Intelligent analysis system for public security information based on dynamic closed-loop regulation

Abstract

This invention relates to the field of social security management data processing, and discloses a social security information intelligent analysis system based on dynamic closed-loop control. The system includes: an analysis and prediction unit that calculates risk scores based on feature weights; a subthreshold residual interception unit that captures weak score data below a trigger threshold; a spatiotemporal entropy aggregation unit that calculates the Shannon entropy value of the weak data; an agility bias unit that lowers the trigger threshold based on the low entropy value aggregation state; and an adaptive damping correction unit that uses the difference vector between prediction and feedback to inversely correct the feature weights. This invention calculates the spatiotemporal entropy value of the discarded residual data to sense environmental background pressure, dynamically biases the trigger threshold accordingly, and introduces a damping mechanism correction model in the feedback closed loop to solve the problem of missed weak risk detection caused by a fixed threshold, achieving an adaptive dynamic balance between system sensitivity and false alarm rate.

Description

Technical Field

[0001] This invention relates to an intelligent analysis system for social security information based on dynamic closed-loop control, belonging to the field of social security management data processing technology. Background Technology

[0002] In current practices of comprehensive social security governance and administrative management data processing, the mainstream technical architecture is based on a preset feature rule base and a linear threshold triggering mechanism. The data processing system collects multi-source heterogeneous basic business data, maps it into standardized feature vectors, compares it with preset risk models or expert rules, and when the calculated risk score exceeds the statically set triggering threshold, the system generates a corresponding early warning instruction and sends it to the administrative execution terminal. Based on the open-loop processing logic of feature matching and threshold determination, under the conditions of relatively stable environmental factors in the early stage of rule deployment, it realizes the automatic screening and prompting of specific types of management objects, thereby improving the timeliness of administrative instruction issuance.

[0003] The increased mobility of administrative management targets and the evolution of risk manifestations expose the fundamental technical limitations of static open-loop architectures in long-term practical operation. The dimensions and weights of security risk characteristics are time-varying, making it difficult for fixed parameter systems to continuously fit the dynamic and changing real security situation. System prediction models inevitably deviate logically from objective facts. While existing technologies introduce mathematical models to optimize the analysis process, they still fundamentally fail to overcome the limitations of open-loop control. For example, Chinese invention patent CN112667965A discloses a method for assessing social security based on entropy value. This method extracts multi-dimensional analytical elements related to the floating population, uses the concept of entropy to quantify the impact of different information on social security, and addresses the scientific nature of weight allocation. It uses entropy to positively evaluate and rank existing data. However, weight establishment initially relies on experience or calculations based on historical static data, lacking an automatic feedback and correction mechanism for deviations between analysis results and actual handling. It also fails to address the issues of false reporting and verification. The feedback information from the end-level execution layer, such as the severity of the case, serves as a control signal for adjusting model parameters. This one-way evaluation mode prevents the system from achieving adaptive convergence of parameters when prediction deviations occur due to changes in environmental background noise or the evolution of new risk characteristics. Long-term operation faces an imbalance between sensitivity and false alarm rate. Under the current technical approach, the front-end execution layer's handling results of early warning instructions, including false alarm cancellation, event confirmation, or handling feedback, are merely archived as independent business records without being structured and fed back to the core analysis algorithm to correct model parameters. This disconnect between analysis prediction and handling results data leaves the system lacking self-calibration capabilities based on factual feedback. Relying on periodic manual rule resets or offline model updates, this non-real-time intervention method suffers from response lag. Without granular feedback data support, it is difficult to accurately balance the weight configuration of a single feature item in different spatiotemporal environments. This leads to a dual dilemma: either a persistently high false alarm rate wastes administrative resources, or excessive suppression of false alarms results in desensitization and underreporting of weak risk signals.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a data processing method that transforms administrative disposal results into structured feedback control signals, and achieves dynamic convergence and adaptive evolution of analysis rule parameters with the actual security situation through negative feedback damping correction and positive feedback reverse compensation mechanisms. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A social security information intelligent analysis system based on dynamic closed-loop control, comprising:

[0006] The analysis and prediction unit is used to receive multi-source public security data and calculate risk scores using a feature weighting model, and only generates warning instructions when the risk score exceeds the dynamic trigger threshold.

[0007] The subthreshold residual interception unit is used to intercept weak scoring data calculated in the analysis and prediction unit but judged as risk-free because it is below the dynamic trigger threshold in real time, and extract the timestamp and spatial coordinates of the weak scoring data.

[0008] The spatiotemporal entropy aggregation unit is used to calculate the Shannon entropy value of weak scoring data within a preset geographic grid and sliding time window. The Shannon entropy value is used to characterize the degree of orderliness of weak signals in spatiotemporal distribution.

[0009] The agility bias unit is used to generate an agility coefficient based on the Shannon entropy value. When the Shannon entropy value is lower than the preset aggregation baseline, the agility coefficient is controlled to decrease non-linearly, and the agility coefficient is fed back to the analysis and prediction unit in real time to reduce the dynamic trigger threshold, so that the analysis and prediction unit switches to a high sensitivity state in the spatiotemporal environment of low entropy value.

