Adaptive security monitoring method and system based on edge-cloud cooperation and artificial intelligence

Abstract

The application provides an adaptive security monitoring method and system based on edge-cloud cooperation and artificial intelligence, which comprises the following steps: acquiring a monitoring video stream on the edge side and generating a perception availability state quantity; determining the operation mode of the current security monitoring through a preset threshold value judgment based on the perception availability state quantity at the current moment and the smooth state quantity saved on the edge side at the last time slice; generating a decision quality summary through weighted summation based on the execution quality statistical result; acquiring the perception stability statistical result and performing normalization to generate a perception stability direction vector, which is applied to the perception availability state construction of the next monitoring period to form a continuous adaptive closed loop. The application can more effectively reduce invalid data interaction and bandwidth occupation, shorten the response path in critical scenarios, improve the long-term stability of the system in a complex dynamic environment, and enable the security monitoring system to have stronger adaptive operation capability under multiple scene deployment conditions.

Description

Technical Field

[0001] This invention belongs to the field of adaptive security monitoring, and particularly relates to an adaptive security monitoring method and system based on edge-cloud collaboration and artificial intelligence. Background Technology

[0002] As the requirements for real-time security capabilities in industrial parks, communities, transportation hubs, and key locations continue to increase, security monitoring systems have gradually evolved from the traditional manual duty and post-event review mode to intelligent systems that rely on video perception, automatic analysis, and coordinated response.

[0003] In existing technologies, a common approach involves front-end camera devices continuously collecting surveillance video, with edge nodes or cloud platforms handling tasks such as target detection, behavior analysis, scene recognition, and alarm decision-making. While this approach improves surveillance automation to some extent, it still presents significant engineering challenges under long-term operation and large-scale deployment conditions. First, surveillance video is characterized by continuous generation, long time spans, and frequent scene changes. If the system relies on fixed edge or cloud processing methods for extended periods, it can easily lead to excessive cloud load and bandwidth consumption during certain periods, or insufficient edge processing capabilities in complex scenarios, thus affecting the overall system stability. Second, security scenarios are not static environments. Factors such as changes in lighting, occlusion, fluctuations in personnel density, construction disturbances, and changes in device viewing angles continuously affect video perception quality. Existing solutions typically lack a unified metric for determining whether current perception conditions are sufficient to support independent edge decision-making. This results in the collaborative switching between edge and cloud relying heavily on preset rules, empirical thresholds, or static strategies, making real-time adjustments difficult based on operational status. Secondly, while existing systems can generate analysis results, they often lack a systematic measurement of the execution quality of a single actual operating cycle. Consequently, they lack an effective mechanism to use real-world results to correct front-end perception criteria, causing a gradual deviation between perception strategies and actual scenarios. This ultimately leads to increased false alarms, higher false negatives, or unreasonable resource scheduling. Furthermore, security videos inherently involve sensitive information. If cross-node interactions are frequently triggered without distinguishing perception states, it will further increase data transmission pressure and compliance risks.

[0004] Therefore, how to form a unified state criterion around the video perception process, drive the collaborative operation of edge and cloud based on the criterion, and periodically correct the criterion according to the actual operation quality, so as to balance real-time performance, resource utilization efficiency, operational stability and data security in long-term deployment, has become a key problem that existing technologies urgently need to solve. Summary of the Invention

[0005] The purpose of this invention is to propose an adaptive security monitoring method and system based on edge-cloud collaboration and artificial intelligence to solve the above-mentioned problems.

[0006] To achieve the above objectives, a first aspect of the present invention provides an adaptive security monitoring method based on edge-cloud collaboration and artificial intelligence, the method comprising the following steps: S1. Obtain the monitoring video stream from the edge side, perform video perception processing on the monitoring video stream to generate target continuity state value, behavior consistency state value and scene stability state value respectively, and then normalize the three state values ​​and weight them to generate a perception availability state quantity. S2. Based on the perceived availability state quantity at the current moment and the smoothed state quantity stored at the edge side in the previous time slice, generate the smoothed state quantity updated for the current time slice; based on the smoothed state quantity updated for the current time slice, combined with the deviation term, determine the current security monitoring operation mode through a preset threshold judgment, and generate the collaborative execution constraint corresponding to the operation mode. S3. Under the collaborative execution constraints, execute a complete security monitoring and analysis cycle, and collect execution quality statistics and perception stability statistics. Based on the execution quality statistics, generate a decision quality summary by weighted summation. The execution quality statistics include the edge-side task completion ratio, the cloud collaborative call usage ratio, and the result return timing consistency ratio. S4. Obtain the statistical results of perceived stability and normalize them to generate a perceived stability direction vector. Combine the decision quality summary and the preset reference operation quality level to adjust the weight parameters of the weighted convergence in S1. Apply the updated weight parameters to the construction of the perceived availability status in the next monitoring cycle to form a continuous adaptive closed loop.

