A subway intelligent security check data processing method based on edge computing

Abstract

The application discloses a subway intelligent security check data processing method based on edge calculation and relates to the technical field of intelligent security check, which comprises the following steps: acquiring multi-modal security check data, learning and matching space-time correlation features through an edge side machine learning mode to form multi-modal security check object data; performing multi-modal feature self-learning on the multi-modal security check object data, extracting multi-modal abnormal risk features, performing conflict determination on risk differences and track synchronization differences between different detection information through a cross-modal game reasoning mode, generating a multi-modal contradiction index; constructing an uncertainty machine learning evaluation by using the multi-modal contradiction index, performing multi-modal risk game determination on risk cognitive deviation conditions between different detection information and behavior consistency conditions between abnormal behaviors and equipment detection information, and obtaining uncertainty evaluation information. The application realizes dynamic adaptation between edge reasoning resources and risk processing requirements.

Description

Technical Field

[0001] This invention relates to the field of smart security inspection technology, and in particular to a data processing method for smart security inspection in subways based on edge computing. Background Technology

[0002] As urban rail transit systems evolve towards intelligence, subway security inspection systems have gradually transformed from traditional single-device detection modes to multi-device collaborative sensing modes. Currently, subway security inspection scenarios typically integrate various heterogeneous security inspection devices and utilize multimodal data fusion technology to comprehensively identify risks related to passengers, luggage, and passage behavior. Simultaneously, edge computing technology is widely applied to subway security inspection data processing. However, with the increasing complexity of passenger flow and the heterogeneity of equipment, traditional security inspection data processing methods are prone to problems such as inconsistent detection results, deviations in behavioral trajectory correlations, and conflicting risk perceptions.

[0003] In existing smart security inspection technologies for subways, on the one hand, most solutions only perform simple fusion processing of the results output by different detection devices, lacking the ability to dynamically correlate and analyze the risk differences and behavioral deviations between different modalities, which easily leads to high false alarm rates or missed detection of abnormal behaviors; on the other hand, existing edge security inspection systems usually adopt a fixed edge inference resource allocation method, which cannot dynamically schedule edge inference resources in combination with risk status and channel operation status, resulting in insufficient real-time processing capabilities in complex risk scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a data processing method for intelligent security inspection in subways based on edge computing to solve the problems of insufficient multimodal risk conflict identification capability and inability to dynamically schedule edge inference resources in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, this invention provides a data processing method for intelligent security checks in subways based on edge computing, comprising: acquiring multimodal security check data, and learning and matching spatiotemporal correlation features through edge-side machine learning to form multimodal security check object data; performing multimodal feature self-learning on the multimodal security check object data, extracting multimodal abnormal risk features, and determining conflicts between risk differences and trajectory synchronization differences between different detection information through cross-modal game reasoning to generate a multimodal contradiction index; constructing an uncertainty machine learning assessment using the multimodal contradiction index, and evaluating the deviation of risk perception and abnormal behavior between different detection information and equipment detection information. The system performs multimodal risk game analysis to determine the consistency of behavior among different security checkpoints, thereby obtaining uncertainty assessment information. Based on the uncertainty assessment information, resource occupancy status, channel operation status, and network transmission status, it constructs dynamic strategy state tokens for AI inference scheduling. It then uses the risk handling requirements corresponding to different security checkpoints to intelligently and hierarchically schedule edge inference resources, forming a multimodal security check data processing strategy. Based on the multimodal security check data processing strategy, multimodal conflict index, and uncertainty assessment information, it determines security check handling information and combines manual review conclusions, changes in modal conflicts, and risk convergence to perform incremental training, updates, and adaptive optimization of the multimodal security check data.

[0008] As a preferred embodiment of the subway smart security inspection data processing method based on edge computing described in this invention, the specific steps for forming multimodal security inspection object data are as follows:

[0009] Acquire multimodal security inspection data, and perform unified spatiotemporal mapping on the multimodal security inspection data based on the collection time, collection location and equipment source information to form a spatiotemporal collection record;

[0010] Passenger trajectory segments and baggage trajectory segments are generated using spatiotemporal data acquisition records, and a spatiotemporal correlation diagram is constructed based on the temporal and spatial correlation between the passenger trajectory segments and baggage trajectory segments;

[0011] Based on the spatiotemporal correlation graph, the correlation between passengers and luggage is learned and matched through edge-side machine association learning, and the multimodal security inspection data is encapsulated at the object level based on the learned matching information to form multimodal security inspection object data.

[0012] As a preferred embodiment of the edge computing-based smart security inspection data processing method for subways described in this invention, the specific steps for performing multimodal feature self-learning on multimodal security inspection object data and extracting multimodal abnormal risk features are as follows:

[0013] Cross-modal semantic mapping is performed on the multimodal detection information in the multimodal security inspection object data, and similarity alignment and difference reconstruction are performed on the correlation features between different modalities to form a unified latent space representation of multimodality;

[0014] We conduct temporal correlation analysis on the abnormal risk information in the multimodal unified latent space representation, and extract the risk change trend, behavior deviation trend and trajectory coupling abnormal trend from the temporal change process of the abnormal risk information to form multimodal abnormal risk features.

[0015] As a preferred embodiment of the subway smart security inspection data processing method based on edge computing described in this invention, the specific steps for generating the multimodal contradiction index are as follows:

[0016] Based on the characteristics of multimodal anomaly risks, we establish the correlation game relationship between different detection information, and conduct counterfactual perturbation analysis on the risk change trend, behavior deviation trend and trajectory coupling anomaly trend to generate modal conflict direction and conflict evolution intensity.

[0017] Cross-modal conflict propagation analysis is performed on the modal conflict direction and conflict evolution intensity, and a multimodal conflict index is calculated based on the correlation changes between the modal conflict direction and conflict evolution intensity.

