Event analysis method and device, equipment and storage medium

Abstract

The invention belongs to the field of big data processing, and particularly relates to an event analysis method, device and equipment and a storage medium, target event information is obtained, the target event information is converted into event feature vectors, the event feature vectors are input into multiple analysis models respectively, and the analysis models output corresponding event types and initial risk levels respectively. Performing similarity matching on the event feature vector and a historical event feature vector of a historical event in a preset knowledge base, screening one or more pieces of historical event information of which the similarity score is greater than a preset threshold, and determining the credibility of each analysis model based on the historical event information and the event type, performing weighted fusion on the initial analysis result based on the credibility of each analysis model, and determining a target risk level corresponding to the target event information; according to the method, by combining multi-model analysis, historical data similarity matching and credibility weighted fusion, the accuracy of risk level judgment of the target event is improved.

Description

Technical Field

[0001] This application relates to the field of big data processing, and in particular to an event analysis method, apparatus, device, and storage medium. Background Technology

[0002] In the era of rapid digital information dissemination, social hotspots and potential risk events often spread rapidly and continue to escalate within a short period of time through online channels such as social media, news platforms, and forums. This significantly shortens the life cycle of public opinion and makes the dissemination paths more complex and volatile. Against this backdrop, public opinion monitoring systems have become a core tool for enterprises and other social organizations to promptly grasp public opinion trends, assess public attitudes, warn of potential crises, and respond quickly. Therefore, higher and more stringent requirements are currently being placed on public opinion monitoring systems in terms of the accuracy of information identification, the timeliness of analysis and processing, and their dynamic adaptability to emerging topics and semantic changes.

[0003] In existing technologies, event analysis methods typically first use natural language processing techniques to transform massive amounts of unstructured data, such as online text, into structured data. Then, based on pre-trained machine learning or deep learning classification models, the structured event information is automatically judged to identify the event type and the risk level of the event.

[0004] However, existing technologies still heavily rely on manually defined rule templates in the data structuring stage. This strong reliance results in poor system flexibility, making it difficult to adapt to the diversity of language expressions and the rapid evolution of online slang. In particular, it performs poorly when facing new events or semantic variations, and is prone to misjudgment or omission, leading to low accuracy in event analysis results. Summary of the Invention

[0005] This application provides an event analysis method, apparatus, device, and storage medium to address the problem of low accuracy in event analysis results.

[0006] In a first aspect, embodiments of this application provide an event analysis method, including:

[0007] Obtain target event information and convert the target event information into an event feature vector;

[0008] The event feature vector is input into multiple analysis models, and each analysis model outputs the corresponding event type and initial risk level.

[0009] The event feature vector is matched with the historical event feature vectors of historical events in the preset knowledge base to filter one or more historical event information with similarity scores greater than a preset threshold, and the trust level of each analysis model is determined based on the historical event information and the event type.

[0010] The initial analysis results are weighted and fused based on the trust level of each analysis model to determine the target risk level corresponding to the target event information.

[0011] As an optional implementation, determining the trust level of each analysis model based on the historical event information and the event type includes:

[0012] Extract historical event data of the same type as the event type corresponding to each of the analysis models from the historical event information;

[0013] Based on the historical judgment accuracy of each of the aforementioned analytical models in the historical data of similar events, the historical judgment accuracy is used as the first evaluation index of the trustworthiness of each of the aforementioned analytical models.

[0014] Calculate the degree of consistency between the initial risk level of each analytical model and the initial risk level of other analytical models, and use the degree of consistency as the second evaluation index of the trustworthiness of each analytical model;

[0015] The first evaluation index and the second evaluation index are weighted and calculated based on preset weights to obtain the trust level of each analysis model.

[0016] As an alternative implementation, the plurality of analysis models include at least two of the following: a large language model, a generative pre-trained transformation model, and a fine-tuned bidirectional encoder representation model.

[0017] As an optional implementation, the target event information includes at least one of text, image, and audio;

[0018] The step of converting the target event information into an event feature vector includes:

[0019] If the target event information is text, then a pre-trained language model is used to encode the text, generating a text embedding vector containing semantic features as the event feature vector;

[0020] If the target event information is an image, then a convolutional neural network or a visual model is used to extract the visual features of the image and generate an image embedding vector as the event feature vector.

[0021] If the target event information is audio, the audio signal is converted into a spectrogram through Mel spectrum conversion, and then audio features are extracted using an audio feature extraction model to generate an audio embedding vector as the event feature vector.

[0022] If the target event information contains multiple types of information such as text, image, and audio, then feature extraction is performed on each type of information to obtain the corresponding type of embedding vector, and the embedding vectors of multiple types of information are merged into the event feature vector through vector fusion.

[0023] As an optional implementation, the method further includes:

[0024] If the target risk level is greater than or equal to the risk level threshold, an early warning mechanism is triggered;

[0025] If the target risk level is less than the risk level threshold, the early warning mechanism will not be triggered.

[0026] As an optional implementation, the triggering warning mechanism includes:

[0027] The target warning level is determined based on the target risk level, and the higher the target risk level, the higher the warning level.

[0028] According to the preset correspondence between warning levels and warning methods, the warning method corresponding to the target warning level is activated, and the warning method includes at least one of sound alarm, light flashing and information push.

[0029] As an optional implementation, the method for constructing the preset knowledge base includes:

[0030] Acquire historical event information and standardize the historical event information;

[0031] The standardized historical event information is transformed into historical event feature vectors, and the corresponding known event types and actual risk levels are extracted.

[0032] The historical event feature vectors, known event types, and actual risk levels are indexed and stored in the database to generate a preset knowledge base.

[0033] Secondly, embodiments of this application provide an event analysis apparatus, comprising:

[0034] The processing module is used to acquire target event information and convert the target event information into an event feature vector;

[0035] The processing module is also used to input the event feature vector into multiple analysis models respectively, and each analysis model outputs the corresponding event type and initial risk level respectively;

[0036] The determination module is used to perform similarity matching between the event feature vector and the historical event feature vector of historical events in the preset knowledge base, filter one or more historical event information with similarity scores greater than a preset threshold, and determine the trust level of each of the analysis models based on the historical event information and the event type.

