Campus physical deception monitoring and early warning system based on multi-source data fusion

Abstract

The invention relates to the technical field of intelligent monitoring and data processing, and discloses a campus physical spoofing monitoring and early warning system based on multi-source data fusion, and the system comprises a multi-source heterogeneous data collection module which obtains objective behavior data and subjective feedback signals; the security semantic space mapping module converts the behavior intensity and reliability willingness feature vectors into behavior intensity and reliability willingness feature vectors; the logic mutual exclusion potential energy verification module calculates a supervision logic divergence index according to the geometric position relation between the vectors; according to the invention, a logic mutual exclusion potential energy calculation rule is introduced, directivity difference between objective behaviors and subjective intentions is accurately quantified, and the technical problem that hidden risks are smoothly covered in a conventional linear model is effectively solved. And automatic locking and active intervention on the logic exception event are realized.

Description

Technical Field

[0001] This invention relates to a campus physical bullying monitoring and early warning system based on multi-source data fusion, belonging to the field of intelligent monitoring and data processing technology. Background Technology

[0002] In the current digital supervision technology system for campus safety, multimodal data fusion is a common technical path to achieve risk perception. Existing mainstream monitoring systems are usually equipped with two independent data acquisition subsystems: one is a physical behavior monitoring subsystem based on image sensors or wearable devices, used to collect objective physical quantities such as personnel trajectory and limb acceleration; the other is an information feedback subsystem based on mobile terminals or interactive interfaces, used to receive subjective text data such as questionnaire records and voice statements from the parties involved. After acquiring the above-mentioned multi-source heterogeneous data, the data processing center usually uses feature vectorization and linear weighted fusion algorithms to calculate a comprehensive risk index, which serves as the basis for triggering early warning signals or generating management and disposal suggestions. In conventional scenarios where the data sources are semantically consistent, this processing logic can improve the confidence of monitoring results by utilizing the positive superposition characteristics of multi-source data.

[0003] However, when dealing with specific risk events involving implicit coercion or complex social relationships, the aforementioned conventional technical architecture based on linear superposition has inherent logical flaws. In practical engineering applications, objective physical monitoring data and subjective feedback data often exhibit logical mutual exclusion due to opposite feature vector directions. Because existing data fusion models lack a dynamic verification mechanism for the logical consistency between data sources and adhere to the linear calculation rules of static weights, when high-risk objective data and low-risk subjective data are input simultaneously, the linear weighting algorithm will force numerical averaging. This mathematical processing inevitably leads to high-risk feature signals being overwhelmed by low-risk ones. Risk data dilution or smoothing causes the overall risk value output by the system to be lower than the actual level of urgency, resulting in distortion of key early warning signals. In addition, when generating disposal instructions, the existing auxiliary decision-making systems usually establish their authority boundaries based on preset static rule tables, which are decoupled from the quality status of the input data. When the input data has the above-mentioned logical mutual exclusion, the system cannot perceive the decline in data reliability and still allows the output of high-level disposal suggestions, or the disposal suggestions are missing due to the smoothing of risk values. This disconnect between the decision space and the data consistency status makes it impossible for the management system to adaptively shrink the discretionary boundaries or trigger mandatory locking logic when faced with complex logical conflicts.

