A public security risk early warning method and system based on multi-source data fusion

Abstract

This invention discloses a public safety risk early warning method and system based on multi-source data fusion, belonging to the field of public safety risk early warning technology. It includes multi-source heterogeneous data acquisition and fusion processing; risk chain identification based on knowledge graphs; dynamic risk assessment and graded early warning; intelligent contingency plan matching and coordinated response; and feedback learning and continuous model optimization. This invention, employing the aforementioned public safety risk early warning method and system based on multi-source data fusion, achieves a comprehensive upgrade in risk early warning from passive response to proactive prevention, from experience-driven to data-driven, from static grading to dynamic profiling, and from isolated processing to collaborative linkage.

Description

Technical Field

[0001] This invention relates to the field of public safety risk early warning technology, and in particular to a public safety risk early warning method and system based on multi-source data fusion. Background Technology

[0002] Current public safety risk early warning technologies suffer from the following shortcomings: Multi-source heterogeneous data are difficult to integrate effectively due to significant differences in format standards, spatiotemporal benchmarks, and semantic expressions, resulting in prominent data silos; risk identification is limited to isolated event detection, lacking the ability to track the cascading transmission of risks along infrastructure networks; risk assessment relies on historical data to define fixed levels, failing to reflect the dynamic impact of real-time factors such as weather changes and population fluctuations on the risk situation; there is a lack of effective coordination between early warning information and emergency plans, available resources, and responsible departments; plans are mostly stored in natural language text, which computers cannot automatically parse and dynamically match, leading to a gap between early warning and response; and the core parameters of the system remain fixed after deployment, lacking an adaptive optimization mechanism that automatically adjusts based on actual operational feedback. To address these deficiencies, a systematic solution is urgently needed that can achieve deep integration of multi-source heterogeneous data, automatic tracking of risk transmission paths, dynamic assessment of risk situations, intelligent matching of plans, and continuous self-optimization based on feedback. Summary of the Invention

[0003] The purpose of this invention is to provide a public safety risk early warning method and system based on multi-source data fusion, aiming to build a closed-loop early warning system covering the entire chain from data perception, risk identification, dynamic assessment to contingency plan matching, coordinated response, and feedback optimization, thereby improving the accuracy, timeliness, and feasibility of public safety risk early warning.

[0004] To achieve the above objectives, this invention provides a public safety risk early warning method based on multi-source data fusion, comprising the following steps: S1. Multi-source heterogeneous data acquisition and fusion processing: Access multi-source heterogeneous data and perform preprocessing. Use a three-level progressive fusion strategy of data layer fusion, feature layer fusion and decision layer fusion to fuse the data. Construct an urban safety knowledge graph based on the fused data. S2. Risk chain identification based on knowledge graph: Construct an event association topology graph, use graph neural network to identify risk transmission paths, and filter and output risk chains through transmission strength assessment; S3. Dynamic Risk Assessment and Graded Early Warning: A dynamic risk assessment model is constructed from four dimensions: hazard intensity, exposure level, vulnerability and response capability. A deep temporal neural network is used to predict the risk development trend and to trigger graded early warning based on the risk type and risk level. S4. Intelligent contingency plan matching and coordinated response: Construct a structured contingency plan library, and use a risk-contingency plan matching engine to screen, verify and simulate contingency plans, and output recommended decision-making solutions and graded early warning signals; S5. Feedback Learning and Continuous Model Optimization: Collect actual execution feedback data and continuously optimize the risk assessment model, risk chain identification model, and contingency plan matching model through reinforcement learning models.

[0005] Preferably, step S1 includes the following: S11. Access multi-source heterogeneous data. Multi-source heterogeneous data includes at least video surveillance data, IoT sensor data, business operation data, spatial geographic data, text and voice data, and online public opinion data.

[0006] S12. Preprocess the multi-source heterogeneous data. Preprocessing includes data cleaning, format conversion, spatiotemporal registration, noise reduction filtering, missing value imputation, and outlier removal.

[0007] S13. A three-level progressive fusion strategy is adopted to fuse the preprocessed data: Data layer fusion: Fusion of multi-source raw data of the same type and the same monitoring target to form a unified structured data view; Feature layer fusion: Risk features, spatiotemporal features, and correlation features are extracted from the data layer fusion results, and a unified feature vector is formed through feature alignment; Decision-level fusion: Make comprehensive decisions based on the feature-level fusion results and output a unified risk assessment basis; S14. Based on the fused data, construct an urban safety knowledge graph and form a network of "people-vehicles-places-events-facilities" through entity recognition and relationship extraction.

[0008] Preferably, the data fusion in step S13 adopts an adaptive weighted fusion strategy: assuming a certain monitoring point has a total of Multiple sensors simultaneously measure the same physical quantity. One sensor in The reading at time is Its fusion output is: ; in, For the first One sensor in The weighting coefficients at each time point are determined by using the reciprocal of the deviation between each sensor reading and the sliding median over the past 60 seconds. Sensors with larger deviations have lower weights, thus adaptively suppressing the impact of sensor drift or instantaneous disturbances on the fusion results. Feature layer fusion uses a cross-modal attention mechanism to align and fuse features from different modalities. It uses a feature from a certain modality as a query vector, calculates the correlation weight between features from other modalities, and then aggregates the multimodal features by weight to form a unified feature vector. The decision-making level integrates conclusions from different discriminative channels using the DS evidence theory: setting an identification framework. Each discrimination channel independently outputs the basic probability assignment function for each proposition in the recognition framework. The joint confidence of each proposition is calculated using the Dempster synthesis rule, and the proposition with the highest confidence is taken as the comprehensive decision result.

[0009] Preferably, step S2 contains the following: S21. Based on the knowledge graph constructed in step S14, construct an event association topology graph, with risk events as nodes and the transmission relationship between events as edges; S22. Use graph neural networks to reason about the event association topology and identify risk transmission paths: Graph attention networks are used to calculate the attention weights between event nodes and identify key risk nodes. An adaptive sampling strategy is used to determine the sampling path and discover implicitly related events; The credibility score of the risk chain is calculated by assessing the transmission strength, and the risk chains are then selected and output. S23. Based on the risk chain identification results, generate a risk transmission map and mark the risk source, transmission path and scope of impact.

