A risk dynamic assessment method, system and device for urban operation situation

Abstract

The application provides a risk dynamic evaluation method, system and device for urban operation situation, and relates to the technical field of data processing. The method comprises the following steps: acquiring multi-modal content data and space-time metadata of a target event; inputting the target event into a deep support vector data description model to calculate the square Euclidean distance of the target event to the center of a normal operation hypersphere as a risk benchmark deviation score, wherein the normal operation hypersphere is obtained by training only the data in the period of daily smooth operation of the city; inputting the target event into a context perception fusion network, extracting semantic damage intensity features by a full-element semantic analysis flow, extracting space-time context sensitivity features by a space-time context analysis flow, and generating a context correction score by bidirectional interaction and fusion through a cross attention mechanism; and weighting and fusing the risk benchmark deviation score and the context correction score to obtain a comprehensive risk score. The application is free from dependence on historical abnormal labeled samples, and realizes multi-dimensional context fusion risk evaluation.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a method, system, and equipment for dynamic risk assessment of urban operational status. Background Technology

[0002] With the advancement of urbanization and the increasing demands for refined urban governance, urban operational risk management faces an increasingly complex situation. Traditional urban operational risk management models mainly rely on human experience and comparison with historical cases, which suffers from problems such as delayed response, incomplete coverage, and significant subjective bias. In recent years, risk intelligent assessment methods based on big data and artificial intelligence technologies have gradually become a research hotspot.

[0003] In existing technologies, urban operation risk assessment methods mainly include the following categories: The first type is expert systems based on rule engines. These systems construct judgment rules by manually summarizing the characteristic patterns of historical risk cases, and then perform pattern matching and level determination on events. Their advantages are logical transparency and strong interpretability, but they heavily rely on the prior knowledge of domain experts and historical labeled samples, making it difficult to cope with new types of risk events and anomalies outside the rule coverage.

[0004] The second category is supervised classification methods based on traditional machine learning, including shallow models such as random forests, gradient boosting trees, and support vector machines, as well as deep neural network classifiers. These methods require a sufficient number of positive and negative samples for supervised training, learning the mapping relationship between features and labels from historical cases to predict the risk level of new samples. However, major social risk events are inherently low-frequency, sudden, and non-replicable, and the number of historical samples is extremely limited, far from sufficient to support the training of deep learning models. Furthermore, the severe imbalance between positive and negative samples leads to a high rate of missed detections for truly high-risk events requiring warnings. In addition, the content, form, and dissemination patterns of new risk events often differ fundamentally from historical cases, making it difficult for models to achieve effective cross-type generalization.

[0005] The third category is based on text sentiment analysis and topic modeling. These methods use natural language processing techniques to extract sentiment polarity, topic distribution, and keyword frequency from social media texts to construct a risk sentiment index. They focus on semantic analysis of the content but neglect crucial contextual information such as the time, place, and parties involved in the event.

[0006] The fourth category is methods based on single-dimensional statistical thresholds. These methods determine risk by setting a statistical threshold for a single indicator. They are simple to calculate and easy to deploy, but they cannot handle the fusion of multi-source heterogeneous data, and the threshold setting is highly subjective.

[0007] The aforementioned existing technologies have the following shortcomings: First, existing supervised learning methods heavily rely on historical labeled samples, which fails in low-frequency, high-risk scenarios. Major social risk events are inherently low-frequency and sudden, and the amount of historical samples is insufficient to support supervised learning. Within the supervised learning framework, a severe imbalance between positive and negative samples leads to a strong negative sample bias in the model, resulting in an extremely high rate of missed detections for high-risk events.

[0008] Secondly, existing single-dimensional assessment methods fail to integrate spatiotemporal contextual information, leading to biased and distorted assessment results. For events of the same nature, the risk level is far higher in sensitive areas such as transportation hubs and administrative centers than in ordinary residential areas; risk events occurring during major holidays or other special periods are significantly more harmful than those occurring in a dispersed manner; and the cumulative risk effect of entities with multiple repeated alarm records also needs to be considered. However, most existing technologies focus only on textual analysis of event content, failing to effectively integrate multi-dimensional spatiotemporal contextual factors such as spatial sensitivity, temporal clustering density, and historical frequency accumulation, resulting in discrepancies between assessment results and actual governance needs.

[0009] Finally, existing deep learning models lack interpretability and fail to meet the requirements for transparency and accountability audits in governance decisions. Deep learning models are often considered "black boxes" due to their complex internal decision-making mechanisms. Governance decision-makers need not only to know the risk level but also the basis for determining that risk level and the core influencing factors. Existing black-box models cannot provide fine-grained attribution explanations, making it difficult to meet the requirements for administrative accountability, judicial audits, and public informed consent. Summary of the Invention

[0010] The purpose of this application is to provide a method, system, electronic device, and storage medium for dynamic risk assessment of urban operational status.

[0011] In a first aspect, embodiments of this application provide a method for dynamic risk assessment oriented towards urban operational status, including: Acquire multimodal content data and spatiotemporal metadata of the target event to be evaluated; the multimodal content data includes event text description, on-site images and / or speech-to-text transcription; the spatiotemporal metadata includes the event timestamp, location code and involved entity identifier; The multimodal content data and spatiotemporal metadata of the target event are input into a pre-trained deep support vector data description model. The squared Euclidean distance from the target event to the center of the normal operation hypersphere is calculated in the feature space and used as the risk benchmark deviation score. The normal operation hypersphere is a compact hypersphere that surrounds the normal sample in the feature space, which is trained only using multi-source governance data during the city's daily stable operation period. The multimodal content data and spatiotemporal metadata of the target event are input into a pre-trained context-aware fusion network. The semantic destruction intensity features are extracted by the full-element semantic analysis stream, and the spatiotemporal context sensitivity features, including regional spatial sensitivity features, short-term temporal clustering density features, and historical risk frequency features of the involved subjects, are extracted by the spatiotemporal context analysis stream. The semantic destruction intensity features and the spatiotemporal context sensitivity features are bidirectionally interacted and fused through a cross-attention mechanism to generate a context correction score. The risk benchmark deviation score and the situation correction score are weighted and fused to obtain the comprehensive risk score of the target event.