[0018] As a preferred embodiment of the subway smart security inspection data processing method based on edge computing described in this invention, the specific steps of constructing an uncertainty machine learning evaluation using a multimodal contradiction index are as follows:

[0019] A dynamic risk claim sequence is constructed using the multimodal conflict index, modal conflict direction, and conflict evolution intensity. The temporal correlation of risk change relationships in the dynamic risk claim sequence is then reconstructed to form an evidence game matrix.

[0020] Cross-modal cognitive deviation propagation analysis was performed on the evidence game matrix, and a risk cognitive deviation map was generated based on the deviation changes between risk change trends, behavioral deviation trends and trajectory coupling anomaly trends.

[0021] As a preferred embodiment of the subway smart security inspection data processing method based on edge computing described in this invention, the specific steps for obtaining uncertainty evaluation information are as follows:

[0022] Based on the risk perception deviation diagram, a game theory analysis is conducted on the correlation changes between risk change trends, behavioral deviation trends, and trajectory coupling anomaly trends, and high cognitive uncertainty objects are formed based on the game theory analysis information;

[0023] The risk change trend, behavioral deviation trend and trajectory coupling anomaly trend corresponding to objects with high cognitive uncertainty are perturbed and extrapolated. Based on the changes after perturbed extrapolation, the risk cognition deviation and game change are analyzed in a coordinated manner to generate uncertainty assessment information.

[0024] As a preferred embodiment of the subway smart security inspection data processing method based on edge computing described in this invention, the step of constructing a dynamic strategy status token for artificial intelligence inference scheduling based on uncertainty evaluation information, resource occupancy status, channel operation status, and network transmission status, and intelligently hierarchically scheduling edge inference resources according to the risk processing requirements corresponding to different security inspection objects, is as follows:

[0025] Cross-dimensional state fusion is performed on uncertainty evaluation information, resource occupancy status, channel operation status and network transmission status, and the data is encoded according to the order of different state changes to form dynamic policy state tokens for artificial intelligence inference scheduling.

[0026] Based on the dynamic policy status token, the uncertainty changes, resource capacity and channel operation of different security inspection objects are correlated and analyzed. The edge inference resources are classified according to the processing requirements corresponding to the risk level to form an inference hierarchy configuration.

[0027] As a preferred embodiment of the subway smart security inspection data processing method based on edge computing described in this invention, the specific steps for forming a multimodal security inspection data processing strategy are as follows:

[0028] By utilizing the inference hierarchy configuration, the accuracy level of the inference path corresponding to different security inspection objects is adapted, and the scheduling priority is dynamically adjusted according to the resource consumption and processing requirements corresponding to the risk level of different inference paths.

[0029] Based on the dynamically adjusted inference path and scheduling priority, inference scheduling with corresponding accuracy levels is performed on different security inspection objects to form a multimodal security inspection data processing strategy.

[0030] As a preferred embodiment of the subway smart security inspection data processing method based on edge computing described in this invention, the specific steps for determining security inspection handling information based on multimodal security inspection data processing strategies, multimodal contradiction indices, and uncertainty evaluation information are as follows:

[0031] The multimodal security inspection data processing strategy is used to determine the inference level corresponding to the current security inspection object, and the risk judgment path with the corresponding accuracy level is called according to different inference levels;

[0032] By combining the risk assessment path with the multimodal contradiction index and uncertainty evaluation information, the risk status of the current security check target is assessed and the security check handling information is generated.

[0033] As a preferred embodiment of the subway smart security inspection data processing method based on edge computing described in this invention, the specific steps for incremental training, updating, and adaptive optimization of multimodal security inspection data by combining manual review conclusions, modal conflict changes, and risk convergence are as follows:

[0034] Based on the security check handling information, obtain the correlation changes between the manual review conclusions, modal conflict changes and risk convergence of the current security check object, and conduct feedback analysis to form feedback cognitive information;

[0035] Based on feedback and cognitive information, identify risk perception biases in the current security inspection data processing rules, and perform incremental training, updates, and adaptive optimization of multimodal security inspection data according to these risk perception biases.

[0036] The beneficial effects of this invention are as follows: By using edge computing to perform cross-modal contradiction propagation analysis on the risk change trends, behavioral deviation trends, and trajectory coupling anomaly trends among different detection information, dynamic identification of potential risk conflicts between multimodal security inspection information is achieved, thereby improving the accuracy of risk identification and reducing the false alarm rate; By using edge computing to perform fusion analysis on uncertainty evaluation information, resource occupancy status, channel operation status, and network transmission status, and by hierarchically scheduling edge inference resources according to the risk processing needs corresponding to different security inspection objects, dynamic adaptation between edge inference resources and risk processing needs is achieved, thereby improving real-time response capabilities and security inspection efficiency. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of a data processing method for smart security checks in subways based on edge computing.

[0039] Figure 2 A flowchart for generating multimodal security inspection object data.

[0040] Figure 3 A flowchart for generating the multimodal contradiction index.

[0041] Figure 4 A flowchart for constructing intelligent hierarchical scheduling of edge inference resources. Detailed Implementation

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0044] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0045] Reference Figures 1-4 This is one embodiment of the present invention, which provides a subway smart security inspection data processing method based on edge computing, including the following steps:

[0046] S1. Acquire multimodal security inspection data and learn and match spatiotemporal correlation features through edge-side machine learning to form multimodal security inspection object data.

[0047] Acquire multimodal security inspection data, and perform unified spatiotemporal mapping on the multimodal security inspection data based on the collection time, collection location and equipment source information to form a spatiotemporal collection record.

[0048] The specific steps include collecting image information, human body detection information, behavioral trajectory information and identity verification information of passengers during the passage process through X-ray security inspection equipment, human body detection equipment, video surveillance equipment and identity verification equipment, and summarizing the data generated by different equipment according to the corresponding equipment source to form multimodal security inspection data.