[0007] Preferably, the video perception processing includes: continuous target perception processing, behavior consistency analysis processing, and scene stability analysis processing.

[0008] Preferably, the target continuity state value is generated and normalized by the target continuous perception processing, and its value is determined by the proportion of the target being successfully associated in a continuous time window; the behavior consistency state value is generated and normalized by the behavior consistency analysis processing, and its value is determined by the matching stability between the current trajectory and the basic behavior pattern library; the scene stability state value is generated and normalized by the scene stability analysis processing, and its value is determined by the statistical results of the changes in background block features in adjacent time windows.

[0009] Preferably, the step of generating the updated smooth state quantity for the current time slice based on the perceived availability state quantity at the current moment and the smooth state quantity stored at the edge side in the previous time slice is as follows: The perceived availability state quantity at the current moment and the smoothed state quantity stored in the edge control plane in the previous time slice are obtained. Combined with the smoothing coefficient, the smoothed state quantity updated for the current time slice is generated; wherein, the smoothing coefficient is used to adjust the trade-off between recent history and the current state.

[0010] Preferably, the deviation is the difference between the smoothed state quantity after the current time slice update and the perceived availability state quantity at the current moment; Then, the process of determining the current security monitoring operation mode by using a preset threshold specifically involves: Based on the smoothed state quantity updated after the current time slice, a deviation penalty coefficient is applied and then subtracted from the smoothed state quantity updated after the current time slice to form a decision quantity for segmentation determination; wherein, the deviation penalty coefficient is used to control the impact of deviation terms on the sensitivity of mode switching.

[0011] Preferably, the collaborative execution constraints are pre-configured for each operating mode based on the risk level of the monitored area, network conditions, edge node computing power specifications, and available cloud resources. The collaborative execution constraints include an upper limit on the frequency of cloud-based collaborative calls, a set of information digest types allowed to be carried in a single call, an upper limit on the concurrency of edge-side inference processes, and priority rules for collaborative task queues.

[0012] Preferably, the decision quality summary is obtained by weighting and summing the edge-side task completion rate, cloud-based collaborative call usage rate, and result return timing consistency rate, and then subtracting a risk adjustment term corresponding to the current operating mode; wherein, the value of the risk adjustment term is directly given by the risk level coefficient preset for each mode in the operating mode configuration table; the configuration weight of the weighted sum is used to adjust the relative influence of different statistical items in the comprehensive evaluation.

[0013] Preferably, the perception stability statistics include target continuity stability statistics, behavior consistency stability statistics, and scene stability statistics; Specifically, adjusting the weight parameters of the weighted aggregation in S1 involves: Obtain the perceived stability direction vector; Based on the perceived stability direction vector, the deviation between the decision quality summary and the preset reference operating quality level, and the set of weight parameters for perceived availability state quantities in the current period, and combined with regularization constraints to prevent weights from deviating excessively from the initial configuration, adjusted weight parameters are generated.

[0014] Preferably, after adjusting the weight parameters, the weight components of each weight parameter are normalized so that their sum remains constant, in order to ensure consistency with the weighted aggregation calculation method in S1.

[0015] A second aspect of the invention provides an adaptive security monitoring system based on edge-cloud collaboration and artificial intelligence, the system comprising: The video perception module is used to acquire the monitoring video stream on the edge side. By performing video perception processing on the monitoring video stream, it generates target continuity state value, behavior consistency state value and scene stability state value respectively. After normalizing the three state values, it weights and converges them to generate a perception availability state quantity. The mode decision module is used to generate the smooth state quantity updated for the current time slice based on the perceived availability state quantity at the current moment and the smooth state quantity stored at the edge side in the previous time slice; based on the smooth state quantity updated for the current time slice, combined with the deviation term, the current security monitoring operation mode is determined by a preset threshold judgment, and a collaborative execution constraint corresponding to the operation mode is generated. The collaborative execution module is used to execute a complete security monitoring and analysis cycle under the collaborative execution constraints, and collect execution quality statistics and perception stability statistics. Based on the execution quality statistics, a decision quality summary is generated by weighted summation. The execution quality statistics include the edge-side task completion ratio, the cloud collaborative call usage ratio, and the result return timing consistency ratio. The rule-adaptive module is used to acquire and normalize the statistical results of perception stability, generate a perception stability direction vector, and adjust the weight parameters of the weighted convergence of the video perception module in combination with the decision quality summary and the preset reference operating quality level. The updated weight parameters are then applied to the construction of the perception availability status in the next monitoring cycle, forming a continuous adaptive closed loop.

