CN-121563018-B - Event risk detection method, equipment and medium based on large model

CN121563018BCN 121563018 BCN121563018 BCN 121563018BCN-121563018-B

Abstract

The embodiment of the application discloses an event risk detection method, equipment and medium based on a large model, belongs to the technical field of risk detection, and solves the problems that in the prior art, scene adaptability is poor, samples and manual rules are relied on, so that the risk detection accuracy and efficiency of complex and various events are low. The method comprises the steps of constructing an event risk detection processing template based on acquired industry knowledge and industry scene information, wherein the event risk detection processing template is used for providing detection processing logic for an event risk detection model, acquiring event basic data corresponding to an event to be detected, searching event related knowledge data in a preset vector database based on the event basic data, filling the event basic data and the event related knowledge data into the event risk detection processing template to obtain an input prompt word, calling the event risk detection model, carrying out event risk detection based on the event risk detection model and the input prompt word, and outputting a detection conclusion.

Inventors

LIU GUOQIANG
LI LIANG
HE CHAO
MEI YANZHENG

Assignees

浙江锦智人工智能科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260122

Claims (9)

1. A method for event risk detection based on a large model, the method comprising: constructing an event risk detection processing template based on the acquired industry knowledge and industry scene information, wherein the event risk detection processing template is used for providing detection processing logic for the event risk detection model; acquiring event basic data corresponding to an event to be detected, and retrieving event related knowledge data from a preset vector database based on the event basic data; filling the event basic data and the event related knowledge data into the event risk detection processing template to obtain an input prompt word; invoking the event risk detection model, carrying out event risk detection based on the event risk detection model and the input prompt word, and outputting a detection conclusion, wherein the detection conclusion at least comprises one of an event risk level, a multidimensional detection basis and an event alarm identifier; Before the event basic data and the event related knowledge data are filled into the event risk detection processing template, the method further comprises: Carrying out semantic analysis on the event basic data to extract event semantic units, wherein the event semantic units at least comprise one of entity units, relationship units and attribute units; The event related knowledge data is subjected to semantic analysis, and a knowledge semantic unit is extracted, wherein the knowledge semantic unit at least comprises one of a rule unit, a case unit and a concept unit; According to the semantic similarity, determining an association relation between the event semantic unit and the knowledge semantic unit, and matching the related event semantic unit and the related knowledge semantic unit to a preset semantic slot based on the association relation so as to splice the event basic data and the event related knowledge data and generate an input prompt word.
2. The method for detecting event risk based on the large model according to claim 1, wherein the constructing an event risk detection processing template based on the acquired industry knowledge and industry scene information specifically comprises: The detection system comprises an industry knowledge and industry scene information, a plurality of detection modules, a control module and a control module, wherein the detection modules at least comprise one of a task description module, a processing logic module, a constraint limiting module, a background knowledge module, an output standard module, an event input module, a risk detection instance module, a disthought checking module and a risk alarming module; Wherein the processing logic module comprises at least one of an alarm or cue frequency dimension, a hazard level change dimension, a person related change dimension, a person emotion change dimension, a tool related change dimension, a treatment mode change dimension, an item related or monetary amount related change dimension, and a behavior related change dimension; And constructing the event risk detection processing template based on the multiple modules and the multiple dimensions.
3. The method for detecting event risk based on a large model according to claim 1, wherein the acquiring the event basic data corresponding to the event to be detected, and retrieving the event related knowledge data in a preset vector database based on the event basic data, specifically comprises: determining a mode type of event data corresponding to the event to be detected, wherein the mode type at least comprises one of field data, image data, voice data and time-space track data; Based on the mode type, matching a corresponding feature extraction strategy to extract risk features based on the feature extraction strategy, wherein the risk features at least comprise one of field features, image features, voice features and track features; Inputting the extracted risk features into a preset vector database, screening out co-modal and cross-modal candidate data from the preset vector database through risk feature labels, and determining associated data with the matching degree larger than a preset threshold value from the candidate data based on the feature vector similarity; determining the main dimension of an event, and sequentially mapping the field data and the associated data corresponding to each risk feature to the corresponding main dimension based on a preset feature main mapping relation; And splicing the mapped data in turn based on the sequence of the main body dimensions to obtain the event related knowledge data.
4. A method for detecting risk of event based on big model as claimed in claim 3, wherein said matching a corresponding feature extraction policy based on said modality type to perform risk feature extraction based on said feature extraction policy specifically comprises: If the field data is the field data, semantic analysis and structured information extraction are carried out to construct an event logic map, and field feature extraction is carried out based on the event logic map, wherein the field feature at least comprises one of entity features, logic relationship features, event chain features and contradiction conflict features; if the image data is the image data, determining an image key frame, and extracting image characteristics of the image key frame, wherein the image characteristics at least comprise one of target entity characteristics, scene environment characteristics and image semantic association characteristics; if the voice data is voice data, converting the voice data into text data, and extracting voice characteristics in the conversion process, wherein the voice characteristics at least comprise one of voice text conversion characteristics and emotion intention characteristics; And if the track is space-time track data, track analysis is carried out based on a time sequence, a track time sequence diagram is constructed, and track feature extraction is carried out based on the track time sequence diagram, wherein the track feature at least comprises one of track morphological features, track association features and track abnormal features.
5. The method for detecting event risk based on the large model according to claim 1, wherein the calling the event risk detection model, performing event risk detection based on the event risk detection model and the input prompt word, and outputting a detection conclusion specifically comprises: Invoking an event risk detection model, inputting the input prompt word into the event risk detection model, and driving the event risk detection model to conduct risk analysis deduction based on the prompt word; Based on a preset output specification, outputting an initial detection conclusion through the event risk detection model; and checking the initial detection conclusion, and outputting a final detection conclusion after the checking is passed.
6. The method for detecting event risk based on a large model according to claim 1, wherein after the event risk detection is performed with the input prompt word based on the event risk detection model, and a detection conclusion is output, the method further comprises: based on the multi-dimensional detection basis in the detection conclusion, constructing an event dynamic diagram, simulating risks through a preset time sequence diagram neural network, and generating a risk evolution path; acquiring environmental characteristics and external intervention variables corresponding to the event to be detected so as to construct a simulation environment; inputting the risk evolution paths into different simulation environments to be deduced, and generating a risk evolution scene set; And analyzing each path in the risk evolution scene set, determining a scene branching point, and generating a dominant evolution path based on the path probability, the maximum risk level and the scene branching point so as to determine a risk evolution prediction result based on the dominant evolution path.
7. The method for detecting event risk based on large model according to claim 6, wherein analyzing each path in the risk evolution scenario set, determining a scenario branch point, and generating a dominant evolution path based on a path probability, a maximum risk level and the scenario branch point, specifically comprises: taking nodes which appear in risk evolution paths larger than a preset proportion and the risk vectors corresponding to the risk evolution paths are larger than a preset vector threshold as scene branching points; constructing a path comprehensive detection function based on preset weight, path probability and maximum risk level, and determining risk values corresponding to the paths based on the path comprehensive detection function; Taking a path corresponding to the highest risk value as a reference path; and carrying out path fusion on the basis of the similarity relation between the reference path and other paths after the scene branching point, and generating the dominant evolution path.
8. A large model based event risk detection device, characterized in that the device comprises a memory for storing computer program instructions and a processor for executing the program instructions, wherein the computer program instructions, when executed by the processor, trigger the device to perform the method of any of claims 1-7.
9. A non-transitory computer storage medium storing computer executable instructions, wherein the computer executable instructions are capable of performing the method of any one of claims 1-7.

