Urban safety risk assessment method and system based on big data

Abstract

The invention discloses an urban safety risk assessment method and system based on big data, and relates to the technical field of big data, and the method comprises the steps: constructing a multi-source data collection network, and obtaining data from a government department database, Internet of Things equipment, a social media platform and a traffic monitoring system in real time; preprocessing the collected multi-source data, wherein the preprocessing comprises data cleaning, format standardization, unstructured data semantic analysis and sentiment analysis; a cross-department data security sharing mechanism is established, and the traceability and security of data exchange are ensured through a block chain technology; constructing a dynamic risk assessment model, analyzing multi-source data relevance based on a deep learning algorithm, and dynamically adjusting the weight of each risk factor; and generating a visual risk assessment report, and pushing the visual risk assessment report to related departments in real time through an early warning system. According to the invention, through multi-technology fusion and a dynamic optimization mechanism, the accuracy, real-time performance and cooperation efficiency of urban safety risk assessment are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of big data technology, and in particular to a city safety risk assessment method and system based on big data. Background Art

[0002] With the acceleration of global urbanization, urban management must rely on data-driven decision-making to improve public safety. Big data technology is relatively mature. Advances in distributed computing frameworks like Hadoop and Spark, as well as machine learning algorithms, have enabled real-time processing of massive amounts of heterogeneous data, such as video surveillance, social media, and IoT sensor data. The integration of multi-dimensional data, including GIS, traffic flow, meteorological data, and police records, provides a more comprehensive perspective for risk assessment. Furthermore, governments around the world are promoting the development of "safe cities," such as China's "Xueliang Project" and the United States' "Smart and Safe Cities" initiative.

[0003] There are still problems in existing technologies that need to be solved: traditional risk assessment methods rely on historical statistical data and expert experience, making it difficult to dynamically respond to emergencies such as traffic accidents and natural disasters. They rely on structured data and ignore unstructured data; the data source is single, and data from various departments are not interconnected, resulting in one-sided assessments, and single data is easily affected by noise. Summary of the Invention

[0004] In order to solve the above technical problems, a method and system for urban safety risk assessment based on big data is provided. This technical solution solves the above-mentioned problems of difficulty in dynamically responding to emergencies, reliance on structured data, and ignoring unstructured data; single data source and lack of intercommunication between data of various departments, resulting in one-sided assessment and susceptibility of single data to noise interference.

[0005] In order to achieve the above objects, the technical solution adopted by the present invention is:

[0006] A method and system for urban safety risk assessment based on big data, comprising:

[0007] Build a multi-source data collection network to obtain structured and unstructured data from government databases, IoT devices, social media platforms, and traffic monitoring systems in real time;

[0008] Pre-processing of collected multi-source data, including data cleaning, format standardization, unstructured data semantic parsing and sentiment analysis;

[0009] Establish a cross-departmental data security sharing mechanism and ensure the traceability and security of data exchange through blockchain technology;

[0010] Build a dynamic risk assessment model, analyze the correlation of multi-source data based on deep learning algorithms, and dynamically adjust the weight of each risk factor;

[0011] Generate visual risk assessment reports and push them to relevant departments in real time through the early warning system.

[0012] Preferably, the construction of a multi-source data collection network to obtain structured and unstructured data from government databases, Internet of Things devices, social media platforms, and traffic monitoring systems in real time specifically includes:

[0013] Public safety data from government databases is obtained through the API interface of the government data sharing platform. Infrastructure data is obtained by connecting to the BIM system of the Housing and Urban-Rural Development Bureau and extracting IoT sensor readings. Environmental meteorological data is obtained by spatial interpolation calculation of radar data from the Meteorological Bureau.

[0014] Urban sensing terminals in IoT devices integrate PM2.5 monitoring, noise sensors, cameras, manhole cover water level monitors, and elevator operation status monitors; mobile monitoring equipment includes mobile air quality monitors and drone inspections of illegal construction heat maps;

[0015] Text data from social media platforms is labeled with negative sentiment using distributed crawler clusters and sentiment analysis models. Multimedia data is extracted from user-uploaded disaster scene videos and images located in risk areas, and then detected using YOLOv8 for target detection and structured description storage.

[0016] The video surveillance data in the traffic monitoring system uses the optical flow method to calculate abnormal crowd gatherings and automatically detect traffic accidents; traffic operation data uses OD analysis of bus and subway card swiping records to identify abnormal commuting, cluster analysis of illegally parked shared bicycles, and cross-verify the authenticity of congestion with the navigation software's road condition API.

[0017] Preferably, the preprocessing of the collected multi-source data, including data cleaning, format standardization, unstructured data semantic parsing and sentiment analysis, specifically includes:

[0018] Data cleaning uses methods such as deleting and filling missing values, and processing outliers by removing, pruning, replacing, summarizing, and transforming them; using functions of data processing tools to detect and delete duplicate records;

[0019] Format standardization is achieved by converting data of different data types into a consistent type, converting data of different ranges into the same range, and encoding categorical variables;

[0020] Semantic parsing of unstructured data extracts useful information by performing word segmentation, stop word removal, and stemming operations on text data. It then identifies compound events by combining text, geographic location, and publication time. It also uses visual positioning algorithms to match street view images to high-precision maps.

