Campus security risk identification method and system based on multi-algorithm nested cooperation, and medium

Abstract

The invention discloses a campus security risk identification method and system based on multi-algorithm nested cooperation, and a medium, and relates to the technical field of campus risk identification. Constructing a dynamic rule graph library, matching all dynamic rule graphs corresponding to the initial event signal from the dynamic rule graph library, and creating a dynamic rule graph instance; activating each node according to a topological sequence and an activation condition, and calculating the comprehensive reliability and the likelihood of the potential risk hypothesis at the sink node based on the DS evidence theory to perform risk early warning; deep association and comprehensive research and judgment are carried out on multi-dimensional information, and the problems of information isolated island and low-order misjudgment are solved; according to the method, the time background can be sensed by converting the school rhythm into the time weight factor and embedding the time weight factor into the DS evidence fusion process, the contribution degree of different types of evidences and the sensitivity of risk decision are automatically adjusted, false alarms caused by neglecting the context are reduced, and early warning can be sent out earlier and more accurately in the high-risk time period.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of campus risk identification, and in particular to a campus safety risk identification method and system based on multi-algorithm nested cooperation and a medium. BACKGROUND

[0002] In the field of campus safety protection, using artificial intelligence (AI) technology, especially computer vision algorithms, to automatically analyze monitoring videos has become an important means to improve security efficiency. Existing technical solutions generally use detection models based on a single or a few specific algorithms to identify and alarm specific safety events.

[0003] The current typical technical implementation is to deploy video analysis-capable cameras in key areas of the campus or deploy analysis servers in the back end, running independent algorithm models such as regional intrusion detection, crowd detection, running and falling detection, and fire recognition. When a target or behavior that meets the algorithm's preset features appears in the monitoring picture, such as someone entering a warning area, the system will trigger an alarm event directly corresponding to the algorithm. For example, a system may run A algorithm (detect regional intrusion) and B algorithm (detect people running) simultaneously, and the two work in parallel, each outputting corresponding alarm information.

[0004] However, this single-event-triggered technical architecture has significant limitations and cannot meet the needs of safety risk prevention and control in the complex and dynamic environment of the campus. Its defects mainly manifest in the following three aspects:

[0005] Algorithm island and information fragmentation problem: various AI algorithms run as independent perception islands, only outputting their own independent perception results, and there is a lack of association, fusion, and collaborative analysis between these output results; the system cannot organically integrate multi-dimensional information such as identity, high-risk location, specific behavior, time sequence, and psychological state background. For example, the system may detect someone entering the roof and recognize the person, but cannot automatically associate the two and further combine the person's personal profile to determine that this is a high-risk event.

[0006] Risk misjudgment and risk omission problem: the judgment of a single algorithm relies on limited visual features, and in the complex campus environment of light changes, obstructions, and similar behavior interference, it is easy to produce a high false alarm rate, such as misjudging playful fighting as fighting and fighting. More importantly, many major safety risks are not directly manifested by a single abnormal behavior, but may be manifested by a series of seemingly ordinary but highly correlated behaviors and states in a specific context. The existing technology lacks the ability to analyze the correlation of comprehensive behavior patterns in time, space, and attributes, resulting in weak recognition ability for such high-concealment and complex-cause risks.

[0007] Response delay and lack of decision support: the existing system generates alarm information with coarse granularity and poor operability. A large number of isolated alarms received by security personnel need to rely on personal experience for secondary screening, correlation analysis and priority judgment, which is time-consuming and labor-intensive and prone to errors. In emergency situations, there is a long time delay from receiving simple alarms to manually judging risks and starting the correct response, which may miss the best intervention opportunity.

[0008] Therefore, there is an urgent need for a new technical solution in the specific field of campus safety to break the algorithm island and realize intelligent collaboration and deep research and judgment of multi-dimensional perception information, so as to truly realize the avoidance of campus safety risks. SUMMARY

[0009] In order to solve the problem that the traditional campus security system adopts a single algorithm (such as only face recognition or regional intrusion detection), cannot cope with complex and composite security events, the independent running of the algorithm leads to high false alarm rate and early warning delay, and cannot realize cross-scene risk correlation analysis; the purpose of the present application is to provide a campus security risk identification method, system and medium based on multi-algorithm nested collaboration, by constructing a dynamic rule graph, arranging multi-source perception algorithm data and business data into an executable collaborative reasoning network according to logical and time sequence relationship; it can automatically associate and comprehensively judge multi-dimensional information; automatically link and identify multiple isolated evidence as composite evidence, solving the problems of information island and low-level misjudgment.