[0010] Preferably, the use of graph neural networks to identify risk transmission paths in step S22 specifically involves: Construct an event association topology graph G=(V,E), where V is the set of event nodes, and each node is associated with risk type, occurrence time, spatial location and severity attributes; E is the set of edges, and the edge weights are jointly determined by spatial distance, time interval and historical co-occurrence frequency. Graph attention networks are used for message passing, in the first... Layers, nodes The hidden state is updated in the following way: ; in, For nodes The neighborhood, For the first Layer nodes Passed to the node Attention weights are calculated by considering the nodes simultaneously. and nodes The hidden state of the two and the properties of the edge between them; An adaptive bidirectional sampling strategy is used to determine the propagation path: after each hop of propagation, the information gain ratio of the current propagation path is calculated. ; in, The information entropy of the currently visited set of nodes. The information entropy after adding candidate nodes, The computational cost required to add this node; if the information gain ratio is lower than a preset threshold, propagation stops in this direction; if the information gain ratio continues to be higher than the threshold, the number of sampling steps is automatically increased; forward propagation tracks the consequence events, and backward propagation traces the cause events; The specific method for assessing conduction strength is as follows: For candidate risk chains ,in As the source of the incident, For end-stage risk events, there is a directed transmission relationship between adjacent nodes. The overall credibility score of this risk chain is... for: ; in, The temporal consistency strength represents the occurrence time of each node according to... The degree to which the sequential arrangement and time intervals match the typical propagation delay distribution of similar historical events; Spatial continuity intensity represents the degree to which the spatial locations of each node are continuously distributed along the actually reachable path within a spatial distance threshold; The attribute compatibility strength represents the degree to which the causal relationship between the source event type and the end risk type is supported by the domain knowledge rule base; Only risk chains with a confidence level exceeding the dynamic threshold are output. The dynamic threshold is adaptively adjusted based on historical data of recent false positive and false negative rates: the threshold is raised when the false positive rate increases and lowered when the false negative rate increases.

[0011] Preferably, step S3 contains the following: S31. Based on the fused data from step S1, construct a dynamic risk assessment model from four dimensions: hazard intensity, exposure level, vulnerability, and response capability; the specific calculation method of the dynamic risk assessment model is as follows: The overall risk value R is calculated using the following nonlinear coupling function: ; in, This is the hazard intensity index. For exposure level index, As a vulnerability index, For responsiveness index, For learnable weight parameters in each dimension, It is a natural constant; Danger Intensity Index Exposure level index is calculated based on the degree and rate of deviation of real-time monitoring parameters from the fused data from the normal range. Vulnerability index calculated based on population density and critical infrastructure value in the affected area. Response capability index is calculated based on the service life, maintenance records, and past accident frequency of infrastructure in the target area. Calculated based on the number of available emergency resource types, resource sufficiency, and response time window; S32. Use deep temporal neural networks to dynamically fit the risk development trend and predict the risk spread speed, impact range and duration. The deep temporal neural network is a spatiotemporal attention prediction network, consisting of a spatial feature extraction branch, a temporal feature extraction branch, and a fusion prediction branch. The spatial feature extraction branch takes the spatial adjacency matrix of each risk node in the knowledge graph as input, and uses a graph convolutional network to model the spatial propagation characteristics of risk, capturing the spatial transmission patterns of risk along the pipeline upstream and downstream, and road network radiation. The time feature extraction branch takes the time series of risk indicators of each node as input and uses a long short-term memory network to extract the dynamic pattern of risk evolution over time, and identifies the upward trend, periodic fluctuation and abrupt inflection point of risk indicators. The fusion prediction branch jointly encodes spatial and temporal features, automatically identifies key risk patterns at different spatiotemporal locations through an attention mechanism, and outputs risk prediction results for multiple future time steps.

[0012] S33. Based on a comprehensive assessment of risk type and risk level, risks are classified into four levels: alert, attention, warning, and handling, triggering tiered early warnings. The triggering method for the tiered early warning is as follows: When the comprehensive risk value R is in the first range, a warning level alert is triggered; when it is in the second range, a warning level alert is triggered; when it is in the third range, a warning level alert is triggered; and when it is in the fourth range, a warning level alert is triggered. Different levels of warning correspond to different warning information push frequencies, push scopes, and push channels: the alert level is only pushed to the duty post, the attention level is pushed to the department head, the warning level is pushed to the supervisor in charge and an emergency consultation is initiated, and the response level is pushed to all relevant departments and a full-process emergency response is initiated.

[0013] Preferably, step S4 contains the following: S41. Construct a structured contingency plan library. Each contingency plan consists of multiple atomic operation instructions, and each atomic operation instruction is marked with the triggering preconditions, the required resource status, and the expected execution effect. Atomic operation instructions are represented using a 5-tuple structure: <Operation ID, Operation Type, Executing Entity, Prerequisites Set, Resource Requirements Set, Expected Result Vector> The operation type includes at least one standard action category among evacuation, lockdown, emergency repair, reinforcement, and notification; the executing entity is mapped to a specific department or job role; the precondition set contains a list of conditions that must be met before the operation can be performed; the resource requirement set contains a list of materials, equipment, and personnel required to perform the operation, with each resource labeled with its type, specifications, and quantity; and the expected effect vector is a multi-dimensional numerical vector, representing the expected improvement of the operation on each evaluation dimension and the expected completion time.

[0014] S42. Input the risk level, risk location and key information generated in step S3 into the risk-contingency plan matching engine to generate a recommended decision-making scheme. The matching process of the risk-contingency plan matching engine is as follows: Level 1 – Spatial Filtering: Based on the geographical coordinates of the risk event, select contingency plans from the contingency plan database that cover the applicable jurisdiction of that location; Level 2 – Type Filtering: Based on the risk type label, further filter contingency plans that are suitable for that risk type; Level 3 – Level Screening: Based on the comprehensive risk level, retain contingency plans whose applicable level scope covers the current level; Level 4 – Condition Validation: Check each candidate plan to see if the set of preconditions for its atomic operation instructions is met in the current environment. If the instructions that do not meet the conditions are non-critical instructions, they are marked as skippable. If they are critical instructions, the plan is filtered out. Level 5 – Deduction and Scoring: The deduction engine is invoked to simulate the execution process of the verified plan, the score is calculated through the utility function, and the plan with the highest score is output as the recommended decision solution.

[0015] The utility function for the inferential score is: ; Where S is the probability of successful deduction. To estimate the completion time, The longest acceptable handling time for this type of risk To estimate the value of the protected resources, To estimate resource consumption costs, β are the adjustment weighting coefficients.