[0016] The beneficial technical effects of the present invention are at least as follows: This invention proposes an adaptive security monitoring method and system based on edge-cloud collaboration and artificial intelligence. Its core concept lies in constructing a perception availability state quantity based on intermediate states such as target continuity, behavioral consistency, and scene stability formed during video perception. This state quantity serves as the direct basis for edge-side operation mode selection and collaborative execution constraint generation, transforming the collaborative relationship between the edge and cloud from a fixed division of labor to a dynamically changing operational mechanism that adapts to current perception conditions. Furthermore, this invention organizes information such as execution results, collaborative invocation status, and temporal consistency from a single actual security monitoring cycle into a decision quality summary. This summary is then used to update the rule parameters of the perception availability state quantity, enabling the system to immediately enter the next operating cycle under new rule constraints. This forms a continuous adaptation process driven by real-world operational results. Compared to existing technologies, this invention does not simply add new perception models or overlay edge computing and cloud computing capabilities. Instead, it organizes the intermediate state of video perception, collaborative operation mode, execution quality characterization, and rule parameter updates into a single, continuous engineering link. This allows the system to first determine whether the edge side has the conditions to independently support security analysis in the current scenario, then select a matching collaborative method based on this determination, and subsequently correct the aforementioned determination criteria based on the actual operating results. The corrected rules are then immediately applied to the new monitoring cycle. In this way, this invention can more effectively reduce invalid data interaction and bandwidth consumption, shorten the response path in critical scenarios, improve the long-term stability of the system in complex dynamic environments, and enable the security monitoring system to have stronger adaptive operation capabilities under multi-scenario deployment conditions. Attached Figure Description

[0017] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0018] Figure 1 This is a flowchart of the adaptive security monitoring method based on edge-cloud collaboration and artificial intelligence of the present invention. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] like Figure 1 As shown in the embodiment of the present invention, an adaptive security monitoring method based on edge-cloud collaboration and artificial intelligence is provided. The method includes: S1. Obtain the monitoring video stream from the edge side, perform video perception processing on the monitoring video stream to generate target continuity state value, behavior consistency state value and scene stability state value respectively, and then normalize the three state values ​​and weight them to generate a perception availability state quantity. S2. Based on the perceived availability state quantity at the current moment and the smoothed state quantity stored at the edge side in the previous time slice, generate the smoothed state quantity updated for the current time slice; based on the smoothed state quantity updated for the current time slice, combined with the deviation term, determine the current security monitoring operation mode through a preset threshold judgment, and generate the collaborative execution constraint corresponding to the operation mode. S3. Under the collaborative execution constraints, execute a complete security monitoring and analysis cycle, and collect execution quality statistics and perception stability statistics. Based on the execution quality statistics, generate a decision quality summary by weighted summation. The execution quality statistics include the edge-side task completion ratio, the cloud collaborative call usage ratio, and the result return timing consistency ratio. S4. Obtain the statistical results of perceived stability and normalize them to generate a perceived stability direction vector. Combine the decision quality summary and the preset reference operation quality level to adjust the weight parameters of the weighted convergence in S1. Apply the updated weight parameters to the construction of the perceived availability status in the next monitoring cycle to form a continuous adaptive closed loop.

[0021] Specifically, in S1: This step focuses on characterizing the edge's own perception capabilities in the security monitoring scenario. It systematically integrates various intermediate states generated during real-time video perception to generate a perception availability state quantity that reflects the overall level of current perception conditions. Edge devices continuously receive monitoring video streams from fixed security cameras, typically deployed at park entrances / exits, passageways, or key areas. Video frames arrive sequentially in chronological order. After receiving video frames, the edge devices organize adjacent frames according to a preset time window. This allows perception processing to simultaneously utilize current image information and short-term trends, thus better reflecting the continuous nature of target activity and environmental changes in actual security scenarios. The preset time window is given by the edge node's runtime configuration file during system deployment and remains fixed during subsequent operation, not being reset in this step.

[0022] In video perception processing, edge devices run multiple processing modules targeting different perception dimensions in parallel. These modules share the same video input but focus on different aspects. Specifically, video perception processing involves: The target continuous perception processing uses a perception network composed of a convolutional feature extraction structure and a temporal association structure to locate targets in video frames and establish associations between adjacent frames. This processing forms an intermediate state that reflects whether the target can be stably associated within a continuous time window.