[0010] The above technical purpose of the present application is realized by the following technical scheme:

[0011] The present application provides a campus security risk identification method based on multi-algorithm nested collaboration, which comprises:

[0012] Constructing a dynamic rule graph library: defining a directed acyclic graph containing data nodes, algorithm nodes and logic nodes, the data nodes are used to represent input data or evidence, the algorithm nodes are used to associate preset AI algorithms; the logic nodes are used to perform logical operations or time sequence operations; each node is associated with one or more activation conditions, and the nodes are connected through directed edges to represent the dependency relationship, the dependency relationship includes data flow relationship or control flow relationship;

[0013] Event triggering and graph creation: receiving an initial event signal from a perception device, matching all dynamic rule graphs corresponding to the initial event signal from the dynamic rule graph library, and creating a dynamic rule graph instance for each dynamic rule graph;

[0014] Forward reasoning and evidence generation: in the dynamic rule graph instance, each node is activated in topological order and according to activation conditions, wherein, when the current node is an algorithm node, an associated preset AI algorithm is called to process input data of the current node to generate evidence; when the current node is a logic node, a logic operation or a time sequence operation is performed on the evidence input into the current node to generate composite evidence;

[0015] Evidence fusion: when reasoning reaches a convergence node representing a potential risk hypothesis, all target evidence pointing to the potential risk hypothesis is collected, and the basic probability assignment of each target evidence is fused based on DS evidence theory and a calendar rhythm to calculate the comprehensive belief and plausibility of the potential risk hypothesis; the calendar rhythm affects the contribution degree of the fusion process of each target evidence;

[0016] Risk warning: the comprehensive belief and plausibility of each potential risk hypothesis are compared with a preset threshold to generate a risk warning signal of a corresponding level.

[0017] Further optimization scheme: the logic node includes at least one of an AND node for performing a logical AND operation, an OR node for performing a logical OR operation, and a time sequence logic node for judging whether input evidence occurs within a specific time window.

[0018] Further optimization scheme: when constructing the dynamic rule graph library, for the convergence node representing the potential risk hypothesis, context information related to the calendar rhythm is configured as an activation condition; the context information includes a day rhythm period, a week rhythm period and a semester rhythm period;

[0019] The day rhythm period is represented by a plurality of day periods divided according to the daily routine of a day, and each day period assigns different weights to different potential risk hypothesis types;

[0020] The week rhythm period is represented by a week period of a day of the week; each week period assigns different weights to different potential risk hypothesis types;

[0021] The semester rhythm period is represented by a plurality of semester stages divided according to the learning stages of a semester, and each semester stage assigns different weights to different potential risk hypothesis types.

[0022] Further optimization scheme: the dynamic rule graph instance includes an interface node for accessing external business data, which is used to obtain psychological records, attendance records or leave information related to the target person in real time during the forward reasoning process, and input them as evidence into the subsequent node.

[0023] Further optimization scheme: the calculation method of the comprehensive belief and plausibility of the potential risk hypothesis includes:

[0024] Define the recognition framework Θ = {H k , ¬H k} of the DS evidence theory, wherein H k represents that the potential risk hypothesis r k is true; ¬H k represents that the potential risk hypothesis r k is not true; and basic probability assignments are given to each target evidence, wherein the basic probability assignment of the target evidence i to the potential risk hypothesis k is m ik .

[0025] Calculate a time weight factor based on the type of the potential risk hypothesis and the context information, and normalize the time weight factor;

[0026] Calculate the ratio of the historical accuracy rate of each target evidence to the potential risk hypothesis under the same time condition to the overall average accuracy rate as a time sensitivity coefficient;

[0027] Combine the time sensitivity coefficient, modify the basic probability assignment of the target evidence by taking the time weight factor as a reliability discount of the target evidence, and then fuse the comprehensive belief and plausibility of the potential risk hypothesis; or combine the time sensitivity coefficient, convert the time weight factor into a combination weight of the target evidence, and then fuse the comprehensive belief and plausibility of the potential risk hypothesis.

[0028] Further optimization scheme is that the calculation method of the time weight factor comprises:

[0029] The time weight factor T k of the potential risk hypothesis k is the product of the weight w dk allocated to the potential risk hypothesis k by the daily time period, the weight w wk allocated to the potential risk hypothesis k by the weekly time period, and the weight w sk allocated to the potential risk hypothesis k by the semester stage.

[0030] Further optimization scheme is that the modification of the basic probability assignment of the target evidence by taking the time weight factor as the reliability discount of the target evidence comprises the method:

[0031] Calculate the reliability discount factor B ik of the target evidence i to the potential risk hypothesis k:

[0032] B ik =min(1, T k1 ×α ik / T1);

[0033] Wherein, T k1 represents the normalized time weight factor; and α ikrepresents the time sensitivity coefficient of the target evidence i to the potential risk hypothesis k; T1 represents the normalized reference value;

[0034] based on the reliability discount factor B ik corrects the basic probability assignment of the target evidence i to the potential risk hypothesis k as m ik obtained:

[0035]

[0036] wherein, represents the corrected basic probability assignment of the target evidence i to the potential risk hypothesis k; A represents any subset of the identification framework Θ; represents that A is a true subset of Θ; represents the basic probability assignment of the target evidence i to the potential risk hypothesis k as A; represents the basic probability assignment of the target evidence i to the potential risk hypothesis k as uncertain.

[0037] Further optimization scheme is that the time weight factor is converted into the combination weight of the target evidence, including the method:

[0038] In the weighted combination of each target evidence to the potential risk hypothesis k, the combination weight of the target evidence i to the potential risk hypothesis k is W ik :

[0039] W ik = ;

[0040] wherein, exp() represents the exponential function; represents the temperature parameter, ; represents the time sensitivity coefficient of the target evidence j to the potential risk hypothesis k; N represents the total number of target evidences pointing to the potential risk hypothesis k.