[0021] Sentiment analysis uses a hierarchical sentiment model to perform multi-dimensional sentiment calculation. The hierarchical levels include using Prompt's zero-shot classification to analyze basic sentiment; using the fine-tuned Llama3-8B sentiment analysis adapter to analyze fine-grained sentiment; and using spatial kernel density estimation and sentiment value aggregation to analyze group sentiment.

[0022] Preferably, the establishment of a cross-departmental data security sharing mechanism to ensure the traceability and security of data exchange through blockchain technology specifically includes:

[0023] The consortium chain in the blockchain architecture uses Hyperledger Fabric 4.0, and the nodes are classified into endorsement nodes, sorting nodes, and light nodes;

[0024] Data exchange uses a hierarchical authorization mechanism. Public-level access conditions require that any node can read it. Restricted-level access conditions require signatures from multiple endorsing nodes. Confidential-level access conditions require federated learning + multi-party secure computing.

[0025] Cross-chain interoperability: the judicial evidence chain and the government data chain automatically synchronize judicial and administrative punishment information through a cross-chain oracle, and use zero-knowledge proof to verify data authenticity;

[0026] Quantum security protection uses lattice cryptography algorithms, and quantum-resistant signature keys are updated every quarter.

[0027] Preferably, the construction of a dynamic risk assessment model, analyzing the correlation of multi-source data based on a deep learning algorithm, and dynamically adjusting the weight of each risk factor specifically include:

[0028] The correlation of multi-source data uses spatiotemporal graph neural networks to process the spatial correlation of urban elements, dynamically model the risk propagation path, and introduce causal discovery algorithms to distinguish between correlation and causality;

[0029] A dynamic weight adjustment mechanism divides weights into three dimensions: data credibility, calculated using Bayesian credit scoring; timeliness, calculated using an exponential decay function; and event relevance, obtained using ontology-based semantic matching. A meta-learning framework is used to regularly update the weight strategy, with emergencies triggering incremental learning.

[0030] Risk prediction is divided into multi-scenario prediction. Short-term risks are predicted by predicting risk values, medium- and long-term risks are predicted by combining numerical weather forecasts for physical-informed predictions, and implementation risks are predicted by risk diffusion simulation based on urban digital twins. Uncertainty is quantified, and the output is in the form of confidence intervals. Monte Carlo Dropout is used to evaluate the model confidence.

[0031] Preferably, generating a visual risk assessment report specifically includes:

[0032] Build a multi-dimensional risk map that integrates GIS maps, real-time video streams, and IoT sensor data. The risk heat map supports multi-level penetration and drilling down into administrative districts, streets, and key facilities.

[0033] Cross-system data dashboards, including weather bureau radars in meteorological and hydrological areas, water level sensors at the water bureau, and grid load rates and subway outage information in key infrastructure areas;

[0034] Llama3-Emergency Edition large model is used to automatically generate bilingual reports in Chinese and English. Llama3-Emergency Edition large model is used to automatically generate bilingual reports in Chinese and English. ProphetX algorithm is used to generate future short-term risk trend forecast charts.

[0035] The visualization output includes a three-dimensional emergency command view, which uses Unreal Engine 5.3 to render the city digital twin, with emphasis on marking risk spillover paths and emergency resource distribution.

[0036] Preferably, the real-time push to relevant departments through the early warning system specifically includes:

[0037] A hierarchical warning trigger mechanism: Level 1 is a city-wide emergency broadcast and mandatory pop-up notifications on mobile phones; Level 2 is DingTalk for government affairs and in-car navigation warnings; Level 3 is WeChat service accounts and electronic bulletin boards;

[0038] Track disposal efficiency, establish an early warning response KPI dashboard, and regularly and automatically generate early warning effect evaluation reports, including early warning accuracy and average response time.

[0039] Furthermore, a city safety risk assessment system based on big data is used to implement the above-mentioned city safety risk assessment method based on big data, including:

[0040] Multi-source data acquisition module: real-time acquisition of structured and unstructured data from government databases, IoT devices, social media platforms, and traffic monitoring systems;

[0041] Data preprocessing module: cleans, standardizes and semantically analyzes heterogeneous data to establish a unified spatiotemporal reference coordinate system;

[0042] Blockchain data sharing platform: uses Hyperledger Fabric 4.0 architecture to achieve secure cross-departmental data exchange, and sets three levels of access rights: public, restricted, and confidential;

[0043] Dynamic risk assessment module: built on spatiotemporal graph neural networks and a meta-learning framework, it supports regular dynamic adjustment of risk factor weights;

[0044] Intelligent early warning push module: automatically triggers multi-channel early warnings based on risk levels, has multi-modal early warning forms, and optimizes the closed loop through feedback.

[0045] Optionally, the multi-source data acquisition module specifically includes:

[0046] Structured data units: Primarily based on government databases, public safety data includes alarm records with timestamps, longitude and latitude, and event types, and real-time pedestrian traffic in key areas. Urban infrastructure data includes monitoring of the inclination and vibration frequency of bridges and tunnels, and pressure and flow anomaly alarms for water and power pipelines.