[0023] The behavior consistency analysis process, based on the continuous correlation results of the target, models the target's motion changes within a time window. Through temporal coding and behavior pattern matching, an intermediate state reflecting the consistency of current behavior changes is formed. The existing set of behavior patterns used for behavior pattern matching originates from the scenario calibration process during the initial system deployment phase: during equipment online acceptance or initial debugging, several continuous trajectory segments are extracted from normal traffic videos of the corresponding monitoring area, and trajectory representations are generated by the temporal coding module. These are further clustered to form a basic behavior pattern library for the area, which is used as the matching benchmark during subsequent operation.

[0024] The scene stability analysis and processing revolves around the video background structure. By dividing the background area into blocks and comparing the degree of difference in visual features of each block in adjacent time windows, an intermediate state reflecting the overall stability of the current monitoring environment is formed. The degree of change in the background structure is given by the statistical results of the feature difference of each background block.

[0025] It is important to understand that these intermediate states all originate from the same video perception process, but the perception conditions are described from the perspectives of target, behavior, and environment.

[0026] Furthermore, to utilize these intermediate states for system-level decision-making, edge devices abstract each intermediate state uniformly, constructing a perceptual availability state quantity. This state quantity is obtained by weighted aggregation after normalizing the target continuity state, behavioral consistency state, and scene stability state, and its calculation relationship is as follows: ;

[0027] in, This represents the perceived availability state quantity; It represents the target continuity state value obtained and normalized by the target continuous perception processing, and its value is determined by the proportion of the target that is successfully associated in the continuous time window; This represents the behavior consistency state value obtained and normalized by behavior consistency analysis. Its value is determined by the degree of matching stability between the current trajectory and the basic behavior pattern library. This represents the scene stability state value obtained and normalized by scene stability analysis. Its value is determined by the statistical results of the changes in background block features in adjacent time windows. , , These are the weight parameters that take effect in the current running cycle. This set of parameters is given by the scenario calibration configuration file when the system is first deployed, and is updated by the S4 rule management module and loaded at the beginning of each running cycle in subsequent runs.

[0028] In practical implementation, for example, within a certain time window, the proportion of stable associations obtained from continuous target perception processing is at a high level, corresponding to... The value is close to the configuration limit; in behavioral consistency analysis, the vast majority of target trajectories maintain a stable match with the basic behavioral pattern library of the region, corresponding to... It falls within the medium-to-high range; meanwhile, scene stability analysis shows that the overall difference in background block features is relatively small, corresponding to... It is in a relatively high range. In this case, the perceived availability state is obtained by weighting and aggregating the weight parameters that are in effect in the current period. The current edge-side perception of the monitored scene is at a relatively high level, reflecting good overall perception conditions. Conversely, during time windows with frequent changes in illumination or occlusion, the state values ​​corresponding to target continuity and scene stability decrease, resulting in a weighted calculation. The decrease directly reflects the changing trend of perceived conditions.

[0029] Specifically, in S2: After completing S1, the edge device obtains the perceived availability state quantity. This quantity is obtained by normalizing and weighting multiple intermediate sensing states, and is expressed using the same evaluation scale in engineering, making it convenient for direct use in scheduling decisions. This step... Converted into two types of executable control information: security monitoring operation mode Cooperative execution constraints In implementation, edge nodes will Write to the control plane shared memory or message queue; the run mode decision module reads the latest data at a fixed tick. And update and Then The rules are distributed to the edge-side inference process and the cloud-based collaborative invocation process, enabling subsequent security analysis to be executed according to predetermined patterns and boundaries. The collaborative execution constraint configuration table is established during the system deployment phase. Its establishment method is as follows: based on the risk level of the monitored area, network conditions, edge node computing power specifications, and available cloud resources, a set of corresponding execution constraints is pre-configured for each operating mode, and stored in the edge node's local configuration file or policy database mirror using the mode identifier as an index.

[0030] Furthermore, to achieve robust handling of short-term jitter, this step introduces a smoothing state variable. Its initial source is the exponentially weighted moving average, commonly used in signal processing and time series analysis. When used in this application, the recursive input is selected as the output of S1. The recursive output serves as the main state variable for pattern determination, forming the following update relationship: ;

[0031] in, The perceived availability state of the current time slice is derived from the aggregation result of S1 on the intermediate perceived states of the video within the current time window; This represents the smoothed state quantity stored in the edge control plane in the previous time slice; This represents the smoothed state quantity after the current time slice update, used for subsequent operation mode determination; The smoothing coefficient, derived from the runtime configuration file or policy configuration table of the edge node, is used to adjust the trade-off between recent history and current state.