[0041] The scheme also provides a campus safety risk identification system based on multi-algorithm nested cooperation, which is used to realize the above-mentioned campus safety risk identification method based on multi-algorithm nested cooperation, and the system comprises:

[0042] A dynamic rule graph construction module is used to define a directed acyclic graph comprising data nodes, algorithm nodes and logic nodes, the data nodes are used to represent input data or evidence, the algorithm nodes are used to associate preset AI algorithms, and the logic nodes are used to perform logic operations or time sequence operations; each node is associated with one or more activation conditions, and the nodes are connected through directed edges to represent a dependency relationship, and the dependency relationship comprises a data flow relationship or a control flow relationship;

[0043] ​An event trigger and legend creation module is configured to receive an initial event signal from a sensing device, match all dynamic rule graphs corresponding to the initial event signal from a dynamic rule graph library, and create a dynamic rule graph instance for each dynamic rule graph;

[0044] A forward reasoning and evidence generation module is configured to activate nodes in a dynamic rule graph instance according to a topological order and an activation condition, wherein when a current node is an algorithm node, a preset AI algorithm associated with the current node is invoked to process input data of the current node to generate evidence; and when the current node is a logic node, a logic operation or a timing operation is performed on the evidence input into the current node to generate composite evidence.

[0045] An evidence fusion module is configured to collect all target evidence pointing to a potential risk hypothesis when reasoning reaches a convergence node representing the potential risk hypothesis, and fuse basic probability assignments of the target evidence based on a DS evidence theory and a calendar rhythm to calculate a comprehensive confidence and a plausibility of the potential risk hypothesis; the calendar rhythm affects a contribution degree of the fusion process of the target evidence.

[0046] A risk warning module is configured to compare the comprehensive confidence and the plausibility of each potential risk hypothesis with a preset threshold to generate a risk warning signal of a corresponding level.

[0047] The scheme also provides a computer readable medium having a computer program stored thereon, and the computer program is executable by a processor to implement the campus security risk identification method based on multi-algorithm nested cooperation.

[0048] Compared with the prior art, the present application has the following beneficial effects:

[0049] 1. The present application provides a campus security risk identification method, system and medium based on multi-algorithm nested cooperation, which constructs a dynamic rule graph, arranges multi-source sensing algorithm data and business data into an executable cooperative reasoning network according to a logic and timing relationship, and can automatically associate and comprehensively analyze multi-dimensional information, automatically connect and identify multiple isolated evidence as composite evidence, and solve the problems of information silos and low-level misjudgment.

[0050] 2. The present application provides a campus security risk identification method, system and medium based on multi-algorithm nested cooperation, which converts a calendar rhythm (such as an examination period, a semester start and end, and a day and night period) into a calculable time weight factor and embeds it into a DS evidence fusion process, can dynamically perceive a time background of a current time, and automatically adjusts a contribution degree of different types of evidence and a sensitivity of a risk decision, greatly reduces false positives caused by ignoring context, and ensures that a warning can be issued earlier and more accurately in a high-risk period.

[0051] 3.The application provides a campus security risk identification method, system and medium based on multi-algorithm nested cooperation, which adopts a dynamic rule graph to provide an intuitive risk identification logic expression; a security administrator can quickly adapt to new risk scenarios or optimize existing strategies by adding, deleting or modifying nodes and logical relationships (such as adjusting AND conditions and modifying time windows) in the dynamic rule graph, without the need for reprogramming, thereby greatly enhancing the flexibility and operability of the system.

[0052] 4.The application provides a campus security risk identification method, system and medium based on multi-algorithm nested cooperation, which is based on rules and traceable evidence fusion calculation (credibility and plausibility) throughout the reasoning process, so that the basis for each early warning decision is clear and visible, solving the trust problem caused by the black box decision of traditional AI models, and continuously optimizing the rule parameters, time weights and fusion thresholds through the record of disposal feedback, thereby realizing the iterative evolution of performance. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments, and it should be understood that the following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor. In the drawings:

[0054] Figure 1 It is a flowchart of the campus security risk identification method based on multi-algorithm nested cooperation.

[0055] Figure 2 It is a schematic diagram of a dynamic rule graph library instance.

[0056] Figure 3 It is a structural schematic diagram of the campus security risk identification system based on multi-algorithm nested cooperation. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of the application more clear and obvious, the following will further describe the application in combination with the embodiments and drawings, and the exemplary embodiments of the application and their descriptions are only used to explain the application, and should not be regarded as a limitation on the application.

[0058] The traditional campus security system adopts a single algorithm (such as only face recognition or regional intrusion detection), which cannot cope with complex and composite security events, the independent running of the algorithm leads to a high false alarm rate and delayed early warning, and cannot realize cross-scene risk correlation analysis; in view of this, the present application provides the following embodiments to solve the above technical problems:

[0059] Embodiment 1

[0060] The embodiment provides a campus safety risk identification method based on multi-algorithm nested cooperation, as shown in the following formula: Figure 1 The method comprises the following steps:

[0061] Step 1, constructing a dynamic rule graph library: defining a directed acyclic graph comprising data nodes, algorithm nodes and logic nodes, wherein the data nodes are used for representing input data or evidence, the algorithm nodes are used for associating preset AI algorithms, and the logic nodes are used for performing logical operation or time sequence operation; each node is associated with one or more activation conditions, and the nodes are connected through directed edges for representing a dependency relationship, wherein the dependency relationship comprises a data flow relationship or a control flow relationship.

