Urban public security intelligent supervision method based on multi-source data fusion

Abstract

This invention relates to the field of public safety supervision technology and discloses an intelligent urban public safety supervision method based on multi-source data fusion. The method includes: acquiring the original sequence of crowd density and trajectory dwell time parameters to calculate a comprehensive characteristic value of crowd aggregation in physical space; acquiring the network information text sequence to calculate a characteristic value of the instantaneous negative sentiment propagation rate in cyberspace; constructing a cyber-physical cross-domain coupled calculation model based on the two characteristic values ​​and calculating a comprehensive public safety supervision risk index; comparing and determining multi-level control boundary points based on the comprehensive public safety supervision risk index and executing anti-shake and fault-tolerant logic to generate administrative intervention instructions; and finally converting these instructions into standardized two-way supervision strategy messages and distributing them to terminal execution units. This invention establishes a heterogeneous data coupling mechanism to avoid response delays in sudden situations, eliminate false alarms caused by data fluctuations, and achieve an adaptive, end-to-end supervision and control closed loop.

Description

Technical Field

[0001] This invention relates to the field of public safety supervision technology, specifically to an intelligent urban public safety supervision method based on multi-source data fusion. Background Technology

[0002] Urban public safety supervision is a crucial component of modern urban governance. With accelerating urbanization and increased population mobility, abnormal clustering phenomena and their associated risks within urban grids are becoming increasingly complex. Intelligent public safety supervision aims to utilize information technology for real-time monitoring, risk assessment, and intervention scheduling of urban operations. This field encompasses physical space IoT sensor network sensing, such as acquiring crowd density and dwell time through video surveillance and radio frequency identification (RFID), as well as cyberspace data mining, such as extracting sentiment characteristics from online information using natural language processing (NLP). Public safety supervision is gradually evolving from a single-dimensional, passive response to proactive intervention through cross-domain cyber-physical interaction.

[0003] Current technologies in this field primarily employ single-domain monitoring schemes that separate physical and cyberspace monitoring. Physical space monitoring mainly relies on video surveillance equipment and radio frequency (RF) probes. Target detection algorithms analyze video streams to calculate crowd density, or RF signals are used to count the dwell time of mobile terminals, triggering a gridded early warning strategy based on a set single physical threshold. Cyberspace monitoring involves deploying data acquisition engines to obtain text sequences from social media platforms, using natural language processing (NLP) algorithms to extract sentiment words and score their emotional polarity, thus monitoring the instantaneous spread of negative emotions. Some comprehensive regulatory platforms combine these two independent monitoring results at the display layer to assist administrative personnel in scheduling decisions.

[0004] However, existing technologies still have shortcomings in processing multi-source heterogeneous data. On the one hand, because existing solutions mostly focus on data display layer stitching, they fail to construct a cross-domain coupled calculation model at the underlying feature dimension, combining the comprehensive feature values ​​of physical space crowd gathering with the feature values ​​of instantaneous negative emotion propagation rate in cyberspace. Insufficient cross-domain data fusion between physical space monitoring and cyberspace early warning leads to insufficient lead time for comprehensive public safety risks. This can easily cause response delays when facing nonlinear abrupt changes in the simultaneous occurrence of crowd gathering and negative emotions, or cause false alarms due to instantaneous data fluctuations. On the other hand, the single-dimensional static threshold determination mechanism used in existing technologies is difficult to adapt to the fault tolerance requirements in complex and non-stationary environments. Therefore, it is urgent to establish an administrative supervision hierarchical triggering mechanism based on multi-source heterogeneous data tension calculation and anti-shake fault tolerance logic to achieve a closed-loop public safety supervision and control system with adaptive execution evolution. Summary of the Invention

[0005] To address the problems in related technologies, this invention provides an intelligent urban public safety supervision method based on multi-source data fusion, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0006] To solve the aforementioned technical problem, the present invention is achieved through the following technical solution: This invention provides an intelligent urban public safety monitoring method based on multi-source data fusion, specifically including: Obtain the original sequence of crowd density and trajectory dwell time parameters, and calculate the comprehensive characteristic value of crowd aggregation in physical space; obtain the network information text sequence and extract the text that has a geographic identifier mapping relationship with the target city grid, and calculate the characteristic value of the instantaneous negative emotion propagation rate in cyberspace; Based on the comprehensive characteristic value of crowd gathering in the physical space and the characteristic value of the instantaneous negative emotion propagation rate in the cyberspace, a cyber-physical cross-domain coupled calculation model is constructed to calculate the comprehensive public safety supervision risk index. Based on the comprehensive public safety supervision risk index, multi-level control boundary point comparison and judgment are performed, and anti-shake fault-tolerant logic is executed to generate multi-level boundary point comparison and judgment results and corresponding administrative intervention instructions. The administrative intervention instructions are converted into standardized two-way regulatory strategy messages and distributed to the corresponding terminal execution units.

[0007] As a preferred embodiment of the intelligent urban public safety monitoring method based on multi-source data fusion described in this invention, the acquisition of the original population density sequence, trajectory dwell time parameters, and network information text sequence includes: Extract the number of human target bounding boxes from the real-time video stream returned by the video surveillance equipment as the original sequence of crowd density. Collect the timestamp of the first detection and the last detection of the same physical address, calculate the time difference and perform an arithmetic average to obtain the trajectory dwell time parameter; Geographical entity information and GPS coordinate labels contained in the network information text stream are extracted and compared with the electronic fence coordinate boundaries of the target city grid to obtain the network information text sequence.

[0008] As a preferred embodiment of the intelligent urban public safety monitoring method based on multi-source data fusion described in this invention, the calculation of the comprehensive characteristic value of physical space crowd aggregation includes: Extract the physical upper limit and lower limit of population density, process the original population density sequence based on the maximum and minimum value normalization method with boundary truncation, and obtain density normalized feature values; The upper and lower limits of trajectory dwell time are extracted, and the trajectory dwell time parameters are processed based on the maximum and minimum value normalization method with boundary truncation to obtain the dwell time normalized feature value. A linear combination mechanism is used to fuse density-normalized eigenvalues ​​and dwell time-normalized eigenvalues ​​to calculate the comprehensive eigenvalues ​​of crowd aggregation in physical space.

[0009] As a preferred embodiment of the intelligent urban public safety supervision method based on multi-source data fusion described in this invention, the calculation of the characteristic value of the instantaneous negative emotion propagation rate in cyberspace includes: A joint sentiment dictionary integrating a general sentiment lexicon and a proprietary sentiment lexicon, as well as a negation word list containing core negation words, are constructed. The original sentiment weighted score of the network information text sequence is calculated based on the preset word window length and the polarity reversal factor based on the number of times the negation word is hit. The negative polarity score of a single text is obtained by combining an asymmetric sentiment excitation mapping function for directional nonlinear transformation. Texts with a negative polarity score greater than the threshold for determining negative emotion are considered as valid negative information samples. The negative polarity scores of individual texts of valid negative information samples are accumulated to obtain the instantaneous absolute increment of negative emotion in the network space. We extract the upper limit of the instantaneous propagation capacity of negative emotions in the network, use a normalization method with boundary truncation to process the absolute increment of instantaneous new negative emotions in the network space, and calculate the characteristic value of the instantaneous negative emotion propagation rate in the network space.

[0010] As a preferred embodiment of the intelligent urban public safety supervision method based on multi-source data fusion described in this invention, the method includes constructing the cyber-physical cross-domain coupled computational model to calculate the comprehensive public safety supervision risk index, including: A fixed-length sliding historical queue based on the maximum physical boundary is constructed, and the smoothed cumulative amount of the feature value of the instantaneous negative emotion propagation rate in the network space within a specific time window is extracted by discrete time series operation. A cyber-physical cross-domain coupled computational model is constructed, and the numerical sum of the physical space basic cumulative term, the network sentiment smoothing cumulative term, and the nonlinear coupling term is combined to output a comprehensive public safety supervision risk index. The physical space basic accumulation term is determined by combining the physical space crowd aggregation comprehensive feature value with the physical aggregation basic weight coefficient. The network emotion smoothing accumulation term is determined by combining the smoothing accumulation amount with the network emotion basic weight coefficient. The nonlinear coupling term is determined by combining the physical space crowd aggregation comprehensive feature value with the network space instantaneous negative emotion propagation rate feature value with the coupling mutation weight coefficient and mutation sensitivity calibration parameter through exponential function calculation.

[0011] As a preferred embodiment of the intelligent urban public safety supervision method based on multi-source data fusion described in this invention, after constructing the cyber-physical cross-domain coupled computation model, a weighted composite loss function is constructed using a historical concurrent security event dataset for multi-objective joint optimization, including: An adaptive feature weight based on information entropy is used to construct a parameter-free baseline intensity sequence, which maps discrete treatment levels to continuous target values, generating a real target value sequence containing discrete time steps. The mean squared error algorithm is used to quantify the mean squared error between the true target numerical sequence and the predicted risk index sequence. The recall rate of sudden high-risk events is calculated by comparing the number of true positive samples with the total number of positive samples, and the prediction accuracy of normal monitoring periods is calculated by comparing the number of true negative samples with the total number of normal negative samples. The mean squared error, recall, and prediction precision are combined, and a weighted composite loss function is constructed by combining preset error penalty weight coefficients, recall weight coefficients, and precision weight coefficients. A dataset is constructed by extracting positive and negative samples using an event peak alignment strategy based on the physical dissipation time constant. The optimal parameter set of the proposed cyber-physics cross-domain coupled computation model is then solved by combining reparameterization techniques with the stochastic gradient descent algorithm with momentum.