[0004] Existing technologies suffer from the following shortcomings: 1. Multi-source data processing algorithms employ linear weighted logic, lacking the ability to quantify the degree of logical mutual exclusion between heterogeneous data, inevitably leading to numerical attenuation of high-risk signals when data semantic conflicts occur; 2. Risk assessment models lack mode switching functionality based on data logical consistency, failing to trigger nonlinear compensation or state locking when high logical divergence characteristics are detected; 3. The authority boundary settings of administrative discretion auxiliary systems are static, lacking a negative correlation mapping mechanism between the effective volume of the decision space and the data mutual exclusion potential energy, resulting in a lack of dynamic constraints on decision outputs in low-reliability data environments. Therefore, how to construct a data processing mechanism capable of real-time calculation of the logical mutual exclusion potential energy between heterogeneous data, and dynamically blocking linear weighted paths, triggering nonlinear risk locking, and shrinking discretionary boundaries based on this physical quantity, becomes the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A campus physical bullying monitoring and early warning system based on multi-source data fusion, comprising: The multi-source heterogeneous data acquisition module is configured to acquire in real time the objective behavior time-series data stream within the target monitoring area and simultaneously acquire the subjective interactive feedback signals of the monitored object; The security semantic space mapping module, connected to the multi-source heterogeneous data acquisition module, is configured to construct a unified-dimensional security management semantic space. Based on the preset action feature extraction algorithm, it transforms the objective behavior time-series data stream into behavior intensity feature vectors mapped in the security management semantic space. At the same time, based on the preset semantic tendency analysis algorithm, it transforms the subjective interaction feedback signals into confidence and willingness feature vectors mapped in the security management semantic space. The logical mutual exclusion potential energy verification module, connected to the security semantic space mapping module, is configured to execute a logical mutual exclusion potential energy calculation procedure based on the geometric positional relationship between the behavior intensity feature vector and the confidence intention feature vector in the security management semantic space, in order to generate a regulatory logic divergence index that characterizes the degree of logical contradiction between objective physical behavior and subjective feedback information; wherein, the magnitude of the regulatory logic divergence index not only responds to the magnitude of the behavior intensity feature vector, but is also independently controlled by the difference in the direction of the vector angle between the behavior intensity feature vector and the confidence intention feature vector; The graded early warning execution module, connected to the logical mutual exclusion potential energy verification module, is configured to compare the regulatory logic divergence index with the preset risk mutual exclusion threshold, and generate and output a physical intervention instruction indicating that the system is in a state of hidden risk when the regulatory logic divergence index is greater than the risk mutual exclusion threshold.

[0006] Preferably, the multi-source heterogeneous data acquisition module includes: a panoramic vision capture submodule, configured to acquire high-resolution on-site environmental image sequences and extract human skeletal key point coordinate data from them as the spatial feature source of the objective behavior time-series data stream; a non-contact vital sign monitoring submodule, configured to simultaneously acquire the heart rate variability parameters and skin conductivity parameters of the monitored object and use the parameters as physiological auxiliary dimension data of the objective behavior time-series data stream; and an interactive feedback acquisition submodule, configured to generate a control signal to start an inquiry program after detecting a motion event with acceleration exceeding a preset threshold, in order to obtain subjective interactive feedback signals.

[0007] Preferably, the security semantic space mapping module includes: a spatiotemporal action convolution calculation submodule, configured to perform spatiotemporal feature encoding on the objective behavior time-series data stream, extract physical feature parameters including limb contact frequency, center of gravity displacement amplitude and action intensity, and normalize the physical feature parameters to generate a behavior intensity feature vector; and a semantic tendency parsing submodule, configured to perform natural language processing on the subjective interaction feedback signal, extract semantic feature values ​​representing the denial, concealment or fear state of the monitored object, and map the semantic feature values ​​into directional components to generate a credibility intention feature vector.

[0008] Preferably, the logic mutual exclusion potential energy calculation procedure executed in the logic mutual exclusion potential energy verification module determines the regulatory logic divergence index based on the following mathematical relationship: ,in, As a divergence indicator of regulatory logic, For behavior intensity feature vectors, This is the feature vector of confidence and willingness. The intensity modulus characterizing objective physical behavior. The cosine value representing the directional consistency between objective physical behavior and subjective feedback information in the semantic space of safety management. and These are the preset weighted balance coefficients.

[0009] Preferably, the logically mutually exclusive potential energy verification module further includes: a dynamic weighted calibration submodule, configured to adjust the weighted balance coefficients in the relation in real time based on the confidence feature components contained in the confidence intention feature vector. The numerical value; specifically configured as follows: when the reliability intention feature vector indicates that the monitored object is in a high stress characteristic state, the weighted balance coefficient is automatically increased. The value is used to improve the sensitivity of the regulatory logic divergence indicator to logical contradictions.