[0012] Secondly, embodiments of this application provide a dynamic risk assessment system for urban operational status, including: The acquisition module is used to acquire multimodal content data and spatiotemporal metadata of the target event to be evaluated; the multimodal content data includes event text description, on-site images and / or speech-to-text transcription; the spatiotemporal metadata includes the timestamp of the event, location code and identifier of the involved parties; The benchmark scoring module is used to input the multimodal content data and spatiotemporal metadata of the target event into a pre-trained deep support vector data description model, and calculate the squared Euclidean distance from the target event to the center of the normal operation hypersphere in the feature space as the risk benchmark deviation score; the normal operation hypersphere is a compact hypersphere that surrounds the normal sample in the feature space, which is trained only using multi-source governance data during the city's daily stable operation period. The correction scoring module is used to input the multimodal content data and spatiotemporal metadata of the target event into a pre-trained context-aware fusion network. The semantic destruction intensity features are extracted by the full-element semantic analysis stream, and the spatiotemporal context sensitivity features, including regional spatial sensitivity features, short-term temporal clustering density features, and historical risk frequency features of the involved subjects, are extracted by the spatiotemporal context analysis stream. The semantic destruction intensity features and the spatiotemporal context sensitivity features are bidirectionally interacted and fused through a cross-attention mechanism to generate a context correction score. The comprehensive scoring module is used to weight and fuse the risk benchmark deviation score and the situation correction score to obtain a comprehensive risk score; The attribution classification module is used to calculate the marginal contribution of the semantic destruction intensity feature, the regional spatial sensitivity feature, the short-term temporal clustering density feature, and the historical risk frequency feature of the involved subject to the comprehensive risk score using the SHAP game theory attribution method. Based on the marginal contribution of each feature, the risk causes of the target event are classified into at least one of the following: content intensity-driven, spatiotemporal context-sensitive, or historical frequency-cumulative, and the classification results are output.

[0013] Thirdly, this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in the first aspect.

[0014] Fourthly, this application provides a computer-readable storage medium having computer-readable instructions stored thereon, which can be executed by a processor to implement the method described in the first aspect.

[0015] Compared to existing technologies, the risk dynamic assessment method for urban operational status provided in this application acquires multimodal content data and spatiotemporal metadata of the target event; inputs the target event into a deep support vector data description model, calculates the squared Euclidean distance from it to the center of a normal operation hypersphere as a risk benchmark deviation score, and the normal operation hypersphere is trained only using data from the city's daily stable operation period; inputs the target event into a context-aware fusion network, extracts semantic destruction intensity features from a full-element semantic analysis stream, extracts spatiotemporal context sensitivity features from a spatiotemporal context analysis stream, and generates a context correction score through bidirectional interactive fusion using a cross-attention mechanism; and weights and fuses the risk benchmark deviation score and the context correction score to obtain a comprehensive risk score. This application eliminates the dependence on historical anomaly labeled samples, solving the data scarcity problem in low-frequency major risk scenarios; overcomes the one-sidedness of single-dimensional assessment through deep fusion of content semantics and spatiotemporal context; and improves the accuracy of urban operational status risk assessment through the synergistic effect of the unsupervised benchmark of the deep support vector data description model and the context-aware fusion network. Attached Figure Description

[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a risk dynamic assessment method for urban operational status provided in an embodiment of this application is shown; Figure 2 This paper illustrates a flowchart of the process for obtaining a comprehensive risk score for a target event, as provided in an embodiment of this application. Figure 3 This illustration shows a flowchart of the process for attributing and classifying the causes of comprehensive risk scores according to an embodiment of this application; Figure 4 This illustration shows a structural diagram of a risk dynamic assessment system for urban operational status provided in an embodiment of this application; Figure 5A schematic diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0018] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0019] Furthermore, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order. Additionally, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to those processes, methods, products, or apparatuses.

[0020] This invention addresses the three core shortcomings of existing technologies in social risk assessment: strong sample dependence, single and one-sided dimensions, and unexplainable decisions. It proposes a three-in-one dynamic assessment and causal classification technology for social risks, which integrates "unsupervised benchmark establishment, dual-stream fusion assessment, and interpretable causal classification".

[0021] The core design concept of this invention is as follows: First, the Deep Support Vector Data Description (DSVDD) algorithm is employed, utilizing only multi-source governance data from periods of stable urban operation. It learns a compact distribution of normal operating conditions within the feature space encoded by a deep neural network, constructing a "normal operating safety zone" represented by a hypersphere. Normal samples are compressed to a region close to the center of the sphere, and the squared Euclidean distance from any sample deviating from this safety zone to the center is quantified as a risk deviation benchmark score. This method eliminates the need for any historical anomaly sample annotations, fundamentally solving the problem of scarce training data in low-frequency, high-risk scenarios.

[0022] Second, a Context-Aware Fusion Network (CA-FusionNet) is constructed, featuring a dual-channel architecture of a full-element semantic analysis stream and a spatiotemporal context analysis stream. The semantic analysis stream is responsible for parsing the multimodal semantic information of the event content (text, image transcription, and speech transcription are unified into semantic vectors), extracting the destructive intensity features of the event itself. The spatiotemporal context analysis stream is responsible for handling external environmental constraints (regional spatial sensitivity, short-term temporal clustering density, and historical risk frequency of the involved parties), extracting context sensitivity features. The two streams are deeply coupled and linked through a cross-attention mechanism, enabling semantic features to adaptively focus on the most relevant contextual factors, and contextual features to inversely modulate the weight distribution of semantic understanding. The final score is obtained by weighted fusion of the DSVDD benchmark score and the context correction score, taking into account both the content destructive intensity and the sensitivity of contextual handling.

[0023] Third, the SHAP (Shapley Additive Explanations) game-theoretic attribution method is introduced to perform post-hoc interpretability analysis on the risk scores output by CA-Fusion Net. Based on Shapley values, the marginal contribution of each input feature to the final score is calculated, providing both a global-level statistical ranking of feature importance and a heatmap of contribution intensity at the single-sample level. Building upon this, the K-means clustering algorithm is used to summarize the pattern of feature contribution vectors of high-risk samples, refining the complex causal structure into three typical patterns: content intensity-driven, spatiotemporally sensitive, and historical frequency-cumulative. These patterns are further visualized using radar charts, ensuring that the model output is fully transformed into business information that governance managers can understand and use for decision-making.

[0024] Please refer to Figure 1 , Figure 1 This is a flowchart of a dynamic risk assessment method for urban operational status provided in an embodiment of this application, including the following steps S101~S104: S101. Obtain multimodal content data and spatiotemporal metadata of the target event to be evaluated.

[0025] Specifically, multimodal content data and spatiotemporal metadata of the target event to be evaluated are acquired. The multimodal content data includes event text descriptions, on-site images, and / or speech-to-text transcriptions. The spatiotemporal metadata includes the event's timestamp, location code, and identifiers of the parties involved.

[0026] In a specific application, multimodal content data can come from various data sources in the urban governance information platform, including but not limited to: alarm record text in the alarm handling system, text content related to local social governance in social media platforms, event description text and on-site photos reported by grassroots grid administrators through mobile terminals, and abnormal event alarm records generated by video surveillance systems through AI recognition.

[0027] All data undergoes preprocessing after acquisition, including data cleaning, noise reduction, sensitive information anonymization, and format standardization. Text content is encoded into semantic vectors using a pre-trained language model, while image content has visual feature vectors extracted using convolutional neural networks or visual transformers.

[0028] S102. Calculate the risk benchmark deviation score.

[0029] The multimodal content data and spatiotemporal metadata of the target event are input into a pre-trained deep support vector data description model (DSVDD model). The squared Euclidean distance from the target event to the center of the normally operating hypersphere is calculated in the feature space and used as the risk benchmark deviation score.

[0030] The normal operation hypersphere is a compact hypersphere that surrounds normal operation samples in the feature space, obtained by training with multi-source governance data during the normal operation period of the city.