[0016] S43. Automatically generate tiered early warning signals and handling instructions based on the recommended decision-making scheme, and push them to relevant handling departments and responsible persons.

[0017] Preferably, step S5 contains the following: S51. Collect actual implementation feedback data of the early warning decision-making plan, including the handling results, response time, resource consumption and handling effect; S52. Based on feedback data, a reinforcement learning model is used to continuously optimize the risk assessment model, risk chain identification model, and contingency plan matching model, iteratively updating the model parameters. The reward function of the reinforcement learning model is: comprehensive reward. Among them: accurate early warning reward If the warning is confirmed, then If it is a false alarm, then Timely response reward If the actual response time Less than the estimated response time but ,like exist Between 1.0 and 1.3 times ,like Greater than 1.3 times Rewards for effective handling If the risk is successfully eliminated and no secondary events occur, then If the risk is controlled but not completely eliminated, then If a secondary event occurs, Resource efficiency reward If the actual resource consumption Less than the estimated resource consumption but ,like exist Between 1.0 and 1.5 times ,like Greater than 1.5 times The reinforcement learning agent uses a proximal policy optimization algorithm to update the policy. The state space includes risk type, risk level, resource status and environmental conditions. The action space consists of discrete actions that select the optimal solution from candidate solutions. The optimization objective is to maximize the expected cumulative reward.

[0018] Preferably, it also includes fusion confidence assessment and downgrade fusion strategy: after each level of fusion is completed, the fusion confidence score for that level is calculated. ; in, The number of data sources or feature channels participating in the fusion. Let be the probability distribution of the contribution of the j-th data source or feature channel to the fusion result. The variance of each contribution probability distribution; when the fusion confidence score of a certain level is lower than the preset confidence threshold, a downgrade fusion strategy is triggered: if the data layer fusion confidence is lower than the threshold, the single source data with the highest signal-to-noise ratio in the preprocessed original data is used as the output; if the feature layer fusion confidence is lower than the threshold, it reverts to the output result of the data layer fusion; if the decision layer fusion confidence is lower than the threshold, it reverts to the voting results of each channel after feature layer fusion. It also includes an adaptive calibration mechanism for warning thresholds: the system maintains a sliding time window, recording the confirmation result of each warning within the window. When the false alarm rate within the window continuously exceeds a preset upper limit, the system automatically raises the warning threshold for that type of risk; when the false negative rate within the window continuously exceeds a preset upper limit, the system automatically lowers the warning threshold. The adjustment range of the threshold is: ; in, This represents the false alarm rate or false negative rate within the current window. The preset target false alarm rate or false negative rate, The preset adjustment step size coefficient is used. When both the false alarm rate and the false alarm rate are higher than the preset upper limit, the threshold is adjusted first according to the false alarm rate to ensure that the sensitivity to real risks takes precedence over the suppression of false alarms.

[0019] A public safety risk early warning system based on multi-source data fusion includes: The information settings module is used to integrate multi-source heterogeneous data, construct an urban safety knowledge graph, and set permissions and data access standards for each node in the system. The risk chain identification module is used to construct an event association topology graph, use graph neural networks to identify risk transmission paths, and generate risk chains and transmission graphs. The dynamic risk assessment module is used to conduct dynamic risk assessment from four dimensions: hazard intensity, exposure level, vulnerability and response capability, and combines deep temporal neural network to predict risk development trends. The tiered early warning module is used to trigger a tiered early warning mechanism based on the risk level. The intelligent contingency plan matching and coordinated response module is used to screen, verify and simulate contingency plans based on the risk-contingency plan matching engine, and output recommended decision-making solutions and response instructions. The feedback optimization module is used to collect actual execution feedback data and continuously optimize system parameters through a reinforcement learning model.

[0020] Therefore, the present invention employs the above-mentioned public safety risk early warning method and system based on multi-source data fusion, which has the following beneficial effects: (1) A three-level progressive fusion strategy of "data layer - feature layer - decision layer" was designed to improve the semantic level of data abstraction layer by layer, so that different types of data such as video, IoT, and text can be organically integrated under the same framework, effectively eliminating the uncertainty and error of a single data source. In particular, a sensor weighting strategy of "weighting by the inverse of the deviation from the moving median" was designed in the data layer to achieve real-time suppression of occasional sensor drift with extremely low computational cost and avoid false alarms.

[0021] (2) Using graph attention network to perform multi-hop reasoning on event nodes in knowledge graph, implicit related events are discovered through adaptive sampling driven by information gain rate, and the credibility of risk chain is ranked by three-dimensional transmission strength evaluation (temporal consistency, spatial continuity, attribute compatibility), so that the system can track the complete evolution path of risk from initial anomaly to chain amplification, which significantly improves the ability to identify and predict cascading risks.

[0022] (3) The present invention calculates the fusion confidence score after each fusion level. When the confidence score is lower than the threshold, the downgrade fusion strategy is automatically triggered, and the system gradually returns to a more reliable fusion level or single source data. This avoids global misjudgment caused by abnormal data sources and enables the system to output usable results even in abnormal situations such as sensor failure and communication interruption.

[0023] (4) This invention uses the actual effect of each early warning response as a feedback signal, and continuously iterates and optimizes the core parameters through a four-dimensional reward function (early warning accuracy, timely response, response effect, and resource efficiency) and a near-end strategy optimization algorithm. At the same time, an adaptive calibration mechanism for the early warning threshold based on real-time monitoring of false alarm rate / false negative rate is designed, so that the system can automatically adjust its sensitivity in different seasons, different time periods, and different scenarios to avoid performance degradation after deployment.

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of an embodiment of a public safety risk early warning method based on multi-source data fusion according to the present invention; Figure 2 This is a flowchart of the risk chain identification method based on graph neural networks according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an embodiment of a public safety risk early warning system based on multi-source data fusion according to the present invention. Detailed Implementation

[0026] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0028] Example Please see Figures 1-2 This invention provides a public safety risk early warning method based on multi-source data fusion, comprising: Step 1: Acquisition and fusion processing of multi-source heterogeneous data.

[0029] In the target city, a densely populated area of ​​gas pipelines was selected as a monitoring demonstration zone, where various sensing devices were deployed and data was accessed.