[0037] The determining module is used to perform weighted fusion of the initial analysis results based on the trust level of each analysis model to determine the target risk level corresponding to the target event information.

[0038] As an optional implementation, the event analysis device includes: an extraction module;

[0039] The extraction module is used to extract historical event data of the same type as the event type corresponding to each of the analysis models from the historical event information.

[0040] The determining module is further configured to use the historical judgment accuracy rate of each of the analysis models in the historical data of similar events as the first evaluation index of the trust level of each of the analysis models.

[0041] The determining module is further configured to calculate the degree of consistency between the initial risk level of each analysis model and the initial risk level of other analysis models, and use the degree of consistency as a second evaluation index of the trustworthiness of each analysis model.

[0042] The determining module is further configured to perform weighted calculations on the first evaluation index and the second evaluation index based on preset weights to obtain the trust level of each analysis model.

[0043] As an optional implementation, the determining module is also used to include at least two of the plurality of analysis models, including a large language model, a generative pre-trained transformation model, and a fine-tuned bidirectional encoder representation model.

[0044] As an optional implementation, the determining module is further configured to, if the target event information is text, encode the text using a pre-trained language model to generate a text embedding vector containing semantic features as the event feature vector;

[0045] The determining module is further configured to, if the target event information is an image, use a convolutional neural network or a visual model to extract the visual features of the image and generate an image embedding vector as the event feature vector.

[0046] The determining module is further configured to, if the target event information is audio, convert the audio signal into a spectrogram through Mel spectrum conversion, then extract audio features using an audio feature extraction model, and generate an audio embedding vector as the event feature vector.

[0047] The determining module is further configured to, if the target event information contains multiple types of information such as text, image, and audio, extract features for each type of information to obtain the corresponding type of embedding vector, and merge the embedding vectors of multiple types of information into the event feature vector through vector fusion.

[0048] As an optional implementation, the processing module is further configured to trigger an early warning mechanism if the target risk level is greater than or equal to the risk level threshold.

[0049] The processing module is further configured to not trigger the early warning mechanism if the target risk level is less than the risk level threshold.

[0050] As an optional implementation, the determining module is further configured to determine a target warning level based on the target risk level, wherein the higher the target risk level, the higher the warning level;

[0051] The processing module is also used to activate the warning method corresponding to the target warning level according to the preset correspondence between warning levels and warning methods. The warning method includes at least one of sound alarm, light flashing, and information push.

[0052] As an optional implementation, the event analysis device includes: a generation module;

[0053] The processing module is also used to acquire historical event information and to standardize the historical event information.

[0054] The processing module is also used to convert the standardized historical event information into historical event feature vectors and extract the corresponding known event types and actual risk levels.

[0055] The generation module is used to index and store the historical event feature vectors, known event types and actual risk levels in the database to generate a preset knowledge base.

[0056] Thirdly, embodiments of this application provide an event analysis device, including: a receiver, a transmitter, a memory, and a processor;

[0057] Receiver, used to receive instructions and data;

[0058] A transmitter is used to send commands and data;

[0059] The memory stores computer-executed instructions;

[0060] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0061] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0062] The event analysis method provided in this application acquires target event information and converts it into event feature vectors, which are then input into multiple analysis models. This leverages the unique perspectives and algorithmic advantages of different models to analyze events from multiple dimensions, avoiding the limitations of a single model and resulting in a more comprehensive assessment of event type and initial risk level. By performing similarity matching between the event feature vectors and historical event feature vectors in a pre-defined knowledge base, experience from similar past events can be drawn upon, providing a reference for current event analysis. Based on historical event information and event type, the trust level of each analysis model is determined, and the initial analysis results are weighted and fused according to the trust level. This dynamically adjusts the weight of different model results in the final judgment, fully utilizing the role of highly trustworthy models and reducing the influence of untrustworthy models. By integrating multi-model analysis and historical data, the accuracy of target event risk level determination is effectively improved. Attached Figure Description

[0063] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0064] Figure 1 Flowchart of the event analysis method provided in this application Figure 1 ;

[0065] Figure 2 Flowchart of the event analysis method provided in this application Figure 2 ;

[0066] Figure 3 A schematic diagram of the event analysis device provided in this application;

[0067] Figure 4 A schematic diagram of the event analysis device provided in this application.

[0068] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions 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, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0070] In today's environment of rapid digital information dissemination, social hotspots and potential risk events often spread rapidly and continue to escalate through channels such as social media, news platforms, and online forums. The lifecycle of public opinion is significantly compressed, and the dissemination paths are becoming increasingly complex and non-linear. This trend makes public opinion dynamics change rapidly, posing a severe challenge to the response speed and judgment capabilities of organizations. Against this backdrop, public opinion monitoring systems have become a crucial tool for enterprises and various social organizations to grasp public opinion trends, provide early warnings of crisis risks, and implement effective responses.

[0071] Existing public opinion analysis technologies typically employ natural language processing to transform unstructured text data into structured event information. Then, pre-trained machine learning or deep learning models are used to automatically classify and determine the event type and its risk level.

[0072] However, existing technologies still heavily rely on manually designed rule templates in the text structuring process. Faced with the diversity of online language, the rapid evolution of expression methods, and the dynamic semantic changes in emerging events, these rule-driven methods often struggle to effectively identify newly emerging expression patterns, easily leading to misjudgments or omissions. Especially when dealing with sudden events or semantic variations, the system's adaptability impacts the accuracy of event recognition.

[0073] To address the aforementioned issues, the event analysis method provided in this application first acquires target event information and converts it into an event feature vector. This vector is then input into multiple analysis models, each outputting the corresponding event type and initial risk level. Simultaneously, the event feature vector is matched with historical event feature vectors in a pre-defined knowledge base to filter out historical event information with similarity scores exceeding a pre-defined threshold. The trust level of each analysis model is then determined based on this historical event information and the event type. Finally, the initial analysis results are weighted and fused based on the trust levels of each model to determine the target risk level corresponding to the target event information. Accurate judgment is achieved through multi-model analysis, historical data similarity matching, and trust-weighted fusion.