[0012] As a preferred embodiment of the intelligent urban public safety supervision method based on multi-source data fusion described in this invention, the multi-level control boundary point comparison and determination based on the comprehensive public safety supervision risk index includes: Set warning slope thresholds and mutation slope thresholds; use backward difference algorithm to process the comprehensive public safety supervision risk index at the current time and the comprehensive public safety supervision risk index at the previous time, and calculate the risk index change rate; When the rate of change of the risk index is greater than the warning slope threshold but less than or equal to the mutation slope threshold, a pre-mutation warning is triggered, and the continuous sampling judgment time step is dynamically shortened according to the relative position of the rate of change within the threshold range. When the rate of change of the risk index is greater than the threshold of the mutation slope, it is determined that the system has entered a state of instantaneous high risk. The system bypasses the continuous sampling judgment logic and confirms that the corresponding high-level early warning strategy has been triggered.

[0013] As a preferred embodiment of the intelligent urban public safety supervision method based on multi-source data fusion described in this invention, the execution of the anti-shake and fault-tolerant logic to generate administrative intervention instructions includes: Construct a finite state machine for early warning intervals, and set the state codes for normal intervals, first-level early warning intervals, second-level early warning intervals, and third-level early warning intervals. Compare the target status code at the current moment with the historical status code and execute the state transition function. When the target status code meets the conditions for a leap to a higher level or a fall back within a certain range, reset the continuous sampling counter. When the target status code meets the conditions for maintaining the same level, control the continuous sampling counter to perform an increment operation. When the target status code meets the conditions for a normal reset, clear the continuous sampling counter to zero. When the value of the continuous sampling counter is greater than or equal to the continuous sampling judgment time step and the target status code remains within a specific warning range, an administrative intervention instruction corresponding to the gridded degradation warning strategy is output.

[0014] As a preferred embodiment of the intelligent urban public safety supervision method based on multi-source data fusion described in this invention, the method of converting the administrative intervention instruction into a standardized two-way supervision strategy message and distributing it to the corresponding terminal execution unit includes: Extract the event's geographic location coordinates and the unique identifier of the bound controlled object, package them using a serialization protocol, and reconstruct them into a standardized two-way regulatory strategy message containing basic business fields for downgrade warning level code, and add execution status feedback fields, dynamic reconfiguration fields, and execution effect quantification fields to the end of the message structure; Extract the unique identifier of the controlled object from the standardized two-way monitoring policy message, perform device type comparison and address resolution, establish an underlying communication routing channel, and distribute the standardized two-way monitoring policy message to the corresponding terminal execution unit.

[0015] As a preferred embodiment of the intelligent urban public safety supervision method based on multi-source data fusion described in this invention, after the standardized two-way supervision strategy message is distributed to the corresponding terminal execution unit, the method further includes: The terminal execution unit analyzes the status code returned by the execution status feedback field and obtains the actual change in dwell time and the actual change in negative sentiment through differential calculation. Search the baseline feature database of the same historical time period, extract the natural time-series fluctuation sequence data that is not affected by administrative intervention, and calculate the predicted change in dwell time and the predicted change in negative sentiment based on the historical baseline of the same period. The dwell time residual is calculated based on the actual change in dwell time and the predicted change in dwell time. The negative emotion residual is calculated based on the actual change in negative emotion and the predicted change in negative emotion. The dwell time residual and negative emotion residual that meet the confidence requirements are written into the execution effect quantification field. Using natural days as the cycle, when the data in the execution effect quantification field is determined to be non-empty, the execution effect quantification field data and the feature data of the corresponding historical time input to the cyber-physics cross-domain coupled computing model are extracted to construct an incremental fine-tuning dataset. The calibration parameters are incrementally updated using the mini-batch gradient descent method based on online fine-tuning loss function, fine-tuning learning rate and convergence criterion.

[0016] The present invention has the following beneficial effects: 1. This invention calculates the comprehensive characteristic value of physical space crowd gathering by obtaining the original sequence of crowd density and trajectory dwell time parameters, and calculates the characteristic value of instantaneous negative emotion propagation rate in cyberspace by obtaining the network information text sequence. Based on the comprehensive characteristic value of physical space crowd gathering and the characteristic value of instantaneous negative emotion propagation rate in cyberspace, a cyber-physical cross-domain coupled calculation model is constructed, and a comprehensive public safety supervision risk index is output. By establishing a heterogeneous data coupling mechanism of the underlying feature dimension, the problem of insufficient cross-domain integration of physical space monitoring and cyberspace early warning data and data fragmentation is solved.

[0017] 2. This invention calculates the rate of change of the comprehensive public safety supervision risk index at the current moment compared to the historical risk index at the previous moment. This rate of change is then compared to warning slope thresholds and abrupt change slope thresholds. When the rate of change of the risk index exceeds the abrupt change slope threshold, the target city grid is determined to be in a transient high-risk state, and the continuous sampling judgment logic is bypassed. Through these technical means, a dynamic parameter reflecting the speed of situation deterioration is provided for nonlinear abrupt changes characterized by both high physical aggregation and high network sentiment, avoiding response delays in sudden events and addressing the problem of insufficient advance warning time for comprehensive public safety risks.

[0018] 3. This invention compares and judges the comprehensive public safety supervision risk index based on set multi-level control boundary points. It then uses a finite state machine within the early warning interval to execute anti-shake and fault-tolerant logic, generating multi-level boundary point comparison and judgment results and corresponding administrative intervention instructions. These instructions are transformed into standardized two-way regulatory strategy messages and distributed to the corresponding terminal execution units. Furthermore, confidence interval thresholds are extracted based on the Raida criterion to quantify the execution effect and drive online fine-tuning of model parameters. By combining dynamically set multi-level control boundary points with continuous sampling and judgment time steps, it eliminates system false alarms caused by instantaneous data fluctuations, adapts to fault-tolerant requirements in complex and non-stationary environments, and achieves a closed-loop public safety supervision control system with adaptive execution evolution.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0021] Figure 1The flowchart of the intelligent urban public safety supervision method based on multi-source data fusion is provided by the present invention.

[0022] Figure 2 This is a flowchart illustrating step S3 of the present invention. Detailed Implementation

[0023] 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. Example 1:

[0024] Existing urban public safety supervision often faces the challenge of insufficient cross-domain data integration. To address the problem of insufficient early warning time caused by the fragmentation of data between physical and cyberspace, this invention provides a hierarchical triggering mechanism for administrative supervision based on the tension calculation of multi-source heterogeneous data. Example 1 of this invention provides an intelligent urban public safety supervision method based on multi-source data fusion. Figure 1 As shown, in the specific implementation process, the original sequence of crowd density and trajectory dwell time parameters are obtained, the comprehensive feature value of crowd aggregation in physical space is calculated, the network information text sequence is obtained, and the feature value of instantaneous negative emotion propagation rate in cyberspace is calculated. Based on the comprehensive feature value of crowd aggregation in physical space and the feature value of instantaneous negative emotion propagation rate in cyberspace, a cyber-physical cross-domain coupled calculation model is constructed to output a comprehensive public safety supervision risk index. Based on the calculated comprehensive public safety supervision risk index, multi-level control boundary point comparison and judgment are performed, and anti-shake fault-tolerant logic is executed to generate multi-level boundary point comparison and judgment results and corresponding administrative intervention instructions. The generated administrative intervention instructions are converted into standardized two-way supervision strategy messages and distributed to the corresponding terminal execution units, thereby realizing a closed-loop public safety supervision control with adaptive execution evolution. In this invention, cyber-physical refers to a Cyber-Physical system, which covers the deep integration and interaction of cyberspace and physical space.

[0025] Furthermore, the above-mentioned intelligent urban public safety supervision method based on multi-source data fusion is explained in detail, including the following: S1. Obtain multi-source raw security monitoring data within the target city grid and extract features. This includes the following sub-steps: S11. Collect raw crowd density sequences through an IoT sensor network deployed in physical space. With trajectory dwell time parameter By deploying a data acquisition engine in cyberspace, it acquires network information text sequences that have a geographic identifier mapping relationship with the target city grid. It performs time-domain alignment calculations on multi-source data based on an asynchronous acquisition architecture. Specifically, it includes the following sub-steps: S111. The edge gateway connects to video surveillance equipment and wireless RF probes deployed within the target city grid via a preset interface protocol. The edge gateway's built-in target detection algorithm model analyzes the real-time video stream transmitted from the video surveillance equipment frame by frame, extracting the number of human target bounding boxes as the original sequence of crowd density at the current moment. The wireless radio frequency probe periodically scans the medium access control layer physical addresses of mobile terminals within the grid. The system records the timestamp of the first and last detection of the same physical address, and calculates the time difference between the two as the dwell time of a single individual. The system then performs an arithmetic average of the dwell times of all detected individuals within the statistical time window to obtain the trajectory dwell time parameter. .