[0031] The following provides a detailed explanation of the training process of the deep support vector data description model involved in this step. The DSVDD model uses a deep neural network encoder. (·; W) maps the input samples to the feature space and learns a hypersphere with center c as the center and radius R as the boundary in the space.

[0032] The training data uses only multi-source governance data from periods of stable daily urban operation. Specifically, a period of more than six consecutive months without any major social risk events can be selected as the "stable daily operation period." Multi-source governance data includes police report data, social media data, grid patrol reporting data, and video surveillance intelligent recognition data. This data, after preprocessing, forms the normal operation training dataset: X normal = {x1, x2, ..., x N}

[0033] The optimization objective of DSVDD is: ; in, ( x i ; W () is a deep neural network encoder that processes samples from the input spacex i Feature representation mapped to a high-dimensional feature space W Here, c represents the encoder weights, c is the center vector of the hypersphere, and R is the radius of the hypersphere. ν ∈(0,1] is a hyperparameter that controls the tolerance for the proportion of abnormal samples, and N is the total number of normal training samples.

[0034] The optimization process is as follows: First, the encoder part of a pre-trained autoencoder is used as... The initial parameters of (·; W) are used, and the autoencoder is pre-trained on the normal training set using reconstruction loss.

[0035] Secondly, all normal samples are mapped to the feature space through the encoder, and the mean vector of the mapped feature vectors is calculated as the initial value of the sphere center: ; Among them, W init These are the initial weights for the encoder.

[0036] Then, stochastic gradient descent or the Adam optimizer is used to jointly optimize the encoder weights W and the sphere center vector c in an end-to-end manner. In each iteration, a mini-batch of normal samples is sampled, the feature representation is computed through forward propagation, the loss function value is calculated, and the parameters are updated through backpropagation.

[0037] After training, the hypersphere radius R Take the one that can cover at least (1- ν Minimum radius of the proportional training samples: ; in, r To satisfy the minimum radius of the constraints, W* , c* , R* These are the encoder weights, hypersphere center, and hypersphere radius after training convergence, respectively.

[0038] During training, the loss changes on the validation set are monitored, an early stopping strategy is used to prevent overfitting, and L2 regularization constraints are applied to the encoder weights.

[0039] For the target sample x to be evaluated target Its risk benchmark deviation score S DSVDD (x target ) = || (x target ; W) - c||² represents the squared Euclidean distance from the sample feature representation to the center of the sphere. The larger this distance value, the greater the degree to which the target event deviates from its normal operating state.

[0040] To facilitate subsequent fusion and manual interpretation, the risk benchmark deviation score is normalized by using a Sigmoid transformation to map the unbounded distance value to the [0,1] interval: ; in, The score represents the deviation from the normalized risk benchmark; k>0 is the slope control parameter, which determines the sensitivity of the score to changes near the boundary.

[0041] S103, Generate Context Correction Scoring.

[0042] The multimodal content data and spatiotemporal metadata of the target event are input into a pre-trained context-aware fusion network. A full-element semantic analysis stream extracts semantic destruction intensity features, while a spatiotemporal context analysis stream extracts spatiotemporal context sensitivity features, including regional spatial sensitivity features, short-term temporal clustering density features, and historical risk frequency features of the involved entities. A cross-attention mechanism is used to bidirectionally interact and fuse the semantic destruction intensity features and the spatiotemporal context sensitivity features to generate a context correction score.

[0043] The following provides a detailed explanation of each sub-step involved in this process.

[0044] (I) Full-element semantic analysis flow The Full-Element Semantic Analysis Stream aims to extract the semantic damage intensity of the event content itself, answering the content analysis questions of "what did this event say, what did it do, and how serious is it?"

[0045] First, a pre-trained language model is used to perform deep semantic encoding on the event text description and speech-to-text to obtain text semantic vectors. For example, the RoBERTa-wwm-ext-large pre-trained language model (1024 hidden layers) is used to encode the event text description and speech-to-text, and the mean pooling or [CLS] position vector of the last hidden state is taken as the sentence-level semantic representation v. text .

[0046] Secondly, visual semantic encoding is performed on the scene images using a visual transformer or convolutional neural network to obtain image semantic vectors. For example, a ViT-B / 16 visual transformer (768-dimensional hidden layer) is used to extract visual semantic features v from the scene images. img If the image contains identifiable specific elements (such as the size of a crowd gathering, a conflict scene, etc.), a structured attribute vector is further extracted through the object detection submodule.

[0047] Then, the text semantic vector and the image semantic vector are fused across modally to generate a unified multimodal semantic representation: v semantic = Fusion(v text , v img ) = W fuse ·Concat(v text , v img ) + b fuse ; Among them, W fuse b is the fusion projection weight matrix; fuse This is the bias vector for the fusion layer.

[0048] Finally, the multimodal semantic representation is mapped to a dedicated destruction intensity subspace for risk assessment through a task-related projection layer to generate the semantic destruction intensity feature: f semantic =ReLU(W risk ·v semantic +b risk ); Among them, W risk For the projected weights of the risk task, b risk This is the risk projection bias.

[0049] (II) Spatiotemporal Context Analysis Flow The spatiotemporal context analysis stream aims to extract the external contextual constraints of a target event, answering the contextual analysis questions of "when, where, and who is involved in this event, and how these factors amplify or reduce its actual risk."

[0050] First, based on the spatial sensitivity classification knowledge base of the Geographic Information System (GIS), the sensitivity level of the event location is encoded into an embedding vector, serving as the regional spatial sensitivity feature. The city's jurisdiction is then divided into different spatial units (such as streets, communities, and key locations), with each spatial unit... l Assigned a sensitivity level G ( l )∈{1,2,...,K G The grading criteria include the importance of administrative functions, the concentration of historical events, and population density and mobility. Spatial sensitivity levels are encoded as embedding vectors. f space =Embed(G( l )).

[0051] Secondly, within a preset sliding time window preceding the occurrence of the target event, the number and density of related events occurring within the same region are statistically analyzed to construct a density statistical feature vector containing multiple time granularities, serving as a short-term time-series cluster density feature. For example, cluster density is statistically analyzed within three sliding time windows: 24 hours (h), 72 hours, and 7 days (d). ; ; Where t is the timestamp of the current event, and Δt is the length of the sliding statistical time window. j For historical event sample index, loc j For the first j The region where the historical event is located, time j For the first j The timeline of historical events; For the region l The total number of similar events that occurred within a time window Δt prior to the current time t; A l For the region l Area or baseline population; To normalize the event cluster density under unit window and unit baseline population / area, the dimensional differences caused by region size and statistical duration are eliminated.

[0052] This results in a density statistical feature vector containing multiple time granularities: ; If a pattern of similar events occurring in a concentrated manner appears in the near future (such as similar alarms occurring in a certain area for several consecutive days), the cluster density characteristic will show a significant upward trend, indicating the possibility of systemic risk accumulation or chain reactions.

[0053] Then, for the subject s (natural person, organization, or group) involved in the target event, the historical database is queried to statistically analyze their performance during the observation period. tT hist , t Frequency of risk-related records within: ;in, T hist This represents the total duration of the historical observation window. subject j For the first j The entities involved corresponding to each historical record time j For the first j The time when a historical record occurred.