[0010] Preferably, the system further includes: a digital rule base storage module for storing digital entries of management rules and mapping each entry to a rule constraint boundary vector in the security management semantic space; the logical mutual exclusion potential energy verification module is also configured to calculate the Euclidean distance between the behavior intensity feature vector and the rule constraint boundary vector, and to superimpose the Euclidean distance as a correction factor into the regulatory logic divergence index.

[0011] Preferably, the graded early warning execution module includes: a hidden risk locking submodule, configured to generate a hidden risk event marker when the regulatory logic divergence index is greater than the risk mutual exclusion threshold and the magnitude of the behavior intensity feature vector is less than the preset violence threshold; and an evidence chain closed-loop generation submodule, configured to automatically associate and generate the original image sequence, subjective interaction feedback signal and divergence calculation parameters corresponding to the time period of the regulatory logic divergence index in response to physical intervention instructions, and digitally sign them to generate a digital evidence package.

[0012] Preferably, the multi-source heterogeneous data acquisition module further includes an environmental interference filtering submodule, which is configured to monitor the ambient background noise decibel value and the flow density of irrelevant targets in real time, and perform adaptive noise reduction processing on the data stream based on the ambient background noise decibel value only before the objective behavior time-series data stream is input into the security semantic space mapping module, so as to eliminate interference features generated by non-target events.

[0013] Preferably, the system further includes: a historical behavior baseline construction module, used to statistically analyze the average behavior intensity feature value of the group within the target monitoring area based on historical records within a sliding time window, and construct a normalized behavior baseline model; and a logical mutual exclusion potential energy verification module configured to calculate the statistical deviation of the behavior intensity feature vector relative to the normalized behavior baseline model, and to use the statistical deviation exceeding a preset benchmark value as a prerequisite trigger condition for starting the logical mutual exclusion potential energy calculation procedure.

[0014] Preferably, the graded early warning execution module further includes: a multi-level response communication submodule, configured to selectively activate different physical response strategies based on the preset value range into which the regulatory logic divergence index falls; specifically configured as follows: when the index is in the first preset range, an activation signal is generated to drive the on-site audible and visual warning device; when the index is in the second preset range greater than the first preset range, an emergency distress signal containing location coordinate data and on-site summary data is simultaneously sent to the remote management center via an encrypted wireless channel.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In campus physical bullying monitoring, a nonlinear risk locking mechanism based on logically mutually exclusive potential energy coefficients effectively solves the signal dilution problem of traditional linear weighted models when heterogeneous data conflicts. This invention does not adopt the conventional multi-sensor data linear weighted fusion method, but instead uses a regulatory logic divergence calculation unit built inside the system to calculate the logical distance between the objective behavioral feature vector and the subjective confidence intention feature vector in real time, and generates a logically mutually exclusive potential energy coefficient representing the consistency state of the data source. When it exceeds a preset logical threshold, the system can automatically identify that it is currently in a high regulatory entropy state and trigger a nonlinear compensation operator to cut off the conventional weighted average calculation path and forcibly lock the high-risk level determined by the objective behavioral characteristics. This modal switching mechanism based on the degree of data logical mutual exclusion eliminates the weighted smoothing effect caused by the subjective concealment or denial of the parties involved on the objective high-risk physical data from the underlying logic of data processing, ensuring that the system can still maintain the accurate representation and response capability of objective risks under extreme data mutual exclusion conditions such as implicit coercion.