[0049] When performing unified spatiotemporal mapping of multimodal security inspection data using information on collection time, collection location, and device source, the collection time in the multimodal security inspection data is converted to the same time sequence, the collection location in the multimodal security inspection data is mapped to the same security inspection area, and the source of the collection device corresponding to the multimodal security inspection data is determined according to the device source information. This allows the multimodal security inspection data generated by different devices during the passage of the same passenger to be arranged in accordance with time sequence, spatial location, and device source, forming a spatiotemporal collection record.

[0050] Passenger trajectory segments and baggage trajectory segments are generated using spatiotemporal data collection and recording. A spatiotemporal correlation diagram is constructed based on the temporal and spatial correlation between the passenger trajectory segments and baggage trajectory segments.

[0051] The specific steps include using the collection time in the spatiotemporal acquisition records to associate the passage order of the same passenger at different acquisition locations to generate passenger trajectory segments, and using the collection time in the spatiotemporal acquisition records to associate the movement order of the same luggage at different acquisition locations to generate luggage trajectory segments.

[0052] Based on the collection time in passenger trajectory segments and baggage trajectory segments, the temporal correspondence between passengers and baggage during the processes of entering the security check area, passing through security equipment, and leaving the security check area is determined. Based on the collection location in passenger trajectory segments and baggage trajectory segments, the spatial correspondence between passengers and baggage in the same security check area is determined. Passenger trajectory segments and baggage trajectory segments with temporal and spatial correspondences are associated and matched to form a corresponding association between the passenger movement process and the baggage movement process, and a spatiotemporal association graph is constructed.

[0053] It should be noted that the spatiotemporal correlation diagram is a correlation diagram used to represent the temporal and spatial correspondence between passenger trajectory segments and baggage trajectory segments, including passenger trajectory segments, baggage trajectory segments, and their corresponding correlation relationships.

[0054] Based on the spatiotemporal correlation graph, the correlation between passengers and luggage is learned and matched through edge-side machine association learning, and the multimodal security inspection data is encapsulated at the object level based on the learned matching information to form multimodal security inspection object data.

[0055] The specific steps include arranging the passenger trajectory segments and baggage trajectory segments according to the collection time based on the passenger trajectory segments, baggage trajectory segments and corresponding relationships in the spatiotemporal correlation diagram, and comparing the order of appearance, stopping position and movement direction of the passenger trajectory segments and baggage trajectory segments in the same security check area. When the passenger trajectory segments and baggage trajectory segments maintain temporal and spatial correspondence in multiple consecutive collection positions, it is determined that there is a correlation between the passenger and the baggage.

[0056] The passenger information and baggage information corresponding to the learning and matching information are bound together. The image information, behavior trajectory information and identity verification information of the corresponding passenger generated in different security inspection devices are matched with the detection information and movement trajectory information of the corresponding baggage generated in different security inspection devices according to the collection time. The multimodal security inspection data are uniformly classified according to the correspondence between passengers and baggage. Multimodal security inspection data belonging to the same passenger and the same baggage are encapsulated at the object level to form multimodal security inspection object data.

[0057] It should be noted that edge-side machine association learning is a processing method that performs association analysis and learning matching on the temporal and spatial correspondences between passenger trajectory segments and baggage trajectory segments in edge computing devices.

[0058] S2. Perform multimodal feature self-learning on the multimodal security inspection object data, extract multimodal abnormal risk features, and use cross-modal game reasoning to determine the conflict between risk differences and trajectory synchronization differences between different detection information, and generate a multimodal contradiction index.

[0059] Cross-modal semantic mapping is performed on the multimodal detection information in the multimodal security inspection object data, and similarity alignment and difference reconstruction are performed on the correlation features between different modalities to form a unified latent space representation of multimodality.

[0060] The specific steps include associating image information, behavioral trajectory information, identity verification information, and detection information in the multimodal security inspection object data; comparing the behavioral content in the image information, the movement content in the behavioral trajectory information, the identity content in the identity verification information, and the detection content in the detection information; and uniformly mapping the associated content representing the same passage behavior, the same risk behavior, and the same security inspection status in different modal detection information; forming semantic associations between different modal detection information; and performing cross-modal semantic mapping.

[0061] The system matches associated features representing the same passage behavior, the same risk behavior, and the same security check status from different modal detection information, and maintains consistency in the same expression position for associated features with the same meaning. When reconstructing the associated features between different modalities, the associated features with differences in the detection information of different modalities are reorganized according to the corresponding risk meanings. The same and different parts in the detection information of different modalities can be represented simultaneously in a unified expression space, forming a multimodal unified latent space representation.

[0062] It should be noted that cross-modal semantic mapping arranges the behavioral content, location content, and identity content in different modal detection information with the detection content according to the collection time and collection location, and matches the associated content representing the same passage behavior, the same risk behavior, and the same security check status item by item to form a semantic correspondence between different modal detection information; the associated features are the feature content that can characterize the correspondence between different detection information and the same security check object, including time association features, spatial association features, behavioral association features, trajectory association features, and risk association features.

[0063] We conduct temporal correlation analysis on the abnormal risk information in the multimodal unified latent space representation, and extract the risk change trend, behavior deviation trend and trajectory coupling abnormal trend from the temporal change process of the abnormal risk information to form multimodal abnormal risk features.

[0064] The specific steps include arranging the abnormal risk information in the multimodal unified latent space representation in chronological order according to the collection time, comparing the risk status change information in different time stages, performing correlation analysis on the changes in behavior change information and trajectory change information in the same time stage, continuously recording the direction, magnitude and correspondence of abnormal risk information in the continuous time process, and forming the temporal correlation between abnormal risk information.

[0065] The risk state change direction of abnormal risk information in a continuous time period is extracted to form a risk change trend; the deviation of abnormal risk information in behavior in a continuous time period is extracted to form a behavior deviation trend; the abnormal situation corresponding to the trajectory of abnormal risk information in a continuous time period is extracted to form a trajectory coupled abnormal trend. The risk change trend, behavior deviation trend and trajectory coupled abnormal trend are combined in a unified way to form a multimodal abnormal risk feature.