[0074] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0075] Figure 1 Flowchart of the event analysis method provided in this application Figure 1 The executing entity in this embodiment is, for example, a control system. Figure 1 As shown, the method includes:

[0076] S101: Obtain target event information and convert the target event information into an event feature vector.

[0077] Among them, target event information refers to various types of data and content collected and extracted in specific scenarios (such as public opinion perception of high-risk accidents, event early warning analysis, etc.) for target events (i.e. core events that need to be monitored, analyzed or warned, such as high-risk accidents such as chemical explosions, building collapses, major traffic accidents, etc.).

[0078] The key content that is strongly related to the target event is accurately screened and extracted from multi-source data. Taking the public opinion scenario of high-risk accidents as an example, the target event may be a specific accident such as a chemical leak or a building collapse. The information sources need to cover official reports (such as the preliminary report of the event released by the emergency management department), authoritative media reports (including the core facts such as the time, location and parties involved in the event), and social media and forum discussions (the focus, demands and emotional tendencies of public attention).

[0079] Encoding techniques are used to transform unstructured target event information into computable numerical vectors. Pre-trained models such as BGE (BERT-like General Embedding, a pre-trained language model) can be used to map textual content into embedding vectors (i.e., embedding vectors) that contain semantic features, thus preserving the core meaning of the target event while also reflecting the relationships between different pieces of information.

[0080] This conversion process eliminates the differences in the original information format, allowing the key attributes of the target event (such as event type, severity, public opinion characteristics, etc.) to be presented in a quantitative form.

[0081] S102: Input the event feature vector into multiple analysis models, and each analysis model outputs the corresponding event type and initial risk level.

[0082] Among them, event type is a category label assigned after classifying specific events based on their core attributes (such as the area of ​​occurrence, essential characteristics, and manifestation). For example, high-risk events are divided into specific categories such as chemical leaks, building collapses, and major traffic accidents.

[0083] The initial risk level is an initial risk assessment of an event based on the input event feature vector and pre-defined assessment logic (such as the potential scope of harm, degree of impact, urgency, etc.).

[0084] Accurate event analysis is achieved through multi-dimensional modeling. Multiple analysis models have different functional focuses; for example, some models concentrate on fine-grained event classification (e.g., classifying high-risk accidents into chemical leaks, building collapses, major traffic accidents, etc.), while others emphasize quantitative risk level assessment. The core features of an event (such as semantic information, spatiotemporal attributes, and scope of impact) are extracted by each model from the event feature vector. Combined with the judgment logic used during training, each model independently outputs corresponding results—the event type label clarifies the essential attributes of the event, while the initial risk level reflects the urgency and potential harm of the event through numerical values ​​or level classifications (e.g., low, medium, and high).

[0085] S103: Perform similarity matching between the event feature vector and the historical event feature vectors in the preset knowledge base, filter out one or more historical event information with similarity scores greater than the preset threshold, and determine the trust level of each analysis model based on the historical event information and event type.

[0086] By calculating indices such as cosine similarity in vector space, the correlation between current events and historical events can be quantified. When the similarity score exceeds a preset threshold (e.g., 0.7), it indicates that the two events have a high degree of consistency in core features (such as the nature of the event, the scope of impact, and the evolution trend). These selected historical events will serve as reference benchmarks. For example, if the current event is a leak in a chemical industrial park, matching may filter out historical cases of similar parks and leaks of similar substances.

[0087] The event types and initial risk levels output by each model are compared with the actual results of historical events (such as the final determined event types and actual risk levels). If the model's prediction results for similar historical events are highly consistent with the actual situation, the model is given a higher confidence level; if the deviation is large, its confidence level weight is reduced.

[0088] S104: Based on the trust level of each analysis model, the initial analysis results are weighted and fused to determine the target risk level corresponding to the target event information.

[0089] The initial analysis results are weighted and fused based on the confidence levels of each analytical model. During the fusion process, models with high confidence levels are assigned higher weights, while models with low confidence levels are assigned lower weights. This ensures that the final result is more biased towards the output of the model that has performed better in the past. For example, if model A has a confidence level of 0.9 (out of 1.0) in the chemical accident category, and its initial risk level is "medium," while model B has a confidence level of 0.6 and its output level is "low," then model A's result will have a greater influence during the weighted fusion, making the final risk level closer to its judgment.

[0090] For event types with a large number of similar historical events, a "trust level-accuracy" mapping formula can be used to directly calculate the weights based on historical consistency. For new events (with fewer historical cases), expert-preset weights can be appropriately introduced as a supplement. The target risk level is finally obtained through weighted calculation.

[0091] The event analysis method provided in this application acquires target event information, converts the target event information into event feature vectors, inputs the event feature vectors into multiple analysis models, and each analysis model outputs the corresponding event type and initial risk level. The event feature vectors are then matched with historical event feature vectors in a preset knowledge base to determine similarity. One or more historical event information with similarity scores greater than a preset threshold are selected. Based on the historical event information and event type, the trust level of each analysis model is determined. The initial analysis results are then weighted and fused based on the trust levels of each analysis model to determine the target risk level corresponding to the target event information. This method improves the accuracy of target event risk level determination by combining multi-model analysis, historical data similarity matching, and trust-weighted fusion.

[0092] Figure 2 Flowchart of the event analysis method provided in this application Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the examples, the event analysis method is described in detail, which includes:

[0093] S201: Obtain target event information.

[0094] The target event information includes at least one of text, image, and audio.

[0095] S202: If the target event information is text, then a pre-trained language model is used to encode the text and generate a text embedding vector containing semantic features as the event feature vector.

[0096] When using pre-trained language models to encode and generate text embedding vectors, semantic relationships and core information in the text are captured through unstructured processing. Textual information (such as public opinion comments and accident reports) is inherently unstructured, without a fixed format, and contains a large amount of implicit semantics (such as detailed descriptions of events). Pre-trained language models (such as BGE (BAAI General Embedding, an open-source general vector model) and BERT (Bidirectional Encoder Representations from Transformers)) first perform word segmentation and part-of-speech tagging on the text, and then model the contextual relationships of words and sentences through a multi-layer Transformer architecture. For example, the model can identify the semantic relationship between "chemical leak" and "toxic gas diffusion," and finally transform the explicit facts and implicit semantics in these unstructured texts into low-dimensional, continuous numerical vectors, i.e., text embedding vectors.