[0016] 2. Constructing a dynamic permission boundary projection technology based on discretionary legality manifold to achieve adaptive reverse coupling between the decision-making space and the reliability of data evidence. This invention establishes a negative correlation mapping relationship between the validity of decision instructions and the mutual exclusion potential coefficient of data by constructing a digital discretionary legality manifold in a three-dimensional management semantic space. The system dynamically adjusts the effective geometric volume and boundary range of the legality manifold based on the real-time calculated mutual exclusion potential coefficient. When the logical contradiction of the monitored data increases, the system automatically shrinks the manifold boundary and compresses the legal projection area of ​​the input instruction. This achieves dynamic gating and constraint of decision instructions at the technical level, transforming the abstract principle of cautious punishment for doubtful cases into a calculable and verifiable geometric boundary constraint logic. This makes the system's acceptance of user input instructions no longer a static preset permission, but depends on the completeness of the current evidence chain loop, thereby avoiding system mishandling due to excessive human discretion in situations of missing evidence logic or high conflict.

[0017] 3. Establish a semantic space mapping and verification closed loop for multidimensional heterogeneous data to improve the system's feature extraction and regulatory efficiency for complex behavioral logic. This invention maps real-time physical behavior data and subjective feedback data at the psychological level into behavioral intensity feature vectors and confidence intention feature vectors within the same semantic space, respectively, to achieve normalized representation of heterogeneous data in the mathematical dimension. On this basis, the system uses the logical distance calculation between vectors to directly drive the parameter adjustment of the risk assessment model, forming a data processing closed loop of monitoring, verification, and locking. This processing architecture enables the system not only to perceive the frequency of physical collisions in a single dimension, but also to capture the specific logical anomaly feature of intense physical behavior but calm subjective feedback. Thus, it can automatically identify complex and hidden risk events during the underlying data flow process, making up for the technical deficiency of insufficient feature extraction when facing highly disguised behaviors from a single data source. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the multi-source data fusion monitoring and early warning logic of the present invention. Figure 2 This is a multi-dimensional factor correlation diagram for determining the latent risk status of this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A campus physical bullying monitoring and early warning system based on multi-source data fusion includes: The multi-source heterogeneous data acquisition module is configured to acquire in real time the objective behavior time-series data stream within the target monitoring area and simultaneously acquire the subjective interactive feedback signals of the monitored object; The security semantic space mapping module, connected to the multi-source heterogeneous data acquisition module, is configured to construct a unified-dimensional security management semantic space. Based on the preset action feature extraction algorithm, it transforms the objective behavior time-series data stream into behavior intensity feature vectors mapped in the security management semantic space. At the same time, based on the preset semantic tendency analysis algorithm, it transforms the subjective interaction feedback signals into confidence and willingness feature vectors mapped in the security management semantic space. The logical mutual exclusion potential energy verification module, connected to the security semantic space mapping module, is configured to execute a logical mutual exclusion potential energy calculation procedure based on the geometric positional relationship between the behavior intensity feature vector and the confidence intention feature vector in the security management semantic space, in order to generate a regulatory logic divergence index that characterizes the degree of logical contradiction between objective physical behavior and subjective feedback information; wherein, the magnitude of the regulatory logic divergence index not only responds to the magnitude of the behavior intensity feature vector, but is also independently controlled by the difference in the direction of the vector angle between the behavior intensity feature vector and the confidence intention feature vector; The graded early warning execution module, connected to the logical mutual exclusion potential energy verification module, is configured to compare the regulatory logic divergence index with the preset risk mutual exclusion threshold, and generate and output a physical intervention instruction indicating that the system is in a state of hidden risk when the regulatory logic divergence index is greater than the risk mutual exclusion threshold.

[0021] Preferably, the multi-source heterogeneous data acquisition module includes: a panoramic vision capture submodule, configured to acquire high-resolution on-site environmental image sequences and extract human skeletal key point coordinate data from them as the spatial feature source of the objective behavior time-series data stream; a non-contact vital sign monitoring submodule, configured to simultaneously acquire the heart rate variability parameters and skin conductivity parameters of the monitored object and use the parameters as physiological auxiliary dimension data of the objective behavior time-series data stream; and an interactive feedback acquisition submodule, configured to generate a control signal to start an inquiry program after detecting a motion event with acceleration exceeding a preset threshold, in order to obtain subjective interactive feedback signals.