[0066] Based on the characteristics of multimodal anomaly risks, we establish the correlation game relationship between different detection information, and conduct counterfactual perturbation analysis on the risk change trend, behavior deviation trend and trajectory coupling anomaly trend to generate modal conflict direction and conflict evolution intensity.

[0067] The specific steps include: based on the risk change trend, behavioral deviation trend and trajectory coupling anomaly trend in the multimodal anomaly risk characteristics, conducting correlation analysis on the risk change correspondence, behavioral change correspondence and trajectory change correspondence between different detection information, and comparing the corresponding changes between different detection information to form a game relationship of mutual influence and correspondence between different detection information.

[0068] The changes in risk change trends, behavioral deviation trends, and trajectory coupling anomaly trends are perturbed and replaced, and the risk changes, behavioral deviations, and trajectory coupling anomalies before and after the perturbed replacement are compared. When the risk change direction, behavioral deviation direction, and trajectory coupling anomaly direction of different detection information deviate from each other after the perturbed replacement, a modal conflict direction is generated. When the degree of deviation between different detection information after the perturbed replacement continues to increase over time, the conflict evolution intensity is generated.

[0069] It should be noted that modal conflict direction is information used to represent the change direction of risk change trend, behavior deviation trend or trajectory coupling anomaly trend between different detection information, including risk deviation direction, behavior deviation direction and trajectory deviation direction; conflict evolution intensity is information used to represent the intensity of the duration and expansion of conflict change between different detection information, including conflict duration change, conflict change magnitude and conflict diffusion change.

[0070] Cross-modal conflict propagation analysis is performed on the modal conflict direction and conflict evolution intensity, and the multimodal conflict index is calculated based on the correlation changes between the modal conflict direction and conflict evolution intensity. The expression for the index is:

[0071] ;

[0072] in, For multimodal contradiction index, For the total number of modal conflict propagations, This is the sequence number for modal conflict propagation. For the first The risk propagation intensity corresponding to each modal conflict direction For the first The historical stable propagation intensity corresponding to each modal conflict direction For the first The behavioral offset value corresponding to the intensity of each conflict evolution. For the first The trajectory synchronization reference value corresponding to the intensity of each conflict evolution. For the first Risk convergence value corresponding to each modal conflict direction For the first Historical risk convergence reference values ​​corresponding to each modal conflict direction.

[0073] It should be noted that, It is obtained by cross-modal correlation analysis of the risk change trend, behavior deviation trend, and trajectory coupling change trend among different detection information; It was obtained through statistical learning of the changes in risk propagation in the corresponding modal conflict direction during normal security checks in history; It is obtained by conducting a time-series comparative analysis of the differences in behavioral changes between the current abnormal behavior trajectory and the historical normal behavior trajectory; It was obtained through statistical learning of the spatiotemporal synchronization relationship between passenger trajectories and baggage trajectories under normal historical conditions; This was obtained through iterative analysis of the continuous decay of the degree of modal conflict change during the current risk propagation process; It is obtained by statistically learning the risk convergence changes corresponding to the stabilization of modal conflicts during normal security checks in history.

[0074] The specific steps include arranging the modal conflict directions corresponding to different detection information according to the acquisition time, comparing and correlating the conflict change directions between different detection information, continuously analyzing the changes in conflict evolution intensity in a continuous time phase, and forming a continuous propagation process of conflict change relationships between different detection information.

[0075] The modal conflict direction corresponding to different sets of detection information is used as the object of contradiction judgment. The corresponding risk propagation intensity and historical stable propagation intensity are obtained according to the modal conflict direction. The difference between the risk propagation intensity and the historical stable propagation intensity is used as the current risk propagation deviation. The corresponding behavioral offset value, trajectory synchronization reference value, risk convergence value and historical risk convergence reference value are obtained according to the conflict evolution intensity. The difference between the behavioral offset value and the trajectory synchronization reference value is used as the current behavioral trajectory deviation. The difference between the risk convergence value and the historical risk convergence reference value is used as the current risk stability deviation. The current risk propagation deviation, the current behavioral trajectory deviation and the current risk stability deviation are correlated to obtain the degree of contradiction change between different sets of detection information. All contradiction change degrees are normalized to obtain the multimodal contradiction index.

[0076] It should be noted that the risk propagation intensity is an intensity value used to represent the degree of change in the direction of modal conflict propagation between different detection information, and is generated by statistically analyzing the degree of deviation and continuous change in the direction of risk change between different detection information; the historical stable propagation intensity is a reference intensity value used to represent the stable risk propagation between different detection information in a historical time period, and is generated by statistically analyzing the stable change in the direction of risk change in a historical time period; the behavior deviation value is a value used to represent the degree of deviation between the behavior deviation trend and the normal behavior change, and is generated by statistically analyzing the direction of behavior change, the magnitude of behavior change, and the continuous deviation of behavior.

[0077] S3. Construct an uncertainty machine learning assessment using a multimodal contradiction index, and conduct multimodal risk game judgment on the deviation of risk perception between different detection information and the consistency between abnormal behavior and equipment detection information to obtain uncertainty evaluation information.

[0078] A dynamic risk claim sequence is constructed using the multimodal conflict index, modal conflict direction, and conflict evolution intensity. The temporal correlation of risk change relationships in the dynamic risk claim sequence is then reconstructed to form an evidence game matrix.

[0079] The specific steps include associating the degree of contradiction change corresponding to the multimodal contradiction index with the direction of conflict change corresponding to the modal conflict direction, as well as associating and matching the continuous change of conflict corresponding to the conflict evolution intensity with the corresponding modal conflict direction, and forming a continuous arrangement of risk change relationships between different detection information in chronological order to construct a dynamic risk claim sequence.