[0030] The types of data accessed include: (1) IoT sensing data such as methane concentration sensors (including electrochemical and infrared types), pipeline pressure sensors, and flow monitoring instruments installed along the gas pipeline, with a collection frequency of once every 5 seconds; (2) video surveillance images of roads in the area, which are extracted and analyzed every 30 seconds after frame extraction processing; (3) operational status data of water supply networks, power networks, and communication networks in the area; (4) real-time meteorological data provided by the meteorological department, including temperature, humidity, wind force and direction, and precipitation information; (5) operation and maintenance work order records of road occupation and road excavation in the area recently; and (6) population thermal distribution data in the area.

[0031] In the preprocessing stage, the above data is spatiotemporally aligned, mapping all data to a unified geographic coordinate system and time base (accurate to the second). To address the potential for sporadic drift issues from methane concentration sensors, sliding median filtering is used for noise reduction. For video data, inter-frame differencing is employed to extract dynamic foreground targets. For unstructured text in maintenance work orders, named entity recognition is used to extract key fields such as construction location, construction type, and construction time.

[0032] Data layer fusion: This area is equipped with multiple methane concentration monitoring points. Assume a certain monitoring point has n sensors simultaneously measuring methane concentration. One sensor in The reading at time is The fusion output is a weighted sum of the readings from each sensor. The weighting coefficients are determined by the reciprocal of the deviation between each sensor's reading and the sliding median over the past 60 seconds; sensors with larger deviations have lower weights. Taking a gas pressure regulating station as an example, multiple concentration detectors from different manufacturers are deployed within the station. Under normal operating conditions, the readings from each instrument are essentially consistent. One day, one instrument experiences baseline drift due to changes in ambient temperature, resulting in a higher reading than the other instruments. In this situation, the adaptive weighting mechanism automatically and significantly lowers the weighting coefficient of this instrument, and the fusion output is primarily determined by the other instruments, effectively avoiding false alarms caused by equipment drift.

[0033] Feature layer fusion: Statistical features such as rate of change, fluctuation amplitude, and duration of continuous exceedance are extracted from the fused concentration time-series data; semantic features such as the number of construction machines and the extent of enclosures in the area are extracted from video analysis results; and diffusion condition features such as wind direction, wind speed, and atmospheric stability are extracted from meteorological data. These heterogeneous features are aligned and fused through a cross-modal attention mechanism—using the features of one modality as a reference, the correlation weights between features of other modalities are calculated, and then the multi-modal features are weighted and aggregated to form a unified feature vector. Taking a gas leak scenario as an example, the feature of "construction machinery approaching the pipe gallery" detected by the video modality is strongly correlated with the "pressure fluctuation" feature of the sensor modality and the "construction application" feature of the text modality. After attention-weighted aggregation, the semantic components related to construction activities in the unified feature vector are significantly enhanced.

[0034] Decision-making fusion: Feature vectors are input into multiple discrimination channels—channel one judges the likelihood of leakage based on gas concentration trends, channel two judges the risk of external interference based on construction activities and pipeline spatial location, and channel three judges the scope of impact based on meteorological diffusion conditions. A recognition framework is then established. These represent "high risk," "medium risk," and "low risk," respectively. Each channel independently outputs the basic probability allocation for the three propositions, and the joint confidence level is calculated using Dempster's composition rule. If the confidence level of a certain risk level after composition is significantly higher than the output of any single channel, that level is taken as the overall decision result; if the conclusions of a channel seriously conflict with those of other channels, the composition rule automatically reduces the confidence weight of that channel.

[0035] The merged data is synchronized and used to construct a regional safety knowledge graph. Entities in the graph include gas pipelines (including material, diameter, and service life attributes), monitoring sensors, surrounding buildings, construction activities, and historical leak event records. Relationships between entities include types such as "monitoring," "proximity," "impact," and "occurred."

[0036] Step 2: Risk chain identification based on knowledge graph.

[0037] During operation, the knowledge graph receives and updates the node status in real time from the fused data stream. When the methane concentration at a certain pipeline node continues to rise to the warning threshold, the system automatically initiates the risk chain identification process.

[0038] First, starting with the pipe segment node in the knowledge graph, we extract the associated entities and relationships within its first-order neighborhood, including the upstream and downstream pipe segments directly connected to it, all construction records in the area within the last 7 days, and the maintenance history of the pipe segment.

[0039] Then, the graph attention network performs multi-hop reasoning on the starting node. During the reasoning process, the attention weight calculation comprehensively considers the spatial distance, temporal proximity, and historical co-occurrence frequency between entities. Taking a construction activity entity as an example, if a construction location is spatially close to an abnormal pipe section and the construction time overlaps with the time when the abnormality occurred, the attention weight between the construction entity and the abnormal pipe section is significantly increased.

[0040] During propagation, the system performs adaptive bidirectional sampling: forward propagation starts from the starting node and propagates in the positive time direction, tracking "what consequences the current anomaly might have"; backward propagation starts from the starting node and propagates in the negative time direction, tracing "what caused the current anomaly". After each hop, the information gain ratio of the current propagation path is calculated—that is, the ratio of the increase in path information entropy after adding a candidate node to the computational cost required to add that node. If the information gain ratio is lower than a preset threshold, propagation stops; if it remains higher than the threshold, the number of sampling steps is automatically increased.

[0041] Through multi-hop reasoning, the system identified the following implicit correlation path: the ongoing underground pipe jacking construction in the area caused stress redistribution in the surrounding soil, which in turn created uneven settlement forces on the adjacent gas pipeline, leading to a decrease in the sealing performance at the pipeline joints and thus causing minor leaks. In this path, the construction activity is the initial trigger, soil settlement is the intermediate transmission link, pipeline joint sealing failure is the direct cause, and the increase in methane concentration is the final manifestation.

[0042] The system calculates the credibility score of the risk chain through transmission strength assessment, comprehensively considering three dimensions: temporal consistency (whether the occurrence times of each node are arranged sequentially and the time intervals match the historical transmission delay distribution), spatial continuity (whether the spatial locations of each node are continuously distributed along the actually reachable path), and attribute compatibility (whether the causal relationship between the source event and the terminal risk is supported by the domain knowledge rule base). If the final score exceeds a preset screening threshold, it is confirmed as a valid risk chain. The system outputs a risk transmission map, clearly marking the risk source, transmission mechanism, current risk status, and predicted impact range.

[0043] Step 3: Dynamic risk assessment and tiered early warning.