[0026] S112. The cloud-based data acquisition engine periodically pulls data streams from the target social media platform via an application programming interface (API). It calls a named entity recognition algorithm library to extract geographic location entity information and GPS coordinate labels from the text stream. The extracted geographic location entity information and coordinate labels are then compared with the geofence coordinate boundaries of the target city grid using spatial polygon mapping. Data records within the geofence are filtered and retained, while out-of-domain noise data is removed, resulting in a network information text sequence with a strict geographic identification mapping relationship to the target city grid. .

[0027] S113, the IoT sensor network and cyberspace data acquisition engine adopts an asynchronous acquisition architecture, unifying the physical time base of different data sources through a time-domain alignment mechanism. Specifically: Define a globally uniform basic computation step size This involves coordinating the heterogeneous sampling frequencies of video surveillance equipment, wireless RF probes, and cyberspace data acquisition engines. It also addresses the high-frequency sampling frequency of video surveillance equipment. Scan cycle of wireless radio frequency probe When both are equal to the basic calculation step size In case of mismatch, within a single basic computation step size Multiple sample values ​​are extracted within the corresponding time window, and the max pooling aggregation algorithm is used to transform them into a single representative value, controlling the original sequence of population density. With trajectory dwell time parameter Time granularity and basic computation step size Strict alignment.

[0028] Determine the text sequence of network information pulled by the network space data acquisition engine. The state within the interval between two adjacent data collections. Network information text sequence. It has the event increment attribute, if the current basic calculation step size is... No actual data collection occurred within the network information text sequence. Empty sets are filled with zeros; only when new data arrives are the entire data of the current batch assigned the corresponding basic calculation step size. Network information text sequence In this study, a pulsed update mechanism is adopted to replace the forward filling mechanism. This is combined with the integral smoothing operation in the downstream cyber-physics cross-domain coupling calculation model to process the transient pulse and eliminate the cumulative distortion phenomenon in the calculation process of the absolute increment of feature quantities.

[0029] Determine the integrity of the physical sensor network data. If at any basic computation step... No original population density sequence was received. Or trajectory dwell time parameter The newly acquired values ​​are used to complete the missing data by applying a linear interpolation algorithm combined with adjacent historical data. The system accumulates the number of consecutively missing data points and sets the maximum tolerance blind zone time parameter for the target city grid. Calculate the preset hardware alarm threshold The specific calculation formula is as follows: When the number of consecutive missing items exceeds the preset hardware alarm threshold. When this occurs, the device disconnection alarm mechanism of the edge gateway is triggered, and the risk calculation process of the target city grid is suspended.

[0030] S12. The obtained original sequence of population density With trajectory dwell time parameter Perform maximum-minimum normalization with boundary truncation, and calculate density-normalized eigenvalues. Normalized eigenvalues ​​of dwell time It also integrates and calculates the comprehensive characteristic values ​​of crowd aggregation in physical space. Specifically, it includes the following sub-steps: S121. Read the spatial geometric attributes and historical statistical data of the target city grid, and extract the physical upper limit value of population density. Physical lower limit of population density Extract the original sequence of population density within the calculation step size. The density-normalized eigenvalues ​​are calculated using the maximum-minimum normalization method with boundary truncation. The specific calculation formula is as follows: ; In the formula, This indicates the operation of finding the maximum value. This indicates the minimum value operation. This calculation process eliminates the differences in physical dimensions caused by variations in the target city grid size, and ensures that the output characteristic value is strictly limited to a minimum value under extreme congestion conditions through bilateral limit amplitude control. Interval.

[0031] As one possible implementation method, the physical upper limit of crowd density Physical lower limit of population density The calculation method is as follows: obtain a two-dimensional planar map of the target city grid, and extract the total physical area of ​​the grid. Identify and remove non-travelable areas within the grid. Of which, the area of ​​non-traffic areas Includes green belts, buildings, water areas, and fixed obstacles. Calculate the effective accessible area. The specific calculation formula is as follows: Set the physical lower limit value for crowd density. Based on the safety design specifications for urban public places, the constant of the maximum projected area for a single person standing is extracted. Calculate the physical upper limit of population density The specific calculation formula is as follows: .

[0032] S122. Read the target city grid's preset security control configuration table and extract the maximum value of trajectory dwell time. With the lower limit of trajectory dwell time Extract trajectory dwell time parameters within the calculation step. The normalized eigenvalues ​​of dwell time are calculated using the maximum-minimum normalization method with boundary truncation. The specific calculation formula is as follows: ; This calculation process converts physical parameters with time dimensions into dimensionless physical characteristic parameters, while ensuring that the output strictly converges to the upper limit value even under specific extreme conditions such as severe aggregation and stagnation. .

[0033] As one possible implementation method, the security control configuration table is constructed as follows: Extract the historical trajectory dwell time records of all individual entities within the target city grid over the past full calendar year. Based on the planning attributes of the target city grid, classify it into either a through-traffic grid or a lingering grid. For through-traffic grids, set the theoretical shortest travel time as the lower limit of the trajectory dwell time. For persistent grids, the lower limit of trajectory dwell time is set. Set as The probability density function of the historical residence time distribution is constructed using the kernel density estimation method, and the confidence quantile of the danger clustering time is set. Calculate the confidence quantile corresponding to the probability density function. The time boundary value is added to the preset emergency response buffer time. As the upper limit of trajectory dwell time The calculated lower limit of the trajectory dwell time. With the upper limit of trajectory dwell time Write it into the security management configuration table.

[0034] S123. Obtain density-normalized eigenvalues Normalized eigenvalues ​​of dwell time A linear combination mechanism was used to calculate the comprehensive characteristic value of crowd aggregation in physical space. The specific calculation formula is as follows: ; In the formula, This represents density-independent physical calibration parameters. This represents the independent physical calibration parameter for dwell time. This step completes the heterogeneous data alignment at the feature level, constructs a white-box feature extraction structure, and outputs physical space tension parameters that are strictly associated with the cross-domain coupled computational model of cyberphysics.

[0035] S13. Receive the time-domain aligned network information text sequence. A natural language processing algorithm based on matching a sentiment dictionary and a negation word list is used to calculate the sentiment polarity score of the text. The instantaneous negative sentiment propagation rate feature value in cyberspace is then calculated by combining the number of instantaneously added negative information. This involves aligning heterogeneous data at the feature level. Specifically, it includes the following sub-steps: S131. Establish a joint sentiment dictionary that integrates a general sentiment lexicon with a sentiment lexicon specific to the field of urban public safety, and assign an initial sentiment score to each word in the dictionary. The general sentiment lexicon uses the HowNet 2022 general sentiment dictionary, which contains 91,032 basic entries for positive evaluation, negative evaluation, positive sentiment, and negative sentiment. The specialized sentiment lexicon includes vocabulary derived from crisis events specific to the public safety field. The initial sentiment score for negative sentiment vocabulary is set to meet certain criteria. Establish a negation word list, which contains core negation words, specifically including "not", "no", "not", "without", "non-", "absolutely not", and "not".

[0036] Traverse the network information text sequence within the current calculation step. The algorithm extracts sentiment terms from individual texts using natural language segmentation. For each extracted sentiment term, a fixed-scope negation word flipping algorithm is executed: based on the position of the sentiment term in the text string, a preset word window length is truncated towards its beginning. As a negation word's scope (preferably, (Set as 3 word segmentation units); within this scope, scan and match the negative word list, and count the specific number of times a negative word is hit. Introducing a polarity reversal factor Process double and multiple negation grammatical structures in the text, when the number of hits... When the number is odd, polarity is reversed; when the number of hits... When the number is even (e.g., "had to"), the original polarity is maintained, controlling the precise reversal of sentiment polarity. The original sentiment-weighted score for each individual text is calculated. The specific calculation formula is as follows: ; In the formula, This represents the total number of sentiment words that were successfully matched in a single text.

[0037] Furthermore, to adapt to the non-linear burst characteristics of negative emotions in sudden high-risk events under urban public safety supervision scenarios, an asymmetric emotion evoked mapping function is used to weight the original emotion score. Perform directional nonlinear transformation. Set the sensitivity coefficient for emotional arousal during sudden safety events. Calculate the negative polarity score of a single text. The specific calculation formula is as follows: ; In the formula, Represented by natural constant An exponential function with base 0. The value range is [0.5, 2.0]. This calculation mechanism is based on the original sentiment-weighted score. Negative text with a value less than 0 undergoes asymmetric processing, directly truncating non-negative weak noise and positive scores to zero. When cyberspace exhibits extreme negative security sentiment, i.e., the original sentiment-weighted score... When approaching negative infinity, the negative polarity score of a single text. It strictly and monotonically converges to the upper limit of 1, which can achieve exponential amplification of high-risk negative features and filter out the influence of normal emotional fluctuations.

[0038] S132. Set the threshold for judging the extraction of negative emotions. To effectively filter neutral sentiment and weak negative noise data and eliminate false alarms caused by half-scores, a threshold for determining negative sentiment arousal is established. The value of is strictly limited to Preferably, the threshold for determining negative emotion arousal. The system scores negative polarity based on historical routine information text of the target city grid. Quantiles are obtained through offline calibration.

[0039] The system filters the negative polarity scores of individual texts within the current calculation step. Greater than the threshold for determining negative emotion The text is taken as a valid negative information sample, and the negative polarity score of each text in the single text of all valid negative information samples within the current calculation step is calculated. Accumulate the data to obtain the absolute increase in negative sentiment in the current cyberspace at any given moment. .