[0010] The feedback mapping unit is used to receive on-site handling feedback information in response to the early warning command and to transform the feedback information into a real state vector consistent with the dimension of the early warning command.

[0011] The adaptive damping correction unit is used to calculate the difference vector between the predicted state vector and the actual state vector corresponding to the warning command, and to correct the weight value of the corresponding feature that triggers the warning command in the analysis and prediction unit based on the difference vector.

[0012] Preferably, the agility bias unit includes a hysteresis recovery logic module, which is used to block the step reset of the agility coefficient when the Shannon entropy value recovers to above the aggregation baseline, and control the agility coefficient to gradually return to the initial value according to a preset time constant, thereby suppressing the trigger oscillation of the analysis and prediction unit in the critical state.

[0013] Preferably, the adaptive damping correction unit includes a historical confidence constraint module, which is used to configure a damping coefficient for the difference vector, and the value of the damping coefficient is positively correlated with the historical accuracy of the feedback information source; when the adaptive damping correction unit corrects the weight value of the feature, the product of the difference vector and the damping coefficient is used as the step size for weight update to prevent the divergence of the feature weight model parameters caused by a single accidental feedback.

[0014] Preferably, when generating the sensitivity coefficient, the sensitivity bias unit executes the following operation logic including a non-linear mapping relationship: calculate the sensitivity coefficient β using the following formula: , where H is the Shannon entropy value within the current geographical grid output by the spatio-temporal entropy value aggregation unit, H base is the preset aggregation baseline, k is the maximum downward exploration amplitude coefficient and 0 < k < 1, n is the non-linear exponent for adjusting the response sensitivity and n > 1; the sensitivity bias unit applies the calculated sensitivity coefficient β as a multiplication factor to the initial preset threshold to generate a reduced dynamic trigger threshold.

[0015] Preferably, it further includes a reverse traceback compensation unit, which is used to monitor the external case handling input interface, extract the timestamp and geographical coordinates of the independent handling record when detecting an independent handling record not associated with any warning instructions; based on the timestamp and geographical coordinates, retrieve the historical perception state vector within the preset traceback time window from the historical data storage; identify the ineffective feature items in the historical perception state vector where the value is non-zero but the product of the corresponding feature weight and the value does not trigger a life warning instruction; generate a positive wake-up gain coefficient and superimpose the wake-up gain coefficient on the current dynamic weight coefficient of the corresponding ineffective feature item in the analysis and prediction unit.

[0016] Preferably, when generating the wake-up gain coefficient, the reverse traceback compensation unit configures the gain amplitude based on the event severity level recorded in the independent handling record; the higher the event severity level, the larger the modulus of the generated wake-up gain coefficient, so as to apply a higher-intensity weight compensation to the ineffective feature items that cause serious missed reports.

[0017] Preferably, the feedback mapping unit includes a physical semantic orthogonal verification module; the physical semantic orthogonal verification module is used to synchronously retrieve the original physical perception data when the on-site handling feedback information is received; use the preset physical semantic mapping table to verify whether there is a logical conflict between the event category represented by the on-site handling feedback information and the original physical perception data; if there is a logical conflict, generate a blocking signal to the adaptive damping correction unit to forcibly lock the weight value of the corresponding feature from being updated, preventing the feature weight model from being contaminated by incorrect feedback information.

[0018] Preferably, it also includes a resource counter-pressure control unit, which is used to monitor the occupancy status of administrative disposal resources in the target area in real time and calculate the resource saturation index; based on the resource saturation index, it generates a nonlinear threshold floating factor and applies the threshold floating factor to the early warning triggering logic of the analysis and prediction unit in real time; the threshold floating factor is configured to increase the threshold value for early warning triggering when the resource saturation index rises, so as to suppress the generation of low-risk level early warning instructions, thereby maintaining a dynamic load balance between the early warning generation rate and the on-site disposal capacity.

[0019] Preferably, the early warning instruction output by the analysis and prediction unit includes a predicted state vector, which includes risk type dimension components, urgency dimension components, and impact range dimension components; the feedback mapping unit transforms the on-site handling feedback information into a real state vector that is completely consistent with the predicted state vector in terms of dimension definition, and supports the adaptive damping correction unit to perform difference calculation based on vector space.

[0020] Preferably, the system is deployed in a hierarchical distributed computing architecture; the subthreshold residual interception unit and the spatiotemporal entropy aggregation unit are deployed on edge computing nodes to perform localized residual data cleaning and initial entropy calculation; the analysis and prediction unit and the adaptive damping correction unit are deployed on the central cloud server to perform global feature weight model maintenance and dynamic trigger threshold coordination.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. In the intelligent analysis of public security information, the feedback mapping unit transforms the on-site handling facts into a real state vector isomorphic to the predicted state. The difference vector directly drives the feature weights to make reverse corrections. The adaptive weighting unit introduces a damping constraint based on historical confidence, so that the model parameter update follows a small-step trend evolution logic. The negative feedback control loop enables the system to filter out data noise caused by single incidents or atypical cases when facing massive dynamic public security data, and avoids drastic oscillations in model weights. The system automatically fine-tunes the feature weights to fit the changes in the real public security situation based on the statistical trend of handling feedback, solving the problem that the fixed threshold rule of the existing technology will inevitably lead to the decay of timeliness and separation of fitting degree over time.