[0097] Structured processing requires manually defining fixed fields such as "accident type" and "scope of impact" and extracting information. In contrast, pre-trained language models can directly encode the original text end-to-end. Even if the text is not grammatically correct (e.g., colloquial public opinion commentary) or the information is scattered (e.g., a text simultaneously contains the event's time, location, and public demands), the model can still transform the text into an embedded vector containing complete semantic features through its semantic understanding acquired during training on massive amounts of corpus data. This approach preserves the information integrity of unstructured text while avoiding the limitations of manual feature design, making the generated event feature vectors more closely match the true semantics of the text.

[0098] S203: If the target event information is an image, then a convolutional neural network or visual model is used to extract the visual features of the image and generate an image embedding vector as the event feature vector.

[0099] Visual features are extracted directly from raw images using Convolutional Neural Networks (CNNs) or visual models. Images, as unstructured data, contain information (such as smoke, damaged buildings, and rescue equipment at an accident scene) in the form of pixel combinations of pixels, colors, and textures, without the need for manually pre-defined structured labels (such as "fire severity" or "building damage type"). The model processes these features through multiple convolutional and pooling layers, first capturing basic visual elements such as edges and color blocks, then combining them to form higher-order features (such as "smoke-filled areas" or "collapsed structures"). Finally, these abstract visual features are mapped into low-dimensional numerical vectors, i.e., image embedding vectors, enabling the machine to understand image content and transform it into computable event feature vectors.

[0100] This unstructured processing method preserves the integrity of the original image information while avoiding the limitations of manual feature design. For example, when faced with an image of a chemical accident scene, traditional structured processing requires manual annotation of fixed dimensions such as "whether there is a flame" and "the location of the leak," while models such as CNNs can autonomously learn more subtle key features (such as smoke concentration gradients and the degree of equipment deformation), features that are often difficult to exhaustively capture through manual pre-setting. The generated image embedding vectors not only contain explicit visual elements but also implicitly contain the spatial relationships between elements (such as "the relative position of the flame and the liquid storage tank"), making the machine's analysis of image-based event information closer to human visual cognitive logic.

[0101] S204: If the target event information is audio, the audio signal is converted into a spectrogram through Mel spectrum conversion, and then audio features are extracted using an audio feature extraction model to generate an audio embedding vector as the event feature vector.

[0102] Mel-frequency conversion (MFC) transforms raw audio signals into machine-readable spectrograms. Audio, as unstructured data, exists as continuous sound wave vibrations (such as explosions at an accident scene, shouts from rescuers, and abnormal noises from equipment malfunctions), making it impossible for models to directly recognize. MFC first slices the audio signal by time, calculating the frequency components and energy of each time segment. Then, by simulating the sensitivity of the human auditory system to different frequencies (more sensitive to low and mid-frequency frequencies, relatively less sensitive to high frequencies), it maps linear frequencies onto a Mel frequency scale, ultimately generating a two-dimensional spectrogram (where the horizontal axis represents time, the vertical axis represents Mel frequency, and color depth represents the energy intensity of the corresponding frequency), transforming the unstructured sound wave signal into visually appealing grid data.

[0103] Audio feature extraction models (such as CNN-based audio classification models, VGGish (Visual GeometryGroupish), YAMNet (YouTube Audio Mobile Network), etc.) are used to further extract features and generate audio embedding vectors, without requiring manual pre-setting of structured labels (such as "explosion sound type" or "urgency level of shouts"). The model automatically captures key audio features from the spectrogram: for example, identifying energy peaks in specific frequency ranges (corresponding to explosion sounds in accidents), temporal changes in sound waves (corresponding to continuous voice commands for rescue), and the spectral distribution of background noise (corresponding to the degree of chaos in the scene). These abstract features are then mapped into low-dimensional numerical vectors, i.e., audio embedding vectors, through a multi-layer network structure. This unstructured processing method not only preserves the complete temporal and frequency information of the audio signal but also captures subtle features that are difficult to define manually (such as the frequency fluctuation patterns of abnormal equipment noises) through model autonomous learning, enabling the generated event feature vectors to accurately reflect the event scene information contained in the audio.

[0104] S205: If the target event information contains multiple types of information such as text, image, and audio, then feature extraction is performed on each type of information to obtain the corresponding type of embedding vector, and the embedding vectors of multiple types of information are merged into an event feature vector through vector fusion.

[0105] When the target event information contains various unstructured data such as text, images, and audio, feature extraction must be performed separately for each type of information. For text information (such as accident scene descriptions and public opinion comments), pre-trained language models such as BGE are used to generate text embedding vectors containing semantic features to capture the factual details and sentiment of the event. For image information (such as accident scene photos and surveillance footage), CNNs or visual models are used to extract visual features and transform them into image embedding vectors reflecting scene elements (such as damaged equipment and rescue scenes). For audio information (such as on-site recordings and eyewitness accounts), it is processed by Mel-spectrum transformation and models such as YAMNet to generate audio embedding vectors containing sound features (such as explosion sounds and cries for help). Each type of information is processed directly in an unstructured manner without the need for manually pre-setting a uniform format, thus preserving the integrity and uniqueness of the original data to the greatest extent.

[0106] After extracting single-type embedding vectors, vector fusion technology is used to merge multi-dimensional vectors into a unified event feature vector, achieving synergistic effects from multi-source information. Vector fusion is not a simple concatenation, but rather combines the weights of each type of information (e.g., text information, which is more critical in describing the facts of the event, can be given higher weights; images and audio, which more intuitively restore the atmosphere of the scene, can have their weights dynamically adjusted according to the scene), and uses attention mechanisms, weighted averaging, or feature concatenation followed by optimization through a fully connected layer to deeply integrate features such as the semantic logic of text, the visual details of images, and the scene sound effects of audio. For example, in the multi-source information of a chemical accident, the text embedding vector contains the key semantics of "benzene leakage," the image embedding vector presents visual evidence of "damaged storage tank," and the audio embedding vector captures the sound features of "gas leakage hissing sound." The fused event feature vector can simultaneously cover the attributes of these three types of information.