[0022] Preferably, the security semantic space mapping module includes: a spatiotemporal action convolution calculation submodule, configured to perform spatiotemporal feature encoding on the objective behavior time-series data stream, extract physical feature parameters including limb contact frequency, center of gravity displacement amplitude and action intensity, and normalize the physical feature parameters to generate a behavior intensity feature vector; and a semantic tendency parsing submodule, configured to perform natural language processing on the subjective interaction feedback signal, extract semantic feature values ​​representing the denial, concealment or fear state of the monitored object, and map the semantic feature values ​​into directional components to generate a credibility intention feature vector.

[0023] Preferably, the logic mutual exclusion potential energy calculation procedure executed in the logic mutual exclusion potential energy verification module determines the regulatory logic divergence index based on the following mathematical relationship: ,in, As a divergence indicator of regulatory logic, For behavior intensity feature vectors, This is the feature vector of confidence and willingness. The intensity modulus characterizing objective physical behavior. The cosine value representing the directional consistency between objective physical behavior and subjective feedback information in the semantic space of safety management. and These are the preset weighted balance coefficients.

[0024] Preferably, the logically mutually exclusive potential energy verification module further includes: a dynamic weighted calibration submodule, configured to adjust the weighted balance coefficients in the relation in real time based on the confidence feature components contained in the confidence intention feature vector. The numerical value; specifically configured as follows: when the reliability intention feature vector indicates that the monitored object is in a high stress characteristic state, the weighted balance coefficient is automatically increased. The value is used to improve the sensitivity of the regulatory logic divergence indicator to logical contradictions.

[0025] Preferably, the system further includes: a digital rule base storage module for storing digital entries of management rules and mapping each entry to a rule constraint boundary vector in the security management semantic space; the logical mutual exclusion potential energy verification module is also configured to calculate the Euclidean distance between the behavior intensity feature vector and the rule constraint boundary vector, and to superimpose the Euclidean distance as a correction factor into the regulatory logic divergence index.

[0026] Preferably, the graded early warning execution module includes: a hidden risk locking submodule, configured to generate a hidden risk event marker when the regulatory logic divergence index is greater than the risk mutual exclusion threshold and the magnitude of the behavior intensity feature vector is less than the preset violence threshold; and an evidence chain closed-loop generation submodule, configured to automatically associate and generate the original image sequence, subjective interaction feedback signal and divergence calculation parameters corresponding to the time period of the regulatory logic divergence index in response to physical intervention instructions, and digitally sign them to generate a digital evidence package.

[0027] Preferably, the multi-source heterogeneous data acquisition module further includes an environmental interference filtering submodule, which is configured to monitor the ambient background noise decibel value and the flow density of irrelevant targets in real time, and perform adaptive noise reduction processing on the data stream based on the ambient background noise decibel value only before the objective behavior time-series data stream is input into the security semantic space mapping module, so as to eliminate interference features generated by non-target events.

[0028] Preferably, the system further includes: a historical behavior baseline construction module, used to statistically analyze the average behavior intensity feature value of the group within the target monitoring area based on historical records within a sliding time window, and construct a normalized behavior baseline model; and a logical mutual exclusion potential energy verification module configured to calculate the statistical deviation of the behavior intensity feature vector relative to the normalized behavior baseline model, and to use the statistical deviation exceeding a preset benchmark value as a prerequisite trigger condition for starting the logical mutual exclusion potential energy calculation procedure.

[0029] Preferably, the graded early warning execution module further includes: a multi-level response communication submodule, configured to selectively activate different physical response strategies based on the preset value range into which the regulatory logic divergence index falls; specifically configured as follows: when the index is in the first preset range, an activation signal is generated to drive the on-site audible and visual warning device; when the index is in the second preset range greater than the first preset range, an emergency distress signal containing location coordinate data and on-site summary data is simultaneously sent to the remote management center via an encrypted wireless channel.