[0040] S133. Read the local storage network space historical statistical extreme value parameter library to obtain the upper limit value of the instantaneous propagation capacity of negative sentiment in the network. This upper limit is determined offline by the 99th percentile of the instantaneous increase in absolute negative sentiment within the historical statistical period of the target city's grid, extracted by the system, to filter out noise from extreme and anomalous data. A linear normalization method with boundary truncation is used to calculate the characteristic value of the instantaneous negative sentiment propagation rate in cyberspace. The specific calculation formula is as follows: ; The physical space crowd aggregation comprehensive characteristic value calculated and output in step S12 Compared with the feature value of instantaneous negative sentiment propagation rate in cyberspace output in this step Extracted to the underlying shared memory pool. And the comprehensive feature value of physical space crowd aggregation. Characteristic value of instantaneous negative emotion propagation rate in cyberspace All values ​​are strictly limited to Dimensionless standardized characteristic parameters of the interval.

[0041] S2. Based on the extracted comprehensive feature values ​​of crowd aggregation in physical space Characteristic value of instantaneous negative emotion propagation rate in cyberspace A cyber-physical cross-domain coupled computational model is constructed to output a comprehensive public safety supervision risk index. Specifically, it includes the following sub-steps: S21. Construct a cyber-physical cross-domain coupled computational model to calculate and output the comprehensive public safety supervision risk index. Specifically, it includes the following sub-steps: S211. Synchronously extract the current basic computation step size from the underlying shared memory pool. Comprehensive characteristics of physical space crowd aggregation Characteristic value of instantaneous negative emotion propagation rate in cyberspace A fixed-length sliding history queue is constructed in memory based on pre-allocation of the maximum physical boundary, and a maximum tolerable time delay boundary is set. Calculate the maximum historical queue length The specific calculation formula is as follows: In the formula, This indicates rounding to the nearest integer. The system assigns each basic calculation step size... The instantaneous negative sentiment propagation rate characteristic value in cyberspace Push it into the fixed-length sliding history queue.

[0042] Discrete-time series computation is used to extract the smoothed cumulative value of network sentiment features within a specific time window. Under online real-time monitoring, the optimal time hysteresis constant, which is fixed through offline solution, is extracted. Calculate the actual operation window length The specific calculation formula is as follows: Extract the most recent history from the fixed-length sliding history queue. The arithmetic mean is calculated for each sample data; in offline multi-objective joint optimization training, the time lag constant variable within the current iteration period is used as the basis. The generated length is The continuous truncation mask vector is used to perform dot product and integral operations with a fixed-length sliding historical queue to dynamically fit the smooth cumulative amount of network sentiment features within a specific time window.

[0043] S212, Comprehensive characteristic values ​​of crowd aggregation based on physical space The characteristic value of the instantaneous negative emotion propagation rate in cyberspace And based on the smoothed cumulative amount of network sentiment features obtained from discretization calculation, the information-physics coupling resonance mechanism from sociophysics is introduced to construct a cyber-physics cross-domain coupling calculation model. The specific expression is: ; In the formula, express The comprehensive public safety supervision risk index for a specific grid area at any given time. This represents the basic weighting coefficient of physical aggregation. This represents the basic weighting coefficient of online sentiment. Represents the coupling mutation weight coefficient. Represents the time hysteresis constant. This represents the time scale variable used for integration. This represents the mutation sensitivity calibration parameter.

[0044] This is a physical space-based cumulative item, reflecting the physical security risk benchmark formed by the state of population gathering within the target city grid. This is a smoothed cumulative term for online sentiment, combined with the basic weighting coefficient of online sentiment. Eliminate fluctuation noise in instantaneous data.

[0045] This is a nonlinear coupling term, constructed based on group chain reaction dynamics. In public safety scenarios, the high density of physical space shortens the topological distance of information propagation, increasing the probability of collisions in group emotional contagion; negative emotions in cyberspace rapidly penetrate dense crowds through mobile terminals, significantly reducing the group's physical safety tolerance threshold, with both acting as catalysts. Therefore, [the following is a continuation of the previous sentence, but the context is unclear]. and product term Established as the effective interaction frequency in cyber-physics, it is used to characterize the propagation kinetic energy of negative security information in dense physical networks, combined with the natural constant. Mapping to an exponential function with a base can accurately depict the avalanche-like physical process in which mass panic breaks through the system's critical point due to local resonance in a crowded environment, thus triggering qualitative changes (such as stampedes or mass riots).

[0046] The sum of the physical space basic accumulation term, the network sentiment smoothing accumulation term, and the nonlinear coupling term is combined and calculated, and the current basic calculation step size is output. Comprehensive public safety supervision risk index And this comprehensive public safety supervision risk index Write to the real-time decision buffer channel to drive the dynamic comparison and determination of multi-level control boundary points in the downstream administrative regulatory decision tree.

[0047] S22. Joint fitting and calibration uses historical data of security events from the same period to construct a weighted composite loss function. Perform multi-objective joint optimization. This includes the following sub-steps: S221. Extract real handling records from the historical data set of security incidents from the same period, and construct a mapping rule from discrete handling levels to continuous target values. Map records of normal periods without triggering any control measures to normal target values. Event records that trigger Level 1, Level 2, and Level 3 control measures will be mapped to preset level threshold constants, respectively. , , The system iterates through all security event records in the historical dataset and generates a true target numerical sequence containing all discrete time steps. Establish a benchmark reference system for model fitting.

[0048] As one possible implementation method, the grade threshold constant , , The method for determining it is as follows: S2211, The system extracts density-normalized feature values ​​from historical datasets. , normalized eigenvalues ​​of dwell time and the characteristic value of the instantaneous negative sentiment propagation rate in cyberspace. An adaptive weight allocation algorithm based on information entropy is introduced to dynamically calculate feature weights according to the dispersion of the historical distribution of various data within the target city grid. For the historical all-weather time series, the total number of historical data samples is set to... Set feature index Corresponding to , , Calculate the first... Item feature in the first Feature proportions at each time step The specific calculation formula is as follows: ;In the formula, Indicates the first Item feature in the first The specific values ​​at each time step, and to avoid meaningless calculations in logarithmic operations, when At that time, the settings are smoothly configured according to the limit arithmetic rules. Feature extraction weight Calculate the first Information entropy of a feature With adaptive feature weights The specific calculation formula is as follows: , Adaptive feature weights obtained through dynamic calculation are used. , , Constructing a parameter-free baseline intensity sequence The specific expression is: In the formula, , , All values ​​are within the range of The dimensionless eigenvalues, after adaptive weight combination calculation, output a parameterless benchmark intensity sequence. Both are dimensionless characteristic parameters, and their value range is strictly limited to... Interval. For the calculated annual parameter-free baseline intensity sequence. Extract the historical time series data sequence set from it. Constructing the empirical cumulative distribution function Statistical analysis is performed on the feature benchmarks under historical unsupervised states, and the specific expression is as follows: In the formula, Represents independent real variables, This indicates an indicator function that takes the value when the condition is true. Otherwise, the value is .

[0049] S2212. To ensure that the regression target of the model training is strictly aligned with the set gridded control level probability in terms of probability distribution, the system extracts the parameterless baseline intensity empirical cumulative distribution function. The 75th, 90th, and 99th percentiles were then used to assign the extracted quantile values ​​to the grade threshold constants. , , The specific expression is: , , In the formula, This indicates that the infimum of the set is taken. Represents the set of real numbers; , , satisfy The natural increasing law is used to calculate the output level threshold constant. , , The range of values ​​for is also strictly limited to Interval.

[0050] Since the regression objective value in multi-objective joint optimization training comes from the range of values ​​within... The parameterless baseline intensity sequence, during model training, drives the parameters of the cyber-physics cross-domain coupled computational model to converge naturally. Under normal monitoring and ordinary grid-based downgraded early warning conditions, the comprehensive public safety supervision risk index... The typical range of values ​​is stable at Interval. When the comprehensive characteristic value of crowd gathering in physical space. Characteristic value of instantaneous negative emotion propagation rate in cyberspace When the synchronization is at a high level, the exponential nonlinear coupling term is activated, driving the comprehensive public safety regulatory risk index. Dynamic Breakthrough The upper limit of the normal range accurately reflects the exponential mutation risk under extreme high-risk conditions. This method is applicable to target city grids that have accumulated at least one year of historical data. For newly deployed grids lacking historical data, the system follows the cold-start mapping rule, employing migration parameters for similar functional areas within the same city or a conservative fixed threshold. To eliminate model computational crashes caused by excessively large absolute values, the system sets the conservative fixed threshold based on the principle of maximum entropy with uniform distribution. , , This setting ensures that the regression target sequence remains consistent during a cold start. The basic scale space.

[0051] S222. Input the multi-source raw monitoring sequences from the historical dataset into the cyber-physical cross-domain coupled computation model to calculate the predicted risk index sequence under the current parameter state. Extract the true target numerical sequence. The mean squared error algorithm is used to quantify the overall regression accuracy of the model in continuous numerical space, and the mean squared error is calculated. The specific calculation formula is as follows: ; In the formula, This indicates the total number of historical data samples within the input training batch. Indicates the first The predicted risk index for each sample Indicates the first The true target value of each sample.