[0054] Simultaneously, a time-decay-weighted version of the frequency is calculated, giving more weight to recent records: Where λ>0: time decay coefficient, which controls the decay rate of historical record weights; The historical frequency feature vector of the entities involved is as follows: ;in, risk j For the first j The risk score / risk level corresponding to each historical event; the third term in the formula is the highest risk level in the subject's historical records, used to characterize its historical "peak destructive power".

[0055] Original frequency Freq Reflects the total amount of long-term risks occurring in the main entity; attenuation weighted average. Freq decay It focuses on recent active risk behaviors and downplays the impact of long-standing historical events; the peak risk item captures the most serious event that the subject has ever caused, and is used to predict the potential for escalation of the current event; the three together serve as spatiotemporal context flow features.

[0056] Finally, the regional spatial sensitivity characteristics, the short-term temporal cluster density characteristics, and the historical risk frequency characteristics of the involved entities are concatenated: ; The spatiotemporal context sensitivity features are obtained by performing nonlinear transformation and dimensionality reduction through a context encoding layer: ;in, W ctx Encode the weight matrix for context. b ctx The context encoding bias vector is LayerNorm(), which is the layer normalization operation.

[0057] First, the three independent contextual sub-features—spatial, temporal, and subject-specific—are directly concatenated and fused, preserving all external environmental constraint information. Then, a learnable linear transformation is used for dimensionality compression, and ReLU is applied to introduce non-linear expressive power. Finally, a layer normalization is applied to stabilize the feature distribution, outputting a unified and standardized spatiotemporal contextual feature. h context This is used for subsequent cross-attention fusion with semantic features.

[0058] (III) Fusion of cross-attention mechanisms Cross-attention mechanism is a core technology for achieving deep coupling between semantic stream and context stream. Unlike traditional simple concatenation or weighted summation fusion methods, cross-attention allows the establishment of fine-grained element-level correspondences between two feature streams, realizing bidirectional information interaction of "semantic guidance for context focusing" and "context-modulated semantic understanding".

[0059] The specific implementation of the cross-attention mechanism is as follows: First, the semantic violation intensity feature and the spatiotemporal context sensitivity feature are linearly projected into query, key, and value vectors respectively: ; ; in, W Q 、W K 、W V 、W' Q 、W' K 、W' V All are linear projective weight matrices, and all are trainable parameters. Q sem Semantic query vector, K ctx V ctx Contextual side keys, value vectors; Q ctx Context-side query vector, K sem V sem Semantic side keys, value vectors.

[0060] Secondly, calculate the attention weight matrix for contextual features guided by semantic features: ; Generative context-focused representation: ; Where, d attn Single-head attention feature dimension, scaling factor This is used to eliminate the vanishing gradient problem caused by excessively large inner product magnitudes.

[0061] The meaning of this process is: based on the query vector Q of the event's semantic content... sem In the context feature key vector K ctx Similarity matching is performed to determine the relevance of each contextual factor to the current semantic content, forming an attention weight matrix, and then the contextual value vector V is applied. ctxWeighted aggregation is performed to obtain a context-focused representation guided by semantics. For example, if the semantic content indicates that the event involves "campus safety", the attention mechanism will automatically focus on contextual factors such as the regional sensitivity of the school district and the density of schools, rather than contextual information about the commercial area.

[0062] Simultaneously, the attention weight matrix for semantic features guided by contextual features is calculated: ; Generate context-modulated semantic understanding representations: ; The meaning of this process is: based on the query vector Q of the external context... ctx In the semantic feature key vector K sem The system performs matching to determine the importance of semantic elements in the current context. The same semantic content may be assigned different interpretation weights in different contexts. For example, "a gathering of 10 people" is within the normal range in a business district, but it has a higher risk semantic weight around an administrative center.

[0063] Then, a multi-head attention mechanism is used to execute the above process in parallel, concatenating the multi-head outputs and generating the context correction score through a feedforward network. H independent projection matrices are used to perform the above cross-attention calculation in parallel, with each group focusing on different feature subspace relationships: ; Among them, each head .

[0064] in, H : Total number of attention points; head h : No. h Each attention head outputs an independent attention head; CrossAttn(): Single-head bidirectional cross-attention calculation function; Q (h) ,K (h) V (h) : No. h Head-specific query, key, and value vectors generated by a learnable projection matrix; W O : Multi-head output fusion learnable projection matrix; Concat(): Multi-head output vector concatenation operation.

[0065] Finally, the outputs of the bidirectional cross-attention are aggregated and a context-corrected score is generated through a feedforward network: ; ; Where σ(·) is the Sigmoid function, which compresses the output to the [0,1] interval as the context correction score; FFN is a two-layer fully connected feedforward network; Residual is the residual connection input for stable training; W out , b out The output layer can learn weights and biases.

[0066] First, represent the semantically guided context. Context modulation semantic representation By concatenating and fusing the features, and superimposing multi-head input residual terms to alleviate gradient vanishing in deep networks, and then normalizing the feature distribution through layers, stable fused features are obtained. h fused ; The fused features are fed into a two-layer fully connected feedforward network (FFN) for nonlinear depth transformation, then mapped through a linear output layer, and finally constrained to the 0-1 range by a sigmoid activation to obtain the standardized context correction score. S context .

[0067] S104, weighted fusion yields a comprehensive risk score.

[0068] The risk benchmark deviation score and the situation correction score are weighted and fused to obtain the comprehensive risk score of the target event.

[0069] The fusion formula is: S final = α·S DSVDD + (1-α)·S context ; Among them, S final For the comprehensive risk score, S DSVDD S is used to score the deviation from the risk benchmark. context The situation is corrected and scored, with α being a weighting coefficient ranging from [0, 1]. S in the formula... DSVDD Normalized risk benchmark deviation score can be used. .

[0070] For ease of understanding, the present invention provides, as follows: Figure 2 The diagram shows the process of obtaining a comprehensive risk score for a target event.

[0071] The weighting coefficient α can be determined in ways including but not limited to the following: Fixed weight mode: Pre-set by the governance department according to business preferences, for example, α=0.55.

[0072] Adaptive weighting mode: Let α be dynamically adjusted based on the risk benchmark deviation score of the current sample. When the α value is already very high (clearly abnormal), reducing α allows situational correction to play a greater role in refining the risk profile; when When in the boundary region, increase α to strengthen the dominant role of objective deviation: ; Where, α( x ):sample x The corresponding dynamic weighting coefficient; α base : Lower bound of the weight; α max : Maximum upper limit of weight.

[0073] (3) Business scenario switching mode: switch different α preset values ​​according to different governance business scenarios (daily monitoring / major event security / emergency response).

[0074] In some embodiments, to improve the readability of the ratings, continuous ratings can be discretized into warning levels: S final ∈ [0, 0.25): Blue level (normal attention); S final ∈ [0.25, 0.50): Yellow level (general warning); S final ∈ [0.50, 0.75): Orange level (severe warning); S final ∈ [0.75, 1.0]: Red level (major warning).