[0023] 2. The reverse backtracking compensation unit establishes a post-hoc correction mechanism triggered by independent handling records. When the system detects a security incident that actually occurs without triggering an early warning, it uses timestamps to locate and reconstruct a snapshot of the historical perception state within the spatiotemporal window. It identifies non-zero value features that fail to activate the early warning due to insufficient weight, applies compensatory gains, and uses missed events as training samples in the positive feedback compensation loop to activate dormant or excessively suppressed long-tail risk features. It logically blocks the drawbacks of conventional analysis systems that continuously raise thresholds to reduce false alarm rates, leading to irreversible degradation of the ability to perceive new or hidden risks. This ensures that the system suppresses false alarms and maintains a keen ability to capture weak risk signals.

[0024] 3. Construct a parallel subthreshold residual analysis channel with the main analysis engine to collect weak scoring data that has not reached the trigger threshold, calculate the distribution density and Shannon entropy value within the spatiotemporal grid, and use the orderliness of data distribution to characterize the tension of the environmental background. When weak signals show an abnormal clustering pattern of low entropy values, the system determines that the environmental pressure has increased and automatically generates an agility bias coefficient. By nonlinearly lowering the trigger threshold, the main engine enters a high-sensitivity state. The feedforward adjustment mechanism transforms the residual data, which is discarded as noise in traditional schemes, into intelligence resources, so that the system can dynamically adjust the early warning threshold according to the overall situation of the environmental field during the incubation stage before the outbreak of explicit conflict, and break through the masking effect of linear threshold judgment on discrete low-intensity risk signals. Attached Figure Description

[0025] Figure 1 This is a block diagram illustrating the logic principle of the dual-closed-loop dynamic control mechanism of the present invention.

[0026] Figure 2 This is a statistical chart comparing the false alarm rate under different signal-to-noise ratio environments according to the present invention.

[0027] Figure 3 This is a diagram of the layered distributed deployment architecture for cloud-edge collaboration in this invention. Detailed Implementation

[0028] This detailed description aims to illustrate the technical solutions defined in the claims, so that those skilled in the art can understand and implement the present invention. However, the following embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0029] This invention provides a social security information intelligent analysis system based on dynamic closed-loop control. Built upon a closed-loop control model of administrative management data flow, it reconstructs the unidirectional linear analysis process into a bidirectional coupled control structure with negative feedback damping correction and positive feedback inverse compensation capabilities. This enables the dynamic convergence of analysis rule parameters with the actual social security situation. The system mainly consists of an analysis and prediction unit, an adaptive damping correction unit, and a resource backpressure control unit deployed at the central cloud server, and subthreshold residual interception unit, spatiotemporal entropy aggregation unit, and agility bias unit deployed at edge computing nodes. The units interact with each other through standardized state vectors and control coefficients. Addressing the problem of decreased fit between the fixed weight system and the actual situation due to the time-dependent decay of social security risk characteristics, this system employs an adaptive damping weight adjustment mechanism based on the difference in handling potential energy. The analysis and prediction unit receives multi-source social security data, including alarm data, IoT sensing data, and data reported by grid workers. It calculates a risk score using a feature weight model. When the risk score exceeds a dynamic trigger threshold, the system generates an early warning command, including a predicted state vector V. pred This vector consists of risk type, urgency, and impact scope components. Taking the handling of potential mass incidents as an example, the generated predicted state vector V pred It can be quantified as a combined vector of type code 03, urgency 0.85, and range 0.6. After the task is completed, the on-site response force provides standardized response results via mobile terminal. The feedback mapping unit receives this feedback information and maps it into a real state vector V that is completely consistent with the dimension definition of the predicted state vector. real The adaptive damping correction unit executes the difference calculation procedure to calculate V. pred With V real The difference vector between them is ΔE, i.e., ΔE = V. real -V pred The adaptive damping correction unit drives the inverse correction of feature weights based on the difference vector ΔE. To prevent system parameter oscillations caused by single, occasional feedback, the unit incorporates a historical confidence constraint module. This module maintains historical accuracy records for different feedback sources and generates a damping coefficient λ. The correction logic follows W... new =W old The step size control strategy is +λ·ΔE·α, where W represents the feature weight that triggers the warning and α represents the feature correlation. When the difference vector ΔE indicates that the predicted value is higher than the true value, the system reduces the weight value of the corresponding feature to suppress false alarms; when ΔE indicates that the predicted value is lower than the true value, the system increases the weight value of the corresponding feature to reduce false alarms.

