Urban public security risk chain identification method and system based on graph neural network

Abstract

The application provides a kind of city public security risk chain identification method and system based on graph neural network, it is related to public security risk management technical field, including by constructing event correlation topological graph, using adaptive bidirectional jump sampling algorithm to determine optimal sampling path, identify logical break point and search potential transition event to build complete risk chain, finally, based on the conduction intensity characteristic calculation credibility score screening risk chain.The application can automatically find the implicit association in risk chain, improve the risk warning accuracy, and provide scientific basis for city public security management decision.

Description

Technical Field

[0001] This invention relates to the field of public safety risk management technology, and in particular to a method and system for identifying urban public safety risk chains based on graph neural networks. Background Technology

[0002] With the acceleration of urbanization and the increase in urban population density, urban public safety incidents occur frequently, posing a serious threat to social management and the safety of people's lives and property. Urban public safety incidents often do not occur in isolation, but rather are interconnected and propagate in a chain of risks, forming a complex risk network. A risk chain refers to a series of interconnected risk events, where the occurrence of one event triggers or exacerbates the risk of subsequent events. Accurately identifying these risk chains is of great significance for preventing and mitigating the harm caused by public safety incidents.

[0003] Traditional methods for identifying urban public safety risks primarily rely on expert experience and statistical analysis, employing techniques such as event frequency statistics, correlation analysis, and case studies for risk assessment. In recent years, with the development of data science and artificial intelligence technologies, some studies have begun to explore the use of data mining and machine learning methods to analyze public safety incident data, aiming to uncover correlation patterns and risk transmission pathways.

[0004] However, existing methods for identifying urban public safety risk chains have the following defects and shortcomings: First, they lack the ability to comprehensively model the complex relationships between risk events, making it difficult to effectively capture the multi-dimensional correlation characteristics and non-linear transmission relationships between events, resulting in inaccurate and incomplete risk chains. Second, existing risk chain identification methods perform poorly when facing data sparsity issues, making it difficult to identify hidden risk chains that are not obvious in historical data but have high actual risks, especially for risk paths that exist logically but are insufficiently represented in the data, making it impossible to effectively complete and reason about them. Third, existing methods lack a scientific evaluation system for the credibility of risk chains, making it difficult to quantify the reliability of the identification results, and failing to provide a reliable basis for risk prioritization in urban public safety management decisions, thus limiting the accuracy and effectiveness of risk prevention and control measures. Summary of the Invention

[0005] This invention provides a method and system for identifying urban public safety risk chains based on graph neural networks, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for identifying urban public safety risk chains based on graph neural networks, comprising:

[0007] Acquire historical data on urban public safety incidents, use different types of public safety incidents as graph nodes, and use the relationships between public safety incidents as edges to construct an initial event association topology graph;

[0008] Based on the initial event association topology, an adaptive bidirectional skip sampling algorithm is adopted to determine the optimal sampling path set through dynamic step size control and time sequence matching; based on the event sequences in the optimal sampling path set, multidimensional association strength is calculated, and an initial candidate risk chain set is obtained by screening.

[0009] Based on the initial candidate risk chain set, a state feature matrix is ​​constructed through eigenvalue decomposition to identify logical breakpoints, search for potential transition events, and iteratively optimize to obtain a logically complete risk chain.

[0010] Based on the logically complete risk chain, the transmission strength features are extracted using transmission coefficient matrix analysis and state transition matrix decomposition, and the credibility score is calculated to obtain a risk chain credibility score set.

[0011] Based on the risk chain credibility score set, risk chains with credibility scores exceeding a preset credibility threshold are selected, and the final identification result is determined and output.

[0012] In an optional embodiment, based on the initial event association topology graph, an adaptive bidirectional skip sampling algorithm is used to determine the optimal sampling path set through dynamic step size control and timing matching, including:

[0013] The impact parameters of public safety event nodes are obtained from the initial event association topology map. The propagation impact is calculated based on the node association degree and distance attenuation factor. The spatial impact is calculated based on the ratio of the affected area of ​​the public safety event node. The temporal impact is calculated based on the event time attenuation function. The propagation impact, spatial impact and temporal impact are weighted and combined to obtain the node influence measure.

[0014] The initial adaptive step size is obtained by calculating the node influence metric and the baseline sampling step size. The adaptive control threshold is obtained based on the historical differences of the node influence metric. The dynamic sampling step size is obtained by combining the initial adaptive step size and the adaptive control threshold.

[0015] Bidirectional sampling is performed using a dynamic sampling step size to calculate the forward propagation probability and the backward verification probability, and then the time-series matching score is obtained by weighted fusion.

[0016] The path continuity index and length constraint index are calculated based on the time sequence matching score. The time sequence matching score is nonlinearly weighted with the path continuity index and length constraint index to obtain the path evaluation score. The optimal sampling path set is obtained by screening public safety event sequences whose path evaluation scores are greater than the preset evaluation threshold.

[0017] In one optional embodiment, the multidimensional association strength is calculated based on the event sequences in the optimal sampling path set, and the initial candidate risk chain set is obtained by screening, including:

[0018] The event sequence is obtained from the optimal sampling path set to construct a time-series evolution map, extract the time interval and event density period of adjacent events, identify key time nodes, calculate the time delay feature of event propagation by comparing the occurrence time of adjacent events, extract the periodicity of event occurrence by Fourier transform, and calculate the temporal correlation strength of adjacent events based on the time delay feature and periodicity.

[0019] A spatial distribution map is constructed based on the geographic coordinate information of each event in the event sequence. The spatial distance between adjacent events is calculated, and the spatial distance is combined with the spatial influence range of the events to obtain the spatial correlation strength.

[0020] The semantic vectors of events are constructed using the descriptive text of each event in the event sequence. The cosine similarity of the semantic vectors of adjacent events is calculated and combined with the event type matching matrix to obtain the semantic association strength.

[0021] Based on the influence metric of each event in the event sequence, the influence correlation between adjacent events is calculated, and the influence correlation is combined with the influence transmission decay function to obtain the influence correlation strength.

[0022] The multidimensional association strength is obtained by weighted fusion of temporal association strength, spatial association strength, semantic association strength, and influence association strength.

[0023] Event sequences with multidimensional correlation strength greater than a preset strength threshold are selected to obtain an initial candidate risk chain set.

[0024] In one optional embodiment, based on the initial candidate risk chain set, a state feature matrix is ​​constructed through eigenvalue decomposition to identify logical breakpoints, potential transition events are searched, and iterative optimization yields a logically complete risk chain, including:

[0025] The multidimensional correlation strength of event sequences in the initial candidate risk chain set is constructed into a state feature matrix, where each element of the state feature matrix corresponds to the correlation state between adjacent events.

[0026] The state feature matrix is ​​decomposed into an eigenvalue diagonal matrix and an eigenvector matrix. The feature state difference between adjacent event pairs is calculated to determine the breakage probability value. Adjacent event pairs with breakage probability values ​​greater than the adaptive breakage threshold are marked as logical breakage points.

[0027] For the events before and after the logical breakpoint, an effective time window is constructed based on the corresponding timestamp, and an reachable range ellipse is constructed based on the corresponding geographical location.

[0028] Search for potential transition events within the effective time window and reachable ellipse, calculate the temporal, spatial, and semantic contributions of the potential transition events to the preceding and following events, and perform a weighted combination to obtain the event transition evaluation value.

[0029] The potential transition event with the highest transition evaluation value is used as a supplementary event and inserted between the events before and after the logical breakpoint to generate an optimized event sequence.

[0030] The state feature matrix is ​​reconstructed for the optimized event sequence, and the process is repeated until there are no adjacent event pairs marked as logical breakpoints in the optimized event sequence, thus obtaining a logically complete risk chain.

[0031] In an optional embodiment, the state feature matrix is ​​decomposed into an eigenvalue diagonal matrix and an eigenvector matrix. The feature state difference between adjacent event pairs is calculated to determine the break probability value. Adjacent event pairs with break probability values ​​greater than an adaptive break threshold are marked as logical break points, including:

[0032] The state feature matrix is ​​decomposed into eigenvalues ​​to obtain the eigenvalue diagonal matrix and eigenvector matrix, which are then constructed into a third-order tensor. Multi-scale tensor decomposition is performed on the third-order tensor to obtain the eigenvalue component matrix, eigenvector component matrix and time scale component matrix at different time scales.

[0033] A multi-scale feature spectrum of an event sequence is constructed based on the eigenvalue component matrix, the eigenvector component matrix, and the time scale component matrix. The energy distribution of the multi-scale feature spectrum at different time scales is mapped to a feature entropy matrix.

[0034] The characteristic state difference degree of adjacent event pairs is calculated based on the eigenvalue diagonal matrix and the eigenvector matrix, and the entropy difference vector of adjacent event pairs is calculated based on the characteristic entropy matrix.

[0035] The local fluctuation features of the characteristic state difference degree and the cross-scale correlation features of the entropy difference vector are extracted respectively. The local fluctuation features and the cross-scale correlation features are fused to obtain the comprehensive fracture features. The fracture probability value of adjacent event pairs is calculated based on the comprehensive fracture features.

[0036] An adaptive break threshold is constructed based on the variance of the feature state difference and the standard deviation of the entropy difference vector of adjacent event pairs. Adjacent event pairs with a break probability value greater than the adaptive break threshold are marked as logical break points.