[0080] The risk change relationships, modal conflict directions, and conflict evolution intensity corresponding to different time stages in the dynamic risk claim sequence are arranged sequentially according to the collection time. Corresponding analysis is conducted on whether the risk change direction remains consistent in adjacent time stages, whether the modal conflict direction deviates, whether the conflict evolution intensity changes continuously, and whether the multimodal contradiction index changes synchronously. The risk change relationships with continuous correspondence between different time stages are correlated and organized. The risk change relationships in the dynamic risk claim sequence form a continuous temporal correspondence structure. Based on the correspondence between different detection information, the continuous temporal correspondence structure is matrix-arranged to form an evidence game matrix.

[0081] It should be noted that the evidence game matrix is ​​a matrix structure that represents the relationship between risk changes, modal conflict direction and conflict evolution intensity among different detection information, including the relationship between risk changes, modal conflict direction, conflict evolution intensity and corresponding time correlation.

[0082] Cross-modal cognitive deviation propagation analysis was performed on the evidence game matrix, and a risk cognitive deviation map was generated based on the deviation changes between risk change trends, behavioral deviation trends and trajectory coupling anomaly trends.

[0083] The specific steps include: correlating the risk change relationships, modal conflict directions, and conflict evolution intensity in the evidence game matrix; comparing and correlating the changes in risk change trends, behavioral deviation trends, and trajectory coupling anomaly trends among different detection information in continuous time stages; continuously recording the deviation direction, degree of deviation, and duration of deviation when there is a continuous deviation between the risk change direction, behavioral deviation direction, and trajectory anomaly direction in adjacent time stages; and correspondingly connecting the deviation change content in the previous time stage with the deviation change content in the next time stage to obtain the deviation change propagation relationship between different detection information and conducting cross-modal cognitive deviation propagation analysis.

[0084] The deviation direction, duration, and corresponding relationship of risk change trends, behavioral deviation trends, and trajectory coupling anomalies within the same time period are classified and organized. Corresponding deviations in risk change, behavioral change, and trajectory change between different detection information are recorded. Different detection information with corresponding deviations are used as association objects, and the deviation change direction, duration, and degree of deviation between different detection information are used as association content. The deviation change relationship between different detection information is connected and arranged, forming a graph-like association structure of risk perception deviation relationship between different detection information, thus forming a risk perception deviation graph.

[0085] It should be noted that the risk perception deviation map is used to represent the deviation propagation relationship between different detection information trends, behavioral deviation trends, and trajectory coupling anomalies. The risk perception deviation map can identify the continuous risk perception deviation between different detection information, providing a basis for the identification of objects with high cognitive uncertainty and the generation of uncertainty evaluation information.

[0086] Based on the risk perception deviation diagram, game theory analysis is conducted on the correlation changes between risk change trends, behavioral deviation trends, and trajectory coupling anomaly trends, and high cognitive uncertainty objects are formed based on the game theory analysis information.

[0087] The specific steps include: when performing a correlation analysis on the direction, duration, and degree of deviation in the risk perception deviation map, the deviation trends of risk change, behavioral deviation, and trajectory coupling anomaly trends among different detection information are arranged according to the collection time, and the direction, duration, and degree of deviation are continuously compared in different time stages; the mutual influence between the risk change trends, behavioral deviation trends, and trajectory coupling anomaly trends among different detection information is analyzed; when the deviation relationship between different detection information persists and the degree of deviation continues to increase, the corresponding risk change trends, behavioral deviation trends, and trajectory coupling anomaly trends are correlated and organized to form game analysis information, and based on the game analysis information, security inspection objects with persistent cognitive deviation relationships are identified, forming objects with high cognitive uncertainty.

[0088] It should be noted that objects with high cognitive uncertainty are security inspection objects in which there is a continuous cognitive deviation relationship between different detection information, including risk change trends, behavioral deviation trends, trajectory coupling anomaly trends and corresponding deviation changes.

[0089] The risk change trend, behavioral deviation trend and trajectory coupling anomaly trend corresponding to objects with high cognitive uncertainty are perturbed and extrapolated. Based on the changes after perturbed extrapolation, the risk cognition deviation and game change are analyzed in a coordinated manner to generate uncertainty assessment information.

[0090] The specific steps include: changing and replacing some risk change directions in the risk change trend corresponding to the high cognitive uncertainty object; adjusting some behavioral deviations in the behavioral deviation trend; adjusting some trajectory correspondences in the trajectory coupling anomaly trend; comparing the risk changes, behavioral changes, and trajectory changes after the disturbance adjustment with the corresponding changes before the disturbance adjustment item by item; continuously recording the differences in changes between different detection information to form disturbance inference information.

[0091] The changes in risk, behavior, and trajectory before and after the disturbance adjustment are arranged according to the collection time. The direction of risk change, behavioral deviation, and abnormal trajectory changes before and after the disturbance adjustment are compared item by item. Based on the differences in changes between different detection information, it is determined whether the risk perception deviation has weakened, continued, or expanded. Correlation analysis is performed on the correspondence between risk changes, behavioral changes, and trajectory changes between different detection information to determine whether the game changes tend to stabilize. When the risk perception deviation after the disturbance simulation still exists in multiple consecutive time stages and the game changes have not stabilized, the corresponding risk changes, behavioral changes, and trajectory changes are correlated and organized to form uncertainty assessment information.

[0092] S4. Based on uncertainty evaluation information, resource occupancy status, channel operation status and network transmission status, construct a dynamic strategy status token for artificial intelligence inference scheduling, and use the risk processing requirements corresponding to different security inspection objects to intelligently hierarchically schedule edge inference resources to form a multimodal security inspection data processing strategy.

[0093] Cross-dimensional state fusion is performed on uncertainty evaluation information, resource occupancy status, channel operation status and network transmission status, and the data is encoded according to the order of different state changes to form dynamic policy state tokens for artificial intelligence inference scheduling.