[0032] Furthermore, after obtaining This step further incorporates the deviation of the current state from recent trends into the pattern determination, enabling the system to enhance coordination strength earlier in short-term unstable situations such as sudden occlusion, abrupt changes in illumination, and drastic background changes. Deviation terms are then... and with coefficient After weighting from The deduction from the middle value forms the decision quantity used for segmentation; then, based on two threshold parameters, the discretization selection of the three operating modes is realized, as follows: ;

[0033] in, This indicates the security monitoring operation mode output in the current time slice; , , The pattern identifiers representing different levels of coordination intensity are derived from the pattern enumeration table pre-set by the edge nodes; and These originate from the output of S1 and the recursive update result of the above formula, respectively. This represents the deviation penalty coefficient, which comes from the edge node policy configuration table and is used to control the impact of deviation items on mode switching sensitivity. , The mode boundary parameter is derived from the strategy configuration table of the deployment scenario and is used to divide the three-level coordination intensity range.

[0034] Understandably, to demonstrate the operability of the above calculation, an example of parameter substitution and calculation process is given. Assume that an edge node stored in the previous time slice... S1 outputs in the current time slice and using configuration parameters Then, first update according to the recursive formula to obtain Further adoption , , Then the deviation term is The penalty amount is Decision quantity is .because Segmentation determination In this mode, edge nodes select the corresponding collaborative execution constraints from a pre-defined collaborative execution constraint configuration table. And issued, coordinated execution constraints It can be implemented using structured fields, such as parameters including the upper limit of cloud collaborative call frequency, the set of information digest types allowed to be carried in a single call, the upper limit of edge inference process concurrency, and collaborative task queue priority rules.

[0035] Therefore, the output of this step is and It directly becomes the operational control input for subsequent security monitoring and analysis, enabling the system to adopt matching edge-cloud collaborative behavior under the current perception state and fluctuation level.

[0036] Specifically, in S3: The security monitoring operation mode has been obtained in S2. Cooperative execution constraints Subsequently, the edge node initiates a complete security monitoring and analysis cycle under these constraints, and quantifies the analysis execution process itself as the primary observation object. The analysis link used in this step follows the video perception processing link already established in S1, namely, three modules: continuous target perception processing, behavior consistency analysis processing, and scene stability analysis processing. The difference is that these modules are no longer used solely for generating perception availability state quantities in this step, but rather in the operational mode... Cooperative execution constraints Within the defined boundaries of responsibility allocation and scheduling, it is organized into a complete edge-cloud collaborative analysis and execution process. The edge side is responsible for... The execution constraints in the cloud perform basic analysis, information summary organization, and necessary local judgment. When triggered, the cloud side performs supplementary analysis on the specified information summary and sends the results back to the edge side.

[0037] In the specific execution process, the edge nodes operate in the following mode. The above analysis chain is invoked under the specified division of responsibilities, and strictly in accordance with the collaborative execution constraints. Scheduling is performed. During execution, the inference process continuously generates scheduling logs and cooperative call records. Edge nodes accumulate these records over a single runtime cycle, generating two sets of statistical results: The first set consists of execution quality statistics, used to construct a decision quality summary, including the percentage of tasks completed on the edge side, the percentage of collaborative calls on the cloud, and the percentage of consistent result return timing.

[0038] The second set consists of statistical results on perception stability, used to characterize the stability of the three types of perception states in S1 within this period, including statistics on target continuity stability, statistics on behavioral consistency stability, and statistics on scene stability; among them, the statistics on target continuity stability are derived from the data within this period. The fluctuation records and consistent, stable statistics are derived from the current period. The fluctuation records and scenario stability statistics are derived from the current period. The fluctuation records are used by the rule management module to read these three types of statistical results at the end of the cycle and normalize them to form the perceptual stability direction vector used by S4. .

[0039] Furthermore, with operating mode Corresponding risk adjustment items Derived from the same operating mode configuration table used by S2, each mode corresponds to a preset risk level coefficient during deployment, therefore the risk adjustment item The current operating cycle is characterized by the operating mode. A uniquely determined constant.