[0037] In an optional embodiment, based on the logically complete risk chain, transmission strength features are extracted using transmission coefficient matrix analysis and state transition matrix decomposition, and a credibility score is calculated to obtain a risk chain credibility score set including:

[0038] For a logically complete risk chain, a transmission coefficient matrix is ​​constructed based on the fluctuation amplitude, propagation distance, and impact range of adjacent events. The risk transmission intensity sequence is obtained by calculating the temporal evolution of the transmission coefficient matrix.

[0039] A state transition matrix is ​​constructed for the risk transmission intensity sequence. The state transition matrix is ​​decomposed into eigenvalues, and eigenvalues ​​are calculated and determined. Events whose product of eigenvalue and influence range exceeds a preset transmission threshold are identified as diffusion transmission nodes. The energy diffusion index of the diffusion transmission nodes is calculated.

[0040] The risk transmission intensity sequence is divided into multiple time scales, and the fluctuation feature sequence is calculated at each time scale. The fluctuation feature sequence is reconstructed in segments to obtain a fractal sequence. The complexity index of risk transmission is calculated through the self-similarity of the fractal sequence.

[0041] Based on the risk transmission intensity sequence and the fluctuation characteristic sequence, a three-dimensional phase space containing transmission intensity, fluctuation amplitude and time delay is constructed. The phase space trajectory is extracted and the Lyapunov exponent of the phase space trajectory at different time scales is calculated.

[0042] A three-dimensional feature vector is constructed based on the energy diffusion index, complexity index, and Lyapunov index. A pre-trained adaptive weight network is used to calculate the transmission stability score of the logically complete risk chain. The transmission stability score is then transformed by probability distribution to obtain a set of risk chain credibility scores.

[0043] In one optional embodiment, segmented iterative reconstruction of the wave characteristic sequence to obtain a fractal sequence includes:

[0044] A fluctuation feature sequence is obtained, and the fluctuation feature sequence is processed by an overlapping segmentation method. The number of overlapping points between adjacent segments is determined by a preset segment length and overlap rate, and multiple overlapping sequence segments are generated.

[0045] Extract the fluctuation features of each of the overlapping sequence segments, the fluctuation features including fluctuation amplitude and fluctuation trend, and pair the overlapping sequence segments together according to the fluctuation features;

[0046] For each pair of adjacent overlapping sequence segments, a segmented mapping rule is constructed based on the fluctuation characteristics. The fluctuation amplitude ratio of adjacent overlapping sequence segments is used as the amplitude ratio parameter, and the fluctuation trend difference is used as the trend difference parameter to generate a segmented linear mapping function.

[0047] Overlapping sequence segments are iteratively reconstructed. In each iteration, the amplitude ratio parameter and trend difference parameter are dynamically adjusted based on the characteristic changes of the current overlapping sequence segment, and applied to the next iteration to generate reconstructed sequence segments.

[0048] The reconstructed sequence segments are sequentially spliced ​​together and smoothed to obtain a fractal sequence.

[0049] A second aspect of this invention provides a system for identifying urban public safety risk chains based on graph neural networks, comprising:

[0050] The first unit is used to acquire historical data on urban public safety incidents. It uses different types of public safety incidents as graph nodes and the relationships between public safety incidents as edges to construct an initial event association topology graph.

[0051] The second unit is used to determine the optimal sampling path set based on the initial event association topology graph by adopting an adaptive bidirectional skip sampling algorithm, through dynamic step size control and time sequence matching; and to calculate the multidimensional association strength based on the event sequences in the optimal sampling path set, thereby obtaining an initial candidate risk chain set.

[0052] The third unit is used to identify logical breakpoints by constructing a state feature matrix through eigenvalue decomposition based on the initial candidate risk chain set, search for potential transition events, and iteratively optimize to obtain a logically complete risk chain.

[0053] The fourth unit is used to extract transmission strength features based on the logically complete risk chain by using transmission coefficient matrix analysis and state transition matrix decomposition, calculate the credibility score, and obtain a risk chain credibility score set.

[0054] The fifth unit is used to select risk chains whose credibility scores exceed a preset credibility threshold based on the risk chain credibility score set, and to determine and output the final identification result.

[0055] A third aspect of the present invention provides an electronic device, comprising:

[0056] processor;

[0057] Memory used to store processor-executable instructions;

[0058] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0059] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0060] In this embodiment of the invention, by constructing an initial event association topology graph and using an adaptive bidirectional skip sampling algorithm to determine the optimal sampling path, the complex association patterns between urban public safety events can be effectively captured. This avoids the blind path search and wasted computational resources problems in traditional methods, improving the accuracy and efficiency of risk chain identification. The method of constructing a state feature matrix based on eigenvalue decomposition to identify logical breakpoints and search for potential transition events solves the problem of logical incoherence in risk chains, making the identified risk chains more complete and reasonable, and avoiding the common problems of risk chain breaks and logical omissions in traditional methods. The mechanism of extracting transmission strength features by using transmission coefficient matrix analysis and state transition matrix decomposition, and calculating credibility scores, provides a quantitative basis for risk chain assessment, effectively reducing the uncertainty in the risk chain identification process, improving the reliability and practical value of the identification results, and providing more accurate decision support for urban public safety management and risk prevention and control. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the urban public safety risk chain identification method based on graph neural networks according to an embodiment of the present invention.

[0062] Figure 2 A schematic diagram of the system architecture for optimizing the logical completeness of risk chains based on eigenvalue decomposition;

[0063] Figure 3 This diagram illustrates the performance comparison between multi-scale feature analysis and traditional methods. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0066] Figure 1 This is a flowchart illustrating the urban public safety risk chain identification method based on graph neural networks according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0067] Acquire historical data on urban public safety incidents, use different types of public safety incidents as graph nodes, and use the relationships between public safety incidents as edges to construct an initial event association topology graph;

[0068] Based on the initial event association topology, an adaptive bidirectional skip sampling algorithm is adopted to determine the optimal sampling path set through dynamic step size control and time sequence matching; based on the event sequences in the optimal sampling path set, multidimensional association strength is calculated, and an initial candidate risk chain set is obtained by screening.

[0069] Based on the initial candidate risk chain set, a state feature matrix is ​​constructed through eigenvalue decomposition to identify logical breakpoints, search for potential transition events, and iteratively optimize to obtain a logically complete risk chain.

[0070] Based on the logically complete risk chain, the transmission strength features are extracted using transmission coefficient matrix analysis and state transition matrix decomposition, and the credibility score is calculated to obtain a risk chain credibility score set.

[0071] Based on the risk chain credibility score set, risk chains with credibility scores exceeding a preset credibility threshold are selected, and the final identification result is determined and output.

[0072] In one optional implementation, based on the initial event association topology graph, an adaptive bidirectional skip sampling algorithm is used to determine the optimal sampling path set through dynamic step size control and timing matching, including:

[0073] The impact parameters of public safety event nodes are obtained from the initial event association topology map. The propagation impact is calculated based on the node association degree and distance attenuation factor. The spatial impact is calculated based on the ratio of the affected area of ​​the public safety event node. The temporal impact is calculated based on the event time attenuation function. The propagation impact, spatial impact and temporal impact are weighted and combined to obtain the node influence measure.

[0074] The initial adaptive step size is obtained by calculating the node influence metric and the baseline sampling step size. The adaptive control threshold is obtained based on the historical differences of the node influence metric. The dynamic sampling step size is obtained by combining the initial adaptive step size and the adaptive control threshold.

[0075] Bidirectional sampling is performed using a dynamic sampling step size to calculate the forward propagation probability and the backward verification probability, and then the time-series matching score is obtained by weighted fusion.

[0076] The path continuity index and length constraint index are calculated based on the time sequence matching score. The time sequence matching score is nonlinearly weighted with the path continuity index and length constraint index to obtain the path evaluation score. The optimal sampling path set is obtained by screening public safety event sequences whose path evaluation scores are greater than the preset evaluation threshold.

[0077] In one specific implementation, when obtaining the parameters for calculating the impact of public safety incident nodes, the characteristics of the incident nodes are extracted from the initial incident association topology, including the association relationships between nodes, spatial distribution information, and temporal attributes. For the node association degree calculation, the number of direct connections between each node and other nodes is counted and the connection strength is analyzed. For example, an explosion incident node is connected to 12 subsequent events, with an association strength value of 0.85. The distance attenuation factor adopts an exponentially decreasing model; when the physical distance between two incident nodes is 5 kilometers, the distance attenuation value is approximately 0.72. The propagation impact degree is calculated by multiplying the association degree by the attenuation factor; in the example above, the propagation impact degree is 0.61.

[0078] Spatial impact is calculated based on the event-affected area ratio, using the ratio of the event's affected area to the reference area as the fundamental indicator. For example, if a chemical leak event has an affected area of ​​15 square kilometers and a reference area of ​​60 square kilometers, the affected area ratio is 0.25. After considering a regional importance weighting factor of 1.2, the final spatial impact value is 0.30.

[0079] The time-dependent impact is calculated using a time decay function, employing a half-life model to assess how the event's impact changes over time. For example, the impact remains above 0.9 within 24 hours of an event, drops to 0.65 after 48 hours, and decreases to 0.32 after 96 hours. The time decay function parameters are adjusted according to the characteristics of different event types; for example, the decay rate is 0.08 for sudden accidents and 0.04 for cumulative events.