[0094] The specific steps include arranging the risk changes in uncertainty assessment information, resource changes in resource occupancy status, channel changes in channel operation status, and network changes in network transmission status in a corresponding manner; comparing and correlating the direction and degree of state changes with the corresponding state relationships in the same time period; uniformly classifying the state change content with corresponding relationships; and forming cross-dimensional correlation state relationships between different states.

[0095] The cross-dimensional correlation relationships corresponding to uncertainty evaluation information, resource occupancy status, channel operation status, and network transmission status are arranged sequentially according to the collection time, and the risk changes, resource changes, channel changes, and network changes in different time stages are marked respectively. The risk changes, resource changes, channel changes, and network changes corresponding to each other in the same time stage are uniformly numbered, and the numbering information of each time stage is arranged continuously according to the chronological order of collection time to obtain a dynamic strategy status token that can characterize the process of risk status, resource status, channel status, and network status changing over time for artificial intelligence inference scheduling.

[0096] It should be noted that dynamic policy status tokens are status-coded data used to represent the correspondence between uncertainty assessment information, resource occupancy status, channel operation status, and network transmission status, including the content of risk status changes, resource status changes, channel status changes, network status changes, and their corresponding time sequence.

[0097] Based on the dynamic policy status token, the uncertainty changes, resource capacity and channel operation of different security inspection objects are correlated and analyzed. The edge inference resources are classified according to the processing requirements corresponding to the risk level to form an inference hierarchy configuration.

[0098] The specific steps include extracting the risk status change content, resource status change content, and channel status change content corresponding to different time stages from the dynamic strategy status token, matching the risk changes and resource changes in the same time stage, associating the resource changes and channel operation changes, and continuously comparing the corresponding changes in different time stages to obtain the correlation between the uncertainty changes, resource carrying capacity, and channel operation for different security inspection objects.

[0099] A comprehensive analysis is conducted on the risk variation, duration, and deviation from risk perception for different security inspection targets. Based on this analysis, low-risk, medium-risk, and high-risk levels are determined. When matching the processing requirements corresponding to each risk level with resource availability, the resource occupancy, remaining resources, and processing capacity are analyzed. Based on the processing requirements for each risk level, available edge inference resources are determined. Low-precision, fast edge inference resources are allocated to security inspection targets at low risk levels, enhanced edge inference resources are allocated to those at medium risk levels, and high-precision edge inference resources are allocated to those at high risk levels. The order of edge inference resource allocation, resource occupancy, and processing priority for each risk level are then uniformly organized to form an inference hierarchy configuration.

[0100] It should be noted that the corresponding risk value is obtained based on quantitative information, and the risk value ranges from zero to one. When the risk value is between zero and 0.3, it is determined to be a low risk level. When the risk value is higher than 0.3 and less than or equal to 0.7, it is determined to be a medium risk level. When the risk value is higher than 0.7, it is determined to be a high risk level. The risk value is obtained by weighting the magnitude of risk change, the duration of risk, and the degree of deviation from risk perception.

[0101] By utilizing the inference hierarchy configuration, the accuracy level of the inference path corresponding to different security inspection objects is adapted, and the scheduling priority is dynamically adjusted according to the resource consumption and processing requirements corresponding to the risk level of different inference paths.

[0102] The specific steps include: using the risk level, edge inference resources, and processing order in the inference hierarchy configuration, analyzing the processing requirements corresponding to different security inspection objects, and selecting the inference path with the corresponding accuracy level according to the processing requirements corresponding to different risk levels. For security inspection objects corresponding to low risk levels, a low-precision fast inference path is matched; for security inspection objects corresponding to medium risk levels, an enhanced inference path is matched; and for security inspection objects corresponding to high risk levels, a high-precision inference path is matched, thus adapting the accuracy level of the inference path.

[0103] The system analyzes the computational, cache, and transmission resources used by different inference paths and matches the resource usage with the risk level of the security check objects. When the processing demand of a high-risk level is higher than the scheduling order of the current inference path, the scheduling priority of the corresponding security check object is increased. When the processing demand of a low-risk level is lower than the resource usage demand of the current inference path, the scheduling priority of the corresponding security check object is decreased. The processing order of different security check objects is then rearranged to dynamically adjust the scheduling priority.

[0104] Based on the dynamically adjusted inference path and scheduling priority, inference scheduling with corresponding accuracy levels is performed on different security inspection objects to form a multimodal security inspection data processing strategy.

[0105] The specific steps include: analyzing the risk level, edge inference resources, and processing order of different security inspection objects according to the dynamically adjusted inference path and scheduling priority; determining the inference execution order of different security inspection objects based on the dynamically adjusted scheduling priority; calling edge inference resources of corresponding accuracy levels according to the inference path of different security inspection objects; performing low-precision fast inference scheduling for security inspection objects corresponding to low-risk levels; performing enhanced inference scheduling for security inspection objects corresponding to medium-risk levels; performing high-precision inference scheduling for security inspection objects corresponding to high-risk levels; and organizing the resource call status, inference execution status, and processing information in different inference scheduling processes to form a multimodal security inspection data processing strategy.

[0106] S5. Based on the multimodal security inspection data processing strategy, multimodal conflict index and uncertainty evaluation information, determine the security inspection handling information, and combine the manual review conclusion, modal conflict change and risk convergence to perform incremental training update and adaptive optimization of the multimodal security inspection data.

[0107] The multimodal security inspection data processing strategy is used to determine the inference level corresponding to the current security inspection object, and the risk judgment path with the corresponding accuracy level is called according to different inference levels.

