicious behaviors such as traveling to sensitive areas late at night, abnormal loitering, and sudden high-speed movement are identified, thus extracting abnormal trajectory features.

[0051] Furthermore, the obtained feature vectors of each modality are used as query conditions and input into a pre-set vector database. This database pre-stores different multimodal historical data and their corresponding vectors. The similarity between vectors is calculated using a cosine similarity algorithm, thereby filtering semantically relevant candidate data from the database. For example, images of chaotic scenes in the current event may retrieve images and report texts of similar group events in the past. Further, the event subject dimension is determined. For example, taking a police report scenario as an example, based on predefined rules such as names, ID numbers, and mobile phone numbers, the core subjects involved, such as the person reporting the incident, and related persons, can be identified from the event data. Subsequently, a mapping network between features and subjects is established, and all extracted features and retrieved related data are respectively merged into their corresponding subject dimensions. For example, a person's facial features, voiceprint features in a call recording, and the trajectory features of their vehicle are all aggregated into the subject dimension of person A. This embodiment of the application, by mapping features to subjects, can integrate scattered multimodal information into a structured archive rich in information centered on each subject.

[0052] Furthermore, following the order of subjects determined by business logic, such as the order of primary and secondary relationships: parties involved, victims, and witnesses, all the original fields, extracted features, and associated case texts aggregated under each subject dimension are read in sequence, and this information is then pieced together into a coherent text description. Finally, event-related knowledge data is generated.

[0053] Step 103: Fill in the basic event data and event-related knowledge data into the event risk detection and processing template to obtain input prompt words.

[0054] In one embodiment of this application, semantic parsing is performed on basic event data to extract event semantic units; wherein, an event semantic unit includes at least one of entity units, relation units, and attribute units. Semantic parsing is also performed on event-related knowledge data to extract knowledge semantic units; wherein, a knowledge semantic unit includes at least one of rule units, case units, and concept units. Based on semantic similarity, the association between event semantic units and knowledge semantic units is determined, and based on the association, related event semantic units and related knowledge semantic units are matched to preset semantic slots to concatenate the basic event data and event-related knowledge data, generating input prompt words.

[0055] Specifically, this application embodiment uses natural language processing technologies, such as named entity recognition, relation extraction, and attribute recognition, to parse basic event data and extract event semantic units, including core entity units such as person A and location B; relation units such as driving and occurring in; and attribute units such as emotional excitement and large sums of money. Simultaneously, event-related knowledge data retrieved from a vector library is parsed to identify and extract knowledge semantic units, such as rule units (e.g., carrying a weapon increases the risk level), case units (e.g., similar events in a certain year and month), and concept units (e.g., the definition of a group fight). Next, the semantic similarity between event semantic units and knowledge semantic units is calculated. For example, the vector similarity of the event entity unit "knife" with the knowledge concept unit "carrying a weapon" is calculated, or the event relation unit "multiple people gathering" is associated with the knowledge rule unit "multiple people gathering requires an escalation of the response."

[0056] Furthermore, this embodiment of the application also includes semantic slots containing multiple roles. These slots constitute the basic framework for prompt word generation, including, for example, core entity slots, key behavior slots, applicable rule slots, and reference case slots. Based on the established relationships, related event semantic units and knowledge semantic units are treated as a group of information and mapped together to the most suitable preset semantic slots. Following the preset logical reasoning order of the slots, the content already filled in each semantic slot is read sequentially, and the structured units within each slot are concatenated to obtain the input prompt word.

[0057] Step 104: Call the event risk detection model, perform event risk detection based on the event risk detection model and input prompt words, and output the detection conclusion.

[0058] In one embodiment of this application, an event risk detection model is invoked, input prompts are fed into the event risk detection model, and the model is driven to perform risk analysis and deduction based on the prompts. Based on preset output specifications, the event risk detection model outputs an initial detection conclusion, which is then validated. Upon successful validation, a final detection conclusion is output. The detection conclusion includes at least one of the following: event risk level, multi-dimensional detection criteria, and event alarm identifier.