[0107] S206: Input the event feature vector into multiple analysis models, and each analysis model outputs the corresponding event type and initial risk level.

[0108] Optionally, multiple analysis models may include at least two of the following: a large language model, a generative pre-trained transformation model, and a fine-tuned bidirectional encoder representation model.

[0109] By inputting event feature vectors into multiple analysis models, the comprehensiveness and reliability of event judgment are improved through parallel analysis of multiple models. The large language model, with its powerful semantic understanding capabilities, extracts contextual relationships from the feature vectors, outputting event types (e.g., "toxic gas leak at a chemical plant") and risk levels that fit the contextual logic. The generative pre-trained transformation model excels in generating structured analysis results, combining event features to generate initial risk assessments containing details such as causes and scope of impact. The finely tuned bidirectional encoder, optimized for specific event types, demonstrates greater accuracy in type identification and risk quantification in subdivided fields (e.g., construction accidents, traffic disasters). The independent outputs of each model include both the classification of the event's essence (event type) and a preliminary quantification of its potential hazards (initial risk level), forming a multi-dimensional analytical perspective.

[0110] For example, the large language model is Qwen3, the generative pre-trained transformation model is GLM-4 (General Language Model 4), or the fine-tuned bidirectional encoder representation model is the BERT small model (Small Bidirectional Encoder Representations from Transformers Model).

[0111] S207: Perform similarity matching between the event feature vector and the historical event feature vectors in the preset knowledge base, and filter one or more historical event information with similarity scores greater than the preset threshold.

[0112] Whether it's text semantic vectors, image visual vectors, or multimodal fusion vectors, they are all compared with standardized historical event feature vectors in the knowledge base (such as feature vectors from past chemical accidents and traffic disasters) using algorithms like cosine similarity and Pearson correlation coefficient to calculate the degree of association. For example, if the target event is a "residential building fire," its feature vector includes information such as the location of the fire (e.g., kitchen, bedroom), combustible materials (e.g., wooden furniture, appliances), the extent of the fire spread (e.g., single household, entire building), and whether anyone is trapped. During matching, it is compared one by one with the vectors of all fire-related historical events in the knowledge base, converting implicit associations such as "whether the causes of the fire are similar," "whether the rescue response time is close," and "whether the scale of property damage is comparable" into a similarity score of 0-1. The higher the score, the stronger the consistency of the two fire events in terms of core features (e.g., events where both are fires caused by kitchen appliances, the fire is confined to a single household, and there are no casualties will have a significantly higher similarity score than a fire caused by the storage of flammable materials that affects the entire building).

[0113] In high-risk accident risk assessment, to ensure the high relevance of reference cases, the threshold is often set at 0.7-0.8 (e.g., only historical events with a score ≥0.75 are retained). In the analysis of new events (where historical cases are fewer), the threshold can be appropriately lowered to 0.5-0.6 to broaden the reference range. For example, if a historical flooding event has a similarity score of 0.81 with the target event, exceeding the threshold of 0.75, information such as its handling process, economic loss data, and rescue resource input will be filtered out; if the score is only 0.68, it will be excluded due to significant differences in core characteristics.

[0114] Optional, preset methods for constructing the knowledge base include:

[0115] Acquire historical event information and standardize the processing of that information;

[0116] The standardized historical event information is transformed into historical event feature vectors, and the corresponding known event types and actual risk levels are extracted.

[0117] The feature vectors of historical events, known event types, and actual risk levels are indexed and stored in the database to generate a pre-defined knowledge base.

[0118] The acquired historical event information may encompass various data types, including text reports, image recordings, and audio files (such as news releases, on-site monitoring footage, and rescue recordings from past chemical accidents). Standardization requires establishing unified rules for different data types: for text information, redundant characters need to be removed, the format of expressions needs to be standardized (e.g., unifying "residential building fire" and "residential building blaze" into "residential building blaze"), and key missing information needs to be supplemented (e.g., unclear event time and location). For image / audio information, basic parameters such as resolution and sampling rate need to be standardized to avoid affecting subsequent feature extraction due to format differences. Through standardization, the originally disorganized historical event information is transformed into standardized data with a unified structure and complete information.

[0119] After standardization, the data needs to be transformed into machine-recognizable historical event feature vectors and indexed and stored by associating them with core labels. For standardized text information, text embedding vectors are generated using models such as BGE; for image / audio information, visual / audio features are extracted using models such as CNN and YAMNet to generate corresponding embedding vectors; if it is multi-type fused information, it is merged into a unified feature vector using vector fusion technology. Simultaneously, known event types (such as "electrical fire in residential buildings" and "flammable material fire in industrial parks") and actual risk levels (such as "high," "medium," and "low") are extracted from the standardized data and used as labels for the feature vectors. Finally, the associated data of "historical event feature vector - known event type - actual risk level" is indexed in a database (such as ChromaDB or Milvus Vector Database) to form a pre-defined knowledge base that can be efficiently retrieved. When new event feature vectors are input, the database can quickly match similar historical events through the index.

[0120] S208: Extract historical event data of the same type as the corresponding event type from historical event information.

[0121] The system extracts historical event data that matches the event types of each analysis model from historical event information, and performs precise semantic matching and data filtering in a pre-defined knowledge base. For example, when Qwen3 outputs "residential building fire," GLM-4 outputs "civil building fire," and the fine-tuned BERT mini-model outputs "residential area fire," the system uses these three highly correlated type tags as filtering conditions to match historical events in databases storing historical events (such as ChromaDB and Milvus) whose event type field contains "fire" and whose scene attributes are "residential building," "civil building," or "residential." The entire process does not require manual definition of matching rules; instead, it is achieved through semantic similarity calculation of event type tags (such as embedding encoding comparison of tags based on the BGE model), ensuring that the extracted data is highly consistent with the output types of each model in terms of core scene and event attributes.

[0122] The extracted historical data not only includes event feature vectors but also links to the output results of various analytical models in handling similar events in the past (such as the accuracy rate of Qwen3 in judging "residential building fires"), the final actual risk level, and the handling results. For example, for "residential building fire" events, the extracted historical data clearly shows that "Qwen3 has an accuracy rate of 92% in this type of event, and GLM-4 has an accuracy rate of 85%", providing a direct basis for the adjudication model to calculate the dynamic trust weights of each model. At the same time, cases in historical events where "two people are injured, corresponding to high risk" can also help verify the rationality of the risk level output by the current model and avoid misjudgments caused by the adaptation bias of a single model to the scenario.