[0062] The logic nodes comprise at least one of the following: an AND node used for performing logical and operation, an OR node used for performing logical or operation, and a time sequence logic node used for judging whether input evidence occurs within a specific time window.

[0063] Specifically, the directed acyclic graph is represented as G=(V,E,Λ,Γ), wherein V represents a node set, which is divided into three categories of data nodes, algorithm nodes and logic nodes; the data nodes represent original data input or algorithm output, that is, evidence nodes; the input directed edges of the algorithm nodes come from the data nodes or other algorithm nodes, and the output directed edges point to the data nodes; the input directed edges of the logic nodes come from the data nodes or the logic nodes, and the output directed edges point to the data nodes or trigger actions (representing a convergence node of potential risk hypotheses); the directed edge set is represented by E; Λ represents a set of activation condition functions on the nodes; Γ represents a set of action functions on the nodes (mainly the convergence node of potential risk hypotheses, and the convergence node is mainly a logic node), and an action (such as generating a warning, calling a new rule graph, etc.) to be performed when the current node output is true. The set of potential risk hypotheses is R={r1,r2,…r i , …,r K}; wherein: the potential risk hypothesis r1 is personal injury / violence conflict (including bullying, fighting, etc.); the potential risk hypothesis r2 is psychological crisis / self-injury risk (including depression, anxiety, self-injury tendency); the potential risk hypothesis r3 is accidental injury (including falling, sports injury, experimental accident); the potential risk hypothesis r4 is property loss / damage; the potential risk hypothesis r i is a network security event (including cyberbullying, privacy leakage); and the potential risk hypothesis r K is a public health event (including infectious disease, food poisoning), etc.

[0064] When the dynamic rule graph library is constructed, for the convergence node representing the potential risk hypothesis, context information related to a school calendar rhythm is configured as an activation condition; the context information comprises a day rhythm period, a week rhythm period and a semester rhythm period.

[0065] The circadian rhythm period is represented by multiple daily periods divided according to the daily learning routine, with each daily period assigning different weights to different potential risk assumption types;

[0066] This embodiment divides a day into: a quiet nighttime period (00:00-06:00), during which students should be in their dormitories and the campus should be quiet; a morning wake-up period (06:00-07:30), during which students wake up, do morning exercises, and have breakfast; a morning teaching period (07:30-12:00), during which students attend classes, conduct experiments, and participate in group teaching; a lunch break period (12:00-14:00), during which students have lunch, rest, and enjoy free time; and an afternoon teaching period (14:00-17:30), during which students attend classes, participate in club activities, etc. Sports; Evening activity days, from 17:30 to 19:00, include student dinner, free time, and social activities; Evening self-study days, from 19:00 to 22:00, include student self-study and library study; Nighttime bedtime days, from 22:00 to 24:00, include student preparation for bed and dormitory management; Based on educational psychology and safety statistics, different weights are assigned to different potential risk assumptions for each time period. For example, the psychological crisis risk r2 has the highest weight on quiet nighttime days because of the strongest sense of loneliness; the accident risk r3 is higher on early morning wake-up days and afternoon teaching days because of frequent activities; the property theft risk r4 has a high weight on quiet nighttime days because of weaker supervision.

[0067] The weekly rhythm period is represented by the week period on which day of the week; each week period is assigned a different weight to different potential risk assumption types; for example, the psychological crisis risk r2 is set highest on Sunday because of pre-school anxiety; the personal injury risk r1 is higher on Friday and Saturday because of more social activities.

[0068] The semester rhythm period is represented by multiple semester stages that divide a semester according to the learning stage, with each semester stage assigning different weights to different potential risk assumption types.

[0069] For example: the school adjustment period, the regular teaching period, exam week, and the period before and after holidays; this embodiment divides a semester into: the initial adjustment period, in the first 1-2 weeks of the semester, characterized by students adapting to the new environment and establishing rules; the regular teaching period, in the 3rd-10th weeks of the semester, characterized by students maintaining a stable teaching pace; the mid-term exam period, in the 11th-12th weeks of the semester, characterized by students experiencing exam pressure and anxiety about grades; the later teaching period, in the 13th-16th weeks of the semester, characterized by students facing heavy workloads and accumulating fatigue; the final exam period, in the 17th-18th weeks of the semester, characterized by students experiencing high pressure and peak anxiety; and the final closing period, in the 19th-20th weeks of the semester, characterized by students preparing for the release of grades and leaving school; different weights are assigned to different potential risk assumptions for each semester stage, among which, the psychological crisis risk r2 The final exam period has the highest weighting and the greatest pressure; the risk of accidents (r3) is higher at the beginning of the semester due to unfamiliarity with the environment; the risk of theft (r4) is higher at the beginning and end of the semester because valuables are often carried during this period.

[0070] Step 2, Event Triggering and Legend Creation: Receive the initial event signal from the sensing device, match all dynamic rule graphs corresponding to the initial event signal from the dynamic rule graph library, and create a dynamic rule graph instance for each dynamic rule graph;

[0071] The dynamic rule graph instance includes interface nodes for accessing external business data. These interface nodes are used to obtain psychological profiles, attendance records, or leave information related to the target personnel in real time during the forward reasoning process and input them as evidence into subsequent nodes.