[0030] To address the issue that weak risk signals with low intensity but high frequency are often filtered out because they do not reach a fixed threshold, the system constructs a parallel subthreshold residual analysis channel. The subthreshold residual interception unit monitors the output of the analysis and prediction unit in real time and intercepts all risk scores S that are less than the dynamic trigger threshold T. trigger We obtain weak scoring data and extract its timestamp t and spatial coordinates (x,y) to generate residual particles P. weak The spatiotemporal entropy aggregation unit performs P analysis within a preset geographic grid and sliding time window. weak Spatial clustering analysis is performed, with the geographic grid set to a 500m x 500m area and the sliding time window set to 30 minutes. The spatiotemporal entropy aggregation unit uses the Shannon entropy principle to quantify the spatiotemporal order of weak signals, calculating the Shannon entropy value H within the current grid. When weak signals exhibit directional clustering, the H value decreases, indicating an increase in environmental background pressure. The agility bias unit dynamically adjusts the system sensitivity based on this entropy value. The system has a preset clustering baseline H. base The maximum downward amplitude coefficient k and the nonlinear exponent n, where k ranges from 0 to 1, and n is greater than 1. When the real-time calculated H is less than H base At that time, the system executes the formula β=1-k·[(H base -H) / H base ] n The agility coefficient β is calculated and fed back to the analysis and prediction unit as a multiplication factor to generate the reduced dynamic trigger threshold T'. trigger =T base •β enables the system to switch to a high-sensitivity state in a low-entropy spatiotemporal environment; to suppress triggered oscillations in the critical state, the agility bias unit includes a hysteresis recovery logic module, which recovers the Shannon entropy value H to the clustering baseline H. base When the above occurs, the module blocks the step reset of the agility coefficient β and controls β to perform exponential decay recovery according to the preset time constant τ, where the time constant τ can be set to 15 minutes.

[0031] To address the issue of sensitivity attenuation that may result from long-term suppression of false alarms, the system integrates a reverse backtracking compensation unit. This unit scans the alarm handling database in real time. When it detects an independent handling record that was manually entered and is not associated with any system warning ID, it determines that a missed event has occurred, and the unit extracts the timestamp T of the record. event and geographic coordinates L event Retrieve the time window [T] from the historical data storage event -Δt,T eventThe system scans the snapshot of the historical perception state vector within the [database name], where Δt can be set to 45 minutes. It identifies failure features with non-zero values ​​that did not trigger an alert due to insufficient weighting coefficients. The system generates a positive wake-up gain coefficient and adds it to the current dynamic weight of the corresponding feature. The magnitude of this coefficient is positively correlated with the severity level of the event. To prevent erroneous feedback from contaminating model parameters, the feedback mapping unit is equipped with a physical semantic orthogonal verification module. When receiving on-site handling feedback information, the module synchronously retrieves the original physical perception data triggered by the alert, such as ambient noise decibels or pedestrian density sensor readings. The module uses a preset physical semantic mapping table to verify whether the event category represented by the feedback information and the original physical perception data are logically mutually exclusive. If a logical conflict exists, the module generates a blocking signal to the adaptive damping correction unit, forcibly setting the damping coefficient λ to zero and locking the corresponding feature weight. Addressing the limited administrative handling resources, the resource counter-pressure control unit executes a global threshold bias based on load capacity. The unit connects to the duty management system to obtain the total number of on-duty police officers N in the target area. total The number of police officers N in a state of emergency busy Calculate the resource saturation index R_ sat =N busy / N total The unit is based on R sat Generate a nonlinear threshold floating factor γ, when R sat As R approaches 1, the factor γ increases, raising the warning trigger threshold; when R... sat When the value is low, the factor γ decreases.

[0032] Example 1: In this example, when the system is applied to a high-density dynamic security management scenario in a complex urban area and faces continuous environmental noise interference from compliant commercial promotions, the analysis and prediction unit triggers a predicted state vector V containing a high-risk urgency component based on acoustic features. pred On-site handling feedback indicated that the signal source was a compliant commercial promotion, and the feedback mapping unit generated a low-risk real state vector V. real The adaptive damping correction unit calculates the difference vector ΔE indicating that the predicted value is higher than the true value, and drives the downlink correction of the acoustic feature weights. This resolves the technical contradiction of the surge in false alarm rate under fixed rules in dynamic noise environments, and establishes a negative feedback mechanism for the automatic convergence of the analysis model to the background noise baseline of the real environment. Under the masking of the high noise environment, the subthreshold residual interception unit continuously captures potential concealed illegal aggregation behaviors that have not reached the dynamic trigger threshold T. trigger Given the weak scoring data, the spatiotemporal entropy aggregation unit calculates the Shannon entropy value H of the residual data within a sliding time window. When this value is lower than the preset clustering baseline H... baseWhen this occurs, it indicates that discrete weak signals exhibit an unnatural tendency to become ordered in the spatiotemporal dimension. Based on this, the agility bias unit generates an agility coefficient β less than 1 and applies nonlinear downward pressure to the initial preset threshold. This logic redefines the boundary of risk perception, so that the long-tail risk characteristics that were originally below the threshold are automatically highlighted and captured by the system in the spatiotemporal background of low entropy.