[0080] The node influence metric combines the dissemination impact, spatial impact, and temporal impact with weights set to 0.4, 0.35, and 0.25, respectively. For example, if the three indicator values ​​for a certain public health event node are 0.65, 0.42, and 0.78, the weighted calculation yields a node influence metric of 0.605.

[0081] The initial adaptive step size is calculated by multiplying the node influence metric by the baseline sampling step size. The baseline sampling step size is set to 2; when the node influence metric is 0.605, the initial adaptive step size is 1.21. The adaptive control threshold is determined based on historical differences in node influence. The changes in the influence metric values ​​of the most recent 10 sampling points are recorded. When the average rate of change for three consecutive points exceeds 15%, the adaptive control threshold is adjusted to 0.18; when the rate of change is less than 5%, the threshold is reduced to 0.08. The dynamic sampling step size is adjusted by comparing the initial adaptive step size with the control threshold. When the initial step size is less than 1.5 times the threshold, the dynamic sampling step size is the initial step size; otherwise, the dynamic sampling step size is a weighted sum of 80% of the initial step size and 20% of the threshold.

[0082] During bidirectional sampling, forward and backward sampling are performed using a dynamic sampling step size. Forward sampling starts from the current node and traverses successor nodes in chronological order of event occurrence, calculating the forward propagation probability. In practical cases, the propagation probabilities of three possible successor events for a given node are 0.72, 0.45, and 0.31, respectively; paths with probabilities greater than 0.5 are preferred. Backward verification sampling verifies the event sequence from the result node, evaluating its reasonableness. The calculation of the backward verification probability considers the strength of the causal relationship between events, typically ranging from 0.2 to 0.85. The forward propagation probability and the backward verification probability are weighted and fused to obtain a temporal matching score. The weights are allocated according to the reliability of the sampling direction, typically with a forward weight of 0.65 and a backward weight of 0.35.

[0083] During the path evaluation phase, a path continuity index is calculated, reflecting the temporal and logical coherence of adjacent events within the path. For a 10-node event sequence, 8 pairs of adjacent nodes were found to have time intervals consistent with the expected pattern, 1 pair showed a slight time difference, and 1 pair exhibited a significant logical jump, resulting in a continuity index value of 0.75. The path length constraint index ensures that the sampled path is neither too long nor too short. An ideal path length range is set based on the event type; for example, for a complete evolution path of a certain type of sudden event, the ideal number of nodes is 8 to 15. When the actual path length is 11 nodes, the length constraint index value is 0.92.

[0084] The final path evaluation score is obtained through a nonlinear weighted calculation of the temporal matching score, path continuity index, and length constraint index. An exponential weighting method is used, with the matching score weighted at 0.5, the continuity index at 0.35, and the length constraint index at 0.15. For example, a path with these three indices of 0.68, 0.75, and 0.92 will yield a path evaluation score of 0.724 after nonlinear weighting. A preset evaluation threshold of 0.65 is used to filter event sequences with evaluation scores greater than the threshold, ultimately generating the optimal sampled path set. In an emergency response simulation, 18 optimal paths were selected from 126 candidate paths, providing diverse response options for decision-making.

[0085] In this embodiment, the importance of public safety event nodes is comprehensively characterized by integrating propagation, spatial, and temporal influence, thereby enhancing the accuracy of path assessment. The sampling step size is dynamically adjusted based on node influence to achieve more precise bidirectional sampling and improve the effectiveness of temporal sequence matching. A nonlinear weighted evaluation is performed by combining temporal sequence matching scores, path continuity, and length constraints to enhance the scientific rigor and robustness of optimal event sequence selection. This approach is applicable to multi-source, heterogeneous, and complex public safety event topologies, improving the efficiency and practicality of event sequence analysis.

[0086] In one optional implementation, the multidimensional association strength is calculated based on the event sequences in the optimal sampling path set, and the initial candidate risk chain set is obtained by screening, including:

[0087] The event sequence is obtained from the optimal sampling path set to construct a time-series evolution map, extract the time interval and event density period of adjacent events, identify key time nodes, calculate the time delay feature of event propagation by comparing the occurrence time of adjacent events, extract the periodicity of event occurrence by Fourier transform, and calculate the temporal correlation strength of adjacent events based on the time delay feature and periodicity.

[0088] A spatial distribution map is constructed based on the geographic coordinate information of each event in the event sequence. The spatial distance between adjacent events is calculated, and the spatial distance is combined with the spatial influence range of the events to obtain the spatial correlation strength.

[0089] The semantic vectors of events are constructed using the descriptive text of each event in the event sequence. The cosine similarity of the semantic vectors of adjacent events is calculated and combined with the event type matching matrix to obtain the semantic association strength.

[0090] Based on the influence metric of each event in the event sequence, the influence correlation between adjacent events is calculated, and the influence correlation is combined with the influence transmission decay function to obtain the influence correlation strength.

[0091] The multidimensional association strength is obtained by weighted fusion of temporal association strength, spatial association strength, semantic association strength, and influence association strength.

[0092] Event sequences with multidimensional correlation strength greater than a preset strength threshold are selected to obtain an initial candidate risk chain set.

[0093] In one specific implementation, after obtaining the event sequence from the optimal sampling path set, a temporal evolution map is constructed to analyze the temporal correlation between events. Specifically, for an event sequence {E1, E2, ..., En}, the occurrence time of each event {T1, T2, ..., Tn} is recorded. By calculating the time interval ΔTi = Ti+1 - Ti between adjacent events, the event concentration period can be identified. For example, when multiple consecutive ΔTi values ​​are less than a preset threshold (such as 24 hours), this period is marked as an event concentration period. The system records the start and end times of these concentration periods as key time nodes. The time delay characteristics of event propagation are calculated by analyzing the occurrence times of adjacent events. For example, if an earthquake event E1 occurs in a certain area, and aftershock events E2 and E3 occur successively in the surrounding areas within 3 hours, then there is a clear time delay relationship between E1 and E2 and E3. Fourier transform is applied to the event sequence to extract periodic patterns. In the specific implementation, discrete Fourier transform is used to analyze the frequency of event occurrence, such as the increased frequency of mass events in a certain area every Friday afternoon. Based on time delay characteristics and periodic patterns, the temporal correlation strength (ST) is calculated. When two events occur in the same period of high frequency and the time delay conforms to historical statistical patterns, the ST value is high, such as ST=0.85; if the time interval is long and there is no obvious periodic correlation, the ST value is low, such as ST=0.2.

[0094] Spatial correlation analysis of event sequences is based on the geographic coordinates of the events. For each event in the sequence, the latitude and longitude coordinates (xi, yi) of its location are recorded. A spatial distribution map is constructed to visually display the distribution of events in geographic space. The spatial distance Dij between adjacent events is calculated using a spherical distance calculation method, taking into account the curvature of the Earth. Different spatial influence ranges Ri are defined for different types of events. For example, the spatial influence range of a natural disaster event may be 50 kilometers, while the spatial influence range of a social security event may be 10 kilometers. The spatial correlation strength SS is calculated as the ratio of spatial distance to spatial influence range. When Dij is less than Ri, it indicates a high degree of spatial correlation, and the SS value is close to 1; when Dij is much greater than Ri, the SS value is close to 0. For example, if the influence range of an explosion event is 2 kilometers, and related secondary disaster events occur within 1.5 kilometers, the calculated SS value is 0.75.

[0095] Semantic association analysis is achieved by processing event description text. Each event description text undergoes preprocessing, including word segmentation, stop word removal, and stemming. Word embedding technology is used to convert the processed text into semantic vectors Vi. The cosine similarity of the semantic vectors of adjacent events is calculated to obtain the degree of semantic similarity. Simultaneously, the system predefines an event type matching matrix M, representing the inherent correlation between different types of events. If two event types are "chemical plant explosion" and "air pollution," the M value might be 0.8, indicating a high correlation between these two types of events. The final semantic association strength SL is obtained by a weighted combination of the cosine similarity and the matching matrix value. In the example, the cosine similarity of the semantic vectors of the two event description texts is 0.6, and the event type matching matrix value is 0.8, resulting in a semantic association strength SL of 0.72.

[0096] Impact correlation analysis is based on the impact metric of events, calculating an impact index I for each event, considering factors such as the number of people affected, economic losses, and social attention. For adjacent events, an impact correlation degree C is calculated, reflecting the ratio of the impact of the subsequent event to that of the preceding event. An impact transmission attenuation function D is defined, considering the attenuation effect of temporal and spatial distance between events on impact transmission. The impact correlation strength SI is calculated by combining the impact correlation degree and the impact transmission attenuation function. For example, if the impact of the preceding event is 85 (out of 100), the impact of the subsequent event is 60, the time interval is 2 days, and the spatial distance is 15 kilometers, the calculated impact correlation degree C = 0.7, the transmission attenuation value D = 0.8, and the final SI = 0.56.