[0072] Specifically, the initial event signal is equivalent to a trigger seed evidence. The system searches for all dynamic rule graphs in the dynamic rule graph library that use the trigger seed evidence as an entry point, and creates an instance for each matching dynamic rule graph.

[0073] Step 3, Forward Reasoning and Evidence Generation: In the dynamic rule graph instance, each node is activated according to the topological order and activation conditions. When the current node is an algorithm node, the associated preset AI algorithm is called to process the input data of the current node to generate evidence. When the current node is a logic node, logical operations or temporal operations are performed on the evidence input to the current node to generate composite evidence.

[0074] Specifically, for an algorithm node, data from all its input directed edges is collected, the corresponding algorithm is called to perform calculations, and the calculated evidence results are passed to the output directed edges. The evidence output by the algorithm node is structured evidence with confidence, represented as a triple X{evidence type, evidence value, confidence}. The confidence is the probability value or score of the AI ​​algorithm, and the confidence value is in the range [0, 1].

[0075] For logical nodes, the evidence of the input directed edges is aggregated according to their logical type. Taking the AND node as an example, its output is the logical AND operation result of all evidence.

[0076] For sequential logic nodes, it is necessary to check whether the timestamp of the evidence for the input directed edge is within the preset time window Δt.

[0077] Step 4, Evidence Fusion: When reasoning reaches the convergence node representing the potential risk hypothesis, all target evidence pointing to the potential risk hypothesis is collected, and the basic probability allocation of each target evidence is fused based on DS evidence theory and calendar rhythm to calculate the comprehensive reliability and similarity of the potential risk hypothesis; the calendar rhythm affects the contribution of each target evidence fusion process.

[0078] The methods for calculating the overall reliability and likelihood of the potential risk assumptions include:

[0079] S41, Define the identification framework of the DS evidence theory: Θ={H k ,¬H k}, where H k Indicates the potential risk assumption r k Established; ¬H k Indicates the potential risk assumption r k This is not valid; assign a basic probability assignment to each piece of target evidence, where the basic probability assignment for target evidence i supporting potential risk hypothesis k is m. ik ;

[0080] Specifically, for each piece of target evidence, based on its evidence type and value, a predefined basic probability allocation function is used to generate a basic probability allocation for the potential risk hypothesis. For example, the basic probability allocation for target evidence i (emotional sadness) to potential risk hypothesis k might be: the probability allocation m for potential risk hypothesis k to be true. ik ({H}) is: m ik ({H})=0.4; Probability assignment m of potential risk assumption k not being true ik ({¬H}) is: m ik ({¬H})=0.1; Probability assignment m of the potential risk assumption k being uncertain. ik ({¬H}) is: m ik (Θ)=0.5.

[0081] S42, calculate the time weight factor based on the type and context information of the potential risk assumption, and normalize the time weight factor.

[0082] Since the time weight factor may have a large value, it needs to be mapped to a reasonable range of evidence weights. Therefore, this scheme normalizes the time weight factor.

[0083] The calculation method for the time weighting factor includes:

[0084] The time weighting factor T of the potential risk assumption k k For: the weight w assigned to the potential risk assumption k for each day period dk The weights w assigned to the potential risk assumption k over a given period of time. wk The weights w assigned to the potential risk assumption k during the semester phase sk The product of.

[0085] Taking the psychological crisis warning on Sunday evening during exam week as an example, the current time is 23:00, which is the nighttime bedtime period, with a weight of w. dk The weekly time period weight w is 1.9, corresponding to Sunday. wk The weight w corresponding to the final exam stage is 1.5. sk The value is 1.8, and the calculated time weighting factor is 1.9 × 1.5 × 1.8 = 5.13.

[0086] S43, calculate the ratio of the historical accuracy of each target's evidence to the overall average accuracy under the same time conditions for the potential risk assumption as the time sensitivity coefficient.

[0087] S44, Combining the time sensitivity coefficient, the time weight factor is used as a reliability discount of the target evidence to correct the basic probability allocation of the target evidence, and then the comprehensive reliability and likelihood of the potential risk hypothesis are fused; or Combining the time sensitivity coefficient, the time weight factor is transformed into a combined weight of the target evidence, and then the comprehensive reliability and likelihood of the potential risk hypothesis are fused.

[0088] In step S44, the basic probability allocation of the target evidence is corrected by using the time weighting factor as a reliability discount for the target evidence, including the following methods:

[0089] S441, Calculate the reliability discount factor B for target evidence i supporting potential risk assumption k. ik :

[0090] B ik =min(1,T) k1 ×α ik / T1);

[0091] Among them, T k1 α represents the normalized time weighting factor. ik T1 represents the time sensitivity coefficient of target evidence i to potential risk hypothesis k; T1 represents the normalized baseline value.