[0047] Hybrid data stream unit: This collects data through IoT devices, including smart light poles, manhole cover sensors, and on-board diagnostic equipment. It collects data such as PM2.5, noise decibels, light intensity, water level, and sudden braking and collision events.

[0048] Unstructured data unit: This unit primarily collects text and multimedia data from social media platforms. Video analysis uses VideoLLaMA-3 to extract key frames, utilizes computer vision to identify risk features, and employs a visual positioning system to match street-level images to a 3D real-world city map.

[0049] Unstructured and structured data conversion unit: This unit enables intelligent video analysis through traffic accident identification and abnormal crowd gathering behavior detection, analyzes real-time data streams of traffic operation data, and cross-validates it with the navigation software's road condition API;

[0050] Special data source unit: collects night-time light index and night-time light index through low-orbit satellites, cooperates with enterprises to obtain the trajectory of food delivery riders and the clustering of online car-hailing sudden braking positions to reflect road anomalies.

[0051] Optionally, the dynamic risk assessment module specifically includes:

[0052] Spatiotemporal graph neural network unit: Models the relationship between urban elements, constructs spatial graphs with nodes including transportation hubs, hospitals, and chemical plants; and dynamically calculates the influence weights of different data sources.

[0053] Meta-learning optimization unit: A rapid adaptation mechanism that performs short-term, regular model fine-tuning. The risk scenario memory database stores historical event patterns and uses a similarity matching algorithm for matching.

[0054] Dynamic weight adjustment unit: Three-dimensional real-time calculation of data credibility, timeliness decay, and event relevance; emergency response strategy includes immediately increasing the weight of relevant factors when a red alert is triggered, and decaying the weight of irrelevant factors to a predetermined baseline value;

[0055] Digital twin simulator unit: Based on the three-dimensional real-life model of the city, the physical engine parameters are set. The Gaussian plume model is coupled with the wind speed field when chemicals leak, and the SEIR model and mobile phone signaling movement trajectory are used when infectious diseases spread.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] This invention proposes to integrate multimodal information such as government data, Internet of Things sensors, and social media public opinion through a heterogeneous data collection network, breaking through the limitations of a single data source in traditional systems; combining NLP sentiment analysis and computer vision technology to achieve automated parsing of unstructured data, improve data coverage dimensions, and solve the problem of information blind spots in risk assessment; a blockchain platform built on Hyperledger Fabric adopts three-level permission management and zero-knowledge proof technology to achieve cross-departmental data sharing while ensuring the security of sensitive data; innovatively integrates spatiotemporal graph neural networks and meta-learning frameworks, ST-GNN models the spatial correlation of traffic nodes and key facilities, supports dynamic deduction of risk transmission paths, and a triple weight adjustment mechanism reduces the model's false alarm rate. Incremental learning triggered by emergencies improves the speed of model parameter updates; a closed-loop system of "assessment-warning-feedback" is constructed, and a hierarchical push strategy significantly improves the public response rate. Automatic optimization of model parameters is achieved through warning code scanning statistics and feedback from grassroots grid workers. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is an internal flow chart of a city safety risk assessment method based on big data;

[0059] Figure 2 To build an internal flow chart of the multi-source data acquisition network;

[0060] Figure 3 An internal flow chart for building a dynamic risk assessment model;

[0061] Figure 4 This is an internal framework diagram of an urban safety risk assessment system based on big data. DETAILED DESCRIPTION

[0062] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0063] Reference Figure 1 As shown in FIG, a method for urban safety risk assessment based on big data includes:

[0064] Build a multi-source data collection network to obtain structured and unstructured data from government databases, IoT devices, social media platforms, and traffic monitoring systems in real time;

[0065] Pre-processing of collected multi-source data, including data cleaning, format standardization, unstructured data semantic parsing and sentiment analysis;

[0066] Establish a cross-departmental data security sharing mechanism and ensure the traceability and security of data exchange through blockchain technology;

[0067] Build a dynamic risk assessment model, analyze the correlation of multi-source data based on deep learning algorithms, and dynamically adjust the weight of each risk factor;

[0068] Generate visual risk assessment reports and push them to relevant departments in real time through the early warning system.

[0069] It should be noted that the basis for the overall coordination of this method is the unification of time and space benchmarks. High-precision time and space alignment includes the mandatory addition of BeiDou-3 timing tags to all data and city-level real-scene three-dimensional coordinates;

[0070] Dynamic geographic grid coding uses H3 Uber hexagonal grids, real-time risk values ​​are aggregated by grid, and cross-departmental data spatial matching is supported.

[0071] The performance guarantee of this method is the edge-cloud collaborative computing architecture, which is divided into five levels of computing nodes:

[0072] L1 nodes are deployed on IoT terminals and their processing task is to perform preliminary data filtering;

[0073] The L2 node is deployed at the 5G edge computing station and processes video key frame extraction / simple semantic analysis.

[0074] The L3 node is deployed in the district-level government cloud, and its processing task is the primary fusion of multi-source data;

[0075] The L4 node is deployed in the main center of the city brain and processes risk assessment model calculations;

[0076] L5 nodes are deployed in provincial disaster recovery centers, and their processing tasks are full data storage / model training.

[0077] Dynamic load balancing is achieved through the Kubernetes federated cluster. When traffic bursts occur, 30% of computing tasks are automatically diverted to nodes in surrounding cities.