[0075] In some embodiments, the above-mentioned risk dynamic assessment method for urban operation status provided in this application further includes the step of using the SHAP game theory attribution method to perform causal attribution and classification.

[0076] The SHAP game theory attribution method is used to calculate the marginal contribution of the semantic destruction intensity feature, the regional spatial sensitivity feature, the short-term temporal clustering density feature, and the historical risk frequency feature of the involved subject to the comprehensive risk score.

[0077] SHAP is an interpretability method based on Shapley values ​​in cooperative game theory. Its core idea is to decompose the model output into the sum of the marginal contributions of each input feature, satisfying four axioms: efficiency, symmetry, dummy elements, and additivity, thus ensuring the fairness and consistency of the attribution results. For the i-th input feature, its Shapley value... iDefined as the average marginal contribution of the feature across all possible permutations of feature subsets: ; in, N For the complete feature set, S For features not included i a subset of f ( S ) indicates that only a subset is used. S The expected output of the model when the features are in the middle (missing features are handled by sampling the background distribution or mask filling). Representing a subset S Add features i Post-model expected output. Weighting factors. denoted as the probability of each subset appearing in the permutation.

[0078] Since the combinatorial complexity of accurately calculating the Shapley value is O(2^3) |N| In practical applications, approximation algorithms such as KernelSHAP or TreeSHAP are used for computation. For the CA-Fusion Net model of this invention, an approximation strategy based on perturbation sampling can be adopted: define a background dataset (representative samples sampled from the normal training set), and estimate the expected marginal contribution of each feature through Monte Carlo sampling.

[0079] After calculating the SHAP value of each feature, based on the marginal contribution of each feature, the risk causes of the target event are classified into at least one of the following: content intensity-driven, spatiotemporal context-sensitive, and historical frequency-cumulative, and the classification results are output.

[0080] Specifically, the SHAP marginal contribution of each input feature of the target event is aggregated according to a preset feature category to generate a four-dimensional causal feature vector: ; in, Indicates the first m A four-dimensional causal feature vector of a high-risk sample. Indicates the first m In the nth sample j SHAP contribution value of each original feature; Semantic: set of all semantic class feature indexes; Spatial: set of all spatial context feature indexes; Temporal: set of all temporal context feature indexes; Subject: set of all historical feature indexes of the involved subjects.

[0081] The four-dimensional causal feature vectors in the formula correspond to the sum of features contributing to semantic content, spatial context, temporal context, and subject history, respectively.

[0082] Calculate the distance between the four-dimensional causal feature vector and the three preset cluster center vectors, and classify the target event into the causal pattern category corresponding to the nearest cluster center.

[0083] The three cluster center vectors are pre-determined by performing K-means clustering on the four-dimensional causal feature vectors of each sample in the historical high-risk sample set, corresponding to three causal pattern categories. The clustering objective is: ; Where K=3: the preset number of clusters; C k For the first k The sample set of each cluster μ k Let be the cluster center vector. A K-means++ initialization strategy is used, selecting the optimal clustering result through multiple random initializations. The silhouette coefficient is used to evaluate the quality of cluster grouping.

[0084] The three causal modes are as follows: The causal pattern corresponding to the center of the first cluster is content intensity-driven, with the total contribution of semantic content features exceeding the total contribution of spatial context, temporal context, and subject-historical features. A typical profile of this type of event is: the event itself possesses strong destructive semantics (physical conflict, extreme emotional expression, radical demands), and even if it occurs in ordinary areas and at ordinary times, its content intensity is sufficient to drive a high score.

[0085] The causal pattern corresponding to the second cluster center is spatiotemporally sensitive, where the sum of contributions from spatial context features and / or temporal context features is higher than the sum of contributions from semantic content features. A typical profile of this type of event is as follows: the event content itself may not be extreme, but it occurs in highly sensitive areas (administrative centers, transportation hubs) or during periods of high concentration (multiple similar events have occurred in the same area within a short period), where contextual factors significantly amplify the actual risk.

[0086] The causal pattern corresponding to the third cluster center is the historical frequency accumulation type, where the total contribution of historical features is higher than the total contribution of semantic content, spatial context, and temporal context features. A typical profile of this type of event is as follows: the involved entity has multiple repeated alarms or risk records; the content of the event itself may be relatively ordinary, but the cumulative effect of historical frequency indicates that the entity or the region has deep-seated systemic problems.

[0087] In some embodiments, the above-mentioned risk dynamic assessment method for urban operation status provided in this application further includes: a result output step, which can output the comprehensive risk score, early warning level label, cause mode classification label, and visualized decision support information such as SHAP contribution heat map and cause mode radar image of the target event.

[0088] Visual presentation methods include, but are not limited to: Risk Situation Card: Basic information about the event, comprehensive risk score (number + progress bar), warning level (color code), score composition ratio, ranking of the top 5 key influencing factors, classification labels of cause patterns and corresponding handling suggestions; Regional risk heat map: The comprehensive risk scores of each event are aggregated according to geographic coordinates, and smoothed using kernel density estimation to generate a risk heat map layer in the GIS spatial view; Time-series trend curve: The average risk score of events within the aggregated area is plotted as a time-series change curve; Causal Pattern Dashboard: The pie chart shows the sample proportion distribution of the three causal patterns, and the radar chart compares the structural differences of causal patterns in different regions or time windows.

[0089] For ease of understanding, this application provides the following: Figure 3 The diagram shows a flowchart of the process for attributing and classifying the causes of comprehensive risk scores.

[0090] The training process of the deep support vector data description model and the context-aware fusion network is described in detail below.

[0091] 1. Data Acquisition and Preprocessing for Routine Operations Acquire multi-source heterogeneous governance data during the city's normal daily operation. This multi-source heterogeneous governance data includes police incident records, social media data, grid patrol reports, and video surveillance intelligent recognition data.

[0092] Specifically, multi-source heterogeneous data is collected from the city governance information platform during its normal operation. Data sources include: alarm records from the emergency response system (alarm time, location, alarm type, alarm text, handling results, etc.); text content related to local social governance from social media platforms (posting time, posting account information, content text, interaction volume); patrol records reported by grassroots grid administrators via mobile terminals (patrol area code, reporting time, event description text, on-site photos, and transcripts of voice descriptions); and abnormal event alarms generated by the video surveillance system through AI recognition (detection time, camera location code, identified event type, confidence score).

[0093] Preprocessing is performed to construct a training dataset for normal operation. Preprocessing includes: (1) Data cleaning: Remove invalid records with a field missing rate of more than 50%; remove duplicate reported records; correct erroneous timestamps and geographic coordinates; filter system test data and artificially simulated data.

[0094] (2) Sensitive information desensitization: Desensitize and replace fields involving citizens' personal privacy (name, ID number, mobile phone number, precise home address) while retaining location information at the administrative division level.

[0095] (3) Multimodal content vectorization: text content is encoded into semantic vectors through a pre-trained language model; image content is extracted into visual feature vectors through a convolutional neural network or a visual transformer; speech-to-text is processed into semantic vectors through the same encoding process as the text content.