[0040] The construction of the decision quality summary employs a multi-indicator weighted evaluation approach, its mathematical foundation derived from the classic weighted sum model. Specifically, the decision quality summary... Calculated according to the following relationship: ;

[0041] in, This indicates the proportion of analysis tasks completed by the edge side within the current running cycle. This proportion is calculated by the number of tasks actually completed by the edge inference process within the cycle and the number of tasks allocated by the scheduling module. This indicates the proportion of cloud-based collaborative call intensity, derived from the number of collaborative calls and collaborative execution constraints. The ratio between the maximum call limits configured in the settings; It indicates the degree of temporal consistency of the collaborative execution results, and its source is the proportion of key intermediate results that are correctly returned in the triggering order within this period; Representation and Operating Mode The corresponding risk adjustment item is directly given by the risk level coefficient preset for each mode in the operation mode configuration table; , , , To configure weights and adjust the relative influence of different statistical items in the comprehensive evaluation, this set of weights is preset during system deployment and remains fixed during subsequent operation.

[0042] Assuming that within a certain operating cycle, the edge side completes the vast majority of analysis tasks, and obtains... The proportion of cloud-based collaborative calls is at a moderate level, which is satisfactory. The proportion of key intermediate results returned correctly in the triggering order is: ,get Current operating mode The corresponding risk adjustment item is taken Configure weights , , , Substituting the values ​​into the calculation, we get: This result serves as a summary of the decision quality for this operational cycle, fully reflecting the execution quality level of the system in completing a security monitoring analysis under the combined influence of the perception state formed in S1 and the operational mode and collaborative constraints determined in S2.

[0043] Specifically, in S4: This step outputs a summary of the decision quality from S3. The system takes as input the target continuity stability statistics, behavior consistency stability statistics, and scene stability statistics recorded by S3 within the current operating cycle. It focuses on how to directly transform the execution quality of a security monitoring operating cycle into an effective perception availability state rule update, and immediately complete the new security monitoring operating cycle under the constraints of the updated rule parameters. Specifically, the target continuity stability statistics are obtained by S3 from the target continuity perception processing and normalized target continuity state values. The fluctuation records within this period are calculated, and the behavioral consistency stability statistics are obtained by S3 from the behavioral consistency analysis and normalized behavioral consistency state values. The fluctuation records within this period are calculated, and the scene stability statistics are obtained by S3 from the scene stability analysis and normalized scene stability state values. The fluctuation records within this cycle are calculated. In system implementation, at the end of each operating cycle, the edge node will... The statistical records formed within this period regarding the continuity of objectives, consistency of behavior, and stability of the scenario are written into the period buffer of the rule management module, and the rule management module uniformly triggers rule update calculations at the period boundary.

[0044] Furthermore, the target object of the rule update is the state quantity in S1 used to construct the perceived availability state. The set of weight parameters, i.e. , , When the system is first deployed, these parameters use the initial values ​​in the scenario calibration configuration file. After each running cycle, the rule management module updates these parameters based on the running quality and perceived stability statistics of the current cycle and loads them in the next cycle.

[0045] Furthermore, in the mathematical modeling of rule updates, this step adopts the proportional adjustment concept derived from classical adaptive control and online optimization, combined with regularization terms from constraint optimization, to form a weight update relationship suitable for long-term operation in security scenarios. Specifically, the weight parameters are represented as vectors. , No. After each running cycle, the weight update relationship is expressed as follows: ;

[0046] in, Indicates the first The quantity used to calculate the perceived availability state in each running cycle The set of weight parameters is derived from the current configuration loaded by the rule management module at the beginning of this period; This represents the decision quality summary output by S3 at the end of the cycle; This represents the preset reference operating quality level in the current security deployment scenario, and its value is given by the scenario calibration file during the deployment phase. This represents the perception stability direction vector for this period. Its three components correspond to target continuity stability statistics, behavior consistency stability statistics, and scene stability statistics, respectively. All three types of statistical results are generated by S3 around this period. , , The fluctuation records are calculated, and the rule management module reads these three types of statistical results at the end of the cycle and normalizes them to form a vector. ; This represents the proportional update coefficient, used to control the magnitude of weight adjustment; This represents the regularization coefficient, used to suppress deviations of weight parameters from their initial configuration. Too far; This represents the initial set of weights written in the scenario calibration configuration file during the initial deployment phase.

[0047] Furthermore, after completing the above update, to ensure that the weight parameters remain consistent with those in S1 in subsequent cycles... The calculation method is consistent. The rule management module normalizes the update results before normalization to keep the sum of the three components constant. The relationship is as follows: ;

[0048] in, , , This represents the three components of the updated result vector before normalization, and the three components after normalization. It is directly used as the set of weight parameters loaded by S1 in the next running cycle.