[0078] Reference Figure 2 As shown in the figure, a multi-source data collection network is built to obtain structured and unstructured data from government databases, IoT devices, social media platforms, and traffic monitoring systems in real time, including:

[0079] Public safety data from government databases is obtained through the API interface of the government data sharing platform. Infrastructure data is obtained by connecting to the BIM system of the Housing and Urban-Rural Development Bureau and extracting IoT sensor readings. Environmental meteorological data is obtained by spatial interpolation calculation of radar data from the Meteorological Bureau.

[0080] Urban sensing terminals in IoT devices integrate PM2.5 monitoring, noise sensors, cameras, manhole cover water level monitors, and elevator operation status monitors; mobile monitoring equipment includes mobile air quality monitors and drone inspections of illegal construction heat maps;

[0081] Text data from social media platforms is labeled with negative sentiment using distributed crawler clusters and sentiment analysis models. Multimedia data is extracted from user-uploaded disaster scene videos and images from WeChat Moments located in risk areas. YOLOv8 is used for object detection and structured description storage.

[0082] The video surveillance data in the traffic monitoring system uses the optical flow method to calculate abnormal crowd gatherings and automatically detect traffic accidents; traffic operation data uses OD analysis of bus and subway card swiping records to identify abnormal commuting, cluster analysis of illegally parked shared bicycles, and cross-verify the authenticity of congestion with the navigation software's road condition API.

[0083] It should be noted that all devices are forced to access the BeiDou-3 PNT service, and implement government data (NTP protocol) and IoT terminals (LoRaWAN gateway) through hardware-level timestamp chips; based on the CIM platform, establish a city-level digital twin coordinate system, BIM model LOD4 accuracy, drone aerial photography real-scene modeling, and social media positioning data through VPS visual positioning correction.

[0084] Real-time data collection is guaranteed, and hierarchical stream processing is implemented. For level one (sensor), the transmission protocol is MQTT+5G URLLC, and the application scenario is bridge tilt exceeding threshold alarm; for level two (video), the transmission protocol is RTMP+FPGA acceleration, and the application scenario is application scenario; for level three (social), the transmission protocol is Kafka+Spark, and the application scenario is public opinion hotspot analysis.

[0085] Sensitive data is encrypted using the CRYSTALS-Kyber quantum-resistant algorithm, and the key is generated by a quantum random number generator.

[0086] Reference Figure 3 As shown, the dynamic risk assessment model is constructed, which analyzes the correlation of multi-source data based on deep learning algorithms and dynamically adjusts the weights of various risk factors, including:

[0087] The correlation of multi-source data uses spatiotemporal graph neural networks to process the spatial correlation of urban elements, dynamically model the risk propagation path, and introduce causal discovery algorithms to distinguish between correlation and causality;

[0088] A dynamic weight adjustment mechanism divides weights into three dimensions: data credibility, calculated using Bayesian credit scoring; timeliness, calculated using an exponential decay function; and event relevance, obtained using ontology-based semantic matching. A meta-learning framework is used to regularly update the weight strategy, with emergencies triggering incremental learning.

[0089] Risk prediction is divided into multi-scenario prediction. Short-term risks are predicted by predicting risk values, medium- and long-term risks are predicted by combining numerical weather forecasts for physical-informed predictions, and implementation risks are predicted by risk diffusion simulation based on urban digital twins. Uncertainty is quantified, and the output is in the form of confidence intervals. Monte Carlo Dropout is used to evaluate the model confidence.

[0090] It should be noted that the spatiotemporal graph neural network includes:

[0091] Heterogeneous node processing: Urban elements such as transportation nodes, weather stations, and population density grids require the design of heterogeneous graph structures and the use of type-aware graph attention mechanisms to distinguish the interaction patterns of different spatial entities.

[0092] Dynamic edge weights: Calculate the strength of the association between nodes through mutual information within a time sliding window, and dynamically update edge weights in combination with Granger causality tests. For example, the causal weight of traffic congestion on accident risk changes over time.

[0093] Causal discovery algorithms, based on the PC algorithm (constrained), are used in sparse causal networks and combined with Do-Calculus verification. Neural causal models such as DAG-GNN handle high-dimensional nonlinear relationships, but require adversarial training to avoid confusion bias.

[0094] Meta-learning weight update mechanism: the initial weight strategy is pre-trained through small sample tasks, such as historical disaster data in different urban areas, and the basic model is updated once a quarter; sudden events (such as earthquakes) activate the online learning module, limiting the gradient update amplitude (learning rate ≤ 0.001) to avoid catastrophic forgetting.

[0095] Multi-scenario forecasting includes:

[0096] For short-term risks (<24 hours), we use a LSTM-Transformer hybrid model, input real-time sensor data, and handle uncertainty through Bootstrap confidence intervals with a coverage rate of 95%;

[0097] For medium- to long-term risks (1-7 days), the WRF meteorological model output is used to generate the PINN physical constraint network, and uncertainty is handled by using the ensemble forecast (ENS) variance as the uncertainty input.

[0098] Real-time risk, using agent-based modeling in digital twins (such as Pathfinder), and uncertainty handling through Monte Carlo Dropout + Latin Hypercube sampling.