[0044] Based on the fused data from step one and the risk chain identification results from step two, a four-dimensional dynamic risk assessment is conducted: Hazard intensity dimension: The current leakage rate is in the trace leakage range; the methane concentration around the leak point has not yet reached 5% of the lower explosive limit, which is the initial stage; the concentration is showing a slow accelerating trend. The hazard intensity index is calculated based on the above indicators.

[0045] Exposure level dimension: Based on population thermal distribution data, the area surrounding the leak point includes residential buildings and commercial premises; power cables are laid in parallel below the leak point. The exposure level index is calculated by combining population density and secondary sensitive targets.

[0046] Vulnerability Dimension: This section of pipeline has a long service life and is made of cast iron. Historically, this type of pipeline has a higher leakage frequency than the regional average under similar geological conditions. The vulnerability index is calculated by considering the degree of material aging, maintenance history, and geological adaptability.

[0047] Response capability dimension: Within a certain range, dispatchable emergency resources include gas company repair teams, fire and rescue stations, etc., with sufficient spare pipe fittings inventory; the area has good road conditions. The response capability index is calculated by combining resource distance and road conditions.

[0048] The comprehensive risk value is calculated using a nonlinear coupling function—the four-dimensional index is input into the Logistic function, and the weight parameters of each dimension are learnable parameters, allowing the evaluation model to continuously optimize the importance of each dimension based on subsequent feedback data. The four-dimensional index is simultaneously input into a spatiotemporal attention prediction network, which consists of a spatial feature extraction branch (using a graph convolutional network to capture the upstream and downstream propagation patterns of risk along the pipeline network), a temporal feature extraction branch (using a long short-term memory network to extract the upward trend, periodic fluctuations, and abrupt inflection points of risk indicators), and a fusion prediction branch, outputting a comprehensive risk evolution trend prediction for the next 30 minutes. The model prediction results show that, without intervention, the leakage rate will increase, and the risk level will rise by one level from the current level. Based on the current risk value and trend prediction, the system determines the comprehensive risk level to be "alert," triggering a level-two warning.

[0049] Step 4: Intelligent contingency plan matching and coordinated response.

[0050] The system inputs the risk level, risk type, leak location, information on affected facilities, and status of surrounding resources into the risk-contingency plan matching engine.

[0051] The structured contingency plan library stores various emergency plans for the area, each of which has been pre-decomposed into a sequence of atomic operation instructions. Each atomic operation instruction is represented by a five-tuple structure: including operation type (evacuation, lockdown, emergency repair, reinforcement, notification, etc.), executing entity (mapped to a specific department or job role), set of preconditions (a list of conditions that must be met before executing the operation), set of resource requirements (a list of required materials, equipment, and personnel, each labeled with type, specifications, and quantity), and expected effect vector (representing the expected improvement of the operation on each evaluation dimension and the expected completion time).

[0052] The engine executes a five-level matching process: Level 1 – Spatial Filtering: Based on the geographical location of the risk event, select contingency plans from the contingency plan database that cover the applicable jurisdiction of that location.

[0053] Level 2 – Type Filtering: Based on the risk type tags, further filter contingency plans that are suitable for the "gas leak" type.

[0054] Level 3 – Level Screening: Based on the comprehensive risk level “Alert”, retain contingency plans whose applicable level range covers “Alert”.

[0055] Level 4 – Condition Validation: Each candidate plan's atomic operation instruction's precondition set is checked against the current environment. Instructions that do not meet the conditions are marked as skippable if they are non-critical; otherwise, the plan is marked as insufficiently conditional and filtered out.

[0056] Level 5 – Simulation Scoring: The simulation engine is used to execute the validated contingency plans. The engine incorporates time-consuming distribution models (based on historical statistics), resource consumption models, and impact range estimation models for various operations. The simulation results are scored using a utility function – considering factors such as the probability of success, the ratio of estimated completion time to the longest acceptable handling time, and the ratio of estimated protection value to estimated resource consumption. When the probability of success is lower than a preset credibility threshold, the contingency plan is marked as unrecommended and filtered. The engine outputs the highest-scoring contingency plan as the recommended decision-making solution, simultaneously generating a sequence of handling instructions, including measures such as dispatching repair teams, notifying surrounding residents, mobilizing firefighters on standby, and traffic control.

[0057] The aforementioned instructions were pushed to the gas company's dispatch center, the local street office, the fire station's duty station, and the traffic police command center via the system platform, and were also pushed to the handheld devices of the responsible persons via mobile terminals.

[0058] Step 5: Feedback learning and continuous model optimization.

[0059] After the incident is handled, the system collects complete execution feedback data, including information such as actual arrival time, actual handling time, actual resource consumption, and handling results.

[0060] The reinforcement learning model uses this data as a training sample and calculates reward signals for each dimension: if the warning is confirmed, the reward for accurate warning is positive, otherwise it is negative; if the actual response time is better than the prediction, the reward for timely response is positive, exceeding a certain proportion results in a negative reward; if the risk is successfully eliminated without secondary events, the reward for effective handling is positive, and secondary events result in a negative reward; if the actual resource consumption is controlled within a reasonable range, the reward for resource efficiency is positive, exceeding a certain proportion results in a negative reward. The comprehensive reward is obtained by summing up the above dimensions.

[0061] The reinforcement learning agent employs a proximal policy optimization algorithm for policy updates. The state space includes features such as risk type, risk level, resource status, and environmental conditions, while the action space consists of discrete actions selected from candidate plans. The model synchronously updates multiple parameters based on comprehensive feedback signals: the transmission probability weights between construction activity factors and gas leaks in the area are appropriately adjusted; the estimated travel time for this road segment is corrected based on actual conditions; and the baseline maintenance time for this type of pipe fitting is adjusted.

[0062] Meanwhile, the adaptive calibration mechanism for warning thresholds continues to operate. The system maintains a sliding time window, recording the confirmation result of each warning within the window. When the false alarm rate continuously exceeds the preset upper limit within the window, the system automatically raises the warning threshold for that type of risk; when the missed alarm rate continuously exceeds the preset upper limit within the window, the system automatically lowers the warning threshold. The adjustment range is determined by the degree to which the false alarm rate / missed alarm rate deviates from the target value, ensuring smooth threshold adjustment and avoiding large fluctuations caused by a single anomaly.

[0063] After multiple rounds of training with positive and negative samples, the model gradually converges to the true distribution of the conditional probability distribution of gas leak risks caused by different seasons, time periods, and construction types in the region. The accuracy of early warning is significantly improved, and the false alarm rate and missed alarm rate are effectively controlled.