[0030] Example 1: In an administrative supervision and safety management scenario deployed in high-frequency activity areas on campus, a multi-source heterogeneous data acquisition module performs continuous monitoring tasks. The panoramic vision capture submodule acquires objective behavioral time-series data streams containing high-frequency physical contact and center of gravity displacement features. The interactive feedback acquisition submodule simultaneously acquires subjective interactive feedback signals containing deniable semantics indicating playful fighting. The system input presents a logically mutually exclusive state where objective physical features display high-risk values ​​and subjective semantic features display low-risk values. The safety semantic space mapping module maps the objective behavioral time-series data stream into behavioral intensity feature vectors based on preset action feature extraction algorithms and semantic tendency analysis algorithms. The subjective interaction feedback signal is mapped into a confidence and willingness feature vector. Behavioral intensity feature vector Having a first modulus length value, the confidence intention feature vector With behavioral intensity feature vector There is a negative correlation angle between them.

[0031] The logical mutual exclusion potential energy verification module is based on the behavior intensity feature vector. With confidence intention feature vector The geometric positional relationship is used to calculate the mutually exclusive potential energy of the execution logic, and the regulatory logic divergence index is determined based on the following mathematical relationship: ,in, As a divergence indicator of regulatory logic, For behavior intensity feature vectors, This is the feature vector of confidence and willingness. The intensity modulus characterizing objective physical behavior. The cosine value representing the directional consistency between objective physical behavior and subjective feedback information in the semantic space of safety management, where α and β are preset weighted balance coefficients. During the calculation process, due to the behavior intensity feature vector... With confidence intention feature vector The opposite direction If the value is negative, the term As the value increases, the dynamic weighted calibration submodule responds to the high-stress features in the confidence intention feature vector by increasing the value of the weighted balance coefficient β. The logical mutual exclusion potential energy verification module outputs a regulatory logical divergence index that is greater than the preset risk mutual exclusion threshold. The tiered early warning execution module will incorporate regulatory logic divergence indicators. Compared with the risk mutual exclusion threshold, in response to the indicator being greater than the threshold, a hidden risk event marker is generated and a physical intervention instruction is output. The system dynamically shrinks the effective boundary of the administrative discretion legality manifold based on the regulatory logic divergence indicator, limiting the weighting influence of low-reliability subjective data on risk assessment.

[0032] Example 2: In the verification scenario of the core algorithm processing performance of the campus safety management system, the test environment was set to process a historical backtracking dataset containing high-intensity unstructured noise. This verified the risk locking capability of the logical mutual exclusion potential energy verification module in the face of objectively high-risk and subjectively low-willing conditions, and its difference in judgment compared to the traditional linear weighted model. The test platform consisted of a cluster of servers configured to perform high-throughput parallel computing. Its core processing unit had floating-point operation capabilities to support vector space mapping and potential energy calculation. The data source used in the experiment was the multimodal campus interaction benchmark dataset (SCID-2025), which included motion capture data recorded by physical sensors and semantic text from synchronized audio recordings. To simulate signal pollution in real engineering sites, Gaussian white noise with a signal-to-noise ratio of 20dB and random video frame packet loss disturbances were actively superimposed on the original data stream to construct a non-ideal input boundary. Before executing the logical mutual exclusion potential energy calculation procedure, the system calibrated the preset weighting balance coefficients α and β according to the least squares principle. The technical considerations for parameter setting lie in balancing the weighted contributions of physical behavior intensity and logical inconsistency in the risk assessment system. When α is too high, the system tends to act as a motion intensity detector, making it susceptible to interference from physical activities; when β is too high, the system is overly sensitive to minor logical inconsistencies. Based on gradient analysis of historical false alarm rate data, the working window is determined to be α in the range of 0.3 to 0.4 and β in the range of 0.6 to 0.7. In this experimental sample group, α=0.35 and β=0.65 are set. The experimental design includes three parallel control groups. The control group adopts a linear weighted risk assessment model, whose calculation logic is the arithmetic weighted average of behavior score and reliability score. Sample group 1 of this invention adopts a logical mutual exclusion potential energy calculation procedure based on fixed coefficients. Sample group 2 of this invention introduces a dynamic weighted calibration mechanism, allowing the β value to dynamically fluctuate with the stress frequency characteristics in the speech signal. The experiment selects coercive denial samples as input. Within a 10-second time window, the frequency of limb contact reaches 4 times, and the center of gravity shifts. After processing by the safety semantic space mapping module, a behavior intensity feature vector is generated. Its mold length The value is 0.85. The synchronous subjective interaction feedback signal is the party's voice recording ("We were just playing around"). After semantic bias analysis, the generated credibility intention feature vector is... In the semantic space, it points to the safe / entertainment area; in the semantic space of security management, it refers to... and They exhibit a divergent state, with their directions exhibiting the same cosine value. The measured value was -0.82.