[0099] Short-term forecast results dynamically correct the boundary conditions of medium- and long-term models; digital twin simulation provides risk heat maps, which are fed back to ST-GNN to optimize spatial weights.

[0100] Reference Figure 4 As shown, a city safety risk assessment system based on big data includes:

[0101] Multi-source data acquisition module: real-time acquisition of structured and unstructured data from government databases, IoT devices, social media platforms, and traffic monitoring systems;

[0102] Data preprocessing module: cleans, standardizes and semantically analyzes heterogeneous data to establish a unified spatiotemporal reference coordinate system;

[0103] Blockchain data sharing platform: uses Hyperledger Fabric 4.0 architecture to achieve secure cross-departmental data exchange, and sets three levels of access rights: public, restricted, and confidential;

[0104] Dynamic risk assessment module: built on spatiotemporal graph neural networks and a meta-learning framework, it supports regular dynamic adjustment of risk factor weights;

[0105] Intelligent early warning push module: automatically triggers multi-channel early warnings based on risk levels, has multi-modal early warning forms, and optimizes the closed loop through feedback.

[0106] The multi-source data acquisition module specifically includes:

[0107] Structured data units: Primarily based on government databases, public safety data includes alarm records with timestamps, longitude and latitude, and event types, and real-time pedestrian traffic in key areas. Urban infrastructure data includes monitoring of the inclination and vibration frequency of bridges and tunnels, and pressure and flow anomaly alarms for water and power pipelines.

[0108] Hybrid data stream unit: This collects data through IoT devices, including smart light poles, manhole cover sensors, and on-board diagnostic equipment. It collects data such as PM2.5, noise decibels, light intensity, water level, and sudden braking and collision events.

[0109] Unstructured data unit: This unit primarily collects text and multimedia data from social media platforms. Video analysis uses VideoLLaMA-3 to extract key frames, utilizes computer vision to identify risk features, and employs a visual positioning system to match street-level images to a 3D real-world city map.

[0110] Unstructured and structured data conversion unit: This unit enables intelligent video analysis through traffic accident identification and abnormal crowd gathering behavior detection, analyzes real-time data streams of traffic operation data, and cross-validates it with the navigation software's road condition API;

[0111] Special data source unit: collects night-time light index and night-time light index through low-orbit satellites, cooperates with enterprises to obtain the trajectory of food delivery riders and the clustering of online car-hailing sudden braking positions to reflect road anomalies.

[0112] The dynamic risk assessment module specifically includes:

[0113] Spatiotemporal graph neural network unit: Models the relationship between urban elements, constructs spatial graphs with nodes including transportation hubs, hospitals, and chemical plants; and dynamically calculates the influence weights of different data sources.

[0114] Meta-learning optimization unit: A rapid adaptation mechanism that performs short-term, regular model fine-tuning. The risk scenario memory database stores historical event patterns and uses a similarity matching algorithm for matching.

[0115] Dynamic weight adjustment unit: Three-dimensional real-time calculation of data credibility, timeliness decay, and event relevance; emergency response strategy includes immediately increasing the weight of relevant factors when a red alert is triggered, and decaying the weight of irrelevant factors to a predetermined baseline value;

[0116] Digital twin simulator unit: Based on the three-dimensional real-life model of the city, the physical engine parameters are set. The Gaussian plume model is coupled with the wind speed field when chemicals leak, and the SEIR model and mobile phone signaling movement trajectory are used when infectious diseases spread.

[0117] It should be noted that the feedback optimization loop is divided into four-dimensional optimization loops, including: data quality, which is based on GAN data enhancement and repair, and its optimization frequency is real-time; model performance, which is optimized every 24 hours through online learning + adversarial training; early warning rules, which are optimized every quarter through expert review meetings + historical event backtracking; system architecture, which is optimized annually through access to quantum computing test nodes.

[0118] The spatiotemporal graph neural network units in the dynamic risk assessment module include:

[0119] Multimodal node expansion: In addition to transportation hubs, hospitals, and chemical plants, it can also include 5G base station density, underground pipe network pressure sensors, social media hotspots, and other nodes to build a heterogeneous network.

[0120] Spatial relationship modeling uses an attention mechanism to dynamically adjust the connection strength between nodes. For example, the association weight between a chemical plant and a residential area downstream of the wind direction changes in real time with meteorological data.

[0121] Time dimension processing uses gated temporal convolution to capture periodic risks, such as the impact of morning and evening peak traffic congestion on emergency routes;

[0122] Dynamic weight calculation: The influence of data source weight is evaluated through adversarial verification. For example, in extreme weather conditions, the weight of meteorological bureau data is higher than that of daily inspection reports. The dynamic graph convolution formula (ST-GCN) is used for calculation:

[0123]

[0124] Where H (l+1) is the output feature matrix of the l+1th layer; υ is the activation function, usually a nonlinear function such as ReLU; K is the number of graph convolution kernels; is the adjacency matrix of the kth type of spatial relationship, such as the connection weight between the chemical plant and the downwind area; is the degree matrix, used for normalization; H (l) is the characteristic matrix of the nodes in the lth layer, including traffic flow, pollutant concentration, etc.; is the trainable parameter matrix; T temp It is a temporal convolution module that uses dilated convolution to capture hourly fluctuations; υ is the spatiotemporal weight coefficient, which defaults to 0.7.