[0064] like Figure 3 This embodiment provides a public safety risk early warning system based on multi-source data fusion. The module structure of the system and its operation in the security scenario of large-scale events are as follows: Information Setup Module: Responsible for configuring multi-source heterogeneous data access, orchestrating the fusion processing workflow, and initializing and building the knowledge graph. During large-scale events, this module additionally accesses security check data, ticketing data, entrance gate data, and traffic control data around the event venue. It overlays event-specific subgraphs on top of a general city knowledge graph, providing a customized data foundation for security risk assessment during the event.

[0065] Risk Chain Identification Module: Continuously monitors anomalous event nodes in the knowledge graph during the event. When a prolonged period of congestion occurs at a venue's security checkpoint, this module uses graph reasoning to identify the following correlation: congestion → decreased entry efficiency → continuous accumulation of people queuing outside the venue → crowd density in the queuing area approaching a threshold → potential stampede risk. This risk chain, from the initial anomaly to the final risk, is fully identified in a short time, providing ample time for response.

[0066] Dynamic Risk Assessment Module: Receives real-time integrated sensing data from inside and outside the venue, dynamically assessing the risk level from four dimensions: hazard intensity (number of people remaining and growth rate), exposure level (area and density of people in queuing areas), vulnerability (width and number of evacuation routes in the area), and response capability (distribution of on-site security forces). The assessment results are updated in real-time according to fluctuations in the flow of people entering the venue; when the growth rate of the flow suddenly increases, the risk level is automatically adjusted upwards within 15 seconds.

[0067] Tiered early warning module: Converts dynamic assessment results into tiered early warning signals. In the event scenario, the early warning signals are simultaneously pushed to the large screen display system of the on-site security command center, the smart bracelets worn by the security team leaders, and the venue broadcast system, achieving multi-channel and multi-level coverage of early warning information.

[0068] The intelligent contingency plan matching and coordinated response module: Multiple contingency plans, including venue emergency evacuation plans, anti-stampede plans, and medical emergency plans, were structurally entered before the event. When a "alert" level warning is triggered, this module automatically performs a five-level filtering process: spatial filtering to lock onto venue area plans, type filtering to match evacuation plans, level filtering to adapt to the alert level, condition verification to confirm all evacuation routes are unobstructed and security personnel are sufficient, and a simulation scoring comparison of different plans to output the optimal solution—opening two backup exits on the west side of the venue, setting up segmented barriers in the queuing area, and deploying 6 additional security personnel to the front of the queuing area for crowd control. After the plan was pushed out, on-site security personnel were deployed within 4 minutes, and the queuing crowd was orderly diverted, with the density returning to a safe range.

[0069] Feedback Optimization Module: After the event, the system comprehensively analyzed all early warning and response records during the large-scale event. A total of 14 early warnings were triggered, of which 12 were valid and 2 were false alarms, with an average response time of 7.3 minutes. The reinforcement learning model incorporated the above data into its training, specifically calibrating the risk assessment threshold for dense crowds in the event scenario, resulting in a reduction of the false alarm rate for subsequent similar events by approximately 40%.

[0070] Fusion confidence assessment mechanism: After each level of fusion is completed, the system calculates the fusion confidence score for that level to measure the reliability of the current fusion result. The confidence score is calculated based on the consistency of the contributions of each data source or feature channel participating in the fusion to the fusion result: if the contribution directions of each source are highly consistent, the confidence score is high; if there is a large discrepancy between the sources, the confidence score is low.

[0071] Taking decision-level fusion as an example, when the confidence levels of the three discriminative channels for the same proposition are 0.7, 0.6, and 0.65, respectively, they are highly consistent, with small variances in their respective contribution probability distributions, resulting in high fusion confidence. However, when the confidence levels of the three channels are 0.8, 0.3, and 0.25, significant discrepancies exist, with large variances in their respective contribution probability distributions, leading to low fusion confidence. Low confidence often indicates problems with the current data quality or that the risk signals themselves are not clear enough.

[0072] Degradation and fusion strategy: When the fusion confidence score of a certain level falls below the preset confidence threshold, the system triggers a downgraded fusion strategy: If the confidence level of data fusion is lower than the threshold, the single source data with the highest signal-to-noise ratio in the preprocessed but unfused original data is directly used as the estimated value for that location. If the confidence score of the feature layer fusion is lower than the threshold, then we revert to the output of the data layer fusion. If the confidence level of the decision layer fusion is lower than the threshold, then it reverts to the voting results of each channel after feature layer fusion (the conclusion of the majority channel shall prevail).

[0073] During a specific operation, multiple methane concentration sensors in the area were simultaneously subjected to electromagnetic interference from a nearby radio tower, causing uncorrelated random fluctuations in sensor readings. At this point, the confidence score for data fusion plummeted from 0.92 during normal operation to 0.31, falling below the preset threshold of 0.50. The system automatically triggered a degradation strategy, abandoning the weighted fusion result and instead using the reading from a reference sensor that had undergone electromagnetic shielding as the concentration estimate for that location. After the interference subsided, the sensor readings returned to normal, the fusion confidence score rose above the threshold, and the system automatically resumed the complete fusion process.

[0074] This mechanism enables the system to still output usable results when faced with abnormal situations such as sensor failure, communication interruption, and environmental interference, significantly improving the system's fault tolerance and operational stability.

[0075] The operation process of the adaptive calibration mechanism for early warning thresholds: Most existing systems use fixed thresholds for early warning triggering, which can easily lead to false alarms or missed alarms when the actual operating environment changes. This invention designs an adaptive calibration mechanism for early warning thresholds, which uses joint modeling based on historical early warning records and real-time feedback data to dynamically adjust the early warning trigger threshold.

[0076] The system maintains a sliding time window (e.g., the last 7 days) and records the confirmation result of each alert within the window (whether it is a real risk). Let the false alarm rate in the current window be Err_fp, and the target false alarm rate be Err_target_fp (e.g., 5%); the false negative rate in the current window be Err_fn, and the target false negative rate be Err_target_fn (e.g., 3%).

[0077] The threshold adjustment logic is as follows: When the false alarm rate within the window continuously exceeds the preset upper limit (e.g., exceeding the target value by 20% for three consecutive days), the system automatically raises the warning threshold for this type of risk. The increase is calculated by multiplying the degree to which the false alarm rate deviates from the target value by a preset adjustment step size coefficient. When the missed alarm rate within the window continuously exceeds the preset upper limit, the system automatically lowers the warning threshold, with the same reduction amount.