[0123] S209: Based on the historical judgment accuracy of each analysis model in historical data of similar events, the historical judgment accuracy shall be used as the primary evaluation indicator of the trustworthiness of each analysis model.

[0124] Specifically, for each analytical model (such as Qwen3, GLM-4, and the fine-tuned BERT mini-model), cases in which the model has participated in the analysis are selected from extracted historical data of similar events. The number of times the model's judgment results match the actual situation of the historical events (such as the actual event type and actual risk level) is counted, and then divided by the total number of analyses to obtain the historical judgment accuracy rate. For example, if Qwen3 outputs event types, risk levels, and actual situations in 92 out of 100 historical events of the "residential building fire" category, its historical judgment accuracy rate for this type of event is 92%; while GLM-4 only correctly judges 68 out of 80 historical events of the same type, with an accuracy rate of 85%. This indicator directly reflects the model's expertise in specific event types; the higher the accuracy rate, the more reliable the model's understanding and judgment of this type of event.

[0125] For example, in public opinion risk assessment scenarios, accurately identifying event types and risk levels is crucial. However, if historical accuracy is low, even if the model responds quickly and its conclusions are consistent with other models, the reliability of the current output cannot be guaranteed. For instance, a fine-tuned BERT model may have a fast response time, but its accuracy in historical events such as "chemical leaks" is only 60%, far lower than Qwen3's 88%. In such cases, Qwen3's output should be prioritized.

[0126] S210: Calculate the degree of consistency between the initial risk level of each analysis model and the initial risk level of other analysis models, and use the degree of consistency as the second evaluation indicator of the trustworthiness of each analysis model.

[0127] For each analysis model (such as Qwen3, GLM-4, and the fine-tuned BERT mini-model), the initial risk levels output by all other models for the same target event are first collected. Then, the percentage of times the model's output level matches that of other models is calculated to obtain the consistency level. For example, if Qwen3 outputs "high risk," and GLM-4 and the fine-tuned BERT mini-model also output "high risk" and "high risk," respectively, then the consistency level of Qwen3 is 100%. If GLM-4 outputs "medium risk," and its level matches only one model (out of two other models), then its consistency level is 50%. This calculation method not only focuses on "whether it is consistent," but also optimizes the results through weighting (such as giving higher comparison weights to models with high historical accuracy), making the consistency level more closely reflect actual trust requirements.

[0128] For example, a model might have high accuracy in most historical events, but exhibit bias in current emerging events (such as "new energy vehicle battery fires") due to insufficient training data. In this case, its output risk level may differ from other models better suited to the scenario. If a model's consistency is significantly lower than other models (e.g., only 30%), even with high historical accuracy, the rationality of its current output should be questioned. Conversely, if a model maintains both high historical accuracy (e.g., 90%) and high consistency (e.g., 95%), it indicates that its output aligns with past experience and the logic of group judgment, significantly increasing its trustworthiness.

[0129] S211: The first and second evaluation indicators are weighted based on preset weights to obtain the confidence level of each analysis model.

[0130] The core step in integrating the model performance of the two dimensions into a unified trust score is to calculate a weighted average of the first assessment indicator (historical judgment accuracy) and the second assessment indicator (consistency of initial risk level) based on preset weights. The key lies in reflecting the importance priority of different indicators through preset weights. Preset weights need to be set according to the needs of the business scenario. For example, in public opinion risk assessment, historical judgment accuracy is directly related to the authenticity of the results and is usually given a higher weight (e.g., α=0.6), while consistency is given a second-highest weight as a supplementary verification indicator (e.g., β=0.4). If facing new types of events (with limited historical data), the weight of consistency can be appropriately increased (e.g., α=0.5, β=0.5). During calculation, the original values ​​of the two indicators (normalized to the 0-1 range) are multiplied by their corresponding weights, and then the products are added together to obtain the model trust score. The formula can be expressed as: Trust Score = (Historical Judgment Accuracy × α) + (Consistency × β). For example, Qwen3 has a historical accuracy rate of 0.92 and a consistency level of 1.0. Calculated with α=0.6 and β=0.4, its confidence level = (0.92×0.6)+(1.0×0.4)=0.552+0.4=0.952. GLM-4 has a historical accuracy rate of 0.85 and a consistency level of 0.5, so its confidence level is (0.85×0.6)+(0.5×0.4)=0.51+0.2=0.71. The numerical differences intuitively reflect the gap in model reliability.

[0131] For example, a fine-tuned BERT model achieved a historical accuracy of 0.88 in "chemical leak" events (higher than GLM-4), but its consistency score was only 0.4 (due to its focus on specific scenarios, which led to significant discrepancies with the general model). Calculated with α=0.6 and β=0.4, the confidence score is (0.88×0.6) + (0.4×0.4) = 0.528 + 0.16 = 0.688. While GLM-4 had a slightly lower accuracy (0.85), its consistency score was higher (0.7), and its confidence score was (0.85×0.6) + (0.7×0.4) = 0.51 + 0.28 = 0.79. In this case, GLM-4 had a higher confidence score due to its superior overall performance.

[0132] S212: Based on the trust level of each analysis model, the initial analysis results are weighted and fused to determine the target risk level corresponding to the target event information.

[0133] To give models with higher trust levels greater influence in the final risk level determination, the impact of single-model bias on the results is reduced by quantitatively integrating the outputs of multiple models. Specifically, the initial risk levels output by each model are first numerically mapped (e.g., "low risk = 1, medium risk = 2, high risk = 3"), and then multiplied by the trust level of the corresponding model to obtain the "weighted risk score" for each model. Subsequently, the weighted risk scores of all models are summed and divided by the sum of the trust levels of each model to obtain the "comprehensive weighted risk score". Finally, based on the range of the comprehensive score, it is back-mapped to the target risk level. For example, for Qwen3 (trust level 0.952, initial risk level "high" = 3), GLM-4 (trust level 0.71, initial risk level "medium" = 2), and the fine-tuned BERT small model (trust level 0.688, initial risk level "high" = 3), the calculation process is as follows: weighted risk score sum = (0.952 × 3) + (0.71 × 2) + (0.688 × 3) = 2.856 + 1.42 + 2.064 = 6.34; trust level sum = 0.952 + 0.71 + 0.688 = 2.35; comprehensive weighted risk score = 6.34 ÷ 2.35 ≈ 2.698. Since 2.698 is closer to the value of 3 corresponding to "high risk", the target risk level is determined to be "high".