[0052] S223. Setting a high-risk threshold for emergency events Statistical prediction risk index series Medium to high risk threshold for emergency events And the actual number of true positive samples labeled as sudden high-risk events. ;Simultaneously count the total number of positive samples in historical data whose actual labels are sudden high-risk events. The system calculates the recall rate for sudden high-risk events. The specific calculation formula is as follows: Statistical prediction risk index series The risk level is below the high-risk threshold for sudden events. And the actual number of true negative samples that did not trigger control measures Simultaneously count the total number of normal negative samples in historical data that did not actually trigger control measures. The system calculates the prediction accuracy for routine monitoring periods. The specific calculation formula is as follows: .

[0053] S224. Based on the current security management priority of the target city grid, extract the preset error penalty weight coefficient. Recall weighting coefficient Precise weighting coefficients ; Fusion mean square error Recall rate Prediction accuracy Calculate the weighted composite loss function The specific expression is: ; The system will use a weighted composite loss function The calculation results are used as global error evaluation parameters and written into the memory unit of the backpropagation computation graph. This process enables deep coupling between regression accuracy and classification tolerance metrics, providing a single composite scalar feedback for gradient solving to downstream stochastic gradient descent algorithm nodes, thus avoiding computational divergence in multi-objective optimization processes.

[0054] S23. A dataset is constructed by extracting positive and negative samples using an event peak alignment strategy. The optimal parameter set is then solved offline using a combination of reparameterization techniques and a stochastic gradient descent algorithm with driving forces, and finally embedded into a cross-domain coupled cyber-physics computational model. This includes the following sub-steps: S231. Training sample labeling is performed using an event peak alignment strategy. The average physical dissipation time constant from the outbreak to the complete evacuation and resolution of historical high-risk events in the target city grid is extracted and set as the half-width of the time interval. Based on the moment when the highest level of control order was actually triggered in a historical event. Take the risk index of the event. The sequence is within a preset time interval defined based on the average physical dissipation time constant. The maximum value within the range is used as the regression target for positive samples. This is derived from the historical all-weather risk index. During the normal periods of the sequence, negative samples are randomly sampled at equal intervals to extract normal samples, and the sampling ratio is controlled to ensure a balance between the number of positive and negative samples in the sample set. The constructed historical dataset is divided into a training set and a validation set. A set of hyperparameters for multi-objective joint optimization training is defined, and a maximum number of iterations is set. As the absolute upper limit of the training cycle, a relative decrease threshold for the loss function is set. Used to quantify convergence conditions and set the early stopping tolerance rounds. This is used to limit the maximum waiting step size to the point where the loss on the validation set no longer decreases significantly, thus establishing a definite termination boundary for subsequent iterations of optimization.

[0055] S232. For optimization variables with physical constraints, a reparameterization technique is used to transform constrained optimization into unconstrained optimization. A mapping relationship is defined in memory, and the physical characteristic coefficients are set... , By constraining non-negativity, the overall physical weights and coupling weights are made equal. , , To provide a smoother gradient flow, the time delay constant is reduced. Increase sensitivity .in , , , , , , These are the free variables used in the actual iterations of the gradient descent algorithm. This process ensures that the optimization space is continuous and differentiable in the real number domain, avoiding computational interruptions caused by parameter out-of-bounds errors.

[0056] S233. Use the stochastic gradient descent algorithm with momentum to iteratively update the free variables on the training set. Extract the weighted composite loss function output from step S224. Calculate the weighted composite loss function Gradients of partial derivatives with respect to each free variable. Introducing the momentum factor. With learning rate Calculate velocity variables The update amount is determined, and the free variables are updated accordingly. The specific expression is: ; ; In the formula, represent , , , , , , Any free variable in, Indicates the number of iterations in the algorithm. This represents the momentum-velocity of the corresponding variable. Represents the weighted composite loss function The gradient of this variable. This process utilizes momentum to accumulate historical gradient directions, improving the algorithm's convergence speed in the parameter space.

[0057] S234. After each training round, the system substitutes the current free variables into the validation set to calculate the weighted composite loss function. When the loss value on the validation set stops decreasing for several consecutive rounds, the system triggers an early stopping mechanism to terminate the iteration and obtains the set of free variables in the convergent state. The system then calls the mapping function defined in step S232 to perform reverse back-substitution and obtain the optimal physical parameters that satisfy the domain requirements. , , , , , and The system will embed the acquired optimal physical parameters into the cyber-physics cross-domain coupled computational model, making the parameters... Actively converge towards the direction of higher sensitivity.

[0058] S3. The calculated comprehensive public safety supervision risk index The system inputs real-time data into the administrative regulatory decision tree model to compare and determine multi-level control boundary points, and executes debouncing and fault-tolerant logic with a dual-threshold determination mechanism. For example... Figure 2 As shown, the specific steps include the following: S31. Preset a set of control boundary points within the system, including primary control boundary points. Secondary control boundary points Level III control boundary point It also completes the initialization setting of the corresponding threshold and offline recalibration, and sets the continuous sampling judgment time step. Specifically, it includes the following sub-steps: S311. Detect the historical running time and accumulated data volume of the target city grid. For newly deployed grids lacking historical data, invoke the cold start configuration strategy, initially using a conservative fixed threshold. , , As the starting value, a data accumulation monitoring thread is established. Once the target city grid has accumulated enough running data, it automatically switches to the recalibration logic based on extreme value theory and physical constraints, avoiding the implementation gap caused by the lack of historical data in newly established grids, which could paralyze the hierarchical supervision system.

[0059] S312. For target city grids with historical data, the system extracts the comprehensive public safety supervision risk index for the past year during the offline maintenance window. Historical time series data sequence To accurately characterize the probability distribution of sudden and rare high-risk events in the field of public safety, a tail risk distribution is constructed by introducing the over-threshold model from extreme value theory, and the empirical cumulative distribution function is used. Quantile extraction of basic cutoff threshold The system extracts data from historical time-series data that exceeds a basic cutoff threshold. Construct a super-threshold sequence set from sample data The maximum likelihood estimation method is used to evaluate the over-threshold sequence set. Fit the distribution to a generalized Pareto model and calculate the distribution shape parameters. With scale parameters We established a risk tail analysis expression specifically designed to characterize the probability of extreme aggregation and negative emotional outbursts.

[0060] S313. Set the tolerable probability parameters for the corresponding Level 1, Level 2, and Level 3 warning frequencies. , , The theoretical boundary point in the statistical dimension is calculated based on the fitted inverse function of the generalized Pareto distribution. , , The specific calculation formula is as follows: ; In the formula, , This represents the total number of samples in the historical time-series data sequence. This represents the number of samples in the sequence set exceeding the threshold. To integrate the theoretically defined statistical boundary with the actual physical carrying capacity of the target city grid, the system extracts the effective passable area of ​​the grid. Total effective evacuation width of emergency exits Calculate the evacuation bottleneck penalty coefficient The specific calculation formula is as follows: ; In the formula, This is the maximum evacuation penalty ratio limit. This represents the standard evacuation demand constant per unit area. The system combines the theoretical boundary point with the evacuation bottleneck penalty coefficient to output the final multi-level control boundary point with physical space constraints. , , The specific calculation formula is as follows: .

[0061] This computational mechanism ensures that the penalty coefficient decreases as grid evacuation conditions worsen and effective exits become narrower. The larger the threshold, the lower the control boundary point is automatically reached, forcing the system to trigger a high-level warning earlier, thus preserving ample lead time for administrative intervention in cases of physical space congestion. The system verifies that the updated parameters meet the requirements. The gradient increment relationship is calculated and loaded into the real-time decision buffer channel, and the initial continuous sampling decision time step is set synchronously. .

[0062] S32. Extract the current comprehensive public safety supervision risk index. Calculate the rate of change of the risk index compared to the historical risk index at the previous moment, and perform a dual-threshold comparison judgment by combining the set warning slope threshold and the mutation slope threshold. Specifically, this includes the following sub-steps: S321. Extract the current time from the real-time decision buffer channel. Comprehensive public safety supervision risk index Compared to the previous basic calculation step time Comprehensive public safety supervision risk index The system uses a backward difference algorithm to calculate the risk index at the current moment. Rate of change compared to the previous time step The specific calculation formula is as follows: ; In the formula, This indicates the basic calculation step size. This step transforms the absolute value of the risk index into a dynamic parameter reflecting the rate of deterioration, providing direct computational input for capturing sudden high-risk situations.

[0063] S322, Set the warning slope threshold With the threshold of mutation slope And satisfy The numerical relationship.

[0064] As one possible implementation, a pre-warning slope threshold With the threshold of mutation slope The determination method is as follows: extract the evolution time series data of historical high-risk emergencies in the target city grid. Statistically analyze the comprehensive public safety supervision risk index from the historical time series data. From the secondary control boundary point The situation continued to deteriorate to the level three control threshold. The physical time consumed is used to extract the lower bound of its distribution and add a preset system response delay redundancy, which is then set as the critical time constant for sudden evolution. ; Comprehensive public safety supervision risk index in historical time series data From the first-level control boundary point The situation continued to deteriorate to the secondary control threshold. The physical time consumed is used to extract the mathematical expectation of its distribution, which is then set as the critical time constant for early warning evolution. For newly deployed grids lacking historical data, the system invokes a cold start configuration strategy, extracting parameters from similar functional areas within the same city and assigning them to the critical time constant for sudden evolution. With the critical time constant of early warning evolution Calculating the abrupt change slope threshold based on the critical time constant of physical evolution. With warning slope threshold The specific calculation formula is as follows: , .