[0092] S442, based on reliability discount factor B ik The base probability assignment for revising target evidence i to support potential risk hypothesis k is m. ik get:

[0093] ;

[0094] in, Let A represent the basic probability assignment that the modified target evidence i supports the potential risk hypothesis k; A represents any subset of the identification framework Θ (the subsets of the identification framework Θ include: the empty set ∅, {H}, {¬H}, {H, ¬H}). Indicate that A is a proper subset of Θ; This indicates that the target evidence i supports the basic probability assignment of the potential risk hypothesis k as A; This indicates that the target evidence i supports the potential risk hypothesis k, which is an uncertain basic probability assignment.

[0095] Then, using Dempster's combination rule, the basic probability assignments are fused pairwise, and the final fusion result m is used. f Calculate the confidence level Bel(H) for the potential risk hypothesis k to be true: Bel(H) = m f ({H}) and similarity Pl(H) = 1 - m f ({H});

[0096] In step S44, the time weighting factor is converted into a combined weight of the target evidence, including the following methods:

[0097] When weighting the support for potential risk hypothesis k from each piece of target evidence, the weight of the combination of support for potential risk hypothesis k from target evidence i is W. ik :

[0098] W ik = ;

[0099] Where exp() represents the exponential function; Indicates temperature parameter, ; denoted as the time sensitivity coefficient of target evidence j to potential risk hypothesis k; N represents the total number of target evidences pointing to potential risk hypothesis k.

[0100] Based on the above formula, the weight vector W(t) of the N target pieces of evidence is determined as [W 1k W 2k , …,W ik …, W Nk ] T ;

[0101] Based on the weight vector W(t) and the basic probability assignments supporting the potential risk hypothesis k of each target evidence, the weighted average basic probability assignment of all target evidence is calculated. This basic probability assignment is then self-combined N-1 times to obtain the final fusion result m. f According to the final fusion result m f Calculate the confidence level Bel(H) for the potential risk hypothesis k to be true: Bel(H) = m f ({H}) and similarity Pl(H) = 1 - m f ({H}).

[0102] Step 5, Risk Warning: Compare the overall reliability and likelihood of each potential risk hypothesis with a preset threshold to generate a risk warning signal of the corresponding level.

[0103] This embodiment transforms school calendar rhythms (such as exam cycles, semester start and end times, and day / night cycles) into calculable time weighting factors and embeds them into the DS evidence fusion process. This allows for dynamic perception of the current time context and automatic adjustment of the contribution of different types of evidence and the sensitivity of risk decisions. For example, during the late night of final exam week, the fusion weight of evidence related to psychological crisis risks (such as low mood or being alone) is significantly enhanced, and the warning threshold is correspondingly lowered, making the system more sensitive during this high-risk period. Conversely, during the afternoon of a regular teaching day, the weight of the same evidence is reduced, making the system more cautious. This greatly reduces false alarms caused by ignoring context and ensures that warnings are issued earlier and more accurately during high-risk periods.

[0104] This embodiment uses the prevention of falls from heights as a potential risk assumption, and its seed evidence e0 is "rooftop area intrusion," such as... Figure 2 As shown, the constructed dynamic rule graph example is as follows:

[0105] Entry node Vd1: Data node, type "Rooftop Area Intrusion", seed evidence e0 provided by the basic motion detection algorithm;

[0106] Algorithm node Va1: Associated with the face recognition algorithm; inputs an image region from seed evidence e0, outputs the person ID and emotional state as evidence e. id and evidence e emo ;

[0107] Data node Vd2: Combines with interface node Vj to query the psychological profile of the person's ID from the business database, and outputs the psychological risk level as evidence e. psy ;

[0108] Algorithm node Va2: Associates a fine-grained behavior recognition algorithm; continuously receives video streams, detects behaviors such as leaning against railings and leaning forward, and outputs behavior sequences as evidence. beh(t) ;

[0109] Logical node Vl1: A time-series logical node that checks whether the target remains within the region after being triggered by seed evidence e0, and outputs the check result as evidence e1.

[0110] Logical node Vl2: AND node; input is evidence e psy (High psychological risk), Evidence e emo (Negative emotions), Evidence e beh(t) (High-risk behavior detected) and evidence e1 (continued presence); activated when all inputs are true;

[0111] Convergence node Vs: Represents the assumption of isolation of fall risk; receives the output of logical node Vl2 and initiates evidence fusion.

[0112] Example 2

[0113] This embodiment provides a campus security risk identification system based on multi-algorithm nesting and collaboration, used to implement the campus security risk identification method based on multi-algorithm nesting and collaboration described in Embodiment 1, such as... Figure 3 As shown, the system includes:

[0114] The dynamic rule graph library construction module is used to define a directed acyclic graph containing data nodes, algorithm nodes, and logic nodes. The data nodes are used to represent input data or evidence, the algorithm nodes are used to associate preset AI algorithms, and the logic nodes are used to perform logical operations or time-series operations. Each node is associated with one or more activation conditions, and the nodes are connected by directed edges to represent dependencies, including data flow relationships or control flow relationships.

[0115] The event triggering and legend creation module is used to receive the initial event signal from the sensing device, match all dynamic rule graphs corresponding to the initial event signal from the dynamic rule graph library, and create a dynamic rule graph instance for each dynamic rule graph.

[0116] The forward reasoning and evidence generation module is used to activate each node in the dynamic rule graph instance according to the topological order and activation conditions. When the current node is an algorithm node, it calls the associated preset AI algorithm to process the input data of the current node to generate evidence; when the current node is a logic node, it performs logical operations or temporal operations on the evidence input to the current node to generate composite evidence.