[0134] For example, if a model outputs an incorrect initial rating due to historical data bias (such as misclassifying "high risk" as "medium risk"), but its confidence level is low (such as 0.5), its impact on the overall score will be significantly weakened during the fusion process; while a model with a high confidence level (such as 0.9) can still dominate the final result even if its conclusions differ from those of a few other models.

[0135] S213: If the target risk level is greater than or equal to the risk level threshold, an early warning mechanism will be triggered.

[0136] Optionally, triggering an early warning mechanism includes:

[0137] The target warning level is determined based on the target risk level; the higher the target risk level, the higher the warning level.

[0138] According to the preset correspondence between warning levels and warning methods, the warning method corresponding to the target warning level is activated. The warning method includes at least one of sound alarm, flashing light, and information push.

[0139] If the target risk level is greater than or equal to a preset risk level threshold, an early warning mechanism is triggered. This threshold setting clearly defines the risk boundaries requiring intervention. Risk level thresholds are typically set in conjunction with the security needs of the business scenario. For example, in urban public opinion risk monitoring, "medium risk" is often set as the threshold. When the target risk level is "medium" or "high," it indicates that the event may cause public panic, property damage, or disruption to social order, requiring an early warning. If it is "low risk" (such as minor negative comments), no triggering is necessary. This threshold setting avoids the waste of resources caused by excessive early warnings for low-risk events while ensuring that medium- and high-risk events are addressed promptly, forming an efficient triggering logic of "responding as soon as the risk threshold is reached."

[0140] After triggering the early warning mechanism, the corresponding early warning level must first be determined based on the target risk level, and then the warning will be activated using a preset method to achieve tiered early warning and precise notification. The classification of early warning levels is directly linked to the target risk level; the higher the level, the higher the warning level. For example, "medium risk" corresponds to "Level 2 early warning," and "high risk" corresponds to "Level 1 early warning." In some scenarios, "high risk" will be further subdivided into "Special Level 1 early warning" (such as events involving major casualties). At the same time, the system will preset the correspondence between early warning levels and early warning methods: low-level early warnings (such as Level 2) may only activate "information push" (sending SMS notifications to relevant personnel); high-level early warnings (such as Level 1) will be combined with "audio alarm" (activation of the monitoring center's audio-visual alarm) and "light flashing" (emergency indicator lights), and the information push scope will be expanded (synchronously pushed to the emergency command group and higher-level regulatory departments). For example, when the target risk level is "high" (corresponding to Level 1 early warning), the system will immediately trigger the monitoring center's audio-visual alarm, push SMS / emails containing event details to all members of the safety management team, and simultaneously pop up an early warning pop-up on the emergency command platform to ensure that relevant personnel receive information and initiate the response process as soon as possible.

[0141] S214: If the target risk level is less than the risk level threshold, the early warning mechanism will not be triggered.

[0142] If the target risk level is less than the preset risk level threshold, the early warning mechanism will not be triggered. This is the core decision-making logic based on the scope of risk impact and the optimal allocation of resources. When the target risk level is lower than the threshold (e.g., the threshold is set to "medium risk" and the target level is "low risk"), it means that the potential harm of the event is small and the scope of its spread is limited (e.g., a single minor negative comment that is not a hot topic, or a small local event with no impact on personnel or property). There is no need to activate early warning actions such as sound and light alarms or information pushes to multiple departments. This approach can avoid "early warning fatigue" (reducing staff's sensitivity to key early warnings) caused by low-value early warnings, and can concentrate emergency resources (human resources, communication channels, etc.) on the handling of medium- and high-risk events. At the same time, the system will store the details and assessment results of the low-risk event in the database.

[0143] Figure 3 A schematic diagram of the event analysis device provided in this application is shown below. Figure 3 As shown, the event analysis device 300 provided in this embodiment includes:

[0144] Processing module 301 is used to acquire target event information and convert the target event information into an event feature vector;

[0145] The processing module 301 is also used to input the event feature vector into multiple analysis models respectively, and each analysis model outputs the corresponding event type and initial risk level respectively;

[0146] The determination module 302 is used to perform similarity matching between the event feature vector and the historical event feature vector of historical events in the preset knowledge base, filter one or more historical event information with similarity scores greater than a preset threshold, and determine the trust level of each analysis model based on the historical event information and event type.

[0147] The determination module 302 is used to perform weighted fusion of the initial analysis results based on the trust level of each analysis model to determine the target risk level corresponding to the target event information.

[0148] As an optional implementation, the event analysis device includes: an extraction module 303;

[0149] Extraction module 303 is used to extract historical event data of the same type as the corresponding event type of each analysis model from historical event information;

[0150] The determination module 302 is also used to take the historical judgment accuracy rate as the first evaluation indicator of the trustworthiness of each analysis model based on the historical judgment accuracy rate of each analysis model in historical data of similar events.

[0151] The determination module 302 is also used to calculate the degree of consistency between the initial risk level of each analysis model and the initial risk level of other analysis models, and to use the degree of consistency as the second evaluation index of the trust level of each analysis model.

[0152] The determination module 302 is also used to perform weighted calculations on the first evaluation index and the second evaluation index based on preset weights to obtain the confidence level of each analysis model.

[0153] As an alternative implementation, the determination module 302 is also used for at least two of a plurality of analysis models, including a large language model, a generative pre-trained transformation model, and a fine-tuned bidirectional encoder representation model.

[0154] As an optional implementation, the determining module 302 is further configured to, if the target event information is text, encode the text using a pre-trained language model to generate a text embedding vector containing semantic features as an event feature vector.

[0155] The determination module 302 is also used to extract the visual features of the image by using a convolutional neural network or a visual model if the target event information is an image, and generate an image embedding vector as an event feature vector.