[0033] Under the boundary conditions of resource shortages caused by sudden large-scale public events, the resource counter-pressure control unit monitors the resource saturation index R in real time. sat As the threshold approaches, the system automatically generates an increased threshold fluctuation factor γ and applies it to the early warning trigger logic, dynamically raising the intervention standard to filter low-priority events. This maintains a load balance between the command system's decision-making bandwidth and on-site handling capabilities. Simultaneously, for occasional, unpredictable, but manually recorded independent handling records, the reverse backtracking compensation unit utilizes the event timestamp T. event By retrieving historical snapshots, identifying the failure characteristics that lead to missed reports, and superimposing wake-up gain coefficients, a data processing closed loop that does not rely on human intervention and has inherent evolutionary capabilities is constructed, solving the problem of adaptability of the analysis system in a complex and nonlinear social governance environment.

[0034] Example 2: To verify the model convergence characteristics and ability to capture weak risk signals of the social security information intelligent analysis system based on dynamic closed-loop control in complex noise environments, this example constructs a discrete event simulation (DES) test platform with multi-source data injection and feedback simulation functions. This platform runs on a computing server equipped with 128GB of memory and a 64-core processor. The Monte Carlo method is used to simulate the random occurrence of urban security events. The data used in the experiment comes from the publicly available City-Pulse urban sensing dataset. Gaussian white noise with a signal-to-noise ratio (SNR) of 15dB is actively superimposed using a signal generator to simulate sensor drift and environmental background interference in the real physical environment. To mitigate interference, the experiment used a traditional analysis system employing static threshold logic as the control group and a system deploying the adaptive damping correction unit and agility bias unit of this invention as the sample group. Both systems ran synchronously for 72 hours under the same input data stream. The setting of the sliding time window, a key parameter, was not based on empirical values ​​but rather determined using an engineering trade-off model between event evolution rate and data statistics. A time window that is too short is insufficient to smooth out spatial noise caused by GPS positioning drift; a time window that is too long leads to a lag in the integral response to sudden clustering events. Based on Shannon's sampling theorem and the statistical distribution of historical event durations, this experiment locked the sliding time window at 30 minutes.

[0035] The experiment was conducted in three phases. The first phase (0-24 hours) involved inputting standard security data. The second phase (24-48 hours) injected high-intensity environmental noise (simulating continuous high-decibel noise and pedestrian disturbance generated by large-scale commercial promotional activities). The third phase (48-72 hours), while maintaining a high-noise background, mixed in low-intensity but spatially directional signals simulating pickpocketing gangs. Throughout the experiment, the false alarm rate (FPR), false negative rate (FNR), and key intermediate variables within the sample group of this invention were recorded for both systems. Data showed that in the initial stage of noise injection in the second phase, the false alarm rate of the control group exhibited a step-like increase, peaking at 18.4%, and subsequently oscillating at a high level above 15.0%, indicating that the false alarm rate of the control group... The fixed-weight model cannot distinguish between effective risk and environmental background noise. In contrast, in the first two hours after noise injection, the difference vector ΔE calculated in this invention consistently indicates that the predicted value is higher than the actual feedback value, i.e., the magnitude of ΔE is negative. The system performs downlink correction based on the damping coefficient λ (dynamically set to 0.65) to drive the weights of acoustic and density features. The false alarm rate of this invention's sample gradually converges from a peak of 12.1% to a baseline level of 3.2% within 4.5 hours, verifying the ability of the negative feedback mechanism to suppress environmental noise. In the third stage, for the mixed low-intensity clustered signal, the control group had a false alarm rate as high as 85.7% because the signal strength did not exceed a fixed hard threshold (set to 0.6), while the spatiotemporal entropy aggregation unit of this invention's sample detected residual particles P. weak The distribution entropy value H decreased from the baseline of 4.8 bits to 2.3 bits, indicating that the weak signal exhibited unnatural ordered clustering. Based on this, the agility bias unit adjusted the agility coefficient β to 0.78, thus reducing the dynamic trigger threshold T'. trigger The temporary reduction enabled the sample group of this invention to successfully capture 88.5% of the simulated latent risks, as shown in Table 1, which records the system state snapshots at key moments in the third phase of the experiment.

[0036] Table 1: Snapshot of Key Data from Phase 3 of the Experiment (High-Noise Background + Latent Risks)

[0037]

[0038] Further stress tests showed that when the simulated administrative resource saturation R... sat When the inflection point exceeds 90%, the resource backpressure control unit intervenes, and the threshold floating factor γ rises exponentially. Data shows that under this high load condition, the early warning generation rate of the sample group of this invention is forcibly reduced by 42.6%, and the intercepted events are all low-priority events with an urgency dimension component of less than 0.4, thereby ensuring that the command issuance channel for high-risk events is not congested.