[0059] Specifically, the assembled complete prompts are used for deep reasoning via a large model API. By adjusting parameters such as temperature, response format, and penalty coefficient of the large model, a satisfactory event risk assessment result is obtained. The output includes three main components: first, the final conclusion of the event risk assessment (high, medium, or low risk); second, the multi-dimensional basis for the event risk assessment (including multiple perspectives such as clue frequency, severity, involved individuals, emotional changes, changes in involved tools, changes in handling methods, changes in involved items or amounts, and changes in involved behaviors); and third, event alarm indicators (including risk escalation alarms, alarm basis, and other suggestions). The output is formatted and checked to form the final event risk assessment result. This risk assessment result can be written to the database as needed or presented directly. If adjustments to the assessment result are required, details and other constraints in the risk detection processing template can be adjusted. Relying on business knowledge input, deep analysis of the large model, and combinations of prompts, risk detection for different scenarios and types of events can be achieved.

[0060] In one embodiment of this application, an event dynamic graph is constructed based on multi-dimensional detection criteria in the detection conclusions, and the risk is simulated using a pre-set time-series graph neural network to generate a risk evolution path. Environmental characteristics and external intervention variables corresponding to the event to be detected are acquired to construct a simulation environment. The risk evolution path is input into different simulation environments for deduction, generating a risk evolution scenario set. Each path in the risk evolution scenario set is analyzed to determine scenario branch points. Based on path probability, maximum risk level, and scenario branch points, a pre-directed path is generated to determine the risk evolution prediction result based on the pre-directed path.

[0061] Specifically, key entities and their dynamic relationships are extracted from the multi-dimensional detection criteria of the detection conclusions, thereby constructing an event dynamic graph with entities as nodes and dynamic relationships as edges. Subsequently, this event dynamic graph and its initial state are input into a pre-set time-series graph neural network model. This time-series graph neural network model has been trained on historical event sequence data and is capable of learning the temporal patterns of risk propagation in the entity network. Through model inference, the diffusion, amplification, or attenuation effects of risk along the relationship network are simulated, thereby generating a baseline risk evolution path based on internal dynamics and predicting the overall risk status of events at future time points.

[0062] Furthermore, the environmental characteristics corresponding to the event to be detected are identified and obtained to construct a multi-agent simulation environment. The generated baseline risk evolution path is placed in this multi-agent simulation environment. By adjusting the type, intensity and timing of intervention variables, the interaction of agents in the simulation environment is driven and multiple rounds of simulation are conducted to generate a risk evolution scenario set covering different external conditions. This set reflects the risk trends under various possible future conditions.

[0063] Furthermore, nodes appearing in risk evolution paths exceeding a preset proportion, and whose corresponding risk vectors exceed a preset vector threshold, are designated as scenario branch points. Based on preset weights, path probabilities, and the maximum risk level, a path comprehensive detection function is constructed to determine the risk value corresponding to each path. The path with the highest risk value is selected as the baseline path. Based on the similarity between the baseline path and other paths after the scenario branch point, path fusion is performed to generate the author-directed programmed path.

[0064] Specifically, firstly, all risk evolution paths are aligned on a timeline. This application defines quantitative rules to identify scenario branch points: a node must simultaneously meet two conditions. The first is a consensus condition, meaning the node appears in more than a preset proportion of paths, indicating it is a prevalent critical stage. The second is a risk condition, meaning that at this node, the magnitude of the risk state vector of each path or the score of a specific dimension is greater than a preset vector threshold, indicating significant risk intensity. A node that simultaneously meets both conditions is considered a scenario branch point; that is, after this node, the event may evolve in multiple different directions. Secondly, a base path is selected from the risk evolution scenario set, and a path comprehensive detection function is constructed based on a preset weight configuration. This function is a weighted calculation formula, whose input variables mainly include the path probability and maximum risk level of each path. The comprehensive risk value of each path is calculated using this function. Based on this comprehensive score, the paths are sorted, and the path with the highest score is determined as the benchmark path, serving as the backbone for subsequent path fusion.

[0065] Furthermore, using the selected baseline path as a benchmark, after each identified scenario branch point, the similarity relationship between the subsequent segments of the baseline path and the subsequent segments of other paths is calculated. If a segment of another path after a branch point differs significantly from the baseline path (i.e., the similarity is low), but the segment itself indicates a higher risk level, then the high-risk segment will be merged into the corresponding position of the baseline path, ultimately generating a master-directed path. The master-directed path in this embodiment retains the main framework of the baseline path while integrating important risk warnings from other paths, making the predicted trajectory more accurate. Finally, based on the master-directed path, the final risk evolution prediction result is generated.