[0033] Under the linear model processing of the comparative sample group, the subjective feedback low-risk score of 0.15 and high-risk behavior score of 0.85 are averaged to calculate an output risk index of 0.50, which is lower than the preset intervention threshold of 0.60. The system determines that the state is risk-free. In sample group 1 of this invention, the logical mutual exclusion potential energy verification module performs the calculation according to the formula. Part 1 The contribution value is 0.2975; in the second part, the cosine value is negative due to the opposite direction, and the mutually exclusive terms... The value is 1.82, calculated using the coefficient β and the modulus. The product amplification, this part contributes 1.0055, and the final output regulatory logic divergence index The value is 1.303, which is higher than the risk mutual exclusion threshold of 0.80. The system identifies the logical anomaly and generates a physical intervention command. In sample group 2 of this invention, the dynamic weighted calibration submodule detects micro-vibration features in the voice signal and temporarily increases the weighted balance coefficient β to 0.85. Under this condition, the calculated regulatory logic divergence index is... Climbing to 1.612, data shows that compared to sample group 1 of this invention, sample group 2 of this invention shortened the response time for risk assessment by 150 milliseconds, and in the recall rate test of similar latent stress samples, it increased from 92.5% of sample group 1 to 96.8%. Further boundary stress tests show that when the β value is set above the upper limit of 0.95, The oversaturation of the indicators led to the system misjudging normal disagreements. When the β value was below the lower limit of 0.20, the gain effect of the mutually exclusive terms was overwhelmed by noise, and the recall rate dropped to 45%, the same level as the comparison sample. The experimental results show that by introducing the direction consistency cosine value into the nonlinear potential energy calculation, this technical solution can generate a risk gain signal by using the logical contradiction itself when dealing with mutually exclusive evidence scenarios. This avoids the risk masking problem caused by the averaging effect of the linear model and achieves quantitative locking of hidden school bullying behavior.

[0034] Example 3: This example addresses the extreme blind spot scenario of non-contact psychological coercion occurring in a hidden corner of a school campus. It details the underlying execution logic of the system's core algorithm at the level of heterogeneous data space mapping and dynamic parameter adaptation. In such scenarios, perpetrators often do not engage in large-scale physical confrontations, but rather exert pressure through continuous close-range encirclement and low-frequency threatening language. This results in low behavioral intensity readings from conventional motion capture systems, creating highly deceptive false safety signals. The system initiates a heterogeneous feature homogenization mapping procedure to resolve the incomparability between physical behavioral data and speech semantic data in their original dimensions. For the video stream acquired by the panoramic visual capture submodule, the action feature extraction algorithm... Instead of directly outputting scalar scores, the algorithm first performs spatiotemporal graph convolution operations to extract kinematic features, including limb extremity velocities, torso tilt angles, and multi-person position topology. These features are then compressed into a 128-dimensional behavioral feature vector through a pre-trained tensor projection layer. Simultaneously, for the audio signals acquired by the interactive feedback acquisition submodule, a semantic sentiment analysis algorithm performs parallel dual-channel processing. The first channel uses Mel-frequency cepstral coefficient analysis to extract acoustic physical features characterizing vocal cord tremors and fundamental frequency jitter. The second channel uses a natural language processing model to parse the semantic negativity of the text. The outputs of both channels are projected into the same 128-dimensional Hilbert space to synthesize a confidence-intention feature vector. This preprocessing mechanism based on isomorphic projection ensures the accuracy of behavior intensity feature vectors. With confidence intention feature vector The mathematical premise of being able to calculate the cosine of the included angle enables the semantic space of safety management to be transformed from a conceptual model into a physical mathematical space with executable Euclidean measurements.