[0039] Example 3: This example combines Figures 1 to 3A description of a social security information intelligent analysis system based on dynamic closed-loop control, such as... Figure 1 As shown, the main logic flow of this system begins with input from multi-source public security data, including alarm data, IoT sensing data, and reports from grid workers, to the analysis and prediction unit. This unit calculates and generates early warning instructions and a predicted state vector V based on a feature weight model and risk score. pred The system's control closed loop comprises two main paths. The first is through on-site handling feedback information, which is input into the standardized results of the mobile terminal and transformed into the real state vector V by the feedback mapping unit. real The data is then transmitted to the adaptive damping correction unit to calculate the difference vector inverse correction weight and output the weight correction parameters to the analysis and prediction unit. Secondly, the subthreshold residual interception unit intercepts weak data below the trigger threshold, and the spatiotemporal entropy aggregation unit calculates the Shannon entropy value to represent the degree of order. Finally, the agility bias unit generates coefficients to dynamically reduce the trigger threshold and implement dynamic threshold bias.

[0040] like Figure 2 As shown in the chart, this graph compares the false negative rate of the control group and the false negative rate of the present invention sample group under different signal-to-noise ratio (SNR) dB environments. The horizontal axis represents five test nodes: 5dB, 10dB, 15dB, 20dB, and 25dB, while the vertical axis represents the false negative rate (%). The legend distinguishes the false negative rate of the control group with horizontal stripes and the false negative rate of the present invention sample group with diagonal stripes. The data shows that as the SNR increases from 5dB to 25dB, the false negative rate of the control group remains at a high level, while the false negative rate of the present invention sample group shows a decreasing trend, exhibiting extremely low false negative characteristics at 15dB and above. Figure 3 As shown, the overall topology of the system consists of a central cloud server, edge computing nodes, and a physical sensing and execution layer. The central cloud server integrates an intelligent analysis and prediction unit, an adaptive damping correction unit, and a global resource backpressure control module. It is connected to edge computing nodes A and B, which are distributed on both sides, through a data aggregation interface. Both edge nodes are equipped with subthreshold residual interception and spatiotemporal entropy initial calculation functions. The bottom layer is the physical sensing and execution layer, which includes environmental noise / people flow sensors, police handling feedback terminals, and security monitoring and acquisition equipment. This layer provides raw data to the upper-layer edge nodes on the one hand, and directly receives early warning instructions from the central cloud server and uploads the results for feedback on the other hand, forming a complete physical-digital image closed loop.

[0041] Example 4: This example aims to construct a complete data processing flow including a feedback orthogonal verification mechanism. In actual public security management scenarios, subjective judgments, data entry errors, or non-standardized task execution by on-site personnel may cause the feedback data to deviate from the actual physical facts. If this deviation directly enters the adaptive damping correction unit, it will lead to the incorrect convergence of the feature weight model. This example introduces a physical-semantic orthogonal verification module to construct a closed-loop verification chain from the original physical perception to the semantic logic judgment. In a typical noise disturbance incident handling scenario, the analysis and prediction unit generates a predicted state vector V based on the data collected by the acoustic monitoring network. pred The vector contains the duration of the high decibel level, spectral characteristics, and the coordinates of the sound source. After the on-site personnel arrive, they will report the handling results through a mobile terminal. The physical semantic orthogonal verification module will intercept the feedback information in real time and simultaneously retrieve the original physical perception data snapshot during the triggering warning period. This snapshot contains the decibel value curve and spectral density distribution of the environmental noise.

[0042] The physical semantic orthogonal verification module executes a logical consistency decision procedure, which maps unstructured feedback text into a semantic state vector V. sem The module uses a pre-defined physical-semantic mapping table to transform raw physical perception data into physical fact vectors V. phy If the ambient noise level in the original data consistently exceeds 90 dB and the spectral characteristics match an argument pattern, then V phy Pointing to a high-risk / conflict state, the system calculates V. sem With V phy If the consistency index α between the feedback information and the physical facts is lower than the preset confidence threshold (e.g., 0.6), the system determines that there is a logical conflict between the feedback information and the physical facts, that is, there may be dirty data injection. In this case, the module generates a blocking signal and sends it to the adaptive damping correction unit. This signal sets the damping coefficient λ of the current update cycle to zero, thereby cutting off the correction path of the erroneous feedback to the feature weight model at the physical level and locking the weight value of the corresponding feature so as not to be updated. If α is higher than the confidence threshold, the system allows the damping coefficient λ to be generated according to normal logic and fine-tunes the model based on the difference vector ΔE to ensure that the adaptive evolution of the system is always based on logically consistent data that has been verified by physical facts.

[0043] Example 5: For the initial deployment phase of the system in different geographical environments, a pre-deployment calibration procedure is performed to eliminate operational deviations caused by differences in environmental baselines. Before the system connects to the real-time data stream, environmental data covering typical weekday and weekend periods is continuously collected through a sensor network deployed in the target area. The system calculates the statistical distribution characteristics of noise power spectral density and pedestrian density within each monitoring grid, and sets the static noise baseline value N for each grid. baseFor fluctuations exceeding the baseline value of 3σ and confirmed as non-security-related routine interference, the system records the characteristic fingerprint into a suppression list for filtering in subsequent operations. Sensitivity calibration of the warning trigger threshold is performed. The system constructs a standard test vector set containing different intensities and types using typical cases confirmed in the historical police database. This test vector set is injected into the system, and the initial trigger threshold T is gradually adjusted. init The system continues until the recall rate of the test vector meets the preset target and the false positive rate is within the tolerance limit. For test vectors that are not recalled, the system analyzes the numerical distribution of the feature dimensions and adjusts the initial weight coefficients of the corresponding feature items until the risk score exceeds the threshold, thereby establishing an initial parameter configuration that is suitable for the current security characteristics of the area.