[0097] Multidimensional association strength is calculated by weighted fusion of the four association strengths mentioned above. Weight coefficients w1, w2, w3, and w4 are set according to different application scenarios, corresponding to temporal, spatial, semantic, and influence association strengths, respectively. The weighted calculation yields the multidimensional association strength S = w1 × ST + w2 × SS + w3 × SL + w4 × SI. In practical applications, the weight values ​​can be dynamically adjusted according to different risk chain identification needs. For example, for natural disaster risk chains, w1 = 0.3, w2 = 0.4, w3 = 0.1, and w4 = 0.2 might be set; while for social security risk chains, w1 = 0.2, w2 = 0.2, w3 = 0.4, and w4 = 0.2 might be set. For the calculated multidimensional association strength values, the system sets a preset strength threshold θ (e.g., θ = 0.6) and filters event sequences where S is greater than θ, forming an initial candidate risk chain set. In one instance, the association strengths of four pairs of adjacent events formed by five consecutive events are 0.75, 0.82, 0.58, and 0.63, respectively. The third pair is below the threshold, so the system divides the event sequence into two candidate risk chains.

[0098] In this embodiment, the correlation strength of four dimensions—time, space, semantics, and influence—is integrated to achieve in-depth and multi-faceted correlation analysis of event sequences. Based on the multi-dimensional correlation strength, highly correlated event sequences are screened, accurately extracting potential candidate risk chains and providing a solid foundation for subsequent risk assessment. The introduction of time delay features and periodic analysis effectively identifies the rhythm and patterns of event propagation, enhancing the understanding of the dynamic evolution trend of events. The fusion of multi-source heterogeneous information makes the model more adaptable to the complex and ever-changing public safety environment, improving the stability and reliability of the analysis results.

[0099] In one optional implementation, based on the initial candidate risk chain set, a state feature matrix is ​​constructed through eigenvalue decomposition to identify logical breakpoints, potential transition events are searched, and iterative optimization yields a logically complete risk chain, including:

[0100] The multidimensional correlation strength of event sequences in the initial candidate risk chain set is constructed into a state feature matrix, where each element of the state feature matrix corresponds to the correlation state between adjacent events.

[0101] The state feature matrix is ​​decomposed into an eigenvalue diagonal matrix and an eigenvector matrix. The feature state difference between adjacent event pairs is calculated to determine the breakage probability value. Adjacent event pairs with breakage probability values ​​greater than the adaptive breakage threshold are marked as logical breakage points.

[0102] For the events before and after the logical breakpoint, an effective time window is constructed based on the corresponding timestamp, and an reachable range ellipse is constructed based on the corresponding geographical location.

[0103] Search for potential transition events within the effective time window and reachable ellipse, calculate the temporal, spatial, and semantic contributions of the potential transition events to the preceding and following events, and perform a weighted combination to obtain the event transition evaluation value.

[0104] The potential transition event with the highest transition evaluation value is used as a supplementary event and inserted between the events before and after the logical breakpoint to generate an optimized event sequence.

[0105] The state feature matrix is ​​reconstructed for the optimized event sequence, and the process is repeated until there are no adjacent event pairs marked as logical breakpoints in the optimized event sequence, thus obtaining a logically complete risk chain.

[0106] Figure 2This diagram illustrates the architecture for optimizing the logical completeness of risk chains based on eigenvalue decomposition. In one specific implementation, an initial candidate risk chain set of data is received. This data contains multiple event sequences, each event including a timestamp, geographical location, and event description. The initial candidate risk chain can be a set of event sequences extracted from a safety monitoring system, such as a sequence consisting of events like E1 (2023-05-01 08:30:00, location coordinates (115.32, 39.45), device startup anomaly) and E2 (2023-05-01 09:15:00, location coordinates (115.35, 39.46), temperature exceeding limits).

[0107] An analysis of the event sequences in the initial candidate risk chain set is performed to construct a state feature matrix. This matrix reflects the multidimensional correlation strength between adjacent events. Assuming there are N events in the risk chain, the constructed state feature matrix has a size of (N-1)×D, where D is the number of correlation dimensions. For each pair of adjacent events (Ei, Ei+1), their correlation strength values ​​are calculated as follows: time dimension RT(i, i+1), spatial dimension RS(i, i+1), and semantic dimension RL(i, i+1). The time dimension correlation strength is calculated by the ratio of the time interval between the two events to the expected time window; the spatial dimension correlation strength is calculated by the ratio of the geographical distance between the two events to the expected reachable distance; and the semantic dimension correlation strength is calculated by the similarity of the event description texts. The correlation strength values ​​of the three dimensions constitute the elements of the i-th row of the state feature matrix, ranging from 0 to 1, with larger values ​​indicating stronger correlations.

[0108] The constructed state feature matrix is ​​subjected to eigenvalue decomposition to obtain an eigenvalue diagonal matrix and an eigenvector matrix. Eigenvalue decomposition is achieved through iterative computation. The eigenvalue diagonal matrix contains the sorted eigenvalues, and the eigenvector matrix contains the corresponding eigenvectors. Based on the decomposition results, the feature state difference between adjacent event pairs is calculated. For each pair of adjacent events (Ei, Ei+1), its feature state difference D(i, i+1) is calculated. A larger difference indicates a weaker logical coherence between the two events. When calculating the difference, the event states are projected onto the feature space using the eigenvectors, and the Euclidean distance between the projected state vectors is calculated. The feature state difference is normalized to obtain the break probability value P(i, i+1).

[0109] An adaptive break threshold θ is set, the value of which is determined through historical data analysis or expert experience; for example, it can be set to 0.75. Adjacent event pairs whose break probability values ​​P(i, i+1) are greater than the threshold θ are marked as logical break points. For example, if the break probability value of event pair (E2, E3) is 0.82, which is greater than the threshold 0.75, then (E2, E3) is marked as a logical break point.

[0110] For each marked logical breakpoint (Ei, Ei+1), an effective time window [Ti+α, Ti+1-β] is constructed based on its timestamp, where Ti and Ti+1 are the timestamps of events Ei and Ei+1, respectively, and α and β are time buffer parameters. For example, if E2 time is 09:15:00 and E3 time is 10:30:00, and α=5 minutes and β=5 minutes, then the effective time window is [09:20:00, 10:25:00]. Simultaneously, an reachable range ellipse is constructed based on geographical location. The two foci of the ellipse are the geographical locations of events Ei and Ei+1, and the semi-major axis length of the ellipse is related to the distance between the two events and their possible movement speed. In the aforementioned example, if the coordinates of E2 are (115.35, 39.46) and the coordinates of E3 are (115.40, 39.49), then the constructed ellipse includes these two points, and its major axis is along the line connecting the two points.

[0111] Within a defined effective time window and reachable ellipse, potential transition events are searched from the event database. For each candidate potential transition event Ec, its correlation with the preceding and following events Ei and Ei+1 is calculated. The temporal dimension contribution CT(Ec) is calculated based on the uniformity of Ec's position on the time axis relative to Ei and Ei+1; the spatial dimension contribution CS(Ec) is calculated based on Ec's spatial proximity to the path from Ei to Ei+1; and the semantic dimension contribution CL(Ec) is calculated based on the semantic continuity of Ec with Ei and Ei+1. The weighted combination of the three dimension contributions yields the event transition evaluation value V(Ec) = w1×CT(Ec) + w2×CS(Ec) + w3×CL(Ec), where the sum of the weights w1, w2, and w3 is 1, which can be adjusted according to specific application scenarios.

[0112] The event with the highest transition evaluation value V(Ec) from the set of potential transition events is selected as the supplementary event E*, and inserted between the events before and after the logical breakpoint to generate an optimized event sequence. For example, if the best transition event found is E*(2023-05-01 09:45:00, location coordinates (115.37, 39.47), pressure fluctuation), it is inserted between E2 and E3 to obtain the optimized sequence [..., E2, E*, E3, ...].

[0113] The optimized event sequence is reconstructed into a state feature matrix. The eigenvalue decomposition, breakpoint identification, transition event search, and sequence optimization steps are repeated until no more pairs of related events marked as logical breakpoints exist, resulting in a logically complete risk chain. In practical applications, a maximum number of iterations (e.g., 10) or a minimum improvement threshold can be set as a termination condition to avoid over-optimization.

[0114] In a case study of an industrial production safety monitoring system implementation, the initial candidate risk chain contained five events: E1 (08:30, equipment startup anomaly), E2 (09:15, temperature exceeding limits), E3 (10:30, pressure anomaly), E4 (11:20, valve failure), and E5 (12:00, system shutdown). After the above processing, (E2, E3) was identified as a logical breakpoint, and the transitional event E* (09:45, pressure fluctuation) was added. In the second iteration, (E3, E4) was identified as a logical breakpoint, and the event E** (10:50, pipeline vibration) was added. The final logically complete risk chain is: E1→E2→E*→E3→E**→E4→E5, making the risk event sequence logically more coherent and complete.

[0115] In this embodiment, by detecting logical breakpoints and inserting transitional events, the problem of missing information or interrupted associations is effectively repaired, and a coherent and continuous event sequence is constructed. By utilizing the eigenvalue decomposition of the state feature matrix, the state differences between adjacent events are accurately quantified, improving the accuracy of logical breakpoint identification. Transitional events are evaluated by combining temporal, spatial, and semantic multidimensional contributions to ensure that the supplementary nodes are highly consistent with the preceding and following events, thereby improving the rationality and credibility of the risk chain. A cyclic iterative optimization mechanism is adopted to continuously improve the event sequence structure, ultimately obtaining a high-quality risk chain with a logical closed loop, providing a more reliable basis for risk prediction and prevention.