[0108] The specific steps include extracting the risk level, edge inference resource call status, and processing order corresponding to different security inspection objects from the multimodal security inspection data processing strategy, and performing a corresponding analysis on the risk changes, resource call requirements, and processing priority requirements corresponding to the current security inspection object. The risk changes are matched with the risk level, the resource call requirements are matched with the edge inference resource call status, and the processing priority requirements are matched with the processing order to determine the inference level corresponding to the current security inspection object.

[0109] The system analyzes the risk level, edge inference resource call requirements, and processing priority of the current security check object. Based on the processing requirements of different inference levels, it selects the corresponding accuracy level of risk assessment path: low inference level corresponds to low-precision fast risk assessment path, medium inference level corresponds to enhanced risk assessment path, and high inference level corresponds to high-precision risk assessment path. The system then calls the corresponding accuracy level of risk assessment path according to the processing order of the current security check object, and processes the inference execution process and resource call process in the risk assessment path accordingly.

[0110] By combining the risk assessment path with the multimodal contradiction index and uncertainty evaluation information, the risk status of the current security check target is assessed and the security check handling information is generated.

[0111] The specific steps include: conducting a correlation analysis on the changes in risk status in the risk assessment path, the degree of change in contradiction in the multimodal contradiction index, and the deviation in risk perception in the uncertainty evaluation information; and comprehensively judging the persistence, change, and deviation of risk corresponding to the current security check object; when the degree of change in contradiction corresponding to the multimodal contradiction index is high and the deviation in risk perception corresponding to the uncertainty evaluation information persists, high-risk handling information for the current security check object is determined; when the degree of change in contradiction corresponding to the multimodal contradiction index is low and the deviation in risk perception corresponding to the uncertainty evaluation information tends to stabilize, low-risk handling information for the current security check object is determined; and corresponding processing, review processing, and interception processing for different risk handling information are organized to form security check handling information.

[0112] Based on the security check processing information, obtain the correlation changes between the manual review conclusions, modal conflict changes and risk convergence of the current security check object, and conduct feedback analysis to form feedback cognitive information.

[0113] The specific steps include obtaining the corresponding manual review conclusion, modal conflict change and risk convergence status of the current security check object based on the risk handling information in the security check handling information, as well as the corresponding review change status in the manual review conclusion and the conflict change direction in the modal conflict change status, and matching the modal conflict change status with the risk stability change status in the risk convergence status to obtain the correlation change relationship.

[0114] The changes in the manual review conclusions are compared item by item with the high-risk and low-risk handling information in the security check handling information. Corresponding correlation analysis is performed on the direction and degree of conflict changes in the modal conflict changes and the risk stability changes in the risk convergence. When there are inconsistencies between the manual review conclusions and the security check handling information, and the modal conflict changes continue to change over a continuous period of time, and the risk convergence has not reached a stable state, the corresponding review changes, conflict changes, and risk stability changes are classified and organized, and linked and recorded according to the collection time to form feedback and cognitive information.

[0115] Based on feedback and cognitive information, identify risk perception biases in the current security inspection data processing rules, and perform incremental training, updates, and adaptive optimization of multimodal security inspection data according to these risk perception biases.

[0116] The specific steps include comparing the changes in the manual review conclusions with the high-risk and low-risk handling information in the security check handling information based on the changes in the review, conflict, and risk stability information in the feedback cognitive information, and recording any inconsistencies in the risk changes. The steps also involve continuously comparing the risk change relationships, behavioral change relationships, and trajectory change relationships in different time periods to determine the risk cognitive biases that exist in the current security check data processing rules.

[0117] The risk perception bias, behavioral change, and trajectory change relationships corresponding to the risk perception bias are extracted from the multimodal security inspection data, and compared item by item with the review changes corresponding to the manual review conclusions. Inconsistent risk perception bias, behavioral change, and trajectory change content are marked and re-incorporated into the corresponding multimodal security inspection data according to the collection time and equipment source information. The marked multimodal security inspection data are supplemented with corresponding risk perception bias and manual review conclusions to obtain the incrementally trained and adaptively optimized multimodal security inspection data.

[0118] In summary, this invention achieves dynamic identification of potential risk conflicts between multimodal security inspection information by using edge computing to analyze cross-modal contradiction propagation trends, behavioral deviation trends, and trajectory coupling anomaly trends among different detection information. This improves the accuracy of risk identification and reduces the false alarm rate. Furthermore, by fusing and analyzing uncertainty evaluation information, resource occupancy status, channel operation status, and network transmission status using edge computing, and by hierarchically scheduling edge inference resources according to the risk processing needs of different security inspection objects, this invention achieves dynamic adaptation between edge inference resources and risk processing needs, thereby improving real-time response capabilities and security inspection efficiency.

[0119] 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, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A data processing method for intelligent security checks in subways based on edge computing, characterized in that, include: Acquire multimodal security inspection data and learn and match spatiotemporal correlation features through edge-side machine learning to form multimodal security inspection object data; Multimodal feature self-learning is performed on multimodal security inspection object data, and multimodal abnormal risk features are extracted. Conflict judgment is made on the risk differences and trajectory synchronization differences between different detection information through cross-modal game reasoning, and a multimodal contradiction index is generated. An uncertainty machine learning assessment is constructed using a multimodal contradiction index, and multimodal risk game judgment is performed on the deviation of risk perception between different detection information and the consistency between abnormal behavior and equipment detection information to obtain uncertainty evaluation information; Based on uncertainty assessment information, resource occupancy status, channel operation status, and network transmission status, a dynamic strategy status token for artificial intelligence inference scheduling is constructed. The edge inference resources are intelligently hierarchically scheduled using the risk processing requirements corresponding to different security inspection objects, forming a multimodal security inspection data processing strategy. Based on the multimodal security inspection data processing strategy, multimodal conflict index and uncertainty evaluation information, security inspection handling information is determined, and the multimodal security inspection data is incrementally trained, updated and adaptively optimized in combination with manual review conclusions, changes in modal conflicts and risk convergence.