[0044] Example 6: This example details the standardized engineering calibration procedures for core control parameters before the system is officially put into operation. The aim is to eliminate nonlinear response deviations caused by regional differences. When determining the nonlinear exponent n and the maximum downward amplitude coefficient k in the agility bias unit, the engineering team does not rely on general empirical values, but performs gradient stress tests based on controlled signal injection. In the test environment, a set of standard weak signal sequences of preset strength is injected into the system. This sequence covers a continuum from background noise to critical risk. The system records response curves under different (n, k) combinations and constructs receiver operating characteristic curves. By calculating the area under the curve, the parameter combination that maximizes the area value is selected as the operating benchmark for the region. When the test data shows that the detection rate of low-intensity clustered events increases in an S-shape with the increase of the n value and reaches saturation at n=2.5, while the false alarm rate remains below the allowable limit, the system locks n to 2.5.

[0045] For the Shannon entropy calculation logic in the spatiotemporal entropy aggregation unit, the system adopts a probability distribution estimation method based on gridded residual density, dividing the target region into square grids with side length L, and counting the number c of residual particles falling into the i-th grid within the sliding time window T. i Define the event probability p of the i-th grid. i =c i / Σc i Therefore, according to the formula H=-Σp i ·log2(p i Calculate the entropy value of the current spatiotemporal slice to determine the clustering baseline H. base The system performs backtesting analysis on historical data during periods without events, calculates the mean μ and standard deviation σ of the entropy values ​​within that period, and then sets H... baseThe value is set to μ-3σ to ensure that the system only makes agile adjustments to statistical anomaly clusters. For the semantic vectorization process of the feedback mapping unit, the embodiment performs discretization calibration based on semantic anchors, pre-defining a set of standardized descriptive texts of the treatment results as semantic anchors, and strictly mapping these texts to the real state vector V. real For specific coordinate points, for non-standardized feedback text, the system calculates the cosine similarity between the text and each semantic anchor point, and uses a weighted interpolation method to generate the corresponding continuous state vector, ensuring that the mapping from unstructured administrative language to structured mathematical vectors is deterministic and reproducible.

[0046] Example 7: This example details the engineering boundary conditions and execution procedures of the system's core computing unit. When the physical semantic orthogonal verification module executes the verification logic, it receives unstructured on-site handling feedback information input based on the police professional dictionary semantic standardization interface. After keyword matching and synonym normalization, it is converted into a standard semantic vector. The cosine similarity between the vector and the preset semantic anchor point is calculated to determine the true state vector V. real Coordinate values, orthogonal verification refers to the independent source logical compatibility comparison between objective feature data from physical sensors and subjective semantic data from human feedback. When the environmental noise decibel value or pedestrian density value in the original physical perception data is in the physically impossible range of the event type indicated by the feedback information, a logical conflict judgment is generated and a blocking signal is triggered, so that the feature weight model is only iteratively corrected in the logically self-consistent data closed loop of physical facts and semantic description. In the spatiotemporal entropy aggregation unit operation procedure, the Shannon entropy value H is calculated using a gridded residual density probability distribution estimation method. The grid unit calculation process is excluded when the residual particle count is zero within the sliding time window to avoid logarithmic divergence, and the entropy value update is initiated when the total number of particles within the window exceeds the preset statistical sample size threshold. The nonlinear exponent n and the maximum downward amplitude coefficient k of the agility bias unit are determined by gradient pressure test. The calibration process includes injecting a set of weak signal sequences covering the background noise to the critical risk intensity standard into the system, recording the system response under different parameter combinations and constructing the receiver operation characteristic curve, and selecting the parameter combination that maximizes the area under the curve under the preset false alarm rate tolerance limit as the operating benchmark.