Description

Event risk detection method, equipment and medium based on large model Technical Field The present application relates to the field of risk detection technologies, and in particular, to a method, an apparatus, and a medium for event risk detection based on a large model. Background In the fields of public safety, emergency management and the like, the rapid and accurate risk detection of the event is a key basis for effective treatment and resource allocation. In the prior art, event risk detection mainly depends on three evaluation modes of expert experience, multi-dimension index weighting-based evaluation and traditional machine learning model-based evaluation. The evaluation of expert experience is highly dependent on personal knowledge and on-site judgment of the field expert, but when dealing with sudden, novel or complex events, the response speed and the judgment accuracy are obviously insufficient. Secondly, the evaluation based on multi-dimension index weighting is to preset a series of risk evaluation indexes and assign weights to the indexes, and obtain a risk score through weighting calculation, however, the construction of an index system and the setting of the weights usually depend on subjective experience, and when the information of the event is incomplete and the index calculation deviates, the accuracy of an evaluation result is rapidly reduced. In addition, although the detection mode based on the traditional machine learning model realizes automation to a certain extent, the model effect is seriously dependent on a large number of high-quality and high-cost labeling samples, in the public safety field, the samples are often scarce and have limited coverage scenes, and once the model is trained, evaluation logic of the model is solidified, so that the model is difficult to adapt to the change of business rules. The events are complex and various, and the problems of poor scene adaptability, severe dependence on samples and manual rules occur in the prior art, so that the risk detection accuracy and efficiency of the complex and various events are low. Disclosure of Invention The embodiment of the application provides an event risk detection method, equipment and medium based on a large model, which are used for solving the technical problems that the prior art often has poor scene adaptability and seriously depends on samples and manual rules, so that the risk detection accuracy and efficiency of complex and diverse events are low. The embodiment of the application adopts the following technical scheme: The embodiment of the application provides an event risk detection method based on a large model. The method comprises the steps of constructing an event risk detection processing template based on acquired industry knowledge and industry scene information, wherein the event risk detection processing template is used for providing detection processing logic for an event risk detection model, acquiring event basic data corresponding to an event to be detected, searching event related knowledge data in a preset vector database based on the event basic data, filling the event basic data and the event related knowledge data into the event risk detection processing template to obtain an input prompt word, calling the event risk detection model, carrying out event risk detection based on the event risk detection model and the input prompt word, and outputting a detection conclusion, wherein the detection conclusion at least comprises one of an event risk level, a multi-dimensional detection basis and an event alarm mark. In one implementation mode of the application, an event risk detection processing template is constructed based on acquired industry knowledge and industry scene information, and the event risk detection processing template specifically comprises a plurality of detection modules constructed based on the industry knowledge and the industry scene information, wherein the plurality of detection modules at least comprise one of a task description module, a processing logic module, a constraint limiting module, a background knowledge module, an output specification module, an event input module, a risk detection instance module, a dislike checking module and a risk warning module, and the processing logic module at least comprises one of an alarm or cue frequency dimension, a hazard degree change dimension, a person emotion change dimension, a tool change dimension, a treatment mode change dimension, an article or amount change dimension and a behavior change dimension. And constructing an event risk detection processing template based on the plurality of modules and the plurality of dimensions. In one implementation mode of the method, event basic data corresponding to an event to be detected are obtained, event related knowledge data are retrieved in a preset vector database based on the event basic data, the method specifically comprises the steps of determining a mode type of t