[0117] The evidence fusion module is used to collect all target evidence pointing to the potential risk hypothesis when reasoning reaches the convergence node representing the potential risk hypothesis, and to fuse the basic probability assignments of each target evidence based on DS evidence theory and calendar rhythm to calculate the comprehensive reliability and similarity of the potential risk hypothesis; the calendar rhythm affects the contribution of each target evidence fusion process.

[0118] The risk warning module is used to compare the overall reliability and likelihood of each potential risk hypothesis with a preset threshold and generate a risk warning signal of the corresponding level.

[0119] Example 3

[0120] This embodiment provides a computer-readable medium storing a computer program. The computer program, when executed by a processor, can implement the campus security risk identification method based on multi-algorithm nesting and collaboration as described in Embodiment 1; specifically, it performs the following steps:

[0121] Step 1, Construct a dynamic rule graph library: Define a directed acyclic graph containing data nodes, algorithm nodes, and logic nodes. The data nodes are used to represent input data or evidence, the algorithm nodes are used to associate preset AI algorithms, and the logic nodes are used to perform logical operations or time-series operations. Each node is associated with one or more activation conditions, and the nodes are connected by directed edges to represent dependencies, including data flow relationships or control flow relationships.

[0122] Step 2, Event Triggering and Legend Creation: Receive the initial event signal from the sensing device, match all dynamic rule graphs corresponding to the initial event signal from the dynamic rule graph library, and create a dynamic rule graph instance for each dynamic rule graph;

[0123] Step 3, Forward Reasoning and Evidence Generation: In the dynamic rule graph instance, each node is activated according to the topological order and activation conditions. When the current node is an algorithm node, the associated preset AI algorithm is called to process the input data of the current node to generate evidence. When the current node is a logic node, logical operations or temporal operations are performed on the evidence input to the current node to generate composite evidence.

[0124] Step 4, Evidence Fusion: When reasoning reaches the convergence node representing the potential risk hypothesis, all target evidence pointing to the potential risk hypothesis is collected, and the basic probability allocation of each target evidence is fused based on DS evidence theory and calendar rhythm to calculate the comprehensive reliability and similarity of the potential risk hypothesis; the calendar rhythm affects the contribution of each target evidence fusion process.

[0125] Step 5, Risk Warning: Compare the overall reliability and likelihood of each potential risk hypothesis with a preset threshold to generate a risk warning signal of the corresponding level.

[0126] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A campus security risk identification method based on multi-algorithm nesting and collaboration, characterized in that, The method includes: Construct a dynamic rule graph library: Define a directed acyclic graph containing data nodes, algorithm nodes, and logic nodes. The data nodes are used to represent input data or evidence, the algorithm nodes are used to associate preset AI algorithms, and the logic nodes are used to perform logical operations or time-series operations. Each node is associated with one or more activation conditions, and the nodes are connected by directed edges to represent dependencies, which include data flow relationships or control flow relationships. Event Triggering and Legend Creation: Receive the initial event signal from the sensing device, match all dynamic rule graphs corresponding to the initial event signal from the dynamic rule graph library, and create a dynamic rule graph instance for each dynamic rule graph; Forward reasoning and evidence generation: In the dynamic rule graph instance, each node is activated according to the topological order and activation conditions. When the current node is an algorithm node, the associated preset AI algorithm is called to process the input data of the current node to generate evidence. When the current node is a logic node, logical operations or temporal operations are performed on the evidence input to the current node to generate composite evidence. Evidence fusion: When reasoning reaches the convergence node representing the potential risk hypothesis, all target evidence pointing to the potential risk hypothesis is collected, and the basic probability assignments of each target evidence are fused based on DS evidence theory and calendar rhythm to calculate the comprehensive reliability and similarity of the potential risk hypothesis; the calendar rhythm affects the contribution of each target evidence fusion process. Risk warning: The overall reliability and likelihood of each potential risk hypothesis are compared with a preset threshold to generate a risk warning signal of the corresponding level.

2. The campus security risk identification method based on multi-algorithm nesting and collaboration as described in claim 1, characterized in that, The logical node includes at least one of the following: an AND node for performing logical AND operations, an OR node for performing logical OR operations, and a timing logic node for determining whether input evidence occurred within a specific time window.

3. The campus security risk identification method based on multi-algorithm nesting and collaboration according to claim 1, characterized in that, When constructing a dynamic rule graph library, for the aggregation node representing potential risk assumptions, contextual information related to the school calendar rhythm is configured as an activation condition. The context information includes daily circadian rhythm periods, weekly circadian rhythm periods, and semester circadian rhythm periods; The circadian rhythm period is represented by multiple daily periods divided according to the daily learning routine, with each daily period assigning different weights to different potential risk assumption types; The weekly rhythm period is represented by the week period on which day of the week; each week period is assigned a different weight to different potential risk assumption types; The semester rhythm period is represented by multiple semester stages that divide a semester according to the learning stage, with each semester stage assigning different weights to different potential risk assumption types.