[0125] The meta-learning optimization unit includes:

[0126] The layered fine-tuning strategy includes short-term fine-tuning (hourly), which uses lightweight gradient updates to adjust only the last layer of the model to address sudden data drift; regular fine-tuning (weekly / monthly), which combines the event pattern clustering results in the risk scenario memory library to reconstruct the feature extractor, such as optimizing the building structure vulnerability assessment after adding an earthquake disaster model;

[0127] The risk scenario memory stores historical event patterns, uses a similarity matching algorithm to match them, quantifies the degree of correlation between current events and historical patterns through mathematical modeling, and combines them with dynamic risk assessment scenarios to develop:

[0128] Weighted graph similarity calculation (applicable to spatiotemporal graph structure matching)

[0129]

[0130] Where, S(G c,G h ) is the graph G c and Figure G h The similarity between G c ,G h is the current event graph and the historical event graph, the nodes are city elements, and the edges are relationship weights; c (i,j),w h (i, j) is the edge weight between nodes i and j in the current graph and the historical graph; α and β are the reconciliation coefficients of graph structure similarity and node attribute similarity, usually α+β=1; NodeSim(V c ,V c ) is the node attribute similarity, calculated using cosine similarity or Euclidean distance;

[0131] When matching a chemical plant leakage event, if the structural similarity S between the current wind speed, plant location, and historical event graph is greater than 0.8, it is determined to be a high-risk reuse mode.

[0132] Dynamic Time Warping Distance (applicable to time series data matching):

[0133]

[0134] Where DTW(Q,C) is the dynamic time warping distance between sequences Q and C; Q and C are the time series of current events and historical events, such as the change in pollutant concentration monitored by sensors; To minimize all possible index mappings π, π is a permutation that maps the index of sequence Q to the index of sequence C; π is the optimal alignment path that allows time axis scaling matching; φ(t i ,t j ) is the time decay function, λ is the attenuation coefficient, the closer the timeliness, the higher the weight;

[0135] When matching the infectious disease spread curve, if the DTW distance is less than the threshold and the mobile phone signaling movement trajectories are similar, the same warning strategy will be triggered.

[0136] Multimodal feature fusion similarity (comprehensive attributes and relationships):

[0137]

[0138] Where Sim is the final similarity score; n is the number of modalities; are the two eigenvectors in the kth mode The similarity between them. is the k-th feature of current and historical events, such as text reports, sensor data, and images; Sim kis the similarity of the k-th feature, TF-IDF cosine similarity is used for text, and SSIM is used for image; γ k is the feature weight, generated by the dynamic weight adjustment unit; D geo is the geographic distance, such as the metric distance between the leakage point and the historical event point; σ is the spatial attenuation coefficient, which is usually set to the average radius of a city block; exp is an exponential function, indicating that the effect of distance on similarity decays exponentially;

[0139] When matching hospital run events, the similarity is calculated based on the bed occupancy rate (numerical feature), public opinion keywords (text feature), and surrounding traffic congestion (graph feature).

[0140] The dynamic weight adjustment unit includes:

[0141] Credibility calculation: The credibility score is determined by the historical accuracy of the data source (e.g., a weather station’s forecast error for the past 30 days is ≤ 5%) and cross-source consistency (when data from multiple sources conflict, the overlapping confidence intervals are used).

[0142] Time-sensitive decay function, designing an exponential decay coefficient, such as the half-life of traffic flow data is 15 minutes, while the half-life of urban planning data is 1 year;

[0143] Event relevance strategy: In red alert response, the weight of related factors (such as wind speed around a chemical plant) is increased to three times the baseline value, and the weight of irrelevant factors (such as the flow of people in shopping malls far away from the accident site) is reduced to 10%; long-tail risk processing, using adversarial generation samples to enhance weight adjustment training for rare events.

[0144] The digital twin simulator unit includes:

[0145] Physical engine parameter calibration: in the case of chemical leaks, a Gaussian plume model superimposed on CFD turbulence simulation is used, and diffusion parameters are dynamically corrected in combination with real-time drone gas concentration detection data. In the case of infectious disease transmission, the SEIR model is coupled with mobile phone signaling data with block-level accuracy, and a super-spreader identification algorithm (based on contact network node centrality) is introduced.

[0146] Multi-hazard linkage simulation supports disaster chain deduction, such as the cascade effect of earthquake → chemical plant leakage → traffic paralysis, and quantifies the risk transmission intensity of each link through discrete event simulation.

[0147] The spatiotemporal graph neural network is deployed on edge nodes, such as street cameras with built-in computing power, to achieve millisecond-level response. The digital twin simulator runs on a cloud-based GPU cluster to handle complex calculations. When a data source fails, it automatically switches to the Bayesian interpolation model to generate temporary data and marks low-confidence areas for manual review.