[0116] In one optional implementation, the state feature matrix is ​​decomposed into an eigenvalue diagonal matrix and an eigenvector matrix. The feature state difference between adjacent event pairs is calculated to determine the break probability value. Adjacent event pairs with break probability values ​​greater than an adaptive break threshold are marked as logical break points, including:

[0117] The state feature matrix is ​​decomposed into eigenvalues ​​to obtain the eigenvalue diagonal matrix and eigenvector matrix, which are then constructed into a third-order tensor. Multi-scale tensor decomposition is performed on the third-order tensor to obtain the eigenvalue component matrix, eigenvector component matrix and time scale component matrix at different time scales.

[0118] A multi-scale feature spectrum of an event sequence is constructed based on the eigenvalue component matrix, the eigenvector component matrix, and the time scale component matrix. The energy distribution of the multi-scale feature spectrum at different time scales is mapped to a feature entropy matrix.

[0119] The characteristic state difference degree of adjacent event pairs is calculated based on the eigenvalue diagonal matrix and the eigenvector matrix, and the entropy difference vector of adjacent event pairs is calculated based on the characteristic entropy matrix.

[0120] The local fluctuation features of the characteristic state difference degree and the cross-scale correlation features of the entropy difference vector are extracted respectively. The local fluctuation features and the cross-scale correlation features are fused to obtain the comprehensive fracture features. The fracture probability value of adjacent event pairs is calculated based on the comprehensive fracture features.

[0121] An adaptive break threshold is constructed based on the variance of the feature state difference and the standard deviation of the entropy difference vector of adjacent event pairs. Adjacent event pairs with a break probability value greater than the adaptive break threshold are marked as logical break points.

[0122] In one specific implementation, eigenvalue decomposition is performed on the state feature matrix to obtain an eigenvalue diagonal matrix and an eigenvector matrix. Assume the state feature matrix is ​​an n×m matrix, where n is the number of events and m is the feature dimension. Eigenvalue decomposition yields an m×m eigenvector matrix U and a diagonal matrix Λ containing m eigenvalues. The eigenvalue diagonal matrix Λ and the eigenvector matrix U are then constructed as a third-order tensor T with dimensions m×m×1.

[0123] A multi-scale tensor decomposition is performed on the third-order tensor T using the Tucker decomposition method, decomposing it into component matrices at different time scales. The decomposition yields an eigenvalue component matrix P (m×r1), an eigenvector component matrix Q (m×r2), and a time scale component matrix R (1×r3), where r1, r2, and r3 are the ranks of each dimension. For example, when m=100, r1=20, r2=30, and r3=5 can be chosen to effectively capture data characteristics at various time scales.

[0124] A multi-scale feature spectrum S of the event sequence is constructed based on component matrices P, Q, and R. These three component matrices are combined using tensor product operations to form a multi-scale feature spectrum of size m×m×r³. For each time scale k (1≤k≤r³), the energy distribution of the feature spectrum at that scale is calculated and mapped to a feature entropy matrix H. The feature entropy matrix H is n×r³, representing the entropy distribution of n events across r³ time scales. For example, the feature entropy value of an event i at time scale k can be obtained by calculating the Shannon entropy of the energy distribution of the feature spectrum at that scale; this value reflects the uncertainty of the event characteristics at that time scale.

[0125] The characteristic state difference D between adjacent event pairs is calculated based on the eigenvalue diagonal matrix Λ and the eigenvector matrix U. For an adjacent event pair (i, i+1), its characteristic state difference D(i, i+1) can be obtained by calculating the Euclidean distance between the two events in the characteristic space. Simultaneously, the entropy difference vector E(i, i+1) of adjacent event pairs is calculated based on the characteristic entropy matrix H. For an adjacent event pair (i, i+1), its entropy difference vector E(i, i+1) is an r3-dimensional vector representing the entropy difference between the two events at various time scales.

[0126] Local fluctuation features W of the feature state difference degree D are extracted using a sliding window method. A window of size 3 slides across the difference degree sequence to calculate the fluctuation of the difference degree within the window. For example, for position i, D(i-1), D(i), and D(i+1) are extracted, and the local fluctuation feature W(i) is obtained by calculating the mean difference between the center position and the two adjacent positions. For the entropy difference vector E, its cross-scale correlation feature C is extracted. Specifically, the correlation of entropy differences between different time scales is calculated to form a correlation matrix of size r3×r3, and the main correlation patterns are extracted as the cross-scale correlation feature C.

[0127] The local fluctuation feature W and the cross-scale correlation feature C are fused to obtain the comprehensive fracture feature F. The fusion method uses weighted summation, and the weights can be determined through training with historical data. For example, if the weight of the local fluctuation feature is set to 0.6 and the weight of the cross-scale correlation feature is set to 0.4, then F = 0.6×W + 0.4×C. The fracture probability value P of adjacent event pairs is calculated based on the comprehensive fracture feature F. The Sigmoid function is used to map F to the interval [0, 1] to obtain the fracture probability of each pair of adjacent events.

[0128] An adaptive breaking threshold θ is constructed based on the variance of the characteristic state difference between adjacent event pairs, var(D), and the standard deviation of the entropy difference vector, std(E). Specifically, it is calculated as θ = α × var(D) + β × std(E), where α and β are coefficients balancing the contributions of the two components, and can be set to α = 0.7 and β = 0.3. This adaptive threshold can dynamically adjust according to the distribution characteristics of the data itself, making it more robust.

[0129] Finally, adjacent event pairs with a break probability value P greater than the adaptive break threshold θ are marked as logical break points. For example, in a sequence containing 1000 events, the break probability values ​​of adjacent event pairs are calculated. Suppose the break probability values ​​of the 150th and 632nd pairs are 0.78 and 0.85 respectively, and the adaptive threshold θ is 0.75. Then these two pairs of events will be marked as logical break points, that is, events 151 and 633 are the starting points of the new logical segment.

[0130] Traditional event sequence analysis techniques typically employ a single time scale, failing to effectively capture the complex cross-scale relationships between events. Common event segmentation methods rely primarily on fixed thresholds or manually preset rules, which often exhibit low adaptability and accuracy when facing complex and ever-changing urban public safety risk scenarios. For example, sliding window-based event segmentation algorithms struggle to handle event sequences of varying lengths; clustering-based methods are sensitive to parameter settings and have limited generalization capabilities; and traditional feature extraction techniques mostly focus on single-dimensional feature changes, ignoring the interactions between multi-dimensional features. This embodiment's method captures feature evolution patterns across different time scales by performing eigenvalue decomposition on the event feature matrix and constructing a third-order tensor. The introduction of feature spectrum analysis and feature entropy calculation allows for simultaneous consideration of local fluctuation features and cross-scale correlation features, providing a more comprehensive characterization of structural changes in the event sequence. Compared to existing technologies, this embodiment's method constructs an adaptive fracture threshold mechanism, dynamically adjusting the threshold based on data distribution characteristics, effectively addressing over-segmentation or under-segmentation issues caused by fixed thresholds. By fusing local fluctuation and cross-scale correlation information, this method can more accurately identify key turning points in the public safety risk chain. Experimental results show that, across different types of urban public safety incident sequences, the proposed method achieves higher accuracy in identifying logical breakpoints and faster processing speed compared to traditional methods. Furthermore, it demonstrates stronger robustness in the face of noise interference and missing data. The application of this technology makes the identification of public safety risk chains more accurate and efficient, providing strong support for the timely discovery of potential risk points and the development of prevention and response strategies.

[0131] Figure 3 This diagram illustrates the performance comparison between multi-scale feature analysis and traditional methods. The data clearly shows that the multi-scale feature analysis method performs best across all evaluation metrics. In terms of recognition accuracy, the multi-scale method achieves 91.5%, significantly higher than the single-scale method's 70.6% and the fixed-threshold method's 60.8%. Regarding noise resistance, the multi-scale method reaches 86.3%, nearly 30 percentage points higher than the single-scale method (58.5%) and over 40 percentage points higher than the fixed-threshold method (43.2%), demonstrating extremely strong noise resistance.

[0132] It is worth noting that in terms of processing speed, the fixed threshold method (72.5%) outperforms the single-scale method (56.3%), but still lags behind the multi-scale method (78.4%). This indicates that multi-scale feature analysis improves performance without sacrificing processing efficiency. Finally, in the comprehensive evaluation metric F1-Score, the multi-scale method also leads with a high score of 89.2%, further confirming its advantage in balancing accuracy and recall.

[0133] Overall, the charts strongly demonstrate the comprehensive advantages of multi-scale feature analysis methods in identifying logical breakpoints in risk chains, particularly their superior performance in handling complex data and resisting noise interference.

[0134] In one optional implementation, based on the logically complete risk chain, transmission strength features are extracted using transmission coefficient matrix analysis and state transition matrix decomposition, and a credibility score is calculated to obtain a risk chain credibility score set including:

[0135] For a logically complete risk chain, a transmission coefficient matrix is ​​constructed based on the fluctuation amplitude, propagation distance, and impact range of adjacent events. The risk transmission intensity sequence is obtained by calculating the temporal evolution of the transmission coefficient matrix.

[0136] A state transition matrix is ​​constructed for the risk transmission intensity sequence. The state transition matrix is ​​decomposed into eigenvalues, and eigenvalues ​​are calculated and determined. Events whose product of eigenvalue and influence range exceeds a preset transmission threshold are identified as diffusion transmission nodes. The energy diffusion index of the diffusion transmission nodes is calculated.