2. The subway smart security inspection data processing method based on edge computing as described in claim 1, characterized in that, The specific steps for generating multimodal security inspection object data are as follows: Acquire multimodal security inspection data, and perform unified spatiotemporal mapping on the multimodal security inspection data based on the collection time, collection location and equipment source information to form a spatiotemporal collection record; Passenger trajectory segments and baggage trajectory segments are generated using spatiotemporal data acquisition records, and a spatiotemporal correlation diagram is constructed based on the temporal and spatial correlation between the passenger trajectory segments and baggage trajectory segments; Based on the spatiotemporal correlation graph, the correlation between passengers and luggage is learned and matched through edge-side machine association learning, and the multimodal security inspection data is encapsulated at the object level based on the learned matching information to form multimodal security inspection object data.

3. The subway smart security inspection data processing method based on edge computing as described in claim 1 or 2, characterized in that, The specific steps for performing multimodal feature self-learning on multimodal security inspection object data and extracting multimodal abnormal risk features are as follows: Cross-modal semantic mapping is performed on the multimodal detection information in the multimodal security inspection object data, and similarity alignment and difference reconstruction are performed on the correlation features between different modalities to form a unified latent space representation of multimodality; We conduct temporal correlation analysis on the abnormal risk information in the multimodal unified latent space representation, and extract the risk change trend, behavior deviation trend and trajectory coupling abnormal trend from the temporal change process of the abnormal risk information to form multimodal abnormal risk features.

4. The subway smart security inspection data processing method based on edge computing as described in claim 3, characterized in that, The specific steps for generating the multimodal contradiction index are as follows: Based on the characteristics of multimodal anomaly risks, we establish the correlation game relationship between different detection information, and conduct counterfactual perturbation analysis on the risk change trend, behavior deviation trend and trajectory coupling anomaly trend to generate modal conflict direction and conflict evolution intensity. Cross-modal conflict propagation analysis is performed on the modal conflict direction and conflict evolution intensity, and a multimodal conflict index is calculated based on the correlation changes between the modal conflict direction and conflict evolution intensity.

5. The subway smart security inspection data processing method based on edge computing as described in claim 4, characterized in that, The specific steps for constructing an uncertainty machine learning evaluation using a multimodal contradiction index are as follows: A dynamic risk claim sequence is constructed using the multimodal conflict index, modal conflict direction, and conflict evolution intensity. The temporal correlation of risk change relationships in the dynamic risk claim sequence is then reconstructed to form an evidence game matrix. Cross-modal cognitive deviation propagation analysis was performed on the evidence game matrix, and a risk cognitive deviation map was generated based on the deviation changes between risk change trends, behavioral deviation trends and trajectory coupling anomaly trends.

6. The subway smart security inspection data processing method based on edge computing as described in claim 5, characterized in that, The specific steps for obtaining the uncertainty evaluation information are as follows: Based on the risk perception deviation diagram, a game theory analysis is conducted on the correlation changes between risk change trends, behavioral deviation trends, and trajectory coupling anomaly trends, and high cognitive uncertainty objects are formed based on the game theory analysis information; The risk change trend, behavioral deviation trend and trajectory coupling anomaly trend corresponding to objects with high cognitive uncertainty are perturbed and extrapolated. Based on the changes after perturbed extrapolation, the risk cognition deviation and game change are analyzed in a coordinated manner to generate uncertainty assessment information.

7. The subway smart security inspection data processing method based on edge computing as described in claim 6, characterized in that, The process involves constructing a dynamic strategy status token for AI inference scheduling based on uncertainty evaluation information, resource occupancy status, channel operation status, and network transmission status. It also utilizes the risk handling requirements corresponding to different security inspection targets to intelligently hierarchically schedule edge inference resources. The specific steps are as follows: Cross-dimensional state fusion is performed on uncertainty evaluation information, resource occupancy status, channel operation status and network transmission status, and the data is encoded according to the order of different state changes to form dynamic policy state tokens for artificial intelligence inference scheduling. Based on the dynamic policy status token, the uncertainty changes, resource capacity and channel operation of different security inspection objects are correlated and analyzed. The edge inference resources are classified according to the processing requirements corresponding to the risk level to form an inference hierarchy configuration.

8. The subway smart security inspection data processing method based on edge computing as described in claim 7, characterized in that, The specific steps for forming the multimodal security inspection data processing strategy are as follows: By utilizing the inference hierarchy configuration, the accuracy level of the inference path corresponding to different security inspection objects is adapted, and the scheduling priority is dynamically adjusted according to the resource consumption and processing requirements corresponding to the risk level of different inference paths. Based on the dynamically adjusted inference path and scheduling priority, inference scheduling with corresponding accuracy levels is performed on different security inspection objects to form a multimodal security inspection data processing strategy.

9. The subway smart security inspection data processing method based on edge computing as described in claim 8, characterized in that, The security inspection handling information is determined based on the multimodal security inspection data processing strategy, multimodal contradiction index, and uncertainty evaluation information. The specific steps are as follows: The multimodal security inspection data processing strategy is used to determine the inference level corresponding to the current security inspection object, and the risk judgment path with the corresponding accuracy level is called according to different inference levels; By combining the risk assessment path with the multimodal contradiction index and uncertainty evaluation information, the risk status of the current security check target is assessed and the security check handling information is generated.

10. The subway smart security inspection data processing method based on edge computing as described in claim 9, characterized in that, The incremental training, updating, and adaptive optimization of multimodal security inspection data, combining manual review conclusions, modal conflict changes, and risk convergence results, are carried out through the following specific steps: Based on the security check handling information, obtain the correlation changes between the manual review conclusions, modal conflict changes and risk convergence of the current security check object, and conduct feedback analysis to form feedback cognitive information; Based on feedback and cognitive information, identify risk perception biases in the current security inspection data processing rules, and perform incremental training, updates, and adaptive optimization of multimodal security inspection data according to these risk perception biases.

Images