[0148] To sum up, the advantages of the present invention are: it has made a breakthrough in integrating multiple types of heterogeneous data sources such as government data, Internet of Things sensors, social media, and video surveillance, and innovatively applied multimodal NLP+CV technology to realize intelligent analysis of unstructured data, improve data coverage dimensions, and eliminate blind spots in risk assessment; the dual-engine architecture of spatiotemporal graph neural network and meta-learning realizes minute-level dynamic weight adjustment of risk factors and improves the response speed to emergencies; it builds a three-level data sharing system based on blockchain, adopts zk-SNARKs to make sensitive data "available but invisible", and improves cross-departmental collaboration efficiency; establishes a multi-level early warning push matrix, innovatively integrates digital twin simulation and real-scene data, and realizes model self-optimization through feedback mechanism.

[0149] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for urban safety risk assessment based on big data, characterized in that: include: Build a multi-source data collection network to obtain structured and unstructured data from government databases, IoT devices, social media platforms, and traffic monitoring systems in real time; Pre-processing of collected multi-source data, including data cleaning, format standardization, unstructured data semantic parsing and sentiment analysis; Establish a cross-departmental data security sharing mechanism and ensure the traceability and security of data exchange through blockchain technology; Build a dynamic risk assessment model, analyze the correlation of multi-source data based on deep learning algorithms, and dynamically adjust the weight of each risk factor; Generate visual risk assessment reports and push them to relevant departments in real time through the early warning system.

2. The urban safety risk assessment method based on big data according to claim 1 is characterized in that: The multi-source data collection network mentioned above is used to acquire structured and unstructured data from government databases, IoT devices, social media platforms, and traffic monitoring systems in real time. Specifically, it includes: Public safety data from government databases is obtained through the API interface of the government data sharing platform. Infrastructure data is obtained by connecting to the BIM system of the Housing and Urban-Rural Development Bureau and extracting IoT sensor readings. Environmental meteorological data is obtained by spatial interpolation calculation of radar data from the Meteorological Bureau. Urban sensing terminals in IoT devices integrate PM2.5 monitoring, noise sensors, cameras, manhole cover water level monitors, and elevator operation status monitors; mobile monitoring equipment includes mobile air quality monitors and drone inspections of illegal construction heat maps; Text data from social media platforms is labeled with negative sentiment using distributed crawler clusters and sentiment analysis models. Multimedia data is extracted from user-uploaded disaster scene videos and images located in risk areas, and then detected using YOLOv8 for target detection and structured description storage. The video surveillance data in the traffic monitoring system uses the optical flow method to calculate abnormal crowd gatherings and automatically detect traffic accidents; traffic operation data uses OD analysis of bus and subway card swiping records to identify abnormal commuting, cluster analysis of illegally parked shared bicycles, and cross-verify the authenticity of congestion with the navigation software's road condition API.

3. The urban safety risk assessment method based on big data according to claim 2 is characterized in that: The pre-processing of the collected multi-source data, including data cleaning, format standardization, unstructured data semantic parsing and sentiment analysis, specifically includes: Data cleaning uses methods such as deleting and filling missing values, and processing outliers by removing, pruning, replacing, summarizing, and transforming them; using functions of data processing tools to detect and delete duplicate records; Format standardization is achieved by converting data of different data types into a consistent type, converting data of different ranges into the same range, and encoding categorical variables; Semantic parsing of unstructured data extracts useful information by performing word segmentation, stop word removal, and stemming operations on text data. It then identifies compound events by combining text, geographic location, and publication time. It also uses visual positioning algorithms to match street view images to high-precision maps. Sentiment analysis uses a hierarchical sentiment model to perform multi-dimensional sentiment calculation. The hierarchical levels include using Prompt's zero-shot classification to analyze basic sentiment; using the fine-tuned Llama3-8B sentiment analysis adapter to analyze fine-grained sentiment; and using spatial kernel density estimation and sentiment value aggregation to analyze group sentiment.

4. The urban safety risk assessment method based on big data according to claim 3 is characterized in that: The establishment of a cross-departmental data security sharing mechanism and the use of blockchain technology to ensure the traceability and security of data exchange specifically include: The consortium chain in the blockchain architecture uses Hyperledger Fabric 4.0, and the nodes are classified into endorsement nodes, sorting nodes, and light nodes; Data exchange uses a hierarchical authorization mechanism. Public-level access conditions require that any node can read it. Restricted-level access conditions require signatures from multiple endorsing nodes. Confidential-level access conditions require federated learning + multi-party secure computing. Cross-chain interoperability: the judicial evidence chain and the government data chain automatically synchronize judicial and administrative punishment information through a cross-chain oracle, and use zero-knowledge proof to verify data authenticity; Quantum security protection uses lattice cryptography algorithms, and quantum-resistant signature keys are updated every quarter.

5. The urban safety risk assessment method based on big data according to claim 4 is characterized in that: The construction of a dynamic risk assessment model, which analyzes the correlation of multi-source data based on a deep learning algorithm and dynamically adjusts the weight of each risk factor, specifically includes: The correlation of multi-source data uses spatiotemporal graph neural networks to process the spatial correlation of urban elements, dynamically model the risk propagation path, and introduce causal discovery algorithms to distinguish between correlation and causality; A dynamic weight adjustment mechanism divides weights into three dimensions: data credibility, calculated using Bayesian credit scoring; timeliness, calculated using an exponential decay function; and event relevance, obtained using ontology-based semantic matching. A meta-learning framework is used to regularly update the weight strategy, with emergencies triggering incremental learning. Risk prediction is divided into multi-scenario prediction. Short-term risks are predicted by predicting risk values, medium- and long-term risks are predicted by combining numerical weather forecasts for physical-informed predictions, and implementation risks are predicted by risk diffusion simulation based on urban digital twins. Uncertainty is quantified, and the output is in the form of confidence intervals. Monte Carlo Dropout is used to evaluate the model confidence.