[0096] (4) Time window division: The cleaned data is divided into windows based on the time series, with each window being a day or week. The period of more than 6 consecutive months without any major social risk events is selected as the "stable daily operation period".

[0097] 2. DSVDD Model Training A deep support vector data description algorithm is adopted. A deep neural network encoder is trained using the normal operation training dataset to map the input samples to the feature space. In the feature space, a compact hypersphere surrounding the normal samples is learned, and the center and radius of the hypersphere are determined to obtain the pre-trained deep support vector data description model.

[0098] The training objective of the deep support vector data description model is to minimize the volume of the hypersphere while constraining normal samples to be enclosed within the hypersphere. The specific optimization objective and training process have been described in detail in step S102 and will not be repeated here.

[0099] 3. Training of Context-Aware Fusion Networks A context-aware fusion network is constructed and trained. This network comprises a parallel full-element semantic analysis stream and a spatiotemporal context analysis stream, as well as a cross-attention fusion module connecting the two streams. The network structure has been described in detail in step S103 and will not be repeated here.

[0100] During training, the weighted fusion result of the risk baseline deviation score output by the deep support vector data description model and the context correction score output by the context-aware fusion network is used as the comprehensive risk score. The pre-trained context-aware fusion network is obtained by end-to-end or phased optimization using a preset loss function.

[0101] Specifically, CA-Fusion Net is trained end-to-end, using a loss function of binary cross-entropy loss (BCELoss) + L2 regularization. During training, the weighted fusion result of the risk baseline deviation score output by the DSVDD model and the context correction score output by CA-Fusion Net is used as the comprehensive risk score. The optimizer used is AdamW, with a learning rate of 2e-5, weight decay of 1e-5, a learning rate warm-up step of 1000, and cosine annealing scheduling. The training epochs are 50, and the batch size is 64.

[0102] To verify the technical effectiveness of this invention, experiments were conducted on a self-constructed SRMC (Social Risk Multi-source Corpus) dataset. This dataset contains approximately 1.2 million police incident records, 850,000 social media data entries, 360,000 grid patrol data entries, and 280,000 video surveillance data entries, spanning 42 months.

[0103] The experimental environment was configured as follows: Intel Xeon Gold 6248R CPU, 512GB DDR4 ECC memory, NVIDIA A100 80GB GPU×2; Ubuntu 20.04 LTS operating system, Python 3.9, PyTorch 2.0 deep learning framework.

[0104] On the SRMC test set, the method of this invention was compared with several baseline methods. The baseline methods included: Support Vector Machine (SVM), Random Forest, Long Short-Term Memory (LSTM), TextCNN convolutional neural network, and Transformer classification model based on self-attention mechanism.

[0105] The evaluation metrics used are AUC (Area Under ROC Curve), F1 score, and MAE (Mean Absolute Error).

[0106] The experimental results are shown in the table below:

[0107] Experimental results show that the CA-Fusion Net of this invention achieves optimal performance across all three evaluation metrics. Compared to the best baseline Transformer, it improves AUC by 10.7%, F1 score by 10.8%, and MAE by 47.7%. All baseline methods are supervised learning models.

[0108] The risk dynamic assessment method for urban operational status provided in this application has the following beneficial effects: First, it eliminates reliance on historical outlier samples. This invention employs a deep support vector data description model, training a normal hypersphere using only data from the city's daily stable operation. The distance from the sample to the center of the sphere is used as the risk benchmark score, eliminating the need for any high-risk event labeled samples. This fundamentally solves the problem of scarce training data in low-frequency major risk scenarios.

[0109] Second, it achieves deep integration of semantics and context. This invention constructs a dual-stream cross-attention fusion network, which extracts the content destruction intensity through the full-element semantic analysis stream and extracts regional sensitivity, temporal clustering degree and subject historical frequency through the spatiotemporal context analysis stream. The two streams are bidirectionally interacted and fused through the cross-attention mechanism to generate a context correction score, overcoming the one-sidedness of single-dimensional evaluation.

[0110] Third, the accuracy of the assessment is significantly improved. This invention integrates unsupervised benchmark scores and context-corrected scores with weighted fusion, taking into account both subjective deviation and context sensitivity. Experimental data show that the AUC reaches 0.928 and the F1 score reaches 0.882, both of which are significantly better than existing baseline methods.

[0111] In the above embodiments, a method for dynamic risk assessment of urban operational status is provided. Correspondingly, this application also provides a system for dynamic risk assessment of urban operational status. The system for dynamic risk assessment of urban operational status provided in this application can implement the above-described method for dynamic risk assessment of urban operational status. This system can be implemented through software, hardware, or a combination of both. For example, the system may include integrated or separate functional modules or units to perform the corresponding steps in the above methods.

[0112] Please refer to Figure 4 As shown, this application provides a dynamic risk assessment system 10 for urban operational status, comprising: The acquisition module 101 is used to acquire multimodal content data and spatiotemporal metadata of the target event to be evaluated; the multimodal content data includes event text description, on-site images and / or speech-to-text transcription; the spatiotemporal metadata includes the timestamp of the event, location code and identification of the involved parties; The benchmark scoring module 102 is used to input the multimodal content data and spatiotemporal metadata of the target event into a pre-trained deep support vector data description model, and calculate the squared Euclidean distance from the target event to the center of the normal operation hypersphere in the feature space as the risk benchmark deviation score; the normal operation hypersphere is a compact hypersphere that surrounds the normal sample in the feature space, which is trained using multi-source governance data during the daily stable operation period of the city. The correction scoring module 103 is used to input the multimodal content data and spatiotemporal metadata of the target event into a pre-trained context-aware fusion network. The semantic destruction intensity features are extracted by the full-element semantic analysis stream, and the spatiotemporal context sensitivity features, including regional spatial sensitivity features, short-term temporal clustering density features, and historical risk frequency features of the involved subjects, are extracted by the spatiotemporal context analysis stream. The semantic destruction intensity features and the spatiotemporal context sensitivity features are bidirectionally interacted and fused through a cross-attention mechanism to generate a context correction score. The comprehensive scoring module 104 is used to weight and fuse the risk benchmark deviation score and the situation correction score to obtain a comprehensive risk score. The attribution classification module 105 is used to calculate the marginal contribution of the semantic destruction intensity feature, the regional spatial sensitivity feature, the short-term temporal clustering density feature, and the historical risk frequency feature of the involved subject to the comprehensive risk score using the SHAP game theory attribution method; based on the marginal contribution of each feature, the risk causes of the target event are classified into at least one of content intensity-dominated, spatiotemporal context-sensitive, or historical frequency accumulation type, and the classification results are output.

[0113] In some embodiments of this application, the system further includes: The attribution classification module is used to calculate the marginal contribution of the semantic destruction intensity feature, the regional spatial sensitivity feature, the short-term temporal clustering density feature, and the historical risk frequency feature of the involved subject to the comprehensive risk score using the SHAP game theory attribution method; based on the marginal contribution of each feature, the risk causes of the target event are classified into at least one of the following: content intensity-driven, spatiotemporal context-sensitive, and historical frequency-cumulative, and the classification results are output.