[0137] The risk transmission intensity sequence is divided into multiple time scales, and the fluctuation feature sequence is calculated at each time scale. The fluctuation feature sequence is reconstructed in segments to obtain a fractal sequence. The complexity index of risk transmission is calculated through the self-similarity of the fractal sequence.

[0138] Based on the risk transmission intensity sequence and the fluctuation characteristic sequence, a three-dimensional phase space containing transmission intensity, fluctuation amplitude and time delay is constructed. The phase space trajectory is extracted and the Lyapunov exponent of the phase space trajectory at different time scales is calculated.

[0139] A three-dimensional feature vector is constructed based on the energy diffusion index, complexity index, and Lyapunov index. A pre-trained adaptive weight network is used to calculate the transmission stability score of the logically complete risk chain. The transmission stability score is then transformed by probability distribution to obtain a set of risk chain credibility scores.

[0140] In one specific implementation, when performing risk transmission analysis on a logically complete risk chain, a transmission coefficient matrix is ​​constructed based on the fluctuation amplitude, propagation distance, and impact range of adjacent events. The fluctuation amplitude is calculated using the rate of change of key indicators before and after the event. For example, if the population density in a public place suddenly increases from 2.3 people / square meter to 4.8 people / square meter, the fluctuation amplitude is 108.7%. The propagation distance is a weighted combination of the geographical distance and time interval between events. For example, if the physical distance between two safety events is 1.2 kilometers and the time interval is 0.5 hours, the normalized propagation distance is 0.65. The impact range is comprehensively assessed based on the size of the geographical area covered by the event and the number of people affected. For example, if the impact range of a public safety event is a circular area with a radius of 2.3 kilometers, affecting approximately 25,000 people, the normalized impact range index is 0.78. These three parameters are combined to form a transmission coefficient matrix with an n×n dimension, where n is the number of events in the risk chain. For a risk chain containing 12 events, a 12×12 transmission coefficient matrix is ​​generated, where each element represents the strength of the transmission relationship between a pair of events.

[0141] The risk transmission intensity sequence is calculated by analyzing the temporal evolution of the transmission coefficient matrix. Specifically, the transmission coefficient matrix is ​​divided into multiple time windows of size 3 with a sliding step of 1. The transmission coefficients within each time window are weighted and accumulated to obtain the transmission intensity value for that window. For example, for a risk chain of 12 events, a sequence of 10 transmission intensity values ​​can be obtained. This sequence is then normalized to ensure its value range is within the [0, 1] interval, facilitating subsequent analysis.

[0142] A state transition matrix is ​​constructed for the risk transmission intensity sequence, discretizing the transmission intensity values ​​into multiple state levels. For example, the interval [0, 1] is divided into 5 states: [0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), and [0.8, 1]. The frequency of state transitions between adjacent transmission intensity values ​​is counted to construct a 5×5 state transition matrix. Eigenvalue decomposition is performed on this matrix to obtain the set of eigenvalues. The eigenvalues ​​are multiplied by the influence range of the corresponding event. If the product exceeds a preset transmission threshold of 0.5, the event is identified as a diffusion transmission node. In the risk chain of 12 events, it is assumed that 3 events are identified as diffusion transmission nodes, with eigenvalue-influence range products of 0.67, 0.58, and 0.72, respectively. For each diffusion transmission node, its energy diffusion index is calculated, which reflects the rate of outward diffusion of event energy. The calculation method is to multiply the rate of change of transmission intensity within a certain time window before and after the node by the rate of change of influence range, and then multiply by a decay factor. For example, for a diffusion-conduction node where the product of the eigenvalue and the influence range is 0.72, its energy diffusion index is calculated to be 0.63.

[0143] The risk transmission intensity sequence was divided into multiple time scales, and multi-scale decomposition was performed using wavelet transform. Volatility characteristic sequences were calculated at each time scale, specifically by calculating the local volatility and frequency of the sequence. For the risk chain of 12 events, it could be divided into three time scales: short-term (1-2 hours), medium-term (6-12 hours), and long-term (over 24 hours). The volatility characteristic sequences at each time scale were reconstructed segmentally and iteratively to obtain fractal sequences. The reconstruction process employed a sliding window R / S analysis method, with the window size starting at 4 and increasing by 2 each time until half the sequence length was reached. The Hearst exponent was obtained by calculating the rescaled range under different window sizes. The complexity index of risk transmission was calculated based on the self-similarity of the fractal sequences, reflecting the complexity and uncertainty of the risk transmission process. For example, the complexity indices for the short-term, medium-term, and long-term time scales were 0.72, 0.58, and 0.41, respectively, indicating that the short-term risk transmission was more complex.

[0144] Based on the risk transmission intensity sequence and fluctuation characteristic sequence, a three-dimensional phase space is constructed, including transmission intensity, fluctuation amplitude, and time delay. The transmission intensity ranges from [0, 1], the fluctuation amplitude ranges from [-1, 1], and the time delay ranges from [0, 24] hours. Phase space trajectories are extracted from this three-dimensional space, with the number of trajectory points equal to the number of events. The Lyapunov exponents of the phase space trajectories at different time scales are calculated, reflecting the system's sensitivity to initial conditions. The calculation method involves selecting adjacent point pairs in the phase space, tracking their trajectory development, and measuring the growth rate of the distance between trajectories. For the risk chain of the above 12 events, the Lyapunov exponents at the short-term, medium-term, and long-term time scales are 0.24, 0.16, and 0.09, respectively, indicating that the risk transmission system is more sensitive to initial conditions in the short term.

[0145] A three-dimensional feature vector is constructed based on the energy diffusion index, complexity index, and Lyapunov index. For the aforementioned risk chain, the feature vector is [0.63, 0.72, 0.24]. A pre-trained adaptive weight network is used to calculate the transmission stability score of the logically complete risk chain. The adaptive weight network consists of a three-layer neural network, with the input layer being the three-dimensional feature vector, the hidden layer containing 8 neurons, and the output layer being a single stability score. The network is trained on 500 historical risk chain cases to learn the mapping relationship between features and stability. After inputting the feature vector [0.63, 0.72, 0.24], the transmission stability score output by the network is 0.37. The transmission stability score is transformed by probability distribution to map it to a confidence score in the interval [0, 1]. The transformation method uses a sigmoid function transformation, and the transformed confidence score is 0.68. The above process is repeated for each event segment in the entire risk chain to obtain a set of risk chain confidence scores, such as [0.68, 0.42, 0.75, 0.83].

[0146] Traditional methods for analyzing urban public safety risk chains primarily rely on statistical models and rule matching, which struggle to effectively capture the nonlinear dynamic characteristics of risk transmission. Common analytical techniques such as Markov chain models and linear regression, while capable of describing the temporal correlation of risk events, have limitations in handling multi-scale transmission characteristics. This embodiment introduces a multi-scale analysis framework, treating the risk transmission process as a complex system with fractal properties. By constructing a three-dimensional phase space and calculating the Lyapunov exponent, it more comprehensively captures the dynamic characteristics of risk transmission. Compared to traditional methods, this embodiment addresses the shortcomings in characterizing scale dependence and nonlinear transmission characteristics in risk transmission assessment. By integrating the energy diffusion index, complexity index, and Lyapunov exponent, a more comprehensive representation of risk transmission characteristics is constructed. Experimental results show that this method improves the accuracy of risk chain identification in complex urban environments compared to traditional methods, enhances the accuracy of transmission path prediction, and significantly improves robustness to data noise and missing data. The application of this technology provides a more reliable risk chain identification tool for urban public safety management, helping to identify potential risk points early and implement targeted preventative measures.

[0147] In one optional implementation, the segmented iterative reconstruction of the wave characteristic sequence to obtain the fractal sequence includes:

[0148] A fluctuation feature sequence is obtained, and the fluctuation feature sequence is processed by an overlapping segmentation method. The number of overlapping points between adjacent segments is determined by a preset segment length and overlap rate, and multiple overlapping sequence segments are generated.

[0149] Extract the fluctuation features of each of the overlapping sequence segments, the fluctuation features including fluctuation amplitude and fluctuation trend, and pair the overlapping sequence segments together according to the fluctuation features;

[0150] For each pair of adjacent overlapping sequence segments, a segmented mapping rule is constructed based on the fluctuation characteristics. The fluctuation amplitude ratio of adjacent overlapping sequence segments is used as the amplitude ratio parameter, and the fluctuation trend difference is used as the trend difference parameter to generate a segmented linear mapping function.

[0151] Overlapping sequence segments are iteratively reconstructed. In each iteration, the amplitude ratio parameter and trend difference parameter are dynamically adjusted based on the characteristic changes of the current overlapping sequence segment, and applied to the next iteration to generate reconstructed sequence segments.

[0152] The reconstructed sequence segments are sequentially spliced ​​together and smoothed to obtain a fractal sequence.