6. The urban safety risk assessment method based on big data according to claim 5 is characterized in that: Generating a visual risk assessment report specifically includes: Build a multi-dimensional risk map that integrates GIS maps, real-time video streams, and IoT sensor data. The risk heat map supports multi-level penetration and drilling down into administrative districts, streets, and key facilities. Cross-system data dashboards, including weather bureau radars in meteorological and hydrological areas, water level sensors at the water bureau, and grid load rates and subway outage information in key infrastructure areas; Llama3-Emergency Edition large model is used to automatically generate bilingual reports in Chinese and English. Llama3-Emergency Edition large model is used to automatically generate bilingual reports in Chinese and English. ProphetX algorithm is used to generate future short-term risk trend forecast charts. The visualization output includes a three-dimensional emergency command view, which uses Unreal Engine 5.3 to render the city digital twin, with emphasis on marking risk spillover paths and emergency resource distribution.

7. The urban safety risk assessment method based on big data according to claim 6 is characterized in that: The real-time push notifications to relevant departments through the early warning system specifically include: A hierarchical warning trigger mechanism: Level 1 is a city-wide emergency broadcast and mandatory pop-up notifications on mobile phones; Level 2 is DingTalk for government affairs and in-car navigation warnings; Level 3 is WeChat service accounts and electronic bulletin boards; Track disposal efficiency, establish an early warning response KPI dashboard, and regularly and automatically generate early warning effect evaluation reports, including early warning accuracy and average response time.

8. A city safety risk assessment system based on big data, characterized by: A method for implementing a city safety risk assessment method based on big data as described in claims 1 to 7, comprising: Multi-source data acquisition module: real-time acquisition of structured and unstructured data from government databases, IoT devices, social media platforms, and traffic monitoring systems; Data preprocessing module: cleans, standardizes and semantically analyzes heterogeneous data to establish a unified spatiotemporal reference coordinate system; Blockchain data sharing platform: uses Hyperledger Fabric 4.0 architecture to achieve secure cross-departmental data exchange, and sets three levels of access rights: public, restricted, and confidential; Dynamic risk assessment module: built on spatiotemporal graph neural networks and a meta-learning framework, it supports regular dynamic adjustment of risk factor weights; Intelligent early warning push module: automatically triggers multi-channel early warnings based on risk levels, has multi-modal early warning forms, and optimizes the closed loop through feedback.

9. The urban safety risk assessment system based on big data according to claim 8 is characterized in that: The multi-source data acquisition module specifically includes: Structured data units: Primarily based on government databases, public safety data includes alarm records with timestamps, longitude and latitude, and event types, and real-time pedestrian traffic in key areas. Urban infrastructure data includes monitoring of the inclination and vibration frequency of bridges and tunnels, and pressure and flow anomaly alarms for water and power pipelines. Hybrid data stream unit: This collects data through IoT devices, including smart light poles, manhole cover sensors, and on-board diagnostic equipment. It collects data such as PM2.5, noise decibels, light intensity, water level, and sudden braking and collision events. Unstructured data unit: This unit primarily collects text and multimedia data from social media platforms. Video analysis uses VideoLLaMA-3 to extract key frames, utilizes computer vision to identify risk features, and employs a visual positioning system to match street-level images to a 3D real-world city map. Unstructured and structured data conversion unit: This unit enables intelligent video analysis through traffic accident identification and abnormal crowd gathering behavior detection, analyzes real-time data streams of traffic operation data, and cross-validates it with the navigation software's road condition API; Special data source unit: collects night-time light index and night-time light index through low-orbit satellites, cooperates with enterprises to obtain the trajectory of food delivery riders and the clustering of online car-hailing sudden braking positions to reflect road anomalies.

10. The urban safety risk assessment system based on big data according to claim 9 is characterized in that: The dynamic risk assessment module specifically includes: Spatiotemporal graph neural network unit: Models the relationship between urban elements, constructs spatial graphs with nodes including transportation hubs, hospitals, and chemical plants; and dynamically calculates the influence weights of different data sources. Meta-learning optimization unit: A rapid adaptation mechanism that performs short-term, regular model fine-tuning. The risk scenario memory database stores historical event patterns and uses a similarity matching algorithm for matching. Dynamic weight adjustment unit: Three-dimensional real-time calculation of data credibility, timeliness decay, and event relevance; emergency response strategy includes immediately increasing the weight of relevant factors when a red alert is triggered, and decaying the weight of irrelevant factors to a predetermined baseline value; Digital twin simulator unit: Based on the three-dimensional real-life model of the city, the physical engine parameters are set. The Gaussian plume model is coupled with the wind speed field when chemicals leak, and the SEIR model and mobile phone signaling movement trajectory are used when infectious diseases spread.

Images