[0114] In some embodiments of this application, the specific processing steps of the full-element semantic analysis stream include: A pre-trained language model is used to perform deep semantic encoding on event text descriptions and speech-to-text transcriptions to obtain text semantic vectors. Visual semantic encoding of on-site images is performed using a visual transformer or a convolutional neural network to obtain image semantic vectors; The text semantic vector and the image semantic vector are fused across modalities to generate a unified multimodal semantic representation; The multimodal semantic representation is mapped to a risk assessment-specific destruction intensity subspace through a task-related projection layer to generate the semantic destruction intensity feature.

[0115] In some embodiments of this application, the extraction of the spatiotemporal context sensitivity features specifically includes: Based on the spatial sensitivity classification knowledge base of geographic information system, the sensitivity level of the location of an event is encoded as an embedding vector, which serves as the regional spatial sensitivity feature. Within a preset sliding time window before the occurrence of the target event, the number and density of related events occurring in the same area are statistically analyzed to form a density statistical feature vector containing multiple time granularities, which serves as a short-term time-series cluster density feature. During the pre-set observation period, the frequency of historical risk-related records of the involved entities is counted, and the frequency after time decay weighting is calculated. Together with the highest historical risk level, it constitutes the historical risk frequency characteristics of the involved entities. The spatiotemporal context sensitivity features are obtained by concatenating the regional spatial sensitivity features, the short-term temporal cluster density features, and the historical risk frequency features of the involved entities.

[0116] In some embodiments of this application, the specific implementation of the cross-attention mechanism includes: The semantic destruction intensity feature and the spatiotemporal context sensitivity feature are linearly projected into query, key, and value vectors, respectively. Calculate the attention weights of context features guided by semantic features to generate context-focused representations; Calculate the attention weights of semantic features guided by contextual features to generate semantic understanding representations modulated by context. The above process is executed in parallel using a multi-head attention mechanism, and the multi-head outputs are concatenated and generated through a feedforward network to produce the context correction score.

[0117] In some embodiments of this application, the formula for weighted fusion of the risk benchmark deviation score and the situation correction score is as follows: S final = α·S DSVDD + (1-α)·S context ; Among them, S final For the comprehensive risk score, S DSVDD S is the deviation score from the aforementioned risk benchmark. context The situation is corrected and scored, with α being the weighting coefficient.

[0118] In some embodiments of this application, the weighted fusion adopts an adaptive weighting mode, and the weighting coefficient is dynamically adjusted according to the risk benchmark deviation score of the current target event; when the risk benchmark deviation score is higher, the weighting coefficient is smaller.

[0119] In some embodiments of this application, the attribution classification module is specifically used for: The SHAP marginal contribution of each input feature of the target event is grouped and aggregated according to a preset feature category to generate a four-dimensional causal feature vector; the four-dimensional causal feature vector corresponds to the sum of features contributing to semantic content, spatial context, temporal context, and subject history. Calculate the similarity or distance between the four-dimensional causal feature vector and the three preset cluster center vectors, and classify the target event into the causal pattern category corresponding to the cluster center with the highest similarity or the closest distance; The three cluster center vectors are pre-determined by performing K-means clustering on the four-dimensional causal feature vectors of each sample in the historical high-risk sample set, and correspond to three causal pattern categories, respectively: The causal pattern corresponding to the center of the first cluster is content intensity-dominated, and the total contribution of its semantic content features is higher than the total contribution of spatial context, temporal context, and subject history features. The causal pattern corresponding to the second cluster center is spatiotemporally sensitive, and the sum of the contributions of spatial context features and / or temporal context features is higher than the sum of the contributions of semantic content features. The causal pattern corresponding to the third cluster center is the historical frequency accumulation type, and the sum of the main historical features is higher than the sum of the features of semantic content, spatial context and temporal context.

[0120] The risk dynamic assessment system for urban operation status provided in this application embodiment and the risk dynamic assessment method for urban operation status provided in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0121] This application also provides an electronic device corresponding to the method provided in the foregoing embodiments. The electronic device may be an electronic device for a client, such as a mobile phone, laptop, tablet computer, desktop computer, etc., to execute the above-mentioned risk dynamic assessment method for urban operation status.

[0122] Please refer to Figure 5 This illustrates a schematic diagram of an electronic device provided by some embodiments of this application. For example... Figure 5As shown, the electronic device 20 includes: a processor 200, a memory 201, a bus 202, and a communication interface 203. The processor 200, the communication interface 203, and the memory 201 are connected via the bus 202. The memory 201 stores a computer program that can run on the processor 200. When the processor 200 runs the computer program, it executes the risk dynamic assessment method for urban operation status provided in any of the foregoing embodiments of this application.

[0123] The memory 201 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 203 (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0124] Bus 202 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The memory 201 is used to store programs. After receiving an execution instruction, the processor 200 executes the program. The risk dynamic assessment method for urban operational status disclosed in any of the foregoing embodiments of this application can be applied to the processor 200, or implemented by the processor 200.

[0125] The processor 200 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 200 or by instructions in software form. The processor 200 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 201. The processor 200 reads the information in memory 201 and, in conjunction with its hardware, completes the steps of the above method.

[0126] The electronic device provided in this application embodiment and the risk dynamic assessment method for urban operation status provided in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0127] This application also provides a computer-readable storage medium corresponding to the risk dynamic assessment method for urban operation status provided in the aforementioned embodiments, wherein a computer program (i.e., a program product) is stored thereon. When the computer program is run by a processor, it executes the risk dynamic assessment method for urban operation status provided in any of the aforementioned embodiments.

[0128] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.

[0129] The computer-readable storage medium provided in the above embodiments of this application and the risk dynamic assessment method for urban operation status provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0130] Finally, it should be noted that 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. Such 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, and they should all be covered within the scope of the claims and specification of this application.

Claims

1. A dynamic risk assessment method for urban operational status, characterized in that, include: Acquire multimodal content data and spatiotemporal metadata of the target event to be evaluated; the multimodal content data includes event text description, on-site images, and / or speech-to-text transcription. The spatiotemporal metadata includes the timestamp of the event, the location code, and the identifier of the involved parties; The multimodal content data and spatiotemporal metadata of the target event are input into a pre-trained deep support vector data description model. The squared Euclidean distance from the target event to the center of the normal operation hypersphere is calculated in the feature space and used as the risk benchmark deviation score. The normal operation hypersphere is a compact hypersphere that surrounds the normal sample in the feature space, which is trained using multi-source governance data during the city's daily stable operation period. The multimodal content data and spatiotemporal metadata of the target event are input into a pre-trained context-aware fusion network. The semantic destruction intensity features are extracted by the full-element semantic analysis stream, and the spatiotemporal context sensitivity features, including regional spatial sensitivity features, short-term temporal aggregation density features, and historical risk frequency features of the involved subjects, are extracted by the spatiotemporal context analysis stream. A context correction score is generated by bidirectional interactive fusion of the semantic destruction intensity feature and the spatiotemporal context sensitivity feature through a cross-attention mechanism. The risk benchmark deviation score and the situation correction score are weighted and fused to obtain the comprehensive risk score of the target event.