[0049] The update process is illustrated with specific examples. Assume an initial weight in a certain security deployment scenario. , No. At the start of each running cycle The decision quality summary output by S3 is: Reference quality level Configuration parameters , And obtained by statistical normalization of the three types of sensing stability in this period. Substituting the update relation, we get: ; ;

[0050] Therefore, the update result before normalization is: The sum of its three components is After normalization, the new weight set is approximately This weight set is directly loaded into S1 in the next runtime cycle to calculate the new perceived availability state quantity. After the rules are updated, the edge nodes immediately start a new security monitoring cycle under the constraints of the updated perceived availability state rule parameters, executing S1 to S3 sequentially to form a new cycle. New operating mode New collaborative execution constraints This will ultimately generate a new summary of decision quality.

[0051] This invention also provides an adaptive security monitoring system based on edge-cloud collaboration and artificial intelligence, the system comprising: The video perception module is used to acquire the monitoring video stream on the edge side. By performing video perception processing on the monitoring video stream, it generates target continuity state value, behavior consistency state value and scene stability state value respectively. After normalizing the three state values, it weights and converges them to generate a perception availability state quantity. The mode decision module is used to generate the smooth state quantity updated for the current time slice based on the perceived availability state quantity at the current moment and the smooth state quantity stored at the edge side in the previous time slice; based on the smooth state quantity updated for the current time slice, combined with the deviation term, the current security monitoring operation mode is determined by a preset threshold judgment, and a collaborative execution constraint corresponding to the operation mode is generated. The collaborative execution module is used to execute a complete security monitoring and analysis cycle under the collaborative execution constraints, and collect execution quality statistics and perception stability statistics. Based on the execution quality statistics, a decision quality summary is generated by weighted summation. The execution quality statistics include the edge-side task completion ratio, the cloud collaborative call usage ratio, and the result return timing consistency ratio. The rule-adaptive module is used to acquire and normalize the statistical results of perception stability, generate a perception stability direction vector, and adjust the weight parameters of the weighted convergence of the video perception module in combination with the decision quality summary and the preset reference operating quality level. The updated weight parameters are then applied to the construction of the perception availability status in the next monitoring cycle, forming a continuous adaptive closed loop.

[0052] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0053] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0054] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0055] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An adaptive security monitoring method based on edge-cloud collaboration and artificial intelligence, characterized in that, The method includes: S1. Obtain the monitoring video stream from the edge side, perform video perception processing on the monitoring video stream to generate target continuity state value, behavior consistency state value and scene stability state value respectively, and after normalizing the three state values, use the weight parameters that are effective in the current running cycle to weight and aggregate to generate the perception availability state quantity. S2. Obtain the perceived availability state quantity at the current moment and the smoothed state quantity stored in the edge control plane in the previous time slice. Combine the smoothing coefficient to generate the smoothed state quantity updated for the current time slice. Based on the smoothed state quantity updated for the current time slice, and combined with the deviation term, determine the current security monitoring operation mode through a preset threshold, and generate the collaborative execution constraint corresponding to the operation mode. The smoothing coefficient is used to adjust the trade-off between recent history and the current state. The deviation term is the deviation between the smoothed state quantity updated for the current time slice and the perceived availability state quantity at the current moment. The process of determining the current security monitoring operation mode through a preset threshold is as follows: Based on the smoothed state quantity updated after the current time slice, a deviation penalty coefficient is applied and then subtracted from the smoothed state quantity updated after the current time slice to form a decision quantity for segmentation determination; wherein, the deviation penalty coefficient is used to control the impact of deviation terms on the sensitivity of mode switching. Specifically, based on the risk level of the monitored area, network conditions, edge node computing power specifications, and available cloud resources, a set of corresponding collaborative execution constraints are pre-configured for each operating mode; the collaborative execution constraints include the upper limit of cloud collaborative call frequency, the set of information digest types allowed to be carried in a single call, the upper limit of edge-side inference process concurrency, and collaborative task queue priority rules. S3. Under the aforementioned collaborative execution constraints, execute a complete security monitoring and analysis cycle, and collect execution quality statistics and perception stability statistics. A decision quality summary is obtained by weighted summation of the execution quality statistics and subtracting the risk adjustment term corresponding to the current operating mode. The execution quality statistics include the edge-side task completion rate, the cloud-based collaborative call usage rate, and the result return timing consistency rate. The value of the risk adjustment term is directly given by the risk level coefficient preset for each mode in the operating mode configuration table. S4. Obtain the perceived stability statistics and normalize them to generate a perceived stability direction vector. Based on the perceived stability direction vector, the deviation between the decision quality summary and the preset reference operating quality level, and the set of weight parameters for the perceived availability state in the current period, and combined with the regularization constraints used to prevent the weights from deviating excessively from the initial configuration, generate adjusted weight parameters. Adjust the weight parameters of the weighted aggregation in S1, and apply the updated weight parameters to the construction of the perceived availability state in the next monitoring period to form a continuous adaptive closed loop.