4. The campus security risk identification method based on multi-algorithm nesting and collaboration according to claim 1, characterized in that, The dynamic rule graph instance includes interface nodes for accessing external business data. These interface nodes are used to obtain psychological profiles, attendance records, or leave information related to the target personnel in real time during the forward reasoning process and input them as evidence into subsequent nodes.

5. The campus security risk identification method based on multi-algorithm nesting and collaboration according to claim 3, characterized in that, The methods for calculating the overall reliability and likelihood of the potential risk assumptions include: Define the identification framework of the DS evidence theory: Θ={H k ,¬H k }, where H k Indicates the potential risk assumption r k Established; ¬H k Indicates the potential risk assumption r k This is not valid; assign a basic probability assignment to each piece of target evidence, where the basic probability assignment for target evidence i supporting potential risk hypothesis k is m. ik ; The time weighting factor is calculated based on the type of potential risk assumption and contextual information, and then the time weighting factor is normalized. The ratio of the historical accuracy of each target's evidence to the overall average accuracy under the same time conditions is calculated as the time sensitivity coefficient. By combining the time sensitivity coefficient, the time weight factor is used as a reliability discount to correct the basic probability allocation of the target evidence, and then the comprehensive reliability and likelihood of the potential risk hypothesis are fused together; or by combining the time sensitivity coefficient, the time weight factor is transformed into a combined weight of the target evidence, and then the comprehensive reliability and likelihood of the potential risk hypothesis are fused together.

6. The campus security risk identification method based on multi-algorithm nesting and collaboration according to claim 5, characterized in that, The calculation method for the time weighting factor includes: The time weighting factor T of the potential risk assumption k k For: the weight w assigned to the potential risk assumption k for each day period dk The weights w assigned to the potential risk assumption k over a given period of time. wk The weights w assigned to the potential risk assumption k during the semester phase sk The product of.

7. The campus security risk identification method based on multi-algorithm nesting and collaboration according to claim 5, characterized in that, The method for correcting the basic probability allocation of target evidence by using a time-weighted factor as a reliability discount for the target evidence includes: Calculate the reliability discount factor B for the evidence i supporting the potential risk hypothesis k. ik : B ik =min(1,T k1 ×α ik / T1); Among them, T k1 α represents the normalized time weighting factor. ik T1 represents the time sensitivity coefficient of target evidence i to potential risk hypothesis k; T1 represents the normalized baseline value. Based on reliability discount factor B ik The base probability assignment for revising target evidence i to support potential risk hypothesis k is m. ik get: ； in, Let A represent the basic probability assignment that the modified target evidence i supports the potential risk hypothesis k; A represents any subset of the identification framework Θ. Indicate that A is a proper subset of Θ; This indicates that the target evidence i supports the basic probability assignment of the potential risk hypothesis k as A; This indicates that the target evidence i supports the potential risk hypothesis k, which is an uncertain basic probability assignment.

8. The campus security risk identification method based on multi-algorithm nesting and collaboration according to claim 5, characterized in that, The method for converting the time weight factor into a combined weight of the target evidence includes: When weighting the support for potential risk hypothesis k from each piece of target evidence, the weight of the combination of support for potential risk hypothesis k from target evidence i is W. ik : W ik = ； Where exp() represents the exponential function; Indicates temperature parameter, ; denoted as the time sensitivity coefficient of target evidence j to potential risk hypothesis k; N represents the total number of target evidences pointing to potential risk hypothesis k.

9. A campus security risk identification system based on multi-algorithm nesting and collaboration, characterized in that: The system is used to implement the campus security risk identification method based on multi-algorithm nesting and collaboration as described in any one of claims 1-8, the system comprising: The dynamic rule graph library construction module is used to define a directed acyclic graph containing data nodes, algorithm nodes, and logic nodes. The data nodes are used to represent input data or evidence, the algorithm nodes are used to associate preset AI algorithms, and the logic nodes are used to perform logical operations or time-series operations. Each node is associated with one or more activation conditions, and the nodes are connected by directed edges to represent dependencies, including data flow relationships or control flow relationships. The event triggering and legend creation module is used to receive the initial event signal from the sensing device, match all dynamic rule graphs corresponding to the initial event signal from the dynamic rule graph library, and create a dynamic rule graph instance for each dynamic rule graph. The forward reasoning and evidence generation module is used to activate each node in the dynamic rule graph instance according to the topological order and activation conditions. When the current node is an algorithm node, it calls the associated preset AI algorithm to process the input data of the current node to generate evidence; when the current node is a logic node, it performs logical operations or temporal operations on the evidence input to the current node to generate composite evidence. The evidence fusion module is used to collect all target evidence pointing to the potential risk hypothesis when reasoning reaches the convergence node representing the potential risk hypothesis, and to fuse the basic probability assignments of each target evidence based on DS evidence theory and calendar rhythm to calculate the comprehensive reliability and similarity of the potential risk hypothesis; the calendar rhythm affects the contribution of each target evidence fusion process. The risk warning module is used to compare the overall reliability and likelihood of each potential risk hypothesis with a preset threshold and generate a risk warning signal of the corresponding level.

10. A computer-readable medium having a computer program stored thereon, characterized in that, The computer program, when executed by a processor, can implement the campus security risk identification method based on multi-algorithm nesting and collaboration as described in any one of claims 1-8.

Images