[0153] In one specific implementation, obtaining the fluctuation characteristic sequence is a key step in analyzing the urban public safety risk chain. This sequence reflects the temporal and spatial fluctuation characteristics of risk events. The fluctuation characteristic sequence can be obtained by processing the original risk indicator data. For example, normalizing the safety risk index of a certain area for 30 consecutive days yields a fluctuation characteristic sequence with a value range of [0, 1]. The fluctuation characteristic sequence is then processed using an overlapping segmentation method. The number of overlapping points between adjacent segments is determined by a preset segment length and overlap rate, generating multiple overlapping sequence segments. In practical applications, a segment length of 10 and an overlap rate of 50% can be set, resulting in 5 overlapping points between adjacent segments. Taking a fluctuation characteristic sequence of 30 points as an example, 5 overlapping sequence segments can be generated: segment 1 contains points 1 to 10, segment 2 contains points 6 to 15, segment 3 contains points 11 to 20, segment 4 contains points 16 to 25, and segment 5 contains points 21 to 30. This overlapping segmentation method maintains the continuity of the sequence and avoids information breaks caused by segmentation.

[0154] The fluctuation characteristics of each overlapping sequence segment are extracted, including fluctuation amplitude and fluctuation trend. The fluctuation amplitude is calculated as the difference between the maximum and minimum values ​​within the sequence segment, reflecting the severity of the fluctuation; the fluctuation trend is obtained by dividing the difference between the first and last points of the sequence segment by the number of points, indicating an overall upward or downward tendency. Taking the first sequence segment as an example, assuming its values ​​are [0.2, 0.3, 0.25, 0.4, 0.35, 0.45, 0.5, 0.48, 0.52, 0.55], its fluctuation amplitude is 0.55-0.2=0.35, and the fluctuation trend is (0.55-0.2) / 10=0.035, indicating an overall upward trend. Based on the fluctuation characteristics, the overlapping sequence segments are paired up into four pairs: segment 1 and segment 2, segment 2 and segment 3, segment 3 and segment 4, and segment 4 and segment 5.

[0155] For each pair of adjacent overlapping sequence segments, a piecewise mapping rule is constructed based on fluctuation characteristics. The ratio of fluctuation amplitudes of adjacent overlapping sequence segments is used as the amplitude ratio parameter, and the difference in fluctuation trends is used as the trend difference parameter to generate a piecewise linear mapping function. Assuming the fluctuation amplitude of segment 1 is 0.35 and that of segment 2 is 0.42, the amplitude ratio parameter is 0.42 / 0.35 = 1.2; the fluctuation trend of segment 1 is 0.035 and that of segment 2 is 0.02, so the trend difference parameter is 0.02 - 0.035 = -0.015. The piecewise linear mapping function uses these two parameters to map segment 1 to the feature space of segment 2, achieving consistent feature transformation. Specifically, for each point in segment 1, the transformation is performed by multiplying by the amplitude ratio parameter and adding the product of the trend difference parameter and the point's position. For example, the 7th point in the first segment has a value of 0.5, and the mapped value is 0.5×1.2+(-0.015)×7=0.6-0.105=0.495.

[0156] Overlapping sequence segments are iteratively reconstructed. In each iteration, the amplitude ratio and trend difference parameters are dynamically adjusted based on the characteristic changes of the current overlapping sequence segment, and applied to the next iteration to generate a reconstructed sequence segment. The number of iterations is usually set to 3 to 5, which is sufficient to achieve good reconstruction results. In the first iteration, the initially calculated amplitude ratio and trend difference parameters are used; from the second iteration onwards, the parameters are fine-tuned according to the degree of difference between the previous iteration result and the target segment. For example, if the fluctuation amplitude of the previous iteration result is 10% smaller than that of the target segment, the amplitude ratio parameter is increased by 0.1; if the fluctuation trend is 0.005 larger than that of the target segment, the trend difference parameter is decreased by 0.005. Through this dynamic adjustment mechanism, the iterative process can gradually approach the target features.

[0157] The reconstructed sequence segments are sequentially concatenated and smoothed to obtain a fractal sequence. During the concatenation process, for data points in overlapping areas, a weighted average method is used to calculate the final value, with the weight related to the point's position within its respective sequence segment. For example, for five overlapping points, the first point has a weight of 0.8 in the preceding segment and 0.2 in the following segment; the last point has a weight of 0.2 in the preceding segment and 0.8 in the following segment. The smoothing process uses a moving average method with a window size of 3 to ensure a natural transition at the concatenation points. The final generated fractal sequence retains the main characteristics of the original wave pattern sequence while exhibiting better self-similarity and structural stability.

[0158] For example, the 30-day security risk fluctuation sequence of a commercial area is [0.25, 0.28, 0.30, 0.35, 0.32, 0.40, 0.45, 0.42, 0.38, 0.48, 0.50, 0.47, 0.52, 0.55, 0.51, 0.58, 0.60, 0.57, 0.55, 0.65, 0.68, 0.64, 0.70, 0.72, 0.68, 0.75, 0.78, 0.76, 0.80, 0.82]. The segment length is set to 10, and the overlap rate is 50%, generating 5 overlapping sequence segments. The fluctuation amplitude of the first segment is calculated to be 0.23, and the fluctuation trend is 0.023; the fluctuation amplitude of the second segment is 0.20, and the fluctuation trend is 0.020. The amplitude ratio parameter is 0.87, and the trend difference parameter is -0.003. Through three iterations of reconstruction, the final generated fractal sequence can more accurately reflect the inherent laws of risk fluctuations, providing a reliable foundation for subsequent risk chain identification.

[0159] Traditional methods for analyzing fluctuation characteristics mainly rely on Fourier transforms or wavelet analysis. While these methods can effectively extract frequency domain features, they struggle to capture the fractal properties and self-similar structures in nonlinear systems. Common time series analysis techniques, such as the Autoregressive Integral Moving Average (ARIMA) model and Long Short-Term Memory (LSTM) networks, often exhibit low prediction accuracy and adaptability when dealing with risk fluctuation sequences with long-range dependencies. This embodiment introduces an overlapping segmented iterative reconstruction method, constructing dynamic mapping rules to more accurately capture the fractal characteristics of risk fluctuation sequences. Compared to traditional methods, this embodiment addresses the scale mixing problem and the difficulty of nonlinear feature extraction in risk sequence analysis. By introducing a dynamic adjustment mechanism for amplitude ratio and trend difference parameters, adaptive extraction of multi-scale features is achieved. Experimental results show that this method improves the prediction accuracy of urban public safety risk fluctuation sequences compared to traditional methods, enhances the efficiency of fluctuation pattern identification, and significantly strengthens robustness against noise and outliers. The application of this technology provides a more precise analytical tool for the early identification and warning of urban public safety risk chains, which helps to improve the foresight and accuracy of risk management and enable timely intervention and effective control of potential risks.

[0160] The urban public safety risk chain identification system based on graph neural networks according to embodiments of the present invention includes:

[0161] The first unit is used to acquire historical data on urban public safety incidents. It uses different types of public safety incidents as graph nodes and the relationships between public safety incidents as edges to construct an initial event association topology graph.

[0162] The second unit is used to determine the optimal sampling path set based on the initial event association topology graph by adopting an adaptive bidirectional skip sampling algorithm, through dynamic step size control and time sequence matching; and to calculate the multidimensional association strength based on the event sequences in the optimal sampling path set, thereby obtaining an initial candidate risk chain set.

[0163] The third unit is used to identify logical breakpoints by constructing a state feature matrix through eigenvalue decomposition based on the initial candidate risk chain set, search for potential transition events, and iteratively optimize to obtain a logically complete risk chain.

[0164] The fourth unit is used to extract transmission strength features based on the logically complete risk chain by using transmission coefficient matrix analysis and state transition matrix decomposition, calculate the credibility score, and obtain a risk chain credibility score set.

[0165] The fifth unit is used to select risk chains whose credibility scores exceed a preset credibility threshold based on the risk chain credibility score set, and to determine and output the final identification result.

[0166] A third aspect of the present invention provides an electronic device, comprising:

[0167] processor;

[0168] Memory used to store processor-executable instructions;

[0169] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0170] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0171] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0172] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying a city public safety risk chain based on a graph neural network, characterized in that, The method comprises the following steps: acquiring city public safety event historical data, taking different types of public safety events as graph nodes, and taking the association relationship between the public safety events as edges to construct an initial event association topology graph; based on the initial event association topology graph, an adaptive bidirectional jump sampling algorithm is adopted to determine an optimal sampling path set through dynamic step length control and time sequence matching; based on the event sequence in the optimal sampling path set, a multi-dimensional association strength is calculated to obtain an initial candidate risk chain set through screening; based on the initial candidate risk chain set, a state feature matrix is constructed through eigenvalue decomposition to identify logical breakpoints, search for potential transition events, and iteratively optimize to obtain a logically complete risk chain, comprising: the multi-dimensional association strength of the event sequence in the initial candidate risk chain set is constructed into a state feature matrix, and each element of the state feature matrix corresponds to the association state between adjacent events; the state feature matrix is subjected to eigenvalue decomposition to obtain an eigenvalue diagonal matrix and an eigenvector matrix, the feature state difference degree of an adjacent event pair is calculated to determine a breakpoint probability value, and the adjacent event pair with a breakpoint probability value greater than an adaptive breakpoint threshold is marked as a logical breakpoint; for the events before and after the logical breakpoint, an effective time window is constructed based on the corresponding time stamp, and a reachable range ellipse is constructed based on the corresponding geographical position; searching for potential transition events within the effective time window and the reachable range ellipse, calculating the time dimension contribution degree, the space dimension contribution degree and the semantic dimension contribution degree of the potential transition events and the events before and after the potential transition events, and performing weighted combination to obtain an event transition evaluation value; the potential transition event with the highest event transition evaluation value is taken as a supplementary event, which is inserted between the events before and after the logical breakpoint to generate an optimized event sequence; the state feature matrix of the optimized event sequence is reconstructed, and the above steps are repeated until there is no adjacent event pair marked as a logical breakpoint in the optimized event sequence, and a logically complete risk chain is obtained; based on the logically complete risk chain, a conduction strength feature is extracted by using a conduction coefficient matrix analysis and a state transition matrix decomposition, a credibility score is calculated, and a risk chain credibility score set is obtained; based on the risk chain credibility score set, a risk chain with a credibility score exceeding a preset credibility threshold is selected to determine and output a final identification result.