[0066] Figure 2 This is a schematic diagram of the structure of an event risk detection device based on a large model, provided as an embodiment of this application. Figure 2 As shown, the event risk detection device 200 based on a large model includes: at least one processor 201; and a memory 202 communicatively connected to the at least one processor 201. The memory 202 stores instructions executable by the at least one processor 201, which, when executed, enable the at least one processor 201 to: construct an event risk detection processing template based on acquired industry knowledge and industry scenario information; wherein the event risk detection processing template provides detection processing logic for the event risk detection model; acquire basic event data corresponding to the event to be detected, and retrieve event-related knowledge data from a pre-set vector database based on the basic event data; fill the basic event data and event-related knowledge data into the event risk detection processing template to obtain input prompts; call the event risk detection model, perform event risk detection based on the event risk detection model and the input prompts, and output detection conclusions; wherein the detection conclusions include at least one of the following: event risk level, multi-dimensional detection criteria, and event alarm identifier.

[0067] This application provides a non-volatile computer storage medium storing computer-executable instructions. These instructions are configured to: construct an event risk detection and processing template based on acquired industry knowledge and industry scenario information; wherein the event risk detection and processing template provides detection and processing logic for the event risk detection model; acquire basic event data corresponding to the event to be detected, and retrieve event-related knowledge data from a preset vector database based on the basic event data; fill the basic event data and event-related knowledge data into the event risk detection and processing template to obtain input prompts; call the event risk detection model, perform event risk detection based on the event risk detection model and the input prompts, and output a detection conclusion; wherein the detection conclusion includes at least one of the following: event risk level, multi-dimensional detection basis, and event alarm identifier.

[0068] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0069] The above descriptions are merely embodiments of this application and are not intended to limit the scope of this application. For those skilled in the art, various modifications and variations can be made to the embodiments of this application. These modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions in the embodiments of this application.

Claims

1. A method for event risk detection based on a large model, characterized in that, The method includes: Based on the acquired industry knowledge and industry scenario information, an event risk detection and processing template is constructed; wherein, the event risk detection and processing template is used to provide detection and processing logic for the event risk detection model; Obtain the basic event data corresponding to the event to be detected, and retrieve event-related knowledge data from a preset vector database based on the basic event data; The basic event data and related event knowledge data are entered into the event risk detection and processing template to obtain input prompt words; The event risk detection model is invoked, and event risk detection is performed based on the event risk detection model and the input prompt words, and the detection conclusion is output; wherein, the detection conclusion includes at least one of the following: event risk level, multi-dimensional detection basis, and event alarm identifier; Before filling the event basic data and event-related knowledge data into the event risk detection and processing template, the method further includes: Semantic parsing is performed on the event basic data to extract event semantic units; wherein, the event semantic unit includes at least one of entity unit, relation unit and attribute unit; Furthermore, semantic parsing is performed on the event-related knowledge data to extract knowledge semantic units; wherein, the knowledge semantic unit includes at least one of rule units, case units, and concept units; Based on semantic similarity, the association between the event semantic unit and the knowledge semantic unit is determined, and the relevant event semantic unit and the relevant knowledge semantic unit are matched to the preset semantic slots based on the association, so as to concatenate the event basic data and the event related knowledge data to generate input prompt words.

2. The event risk detection method based on a large model according to claim 1, characterized in that, The process of constructing an event risk detection and processing template based on the acquired industry knowledge and industry scenario information specifically includes: Based on the industry knowledge and industry scenario information, multiple detection modules are constructed; wherein, the multiple detection modules include at least one of the following: task description module, processing logic module, constraint and restriction module, background knowledge module, output specification module, event input module, risk detection case module, reflection and inspection module, and risk alarm module; The processing logic module includes at least one of the following dimensions: alarm or clue frequency, change in severity, change in involved persons, change in persons' emotions, change in involved tools, change in handling methods, change in involved items or amounts, and change in involved behavior. The event risk detection and processing template is constructed based on multiple modules and multiple dimensions.