[0065] S323, Extracting the rate of change It also executes fast variable channel comparison logic with a dual threshold determination mechanism. Specifically, the system extracts the initial continuous sampling determination time step. This initial value is determined by the system based on the normal monitoring baseline tolerance delay time of the target city grid. Combined with basic calculation step size The offline calculation settings are as follows: ; In the formula, This indicates the rounding operation.

[0066] rate of change Respectively with the warning slope threshold and mutation slope threshold Compare the values: When the system compares and determines the rate of change satisfy At that time, a pre-mutation warning is triggered based on the rate of change. The continuous sampling decision time step is dynamically shortened based on the relative position within the threshold range. Specifically: a minimum sampling step size lower limit is set. To maintain basic image stabilization and fault tolerance, calculate the current continuous sampling decision time step. The specific calculation formula is as follows: ; In the formula, This indicates the operation of finding the maximum value. This calculation mechanism ensures that the faster the situation deteriorates, i.e., the higher the rate of change... The closer to the mutation slope threshold The shorter the continuous sampling step size required for confirmation, the smoother the system transitions to the high-sensitivity monitoring mode.

[0067] When the system compares and determines the rate of change satisfy When the target city grid is determined to be in a sudden high-risk state, the continuous sampling and judgment logic is bypassed directly, and the corresponding high-level early warning strategy is immediately confirmed and triggered to ensure zero-delay response to extreme coupled abrupt changes.

[0068] When the system compares and determines the rate of change satisfy When the risk evolution of the target city grid is determined to be in a stable or slowly changing state, the initially set continuous sampling and determination time step remains unchanged. .

[0069] The system will display the comprehensive public safety supervision risk index at the current moment. Together with the calculated current continuous sampling decision time step The output is synchronized and handed over to the downstream anti-shake and fault-tolerant logic processing module, which is located in the ambiguity range between adjacent threshold levels.

[0070] S33, Receive the comprehensive public safety supervision risk index output in step S323 With continuous sampling and determination time step It executes anti-shake and fault-tolerant logic with state memory and cross-level reset mechanisms, and generates multi-level boundary point comparison and judgment results and corresponding administrative intervention instructions. Specifically, it includes the following sub-steps: S331. The system constructs a complete finite state machine for the early warning interval in memory to perform deterministic anti-jitter fault-tolerant control. Define the set of components of this finite state machine, specifically including: setting the instantaneous target state set. Based on the current moment Comprehensive public safety supervision risk index With dynamically extracted multi-level control boundary points , , Construct state mapping function Get the target status code at the current moment. The specific expression is: ; Declare and maintain two internal state parameters of the finite state machine in memory, including a history state code used to store the output of the previous basic computation step. And a continuous sampling counter used to count the continuous dwell time of the target status code. During system cold start initialization, both of the above parameters are strictly assigned the value. .

[0071] S332. If the system does not trigger a sudden high-risk state, execute the state transition function of the finite state machine. The system extracts the current continuous sampling determination time step. Compare the target status code at the current moment. Historical status codes compared to the previous basic calculation step size The system ensures that the transition paths under any combination of state codes at any given time step are mutually orthogonal and completely exhaustive, specifically by executing the following state transition conditions: When the target status code is satisfied When the high-risk debouncing restart logic is triggered, the state transition function controls the continuous sampling counter to perform a reset operation, i.e., update it to... ; When the target status code is satisfied and When the anti-shake delay accumulation mechanism is triggered, the state transition function controls the continuous sampling counter to perform an auto-increment operation, that is, update it to... ; When the target status code is satisfied and When the debounce restart logic is triggered, a downgrade confirmation is established. The state transition function controls the continuous sampling counter to perform a reset operation, i.e., update it to... ; When the target status code is satisfied When the normal reset logic is triggered, the state transition function controls the continuous sampling counter to perform a clearing operation, that is, update it to... .

[0072] S333, The system extracts the current continuous sampling counter. The value, and the received decision step size. Perform a value comparison and generate corresponding control instructions based on the comparison results and the target status code: When the risk index At the same time, the system maintains a normal monitoring state and records historical parameters, while continuously sampling counters... Reset to zero.

[0073] When the continuous sampling counter satisfies And the target status code remains unchanged. (i.e., continuous) Risk index over a time step All meet When the system confirms that the first-level grid-based downgrade warning strategy has been triggered, it outputs an administrative intervention instruction to increase the frequency of patrols at the edge of specific grid areas and send regional reassurance information.

[0074] When the continuous sampling counter satisfies And the target status code remains unchanged. (i.e., continuous) Risk index over a time step All meet When the system confirms that the secondary grid-based downgrade warning strategy has been triggered, it outputs an administrative intervention instruction to dispatch grid-based management personnel to conduct on-site traffic control and activate high frame rate sampling of surrounding video surveillance equipment.

[0075] When the continuous sampling counter satisfies And the target status code remains unchanged. (i.e., continuous) Risk index over a time step All meet When the system confirms that the three-level grid-based downgrade warning strategy has been triggered, it outputs administrative intervention instructions to block traffic flow at the corresponding intersection, enforce physical isolation of the area, and activate the cross-departmental emergency collaborative response mechanism.

[0076] S4. The generated administrative intervention instructions are parsed and transformed into standardized two-way regulatory strategy messages, which are then distributed to the corresponding terminal execution units, forming a closed loop for end-to-end public safety regulatory control. This includes the following sub-steps: S41. The generated administrative intervention instructions are parsed and converted into standardized two-way regulatory strategy messages, which are then distributed to the corresponding terminal execution units. Different types of terminal execution units execute differentiated blocking actions based on the message parsing results. This includes the following sub-steps: S411. Receive the administrative intervention instruction and warning interval status code output in step S33. Invoke the spatial geographic information database of the target city grid, extracting the event geographic location coordinate set and the unique identifier of the bound controlled object for that grid. Package the above data according to a preset serialization protocol, reconstructing it into a policy message containing the unique identifier of the controlled object, the event geographic location coordinate set, resource scheduling priority parameters, and the basic business fields of the downgrade warning level code. Add an execution status feedback field, a dynamic reconfiguration field, and an execution effect quantification field to the end of the message structure, allocate independent memory space for these three new fields and assign initialization flags, completing the construction of a standardized bidirectional regulatory policy message.

[0077] S412. Extract the unique identifier of the controlled object from the standardized two-way monitoring policy message and perform device type comparison and address resolution. Based on the resolution results, establish a low-level communication routing channel to distribute the standardized two-way monitoring policy message to the corresponding terminal execution unit. The terminal execution unit is divided into field IoT control devices deployed in the physical space and portable devices held by management personnel.

[0078] S413. After receiving standardized two-way monitoring strategy messages, the on-site IoT control equipment runs a local lightweight parsing program to read the basic business fields of the downgrade warning level code to extract the control code. Based on the control code, the on-site IoT control equipment directly calls the underlying mechanical drive module to close the gate channel and execute the area isolation action. The instruction data bypasses the complex upper-layer operating system and is directly written to the hardware register, eliminating the instruction polling delay inside the device.

[0079] S414. After receiving the policy message, the administrative staff's mobile terminal extracts the event geographic location coordinate set and the downgrade warning level code from the message. The administrative staff's mobile terminal simultaneously calls the comprehensive physical space crowd aggregation characteristic value output from the cross-domain coupled computing module. Characteristic value of instantaneous negative emotion propagation rate in cyberspace The electronic map display layer renders a heat map of the situation and suggests routes for emergency resource allocation.

[0080] S42. After completing the action, the terminal execution unit sends back a status code through the execution status feedback field. Based on the actual situation on site, the action of the IoT device is adjusted in reverse through the dynamic reconfiguration field. The execution effect quantification field records the residual between the actual change and the predicted change based on the historical baseline. Only when the residual exceeds the preset confidence interval is the execution action considered to have a significant net effect, and this residual signal is used for online fine-tuning to eliminate the influence of external confounding factors such as weather and holidays. Specifically, it includes the following sub-steps: S421. After completing the physical blocking action, the terminal execution unit writes a status code representing the execution result into the execution status feedback field of the standardized two-way monitoring strategy message, and then sends it back to the system via the underlying communication link. The system parses the execution status feedback field to obtain the actual situation on site. When it is determined that the initial action did not meet the expected boundary conditions, the system extracts the comprehensive characteristic value of the physical space crowd gathering at the same time. Secondary calculations are performed, and remedial control codes are generated by dynamically reconfiguring fields and sent to on-site IoT management and control devices to reverse the actions of IoT devices and build a micro-level bidirectional adjustment link at the physical control level.

[0081] S422. After the intervention is completed, open an observation time window of a preset length and extract the normalized feature values ​​of the dwell time within the window. Characteristic value of instantaneous negative emotion propagation rate in cyberspace The actual change in dwell time is obtained through differential calculation. Compared with the actual change in negative emotions Simultaneously, the baseline characteristic database of the target city grid within the same historical time period is retrieved, and natural temporal fluctuation sequence data unaffected by administrative intervention are extracted. The predicted change in dwell time based on the historical baseline of the same period is then calculated. Change in negative sentiment prediction .