[0078] When both the false alarm rate and the false alarm rate exceed the preset upper limit, the threshold is adjusted first based on the false alarm rate to ensure that sensitivity to real risks takes precedence over the suppression of false alarms. This is based on the business logic in the public safety field that "the consequences of false alarms are usually more serious than those of false alarms."

[0079] In actual operation, this adaptive calibration mechanism enables the system to maintain stable early warning performance across different seasons, time periods, and scenarios. For example, during the high-temperature period in summer, the baseline readings of some sensors tend to rise overall, which could frequently trigger false alarms if a fixed threshold is used. The adaptive calibration mechanism automatically adjusts the threshold to a level suitable for high-temperature environments based on the recent upward trend of the false alarm rate, and then adjusts the threshold back down when the temperature drops, thus maintaining reasonable early warning sensitivity under different climatic conditions throughout the year. After three months of continuous operation, the system's false alarm rate decreased from 14% at the beginning of deployment and stabilized at around 5%, while the missed alarm rate remained below 3%.

[0080] Therefore, this invention adopts the above-mentioned public safety risk early warning method and system based on multi-source data fusion. The three-level progressive fusion strategy effectively integrates heterogeneous data. The risk chain identification driven by graph neural network improves risk perception from isolated event detection to transmission path tracking. The four-dimensional dynamic assessment coupled with time series prediction realizes the leap from static classification to real-time profiling. The structured plan and five-level matching engine bridge the gap between early warning and disposal. The confidence assessment and downgrade strategy enhance the system's robustness. Reinforcement learning and threshold adaptive calibration endow the system with continuous self-optimization capabilities. Blockchain evidence storage ensures the credible traceability of the disposal process.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A public security risk early warning method based on multi-source data fusion, characterized in that, Includes the following steps: S1. Access and preprocess multi-source heterogeneous data, and use a three-level progressive fusion strategy of data layer fusion, feature layer fusion and decision layer fusion to fuse the data, and build an urban safety knowledge graph based on the fused data; S2. Construct an event association topology graph, use a graph neural network to identify risk transmission paths, and filter and output risk chains through transmission strength assessment; S3. Construct a dynamic risk assessment model from four dimensions: hazard intensity, exposure level, vulnerability, and response capability. Combine this with a deep temporal neural network to predict risk development trends and determine the triggering of graded early warnings based on risk type and risk level. S4. Construct a structured contingency plan library, and use a risk-contingency plan matching engine to screen, verify, and simulate contingency plans, and output recommended decision-making solutions and graded early warning signals; S5. Collect actual execution feedback data and continuously optimize the risk assessment model, risk chain identification model, and contingency plan matching model through reinforcement learning models. 2.The public security risk early warning method based on multi-source data fusion according to claim 1, characterized in that, Step S1 includes the following: S11. Multi-source heterogeneous data includes at least video surveillance data, IoT sensor data, business operation data, spatial geographic data, text and voice data, and online public opinion data; S12. Preprocessing includes data cleaning, format conversion, spatiotemporal registration, noise reduction filtering, missing value imputation, and outlier removal. S13. A three-level progressive fusion strategy is adopted to fuse the preprocessed data: Data layer fusion integrates raw data from multiple sources of the same type and the same monitoring target to form a unified structured data view; Feature layer fusion extracts risk features, spatiotemporal features, and correlation features from the data layer fusion results, and forms a unified feature vector through feature alignment; The decision-making level integration process comprehensively decides on the feature-level integration results and outputs a unified risk assessment basis. S14. Based on the fused data, construct an urban safety knowledge graph and form a network of relationships between people, vehicles, places, events, and facilities through entity recognition and relationship extraction.

3. The public safety risk early warning method based on multi-source data fusion according to claim 2, characterized in that: In step S13, the data layer fusion adopts an adaptive weighted fusion strategy: assuming a certain monitoring point has a total of Multiple sensors simultaneously measure the same physical quantity. One sensor in The reading at time is Its fusion output is: ； in, For the first One sensor in Weighting coefficients at different times; Feature layer fusion uses a cross-modal attention mechanism to align and fuse features from different modalities. It uses a feature from a certain modality as a query vector, calculates the correlation weight between features from other modalities, and then aggregates the multimodal features by weight to form a unified feature vector. The decision-making level integrates conclusions from different discriminative channels using the DS evidence theory: setting an identification framework. Each discrimination channel independently outputs the basic probability assignment function for each proposition in the recognition framework. The joint confidence of each proposition is calculated using the Dempster synthesis rule, and the proposition with the highest confidence is taken as the comprehensive decision result.

4. A public safety risk early warning method based on multi-source data fusion according to claim 2, characterized in that, Step S2 is as follows: S21. Based on the knowledge graph constructed in step S14, construct an event association topology graph, with risk events as nodes and the transmission relationship between events as edges; S22. Use graph neural networks to reason about the event association topology and identify risk transmission paths: Graph attention networks are used to calculate the attention weights between event nodes and identify key risk nodes. An adaptive sampling strategy is used to determine the sampling path and discover implicitly related events; The credibility score of the risk chain is calculated by assessing the transmission strength, and the risk chains are then selected and output. S23. Based on the risk chain identification results, generate a risk transmission map and mark the risk source, transmission path and scope of impact.

5. A public safety risk early warning method based on multi-source data fusion according to claim 4, characterized in that, Step S22, which uses a graph neural network to identify the risk transmission path, specifically involves: Construct an event association topology graph G=(V,E), where V is the set of event nodes, and each node is associated with risk type, occurrence time, spatial location and severity attributes; E is the set of edges, and the edge weights are jointly determined by spatial distance, time interval and historical co-occurrence frequency. Graph attention networks are used for message passing, in the first... Layers, nodes The hidden state is updated in the following way: ； in, For nodes The neighborhood, For the first Layer nodes Passed to the node Attention weights are calculated by considering the nodes simultaneously. and nodes The hidden state of the two and the properties of the edge between them; An adaptive bidirectional sampling strategy is used to determine the propagation path: after each hop of propagation, the information gain ratio of the current propagation path is calculated. ； in, The information entropy of the currently visited set of nodes. The information entropy after adding candidate nodes, The computational cost required to add this node; if the information gain ratio is lower than a preset threshold, propagation stops in this direction; if the information gain ratio continues to be higher than the threshold, the number of sampling steps is automatically increased; forward propagation tracks the consequence events, and backward propagation traces the cause events; The specific method for assessing conduction strength is as follows: For candidate risk chains ,in As the source of the incident, For end-stage risk events, there is a directed transmission relationship between adjacent nodes. The overall credibility score of this risk chain is... for: ； in, The temporal consistency strength represents the occurrence time of each node according to... The degree to which the sequential arrangement and time intervals match the typical propagation delay distribution of similar historical events; Spatial continuity intensity represents the degree to which the spatial locations of each node are continuously distributed along the actually reachable path within a spatial distance threshold; The attribute compatibility strength represents the degree to which the causal relationship between the source event type and the end risk type is supported by the domain knowledge rule base.