3. The event risk detection method based on a large model according to claim 1, characterized in that, The process of acquiring the basic event data corresponding to the event to be detected, and retrieving event-related knowledge data from a preset vector database based on the basic event data, specifically includes: Determine the modal type of the event data corresponding to the event to be detected; wherein, the modal type includes at least one of field data, image data, voice data, and spatiotemporal trajectory data; Based on the modality type, a corresponding feature extraction strategy is matched to extract risk features; wherein, the risk features include at least one of field features, image features, speech features, and trajectory features; The extracted risk features are input into a preset vector database. Candidate data of the same modality and cross-modality are filtered out in the preset vector database by risk feature labels. Based on the feature vector similarity, associated data with a matching degree greater than a preset threshold are determined in the candidate data. Once the event subject dimension is determined, based on the preset feature subject mapping relationship, the field data and the associated data corresponding to each of the risk features are sequentially mapped to the corresponding subject dimension; The mapped data are concatenated sequentially based on the order of the subject dimensions to obtain the event-related knowledge data.

4. The event risk detection method based on a large model according to claim 3, characterized in that, The step of matching a corresponding feature extraction strategy based on the modality type, and then extracting risk features based on the feature extraction strategy, specifically includes: If the data is field data, semantic parsing and structured information extraction are performed to construct an event logic graph, and field features are extracted based on the event logic graph; wherein, the field features include at least one of entity features, logical relationship features, event chain features, and conflict features; If the data is an image, keyframes are identified, and image features are extracted from these keyframes. The image features include at least one of the following: target entity features, scene environment features, and image semantic association features. If the data is speech data, it is converted into text data, and speech features are extracted during the conversion process; wherein, the speech features include at least one of speech-to-text conversion features and emotional intent features; If the data is spatiotemporal trajectory data, trajectory analysis is performed based on time series to construct a trajectory time series graph, and trajectory features are extracted based on the trajectory time series graph; wherein, the trajectory features include at least one of trajectory morphology features, trajectory correlation features, and trajectory anomaly features.

5. The event risk detection method based on a large model according to claim 1, characterized in that, The process of invoking the event risk detection model, performing event risk detection based on the event risk detection model and the input prompt words, and outputting detection conclusions specifically includes: The event risk detection model is invoked, the input prompt words are input into the event risk detection model, and the event risk detection model is driven to perform risk analysis and deduction based on the prompt words; Based on the preset output specifications, the event risk detection model outputs the initial detection conclusion. The initial detection conclusion is verified, and the final detection conclusion is output after the verification is passed.

6. The event risk detection method based on a large model according to claim 1, characterized in that, After performing event risk detection based on the event risk detection model and the input prompt words, and outputting the detection conclusion, the method further includes: Based on the multi-dimensional detection criteria in the detection conclusions, an event dynamic graph is constructed, and the risk is simulated through a pre-set time-series graph neural network to generate a risk evolution path; The environmental features and external intervention variables corresponding to the event to be detected are obtained to construct a simulated environment; The risk evolution path is input into different simulation environments for simulation to generate a risk evolution scenario set; Each path in the risk evolution scenario set is analyzed to determine the scenario branch points. Based on the path probability, the maximum risk level, and the scenario branch points, a master-directed path is generated to determine the risk evolution prediction result based on the master-directed path.

7. The event risk detection method based on a large model according to claim 6, characterized in that, The process of analyzing each path in the risk evolution scenario set, determining scenario branching points, and generating a master-directed programmed path based on path probability, maximum risk level, and the scenario branching points specifically includes: The nodes that appear in risk evolution paths that exceed a preset proportion, and whose risk vectors for each risk evolution path are greater than a preset vector threshold, are designated as scenario branch points. Based on preset weights, path probabilities, and maximum risk levels, a comprehensive path detection function is constructed to determine the risk value corresponding to each path. The path corresponding to the highest risk value is used as the baseline path; Based on the similarity between the baseline path and other paths after the scenario branch point, path fusion is performed to generate the director-directed path.

8. An event risk detection device based on a large model, characterized in that, The device includes a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to perform the method described in any one of claims 1-7.

9. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are capable of performing the method described in any one of claims 1-7.

Images