[0082] S423. Calculate the residence time residual based on the actual change and the baseline predicted change. With negative emotional residuals The specific expression is: ; In the formula, This represents the net effect parameter of dwell time after removing natural environmental fluctuations. This represents the net emotional effect parameter after removing the natural calming effect.

[0083] Set the confidence interval threshold for dwell time With the threshold of the negative sentiment confidence interval For the residual of the length of stay With negative emotional residuals Perform an absolute value comparison and determination. When the condition is met... or At that time, it was determined that the current administrative intervention order had a significant net effect. The dwell time residuals that met the confidence level requirements were then considered. With negative emotional residuals The serialized data is written into the execution effect quantification field to eliminate the influence of external confounding factors such as natural weather decline and holiday traffic fluctuations, providing a high-confidence causal correlation signal for the online fine-tuning of calibration parameters in the subsequent cross-domain coupled calculation model.

[0084] As one possible implementation method, the residence time confidence interval threshold With the threshold of the negative sentiment confidence interval The determination method is as follows: Extract the natural time-series fluctuation sequence data obtained from the historical baseline feature database of the same type of time period in step S422, and construct the set of natural fluctuations of historical residence time. Collection of natural fluctuations in historical negative emotions ,in This represents the total number of samples representing historical natural temporal fluctuations. Unbiased standard deviations are calculated to quantify the dispersion of endogenous environmental noise distribution in the target city grid within the same historical time period. The unbiased standard deviation of historical dwell time natural fluctuations is also calculated. Unbiased standard deviation of natural fluctuations in historical negative sentiment The specific calculation formula is as follows: , In the formula, The sample mean representing the natural fluctuations in historical residence time. This represents the sample mean of natural fluctuations in historical negative sentiment. Based on the Laida criterion in mathematical statistics (i.e.... (Guideline), setting the threshold for the confidence interval of the residence time. With the threshold of the negative sentiment confidence interval Strictly set at three times the corresponding unbiased standard deviation, to isolate theoretically... The random fluctuation noise in the normal environment. The specific calculation formula is: , .

[0085] S43. Online fine-tuning operates on a daily cycle, accumulating all quantitative data on execution effects for the day. A small-batch gradient descent method is used to incrementally update the calibration parameters. The updated parameters are automatically used for the next day's risk index calculation, forming a closed-loop public safety supervision and control system encompassing cross-domain data perception, cross-domain coupled calculation, fault-tolerant decision-making, and adaptive execution evolution. Specifically, it includes the following sub-steps: S431. The system uses a calendar day cycle and initiates an online fine-tuning task during the preset business off-peak period each day. It iterates through the interaction logs of all standardized two-way monitoring policy messages for the day, extracting the execution effect quantification field data that meets the confidence requirements, i.e., extracting the dwell time residual output in step S42. With negative emotional residuals The system performs a non-empty state determination on the extracted dataset size. If the extracted execution effect quantification field data for the day is determined to be empty (meaning the target city grid did not trigger administrative intervention instructions or the intervention did not produce a significant net effect residual), the system automatically bypasses the incremental fine-tuning iterative calculation process for that day, maintaining the current calibration parameter set loaded into the cyber-physical cross-domain coupling calculation model, and directly enters the risk feature perception and monitoring cycle for the next natural day. If the extracted execution effect quantification field data is not empty, the system extracts the physical space population aggregation comprehensive feature value of the corresponding historical moment input into the cyber-physical cross-domain coupling calculation model. Characteristic value of instantaneous negative emotion propagation rate in cyberspace The residual signals and input feature data are strictly aligned with and packaged using timestamps to construct the daily incremental fine-tuning dataset, providing a definite data input source for downstream small-batch gradient descent updates.

[0086] S432. The system establishes a mapping relationship between residual signals and parameter corrections, and constructs an online fine-tuning loss function with the optimization objective of minimizing the sum of squared residuals after execution. The specific expression is: ; In the formula, Indicates the number of samples in a mini-batch. Indicates the sample index within the batch. Indicates the first The residence time residuals of each sample Indicates the first The negative sentiment residuals of each sample The residence time residual penalty weighting coefficient, This represents the weighting coefficient for the negative sentiment residual penalty. Fine-tuning the loss function. The design ensures that the direction of parameter fine-tuning is strictly controlled within the net deviation of the actual physical effects on site.

[0087] As one possible implementation method, the residence time residual penalty weighting coefficient Weighting coefficient of negative emotion residual punishment The method for determining it is as follows: S4321. When extracting the predicted change based on the historical baseline, the system simultaneously extracts a natural fluctuation sequence of residence time characteristics within a preset long-term time window from the baseline feature database of the same historical time period. Natural fluctuation sequence of negative emotional characteristics The above sequences are all derived from normal, natural evolution periods unaffected by administrative intervention, ensuring the purity of the sequence data. The preset long-term time window length is set by the system based on the data sampling frequency of the target city grid, and the system sets this time window to be no less than 90 calendar days to ensure the statistical significance of the variance estimation.

[0088] S4322. The system performs unbiased sample variance calculation on the extracted natural fluctuation sequences of dwell time features and negative sentiment features, quantifying the endogenous noise intensity of the two physical or cyber features within the target city grid, and the baseline fluctuation variance of dwell time. Variance of negative sentiment baseline The calculation formula is: ; In the formula, This represents the total number of samples in the historical baseline fluctuation series. The sample mean of the natural fluctuation sequence representing the residence time characteristic. This represents the sample mean of the natural fluctuation sequence of negative emotional characteristics. To prevent pathological degradation of numerical calculations in extremely stable scenarios, the system sets a lower variance protection threshold. ,when Time to take ,when Time to take ,and The system sets the machine precision based on the feature values, typically taking 10. -6 Up to 10 -4 between.

[0089] S4323. The system uses the reciprocal of the variance as the basis for weight allocation, so that features with greater endogenous environmental fluctuations and background noise have smaller weights in the fine-tuning loss function. The final residual penalty weight coefficient is calculated, and the specific expression is as follows: ; The system outputs the calculated residence time residual penalty weight coefficient. Weighting coefficient of negative emotion residual punishment Substitute it directly into the online fine-tuning loss function In the middle. And the two weighting coefficients strictly satisfy The normalized control conditions.

[0090] In addition, the residual penalty weighting coefficient for residence time Weighting coefficient of negative emotion residual punishment Update cycle and control boundary point , , The fixed monthly cycle offline recalibration is kept in sync. That is, when performing offline recalibration each time, the system re-extracts the historical baseline fluctuation sequence of the corresponding time window and executes the above complete calculation process to ensure that the penalty weight coefficient is synchronously adapted to the current environmental fluctuation characteristics of the target city grid.

[0091] S433. The system loads the optimal set of physical parameters currently running in the cyber-physics cross-domain coupled computation model and reads its corresponding underlying free variables. , , , , , , To avoid catastrophic forgetting of pre-trained knowledge caused by online fine-tuning, the system extracts the initial learning rate from the offline multi-objective joint optimization training phase. Combined with a preset conservative fine-tuning attenuation factor Calculate the fine-tuning learning rate The specific calculation formula is as follows: Conservative fine-tuning of attenuation factor The value of is limited to to The scale ensures that the fine-tuning step size is much smaller than the offline training step size. The system sets the convergence criterion for online fine-tuning and extracts the upper limit of the maximum number of fine-tuning iterations. With tolerance for fine-tuning loss variations .

[0092] The system employs a mini-batch gradient descent algorithm to perform incremental iterations on the daily incremental fine-tuning dataset, calculating the online fine-tuning loss function. The learning rate is fine-tuned by introducing the gradient of the partial derivatives of each free variable. To perform incremental parameter updates, the specific expression is: In the formula Represents any free variable to be fine-tuned. Indicates the number of iterations for fine-tuning. This represents the gradient of the fine-tuning loss function with respect to this free variable. After each incremental update, the absolute difference between the current and previous rounds of the fine-tuning loss function is calculated. If the absolute difference is continuously less than the tolerance for changes in the fine-tuning loss for a preset number of rounds, the change is considered complete. Or fine-tune the iteration rounds Reaching the maximum number of fine-tuning iteration rounds When the iteration converges, the update process is forcibly terminated.

[0093] After the system achieves iterative convergence, it calls the preset function. and The mapping function performs a reverse substitution transformation, which enables the model parameters to dynamically approximate the real environment while ensuring that the fine-tuned calibration parameters strictly meet the non-negative physical domain requirements, thus eliminating the potential for model computational collapse caused by extreme residuals.

[0094] S434. Write the updated calibration parameters into the next day's effective area of ​​the system configuration file, and automatically reload the parameters of the cyber-physical cross-domain coupling calculation model at midnight of the next day. The parameters are automatically used for the risk index calculation of the next day. Through daily periodic execution-evaluation-fine-tuning iteration, the non-stationary evolution characteristics within the target city grid are continuously tracked, forming a closed loop of full-link public safety supervision and control that covers cross-domain data perception, cross-domain coupling calculation, fault-tolerant decision-making and adaptive execution evolution.