6. A public safety risk early warning method based on multi-source data fusion according to claim 5, characterized in that, Step S3 is as follows: S31. Based on the fused data from step S1, construct a dynamic risk assessment model from four dimensions: hazard intensity, exposure level, vulnerability, and response capability; the specific calculation method of the dynamic risk assessment model is as follows: The overall risk value R is calculated using the following nonlinear coupling function: ； in, This is the hazard intensity index. For exposure level index, As a vulnerability index, For responsiveness index, For learnable weight parameters in each dimension, It is a natural constant; S32. Use deep temporal neural networks to dynamically fit the risk development trend and predict the risk spread speed, impact range and duration. The deep temporal neural network is a spatiotemporal attention prediction network, consisting of a spatial feature extraction branch, a temporal feature extraction branch, and a fusion prediction branch. The spatial feature extraction branch takes the spatial adjacency matrix of each risk node in the knowledge graph as input, and uses a graph convolutional network to model the spatial propagation characteristics of risk, capturing the spatial transmission patterns of risk along the pipeline upstream and downstream, and road network radiation. The time feature extraction branch takes the time series of risk indicators of each node as input and uses a long short-term memory network to extract the dynamic pattern of risk evolution over time, and identifies the upward trend, periodic fluctuation and abrupt inflection point of risk indicators. The fusion prediction branch jointly encodes spatial and temporal features, automatically identifies key risk patterns at different spatiotemporal locations through an attention mechanism, and outputs risk prediction results for multiple future time steps. S33. Based on a comprehensive assessment of risk type and risk level, risks are classified into four levels: alert, attention, warning, and handling, triggering tiered early warnings. The triggering method for the tiered early warning is as follows: When the comprehensive risk value R is in the first range, a warning level alert is triggered; when it is in the second range, a warning level alert is triggered; when it is in the third range, a warning level alert is triggered; and when it is in the fourth range, a warning level alert is triggered. Different levels of warning correspond to different warning information push frequencies, push scopes, and push channels: the alert level is only pushed to the duty post, the attention level is pushed to the department head, the warning level is pushed to the supervisor in charge and an emergency consultation is initiated, and the response level is pushed to all relevant departments and a full-process emergency response is initiated.

7. A public safety risk early warning method based on multi-source data fusion according to claim 1, characterized in that, Step S4 is as follows: S41. Construct a structured contingency plan library. Each contingency plan consists of multiple atomic operation instructions, and each atomic operation instruction is marked with the triggering preconditions, the required resource status, and the expected execution effect. S42. Input the risk level, risk location and key information generated in step S3 into the risk-contingency plan matching engine to generate a recommended decision-making scheme. S43. Automatically generate tiered early warning signals and handling instructions based on the recommended decision-making scheme, and push them to relevant handling departments and responsible persons.

8. A public safety risk early warning method based on multi-source data fusion according to claim 1, characterized in that, Step S5 is as follows: S51. Collect actual implementation feedback data of the early warning decision-making plan, including the handling results, response time, resource consumption and handling effect; S52. Based on feedback data, a reinforcement learning model is used to continuously optimize the risk assessment model, risk chain identification model, and contingency plan matching model, and iteratively update the model parameters.

9. A public safety risk early warning method based on multi-source data fusion according to claim 1, characterized in that: It also includes fusion confidence assessment and downgrade fusion strategy: After each level of fusion is completed, the fusion confidence score for that level is calculated. ； in, The number of data sources or feature channels participating in the fusion. Let be the probability distribution of the contribution of the j-th data source or feature channel to the fusion result. The variance of each contribution probability distribution; When the fusion confidence score of a certain level is lower than the preset confidence threshold, a downgraded fusion strategy is triggered: if the data layer fusion confidence score is lower than the threshold, the single source data with the highest signal-to-noise ratio in the preprocessed original data is used as the output; if the feature layer fusion confidence score is lower than the threshold, it reverts to the output result of the data layer fusion; if the decision layer fusion confidence score is lower than the threshold, it reverts to the voting results of each channel after feature layer fusion. It also includes an adaptive calibration mechanism for warning thresholds: maintaining a sliding time window, recording the confirmation result of each warning within the window; when the false alarm rate within the window continuously exceeds a preset upper limit, automatically raising the warning threshold for that type of risk; when the false negative rate within the window continuously exceeds a preset upper limit, automatically lowering the warning threshold; the adjustment range of the threshold is: ； in, This represents the false alarm rate or false negative rate within the current window. The preset target false alarm rate or false negative rate, The preset adjustment step size coefficient is used. When both the false alarm rate and the false alarm rate are higher than the preset upper limit, the threshold is adjusted first according to the false alarm rate to ensure that the sensitivity to real risks takes precedence over the suppression of false alarms.

10. A public safety risk early warning system based on multi-source data fusion, comprising a public safety risk early warning method based on multi-source data fusion according to any one of claims 1-9, characterized in that, include: The information settings module is used to integrate multi-source heterogeneous data, construct an urban safety knowledge graph, and set permissions and data access standards for each node in the system. The risk chain identification module is used to construct an event association topology graph, use graph neural networks to identify risk transmission paths, and generate risk chains and transmission graphs. The dynamic risk assessment module is used to conduct dynamic risk assessment from four dimensions: hazard intensity, exposure level, vulnerability and response capability, and combines deep temporal neural network to predict risk development trends. The tiered early warning module is used to trigger a tiered early warning mechanism based on the risk level. The intelligent contingency plan matching and coordinated response module is used to screen, verify and simulate contingency plans based on the risk-contingency plan matching engine, and output recommended decision-making solutions and response instructions. The feedback optimization module is used to collect actual execution feedback data and continuously optimize system parameters through a reinforcement learning model.

Images