2. The adaptive security monitoring method based on edge-cloud collaboration and artificial intelligence according to claim 1, characterized in that, The video perception processing includes: continuous target perception processing, behavior consistency analysis processing, and scene stability analysis processing.

3. The adaptive security monitoring method based on edge-cloud collaboration and artificial intelligence according to claim 2, characterized in that, The target continuity state value is generated and normalized by the target continuous perception processing, and its value is determined by the proportion of the target being successfully associated in a continuous time window; the behavior consistency state value is generated and normalized by the behavior consistency analysis processing, and its value is determined by the matching stability between the current trajectory and the basic behavior pattern library; the scene stability state value is generated and normalized by the scene stability analysis processing, and its value is determined by the statistical results of the changes in background block features in adjacent time windows.

4. The adaptive security monitoring method based on edge-cloud collaboration and artificial intelligence according to claim 1, characterized in that... in, The weights in the weighted summation are used to adjust the relative influence of different statistical items in the comprehensive evaluation.

5. The adaptive security monitoring method based on edge-cloud collaboration and artificial intelligence according to claim 1, characterized in that, The statistical results of perception stability include target continuity stability statistics, behavior consistency stability statistics, and scene stability statistics.

6. The adaptive security monitoring method based on edge-cloud collaboration and artificial intelligence according to claim 5, characterized in that, After adjusting the weight parameters, the weight components of each weight parameter are normalized so that their sum remains constant, ensuring consistency with the weighted aggregation calculation method in S1.

7. An adaptive security monitoring system based on edge-cloud collaboration and artificial intelligence, characterized in that, The system includes: The video perception module is used to acquire the monitoring video stream from the edge side. By performing video perception processing on the monitoring video stream, it generates target continuity state value, behavior consistency state value and scene stability state value respectively. After normalizing the three state values, it uses the weight parameters that are effective in the current running cycle to weight and aggregate them to generate the perception availability state quantity. The mode decision module is used to obtain the perceived availability state quantity at the current moment and the smoothed state quantity stored in the edge control plane in the previous time slice, and generate the smoothed state quantity updated for the current time slice by combining the smoothing coefficient; based on the smoothed state quantity updated for the current time slice, combined with the deviation term, the current security monitoring operation mode is determined by judging through a preset threshold, and a collaborative execution constraint corresponding to the operation mode is generated; wherein, the smoothing coefficient is used to adjust the trade-off between recent history and current state; the deviation term is the deviation between the smoothed state quantity updated for the current time slice and the perceived availability state quantity at the current moment; The process of determining the current security monitoring operation mode through a preset threshold is as follows: Based on the smoothed state quantity updated after the current time slice, a deviation penalty coefficient is applied and then subtracted from the smoothed state quantity updated after the current time slice to form a decision quantity for segmentation determination; wherein, the deviation penalty coefficient is used to control the impact of deviation terms on the sensitivity of mode switching. Specifically, based on the risk level of the monitored area, network conditions, edge node computing power specifications, and available cloud resources, a set of corresponding collaborative execution constraints are pre-configured for each operating mode; the collaborative execution constraints include the upper limit of cloud collaborative call frequency, the set of information digest types allowed to be carried in a single call, the upper limit of edge-side inference process concurrency, and collaborative task queue priority rules. The collaborative execution module is used to execute a complete security monitoring and analysis cycle under the collaborative execution constraints, and collect execution quality statistics and perception stability statistics. A decision quality summary is obtained by weighted summing of the execution quality statistics and subtracting the risk adjustment term corresponding to the current operating mode. The execution quality statistics include the edge-side task completion rate, the cloud-based collaborative call usage rate, and the result return timing consistency rate. The value of the risk adjustment term is directly given by the risk level coefficient preset for each mode in the operating mode configuration table. The rule-adaptive module is used to acquire and normalize the statistical results of perceived stability, generate a perceived stability direction vector, and based on the perceived stability direction vector, the deviation between the decision quality summary and the preset reference operating quality level, and the set of weight parameters for the perceived availability state quantity in the current period, combined with regularization constraints to prevent the weights from deviating excessively from the initial configuration, generate adjusted weight parameters, adjust the weight parameters of the weighted aggregation in the video perception module, and apply the updated weight parameters to the construction of the perceived availability state in the next monitoring period, forming a continuous adaptive closed loop.

Images