[0156] The determination module 302 is also used to convert the audio signal into a spectrogram by Mel spectrum conversion if the target event information is audio, and then use an audio feature extraction model to extract audio features and generate an audio embedding vector as an event feature vector.

[0157] The determination module 302 is also used to extract features from each type of information (text, image, audio) to obtain the corresponding type of embedding vector if the target event information contains multiple types of information, such as text, image, and audio, and to merge the embedding vectors of multiple types of information into an event feature vector through vector fusion.

[0158] As an optional implementation, the processing module 301 is also used to trigger an early warning mechanism if the target risk level is greater than or equal to the risk level threshold.

[0159] The processing module 301 is also used to prevent the early warning mechanism from being triggered if the target risk level is less than the risk level threshold.

[0160] As an optional implementation, the determining module 302 is also used to determine the target warning level based on the target risk level; the higher the target risk level, the higher the warning level.

[0161] The processing module 301 is also used to activate the warning mode corresponding to the target warning level according to the preset correspondence between warning levels and warning modes. The warning mode includes at least one of sound alarm, light flashing and information push.

[0162] As an optional implementation, the event analysis device includes: a generation module 304;

[0163] The processing module 301 is also used to acquire historical event information and perform standardized processing on the historical event information;

[0164] The processing module 301 is also used to convert the standardized historical event information into historical event feature vectors and extract the corresponding known event types and actual risk levels.

[0165] The generation module 304 is used to index and store historical event feature vectors, known event types and actual risk levels in the database to generate a preset knowledge base.

[0166] Figure 4 A schematic diagram of the event analysis device provided in this application. Figure 4 As shown, this application provides an event analysis device 400, which includes a receiver 401, a transmitter 402, a processor 403, and a memory 404.

[0167] Receiver 401 is used to receive instructions and data;

[0168] Transmitter 402 is used to send commands and data;

[0169] Memory 404 is used to store instructions executed by the computer;

[0170] Processor 403 is used to execute computer execution instructions stored in memory 404 to implement the various steps performed by the event analysis method in the above embodiments. For details, please refer to the relevant descriptions in the foregoing event analysis method embodiments.

[0171] Alternatively, the memory 404 can be either standalone or integrated with the processor 403.

[0172] When the memory 404 is set up independently, the electronic device also includes a bus for connecting the memory 404 and the processor 403.

[0173] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the event analysis method performed by the event analysis device described above.

[0174] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0175] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it is readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. The above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. An event analysis method, characterized in that, include: Obtain target event information and convert the target event information into an event feature vector; The event feature vector is input into multiple analysis models, and each analysis model outputs the corresponding event type and initial risk level. The event feature vector is matched with the historical event feature vectors of historical events in the preset knowledge base to filter one or more historical event information with similarity scores greater than a preset threshold, and the trust level of each analysis model is determined based on the historical event information and the event type. The initial analysis results are weighted and fused based on the trust level of each analysis model to determine the target risk level corresponding to the target event information.

2. The method according to claim 1, characterized in that, The process of determining the trust level of each analysis model based on the historical event information and the event type includes: Extract historical event data of the same type as the event type corresponding to each of the analysis models from the historical event information; Based on the historical judgment accuracy of each of the aforementioned analytical models in the historical data of similar events, the historical judgment accuracy is used as the first evaluation index of the trustworthiness of each of the aforementioned analytical models. Calculate the degree of consistency between the initial risk level of each analytical model and the initial risk level of other analytical models, and use the degree of consistency as the second evaluation index of the trustworthiness of each analytical model; The first evaluation index and the second evaluation index are weighted and calculated based on preset weights to obtain the trust level of each analysis model.

3. The method according to claim 1 or 2, characterized in that, The multiple analysis models include at least two of the following: a large language model, a generative pre-trained transformation model, and a fine-tuned bidirectional encoder representation model.

4. The method according to claim 1, characterized in that, The target event information includes at least one of text, image, and audio; The step of converting the target event information into an event feature vector includes: If the target event information is text, then a pre-trained language model is used to encode the text, generating a text embedding vector containing semantic features as the event feature vector; If the target event information is an image, then a convolutional neural network or a visual model is used to extract the visual features of the image and generate an image embedding vector as the event feature vector. If the target event information is audio, the audio signal is converted into a spectrogram through Mel spectrum conversion, and then audio features are extracted using an audio feature extraction model to generate an audio embedding vector as the event feature vector. If the target event information contains multiple types of information such as text, image, and audio, then feature extraction is performed on each type of information to obtain the corresponding type of embedding vector, and the embedding vectors of multiple types of information are merged into the event feature vector through vector fusion.

5. The method according to claim 1, characterized in that, The method further includes: If the target risk level is greater than or equal to the risk level threshold, an early warning mechanism is triggered; If the target risk level is less than the risk level threshold, the early warning mechanism will not be triggered.

6. The method according to claim 5, characterized in that, The triggering early warning mechanism includes: The target warning level is determined based on the target risk level, and the higher the target risk level, the higher the warning level. According to the preset correspondence between warning levels and warning methods, the warning method corresponding to the target warning level is activated, and the warning method includes at least one of sound alarm, light flashing and information push.

7. The method according to claim 1, characterized in that, The method for constructing the preset knowledge base includes: Acquire historical event information and standardize the historical event information; The standardized historical event information is transformed into historical event feature vectors, and the corresponding known event types and actual risk levels are extracted. The historical event feature vectors, known event types, and actual risk levels are indexed and stored in the database to generate a preset knowledge base.

8. An event analysis device, characterized in that, include: The processing module is used to acquire target event information and convert the target event information into an event feature vector; The processing module is also used to input the event feature vector into multiple analysis models respectively, and each analysis model outputs the corresponding event type and initial risk level respectively; The determination module is used to perform similarity matching between the event feature vector and the historical event feature vector of historical events in the preset knowledge base, filter one or more historical event information with similarity scores greater than a preset threshold, and determine the trust level of each of the analysis models based on the historical event information and the event type. The determining module is further configured to perform weighted fusion of the initial analysis results based on the trust level of each analysis model to determine the target risk level corresponding to the target event information.

9. An event analysis device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

Images