2. The method of claim 1, wherein, based on the initial event association topology graph, an adaptive bidirectional jump sampling algorithm is adopted to determine an optimal sampling path set through dynamic step length control and time sequence matching, comprising: obtaining the influence degree calculation parameters of the public safety event nodes from the initial event association topology graph, respectively calculating the propagation influence degree according to the node association degree and the distance attenuation factor, calculating the spatial influence degree according to the public safety event node wave area ratio, calculating the time influence degree according to the event time attenuation function, and weighting and combining the propagation influence degree, the spatial influence degree and the time influence degree to obtain the node influence force measurement; calculating the initial adaptive step length from the node influence force measurement and the benchmark sampling step length, obtaining the adaptive control threshold value from the historical difference of the node influence force measurement, and combining the initial adaptive step length and the adaptive control threshold value to obtain the dynamic sampling step length; The bidirectional sampling is performed by using a dynamic sampling step, forward propagation probability and backward verification probability are calculated respectively, and a time sequence matching score is obtained by weighted fusion; A path continuity index and a length constraint index are calculated according to the time sequence matching score, the time sequence matching score, the path continuity index and the length constraint index are nonlinearly weighted to obtain a path evaluation score, and an optimal sampling path set is obtained by screening a public safety event sequence with a path evaluation score greater than a preset evaluation threshold.

3. The method of claim 1, wherein, Based on the event sequence in the optimal sampling path set, a multi-dimensional correlation strength is calculated, and an initial candidate risk chain set is obtained by screening, including: An event sequence is obtained from the optimal sampling path set to construct a time sequence evolution graph, a time interval and an event intensive period of adjacent events are extracted, a key time node is identified, a time delay feature of event propagation is calculated by comparing adjacent event occurrence times, a periodicity of event occurrence is extracted by Fourier transform, a time correlation strength of adjacent events is calculated according to the time delay feature and the periodicity; A spatial distribution graph is constructed based on geographic coordinate information of each event in the event sequence, a spatial distance between adjacent events is calculated, and a spatial correlation strength is obtained by combining the spatial distance with an event spatial influence range; An event semantic vector is constructed using the description text of each event in the event sequence, a cosine similarity of adjacent event semantic vectors is calculated, and a semantic correlation strength is obtained by combining the cosine similarity with an event type matching matrix; Based on the influence measure value of each event in the event sequence, an influence correlation degree of adjacent events is calculated, and an influence correlation strength is obtained by combining the influence correlation degree with an influence transmission decay function; The time correlation strength, the spatial correlation strength, the semantic correlation strength and the influence correlation strength are weighted and fused to obtain the multi-dimensional correlation strength. The event sequence with a multi-dimensional correlation strength greater than a preset strength threshold is screened to obtain the initial candidate risk chain set.

4. The method of claim 1, wherein, Eigenvalue decomposition is performed on the state feature matrix to obtain an eigenvalue diagonal matrix and an eigenvector matrix, a feature state difference degree of an adjacent event pair is calculated, a fracture probability value is determined, and an adjacent event pair with a fracture probability value greater than an adaptive fracture threshold is marked as a logical fracture point, including: Eigenvalue decomposition is performed on the state feature matrix to obtain an eigenvalue diagonal matrix and an eigenvector matrix, and a three-order tensor is constructed, multi-scale tensor decomposition is performed on the three-order tensor to obtain an eigenvalue component matrix, an eigenvector component matrix and a time scale component matrix under different time scales; A multi-scale feature spectrum of the event sequence is constructed based on the eigenvalue component matrix, the eigenvector component matrix and the time scale component matrix, and the energy distribution of the multi-scale feature spectrum under different time scales is mapped to a feature entropy matrix; The feature state difference degree of the adjacent event pair is calculated according to the eigenvalue diagonal matrix and the eigenvector matrix, and an entropy difference vector of the adjacent event pair is calculated according to the feature entropy matrix; Local fluctuation features of the feature state difference degree and cross-scale correlation features of the entropy difference vector are extracted respectively, the local fluctuation features and the cross-scale correlation features are fused to obtain a comprehensive fracture feature, and a fracture probability value of the adjacent event pair is calculated based on the comprehensive fracture feature; An adaptive breaking threshold is constructed based on variance and entropy difference vector standard deviation of feature state difference of adjacent event pairs, and adjacent event pairs with a breaking probability value greater than the adaptive breaking threshold are marked as logical breaking points.

5. The method of claim 1, wherein, Based on the logical complete risk chain, a conduction intensity feature is extracted by using conduction coefficient matrix analysis and state transition matrix decomposition, a credibility score is calculated, and a risk chain credibility score set is obtained, including: For the logical complete risk chain, a conduction coefficient matrix is constructed based on the fluctuation amplitude, propagation distance and influence range of adjacent events, and a risk conduction intensity sequence is calculated through the time evolution of the conduction coefficient matrix; For the risk conduction intensity sequence, a state transition matrix is constructed, the state transition matrix is subjected to eigenvalue decomposition, the eigenvalues are calculated and determined, events with a product of eigenvalues and influence range exceeding a preset conduction threshold are determined as diffusion conduction nodes, and an energy diffusion index of the diffusion conduction nodes is calculated; The risk conduction intensity sequence is divided into multiple time scales, a fluctuation feature sequence is calculated at each time scale, the fluctuation feature sequence is subjected to segmented iterative reconstruction to obtain a fractal sequence, and a complexity index of risk conduction is calculated through self-similarity of the fractal sequence; Based on the risk conduction intensity sequence and the fluctuation feature sequence, a three-dimensional phase space including conduction intensity, fluctuation amplitude and time delay is constructed, a phase space trajectory is extracted, and Lyapunov exponents of the phase space trajectory at different time scales are calculated; A three-dimensional feature vector is constructed based on the energy diffusion index, the complexity index and the Lyapunov exponents, a conduction stability score of the logical complete risk chain is calculated by using a pre-trained adaptive weight network, and a risk chain credibility score set is obtained by probability distribution transformation of the conduction stability score.

6. The method of claim 5, wherein, Segmented iterative reconstruction of the fluctuation feature sequence to obtain a fractal sequence includes: Obtaining a fluctuation feature sequence, processing the fluctuation feature sequence by using an overlapping segmentation method, determining the number of overlapping points of adjacent segments by a preset segmentation length and an overlap rate, and generating multiple overlapping sequence segments; Extracting fluctuation features of each overlapping sequence segment, the fluctuation features including fluctuation amplitude and fluctuation trend, and pairing adjacent overlapping sequence segments two by two according to the fluctuation features; For each pair of adjacent overlapping sequence segments, constructing a segmented mapping rule based on the fluctuation features, taking the fluctuation amplitude ratio of adjacent overlapping sequence segments as an amplitude ratio parameter, taking the fluctuation trend difference as a trend difference parameter, and generating a segmented linear mapping function; Iterative reconstruction of overlapping sequence segments, in each iteration process, dynamically adjusting the amplitude ratio parameter and the trend difference parameter based on the feature change amount of the current overlapping sequence segment, and applying them to the next iterative reconstruction to generate a reconstructed sequence segment; Sequentially splicing and smoothing the reconstructed sequence segments to obtain a fractal sequence.

7. A city public security risk chain identification system based on a graph neural network, for implementing the method of any one of the preceding claims 1-6, characterized in that, The first unit is configured to obtain city public security event historical data, take different types of public security events as graph nodes, take the correlation between public security events as edges, and construct an initial event correlation topology graph. ​ A second unit is configured to determine an optimal sampling path set by dynamic step control and time sequence matching based on the initial event correlation topology graph and using an adaptive bidirectional skip-sampling algorithm; and calculate a multi-dimensional correlation strength based on an event sequence in the optimal sampling path set, and screen to obtain an initial candidate risk chain set. A third unit is configured to identify a logical breaking point by constructing a state feature matrix through eigenvalue decomposition based on the initial candidate risk chain set, search for a potential transition event, and iteratively optimize to obtain a logically complete risk chain. A fourth unit is configured to extract a conduction strength feature by using conduction coefficient matrix analysis and state transition matrix decomposition based on the logically complete risk chain, calculate a credibility score, and obtain a risk chain credibility score set. A fifth unit is configured to select a risk chain with a credibility score exceeding a preset credibility threshold based on the risk chain credibility score set, and determine and output a final identification result.

8. An electronic device, comprising: The computer program instructions are executed by the processor to implement the method of any one of claims 1-6. The computer program instructions are executed by the processor to implement the method of any one of claims 1-6. ​ ​ 9. A computer-readable storage medium having stored thereon computer program instructions, wherein, ​

Images