[0095] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for intelligent urban public safety supervision based on multi-source data fusion, characterized in that, include: Obtain the original sequence of crowd density and trajectory dwell time parameters, and calculate the comprehensive characteristic value of crowd aggregation in physical space; Obtain the text sequence of network information and extract the text that has a geographic identifier mapping relationship with the target city grid, and calculate the feature value of the instantaneous negative sentiment propagation rate in cyberspace; Based on the comprehensive characteristic value of crowd gathering in the physical space and the characteristic value of the instantaneous negative emotion propagation rate in the cyberspace, a cyber-physical cross-domain coupled calculation model is constructed to calculate the comprehensive public safety supervision risk index. Based on the comprehensive public safety supervision risk index, multi-level control boundary point comparison and judgment are performed, and anti-shake fault-tolerant logic is executed to generate multi-level boundary point comparison and judgment results and corresponding administrative intervention instructions. The administrative intervention instructions are converted into standardized two-way regulatory strategy messages and distributed to the corresponding terminal execution units.

2. The intelligent urban public safety supervision method based on multi-source data fusion according to claim 1, characterized in that, The acquisition of the original population density sequence, trajectory dwell time parameters, and network information text sequence includes: Extract the number of human target bounding boxes from the real-time video stream returned by the video surveillance equipment as the original sequence of crowd density. Collect the timestamp of the first detection and the last detection of the same physical address, calculate the time difference and perform an arithmetic average to obtain the trajectory dwell time parameter; Geographical entity information and GPS coordinate labels contained in the network information text stream are extracted and compared with the electronic fence coordinate boundaries of the target city grid to obtain the network information text sequence.

3. The intelligent urban public safety supervision method based on multi-source data fusion according to claim 1, characterized in that, The calculation of the comprehensive characteristic value of crowd gathering in the physical space includes: Extract the physical upper limit and lower limit of population density, process the original population density sequence based on the maximum and minimum value normalization method with boundary truncation, and obtain density normalized feature values; Extract the upper limit and lower limit of trajectory dwell time, process the trajectory dwell time parameter based on the maximum and minimum value normalization method with boundary truncation, and obtain the dwell time normalized feature value; A linear combination mechanism is used to fuse density-normalized eigenvalues ​​and dwell time-normalized eigenvalues ​​to calculate the comprehensive eigenvalues ​​of crowd aggregation in physical space.

4. The intelligent urban public safety supervision method based on multi-source data fusion according to claim 1, characterized in that, The calculation of the characteristic value of the instantaneous negative sentiment propagation rate in cyberspace includes: A joint sentiment dictionary integrating a general sentiment lexicon and a proprietary sentiment lexicon, as well as a negation word list containing core negation words, are constructed. The original sentiment weighted score of the network information text sequence is calculated based on the preset word window length and the polarity reversal factor based on the number of times the negation word is hit. The negative polarity score of a single text is obtained by combining an asymmetric sentiment excitation mapping function for directional nonlinear transformation. Texts with a negative polarity score greater than the threshold for determining negative emotion are considered as valid negative information samples. The negative polarity scores of individual texts of valid negative information samples are accumulated to obtain the instantaneous absolute increment of negative emotion in the network space. We extract the upper limit of the instantaneous propagation capacity of negative emotions in the network, use a normalization method with boundary truncation to process the absolute increment of instantaneous new negative emotions in the network space, and calculate the characteristic value of the instantaneous negative emotion propagation rate in the network space.

5. The intelligent urban public safety supervision method based on multi-source data fusion according to claim 1, characterized in that, Construct the aforementioned cyber-physical cross-domain coupled computational model to calculate the comprehensive public safety regulatory risk index, including: A fixed-length sliding historical queue based on the maximum physical boundary is constructed, and the smoothed cumulative amount of the feature value of the instantaneous negative emotion propagation rate in the network space within a specific time window is extracted by discrete time series operation. A cyber-physical cross-domain coupled computational model is constructed, and the numerical sum of the physical space basic cumulative term, the network sentiment smoothing cumulative term, and the nonlinear coupling term is combined to output a comprehensive public safety supervision risk index. The physical space basic accumulation term is determined by combining the physical space crowd aggregation comprehensive feature value with the physical aggregation basic weight coefficient. The network emotion smoothing accumulation term is determined by combining the smoothing accumulation amount with the network emotion basic weight coefficient. The nonlinear coupling term is determined by combining the physical space crowd aggregation comprehensive feature value with the network space instantaneous negative emotion propagation rate feature value with the coupling mutation weight coefficient and mutation sensitivity calibration parameter through exponential function calculation.

6. The intelligent urban public safety supervision method based on multi-source data fusion according to claim 1, characterized in that, After constructing the aforementioned cyber-physics cross-domain coupled computational model, a weighted composite loss function is constructed using a historical concurrent security event dataset for multi-objective joint optimization, including: An adaptive feature weight based on information entropy is used to construct a parameter-free baseline intensity sequence, which maps discrete treatment levels to continuous target values, generating a real target value sequence containing discrete time steps. The mean squared error algorithm is used to quantify the mean squared error between the true target numerical sequence and the predicted risk index sequence. The recall rate of sudden high-risk events is calculated by comparing the number of true positive samples with the total number of positive samples, and the prediction accuracy of normal monitoring periods is calculated by comparing the number of true negative samples with the total number of normal negative samples. The mean squared error, recall, and prediction precision are combined, and a weighted composite loss function is constructed by combining preset error penalty weight coefficients, recall weight coefficients, and precision weight coefficients. A dataset is constructed by extracting positive and negative samples using an event peak alignment strategy based on the physical dissipation time constant. The optimal parameter set of the proposed cyber-physics cross-domain coupled computation model is then solved by combining reparameterization techniques with the stochastic gradient descent algorithm with momentum.

7. The intelligent urban public safety supervision method based on multi-source data fusion according to claim 1, characterized in that, Based on the aforementioned comprehensive public safety regulatory risk index, a multi-level control boundary point comparison and determination is performed, including: Set warning slope thresholds and mutation slope thresholds; use backward difference algorithm to process the comprehensive public safety supervision risk index at the current time and the comprehensive public safety supervision risk index at the previous time, and calculate the risk index change rate; When the rate of change of the risk index is greater than the warning slope threshold but less than or equal to the mutation slope threshold, a pre-mutation warning is triggered, and the continuous sampling judgment time step is dynamically shortened according to the relative position of the rate of change within the threshold range. When the rate of change of the risk index is greater than the threshold of the mutation slope, it is determined that the system has entered a state of instantaneous high risk. The system bypasses the continuous sampling judgment logic and confirms that the corresponding high-level early warning strategy has been triggered.

8. The intelligent urban public safety supervision method based on multi-source data fusion according to claim 1, characterized in that, Execute the aforementioned anti-shake fault-tolerance logic to generate administrative intervention instructions, including: Construct a finite state machine for early warning intervals, and set the state codes for normal intervals, first-level early warning intervals, second-level early warning intervals, and third-level early warning intervals. Compare the target status code at the current moment with the historical status code and execute the state transition function. When the target status code meets the conditions for a leap to a higher level or a fall back within a certain range, reset the continuous sampling counter. When the target status code meets the conditions for maintaining the same level, control the continuous sampling counter to perform an increment operation. When the target status code meets the conditions for a normal reset, clear the continuous sampling counter to zero. When the value of the continuous sampling counter is greater than or equal to the continuous sampling judgment time step and the target status code remains within a specific warning range, an administrative intervention instruction corresponding to the gridded degradation warning strategy is output.

9. The intelligent urban public safety supervision method based on multi-source data fusion according to claim 1, characterized in that, The administrative intervention instructions are converted into standardized two-way regulatory policy messages and distributed to the corresponding terminal execution units, including: Extract the event's geographic location coordinates and the unique identifier of the bound controlled object, package them using a serialization protocol, and reconstruct them into a standardized two-way regulatory strategy message containing basic business fields for downgrade warning level code, and add execution status feedback fields, dynamic reconfiguration fields, and execution effect quantification fields to the end of the message structure; Extract the unique identifier of the controlled object from the standardized two-way monitoring policy message, perform device type comparison and address resolution, establish an underlying communication routing channel, and distribute the standardized two-way monitoring policy message to the corresponding terminal execution unit.

10. The intelligent urban public safety supervision method based on multi-source data fusion according to claim 9, characterized in that, After the standardized two-way regulatory policy message is distributed to the corresponding terminal execution unit, the method further includes: The terminal execution unit analyzes the status code returned by the execution status feedback field and obtains the actual change in dwell time and the actual change in negative sentiment through differential calculation. Search the baseline feature database of the same historical time period, extract the natural time-series fluctuation sequence data that is not affected by administrative intervention, and calculate the predicted change in dwell time and the predicted change in negative sentiment based on the historical baseline of the same period. The dwell time residual is calculated based on the actual change in dwell time and the predicted change in dwell time. The negative emotion residual is calculated based on the actual change in negative emotion and the predicted change in negative emotion. The dwell time residual and negative emotion residual that meet the confidence requirements are written into the execution effect quantification field. Using natural days as the cycle, when the data in the execution effect quantification field is determined to be non-empty, the execution effect quantification field data and the feature data of the corresponding historical time input to the cyber-physics cross-domain coupled computing model are extracted to construct an incremental fine-tuning dataset. The calibration parameters are incrementally updated using the mini-batch gradient descent method based on online fine-tuning loss function, fine-tuning learning rate and convergence criterion.

Images