[0047] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A social security information intelligent analysis system based on dynamic closed-loop control, characterized in that, include: The analysis and prediction unit is used to receive multi-source public security data and calculate risk scores using a feature weighting model, and only generates warning instructions when the risk score exceeds the dynamic trigger threshold. The subthreshold residual interception unit is used to intercept weak scoring data calculated in the analysis and prediction unit but judged as risk-free because it is below the dynamic trigger threshold in real time, and extract the timestamp and spatial coordinates of the weak scoring data. The spatiotemporal entropy aggregation unit is used to calculate the Shannon entropy value of weak scoring data within a preset geographic grid and sliding time window. The Shannon entropy value is used to characterize the degree of orderliness of weak signals in spatiotemporal distribution. The agility bias unit is used to generate an agility coefficient based on the Shannon entropy value. When the Shannon entropy value is lower than the preset aggregation baseline, the agility coefficient is controlled to decrease non-linearly, and the agility coefficient is fed back to the analysis and prediction unit in real time to reduce the dynamic trigger threshold, so that the analysis and prediction unit switches to a high sensitivity state in the spatiotemporal environment of low entropy value. The feedback mapping unit is used to receive on-site handling feedback information in response to the early warning command and to transform the feedback information into a real state vector consistent with the dimension of the early warning command. An adaptive damping correction unit is used to calculate the difference vector between the predicted state vector and the actual state vector corresponding to the warning command, and to correct the weight value of the corresponding feature that triggers the warning command in the analysis and prediction unit based on the difference vector. In addition, when generating the sensitivity coefficient, the sensitivity offset unit executes the following operation logic including a non-linear mapping relationship: calculate the sensitivity coefficient β using the following formula: , where H is the Shannon entropy value within the current geographical grid output by the spatio-temporal entropy value aggregation unit, H base is a preset aggregation baseline, k is the maximum downward exploration amplitude coefficient and 0 < k < 1, n is a non-linear exponent for adjusting the response sensitivity and n > 1; the sensitivity offset unit applies the calculated sensitivity coefficient β as a multiplication factor to the initial preset threshold to generate a reduced dynamic trigger threshold; In addition, the system also includes a reverse backtracking compensation unit, which is used to monitor the external case handling input interface. When an independent handling record that is not associated with any warning instruction is detected, the unit extracts the timestamp and geographic coordinates of the independent handling record. Based on the timestamp and geographic coordinates, the unit retrieves the historical perception state vector within the preset backtracking time window from the historical data storage. The unit identifies failure feature items in the historical perception state vector whose values ​​are non-zero, but whose corresponding feature weights and values ​​do not trigger life warning instructions. The unit generates a positive wake-up gain coefficient and superimposes the wake-up gain coefficient onto the current dynamic weight coefficient of the corresponding failure feature item in the analysis and prediction unit. The feedback mapping unit includes a physical semantic orthogonal verification module. When receiving on-site handling feedback information, the physical semantic orthogonal verification module is used to simultaneously retrieve the original physical sensing data when the early warning command is triggered. Using a preset physical semantic mapping table, it verifies whether there is a logical conflict between the event category represented by the on-site handling feedback information and the original physical sensing data. If there is a logical conflict, a blocking signal is generated to the adaptive damping correction unit to forcibly lock the weight value of the corresponding feature and prevent it from being updated.

2. The social security information intelligent analysis system based on dynamic closed-loop control according to claim 1, characterized in that, The agility bias unit includes a hysteresis recovery logic module, which is used to block the step reset of the agility coefficient when the Shannon entropy value recovers to above the clustering baseline, and control the agility coefficient to gradually return to the initial value according to a preset time constant, thereby suppressing the triggering oscillation of the analysis and prediction unit in the critical state.

3. The social security information intelligent analysis system based on dynamic closed-loop control according to claim 1, characterized in that, The adaptive damping correction unit includes a historical confidence constraint module, which is used to configure damping coefficients for the difference vector. The value of the damping coefficients is positively correlated with the historical accuracy of the feedback information source. When correcting the weight values ​​of features, the adaptive damping correction unit uses the product of the difference vector and the damping coefficient as the step size for weight updates, preventing the divergence of feature weight model parameters caused by a single, occasional feedback.

4. The social security information intelligent analysis system based on dynamic closed-loop control according to claim 1, characterized in that, When generating the wake-up gain coefficient, the reverse backtracking compensation unit configures the gain amplitude based on the event severity level recorded in the independent handling record; the higher the event severity level, the larger the magnitude of the generated wake-up gain coefficient, thereby applying a higher intensity of weight compensation to the failure feature items that lead to serious missed reports.

5. The social security information intelligent analysis system based on dynamic closed-loop control according to claim 1, characterized in that, It also includes a resource counter-pressure control unit, which is used to monitor the occupancy status of administrative disposal resources in the target area in real time and calculate the resource saturation index; based on the resource saturation index, a nonlinear threshold floating factor is generated and applied to the early warning triggering logic of the analysis and prediction unit in real time; The threshold fluctuation factor is configured to raise the threshold value for triggering an early warning when the resource saturation index rises.

6. The social security information intelligent analysis system based on dynamic closed-loop control according to claim 1, characterized in that, The early warning instructions output by the analysis and prediction unit include a predicted state vector, which includes risk type dimension components, urgency dimension components, and impact range dimension components. The feedback mapping unit transforms the on-site handling feedback information into a real state vector that is completely consistent with the predicted state vector in terms of dimensional definition, and supports the adaptive damping correction unit to perform difference calculation based on vector space.

7. The social security information intelligent analysis system based on dynamic closed-loop control according to claim 1, characterized in that, The system is deployed in a hierarchical distributed computing architecture; the subthreshold residual interception unit and the spatiotemporal entropy aggregation unit are deployed on edge computing nodes to perform localized residual data cleaning and initial entropy calculation; the analysis and prediction unit and the adaptive damping correction unit are deployed on the central cloud server to perform global feature weight model maintenance and dynamic trigger threshold coordination.

Images