2. The method according to claim 1, characterized in that, The method further includes: The SHAP game theory attribution method is used to calculate the marginal contribution of the semantic destruction intensity feature, the regional spatial sensitivity feature, the short-term temporal clustering density feature, and the historical risk frequency feature of the involved subject to the comprehensive risk score. Based on the marginal contribution of each feature, the risk causes of the target event are classified into at least one of the following: content intensity-driven, spatiotemporal context-sensitive, and historical frequency-cumulative, and the classification results are output.

3. The method according to claim 1, characterized in that, The specific processing steps of the full-element semantic analysis flow include: A pre-trained language model is used to perform deep semantic encoding on event text descriptions and speech-to-text transcriptions to obtain text semantic vectors. Visual semantic encoding of on-site images is performed using a visual transformer or a convolutional neural network to obtain image semantic vectors; The text semantic vector and the image semantic vector are fused across modalities to generate a unified multimodal semantic representation; The multimodal semantic representation is mapped to a risk assessment-specific destruction intensity subspace through a task-related projection layer to generate the semantic destruction intensity feature.

4. The method according to claim 1, characterized in that, The extraction of the spatiotemporal context sensitivity features specifically includes: Based on the spatial sensitivity classification knowledge base of geographic information system, the sensitivity level of the location of an event is encoded as an embedding vector, which serves as the regional spatial sensitivity feature. Within a preset sliding time window before the occurrence of the target event, the number and density of related events occurring in the same area are statistically analyzed to form a density statistical feature vector containing multiple time granularities, which serves as a short-term time-series cluster density feature. During the pre-set observation period, the frequency of historical risk-related records of the involved entities is counted, and the frequency after time decay weighting is calculated. Together with the highest historical risk level, it constitutes the historical risk frequency characteristics of the involved entities. The spatiotemporal context sensitivity features are obtained by concatenating the regional spatial sensitivity features, the short-term temporal cluster density features, and the historical risk frequency features of the involved entities.

5. The method according to claim 1, characterized in that, The specific implementation of the cross-attention mechanism includes: The semantic destruction intensity feature and the spatiotemporal context sensitivity feature are linearly projected into query, key, and value vectors, respectively. Calculate the attention weights of context features guided by semantic features to generate context-focused representations; Calculate the attention weights of semantic features guided by contextual features to generate semantic understanding representations modulated by context. The above process is executed in parallel using a multi-head attention mechanism, and the multi-head outputs are concatenated and generated through a feedforward network to produce the context correction score.

6. The method according to claim 1, characterized in that, The formula for weighted fusion of the risk benchmark deviation score and the situation correction score is as follows: S final = α·S DSVDD + (1-a)·S context ; Among them, S final For the comprehensive risk score, S DSVDD S is the deviation score from the aforementioned risk benchmark. context The situation is corrected and scored, with α being the weighting coefficient.

7. The method according to claim 1 or 6, characterized in that, The weighted fusion adopts an adaptive weighting mode, and the weighting coefficient is dynamically adjusted according to the risk benchmark deviation score of the current target event; the higher the risk benchmark deviation score, the smaller the weighting coefficient.

8. The method according to claim 2, characterized in that, The classification of the risk causes of the target event into at least one of the following: content intensity-driven, spatiotemporal context-sensitive, and historical frequency-cumulative. The SHAP marginal contribution of each input feature of the target event is grouped and aggregated according to a preset feature category to generate a four-dimensional causal feature vector; the four-dimensional causal feature vector corresponds to the sum of features contributing to semantic content, spatial context, temporal context, and subject history. Calculate the similarity or distance between the four-dimensional causal feature vector and the three preset cluster center vectors, and classify the target event into the causal pattern category corresponding to the cluster center with the highest similarity or the closest distance; The three cluster center vectors are pre-determined by performing K-means clustering on the four-dimensional causal feature vectors of each sample in the historical high-risk sample set, and correspond to three causal pattern categories, respectively: The causal pattern corresponding to the center of the first cluster is content intensity-dominated, and the total contribution of its semantic content features is higher than the total contribution of spatial context, temporal context, and subject history features. The causal pattern corresponding to the second cluster center is spatiotemporally sensitive, and the sum of the contributions of spatial context features and / or temporal context features is higher than the sum of the contributions of semantic content features. The causal pattern corresponding to the third cluster center is the historical frequency accumulation type, and the sum of the main historical features is higher than the sum of the features of semantic content, spatial context and temporal context.

9. The method according to claim 1, characterized in that, The training process of the deep support vector data description model includes: The city acquires multi-source heterogeneous governance data during normal daily operation, and preprocesses it to construct a normal operation training dataset. The multi-source heterogeneous governance data includes police incident record data, social media data, grid patrol reporting data, and video surveillance intelligent recognition data. A deep support vector data description algorithm is adopted. A deep neural network encoder is trained using the normal operation training dataset to map the input samples to the feature space. In the feature space, a compact hypersphere surrounding the normal samples is learned, and the center and radius of the hypersphere are determined to obtain the pre-trained deep support vector data description model. The training objective of the deep support vector data description model is to minimize the volume of the hypersphere while constraining the normal samples to be surrounded by the hypersphere. A context-aware fusion network is constructed and trained. The context-aware fusion network includes a parallel full-element semantic analysis stream and a spatiotemporal context analysis stream, as well as a cross-attention fusion module connecting the two streams. During training, the weighted fusion result of the risk baseline deviation score output by the deep support vector data description model and the context correction score output by the context-aware fusion network is used as the comprehensive risk score. The pre-trained context-aware fusion network is obtained by end-to-end or phased optimization using a preset loss function.

10. A dynamic risk assessment system for urban operational status, characterized in that, include: The acquisition module is used to acquire multimodal content data and spatiotemporal metadata of the target event to be evaluated; The benchmark scoring module is used to input the multimodal content data and spatiotemporal metadata of the target event into a pre-trained deep support vector data description model, and calculate the squared Euclidean distance from the target event to the center of the normally operating hypersphere in the feature space as the risk benchmark deviation score. The correction scoring module is used to input the multimodal content data and spatiotemporal metadata of the target event into a pre-trained context-aware fusion network, and the context-aware fusion network extracts semantic destruction intensity features and spatiotemporal context sensitivity features. A context correction score is generated by bidirectional interactive fusion of the semantic destruction intensity feature and the spatiotemporal context sensitivity feature through a cross-attention mechanism. The comprehensive scoring module is used to weight and fuse the risk benchmark deviation score and the situation correction score to obtain a comprehensive risk score; The attribution classification module is used to calculate the marginal contribution of each feature pair to the comprehensive risk score using the SHAP game theory attribution method; based on the marginal contribution of each feature, the risk causes of the target event are classified into at least one of content intensity-driven, spatiotemporal context-sensitive, or historical frequency-cumulative types, and the classification results are output.

11. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the method as claimed in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, It stores computer-readable instructions that can be executed by a processor to implement the method as described in any one of claims 1 to 9.

Images