[0035] Based on this, the logic mutual exclusion potential energy verification module activates the nonlinear stress gain calculation logic to address the engineering challenge that static weights cannot capture implicit risks. The system monitors the acoustic physical feature components in the confidence intention feature vector in real time and calculates the stress fingerprint index characterizing the psychological tension level of the monitored subjects. To avoid abrupt weight changes due to a single noise interference, the dynamic adjustment of the weighted balance coefficient β no longer follows a simple linear rule, but evolves in real time based on the following saturated gain function: ,in, The dynamic weighted balance coefficient within the current time window. Basic weight value, The system allows a maximum weight limit, and γ is a sensitivity adjustment factor. For the normalized stress fingerprint index, the hyperbolic tangent structure of this function ensures that when the stress index... When the value is low, the β value remains at the baseline level, ensuring the stability of the system under normal conditions; while when When it rises, Value rapidly to It approximates but does not diverge infinitely, thus improving sensitivity to logical mutual exclusions while preventing system decision logic from collapsing due to weight overflow; in a specific operational instance, the monitored object is cornered, and although its physical movements are minimal, it leads to... The score was only 0.3, which is considered low-intensity behavior; however, the system detected high-frequency vocal cord micro-vibrations in the speech, leading to... A surge, according to the formula above, The value rapidly and nonlinearly climbs from the baseline of 0.65 to 0.92. At this point, although the behavioral modulus is relatively small, due to... (Static passive) and (Verbal defensive denial) has an essential difference in semantic space direction, leading to a change in the cosine value of directional consistency. Presenting a negative value, thus being amplified. Acting on mutual exclusion terms This generated a high-intensity gain signal, which affected the final calculated regulatory logic divergence index. Exceeding the warning threshold of 0.80.

[0036] Example 4: This example constructs and executes an offline calibration and data filling procedure, initiated before the formal deployment of the system. It selects a standardized video and audio sample set covering four typical scenarios: normal teaching, minor misbehavior, implicit coercion, and overt violence, with a total duration of no less than 200 hours. The system performs full feature extraction on the sample set, using the extracted behavioral intensity feature vector and confidence intention feature vector as input. The expert-annotated real risk level is used as a supervision signal. The convolutional kernel weights and fully connected layer biases in the feature extraction algorithm are trained and solidified through backpropagation. Based on this, the system constructs a dynamic lookup table containing multi-dimensional feature indexes. The table stores the standard coordinate mapping relationships of various typical behavioral patterns in the semantic space, ensuring that in subsequent real-time monitoring, the system can quickly compare and calibrate based on real-time data and the benchmark vectors in the lookup table, thereby guaranteeing the determinism and consistency of heterogeneous feature mapping. Regarding the behavioral intensity feature vector... With confidence intention feature vector In the geometric metric benchmark problem within the same semantic space, the model building phase executes a heterogeneous joint contrastive training procedure, constructing a paired dataset containing time-synchronized video frame sequences and audio waveform segments. Based on logical consistency labels from an expert knowledge base, the training employs a contrastive loss function to synchronize the weight updates of the action feature extraction network and the semantic tendency parsing network. For labeled logically consistent sample pairs, the optimizer performs Euclidean distance minimization to drive the 128-dimensional Hilbert space feature projection points to move closer to each other; for labeled logically mutually exclusive sample pairs, distance maximization is performed until a preset edge threshold is exceeded.

[0037] During training, the system employs a Siamese network architecture. The input layer does not directly interface with the raw data; instead, it uses two independent, pre-trained multilayer perceptron projectors. For positive sample pairs (i.e., actions in the video and semantics in the audio have a temporal overlap exceeding 90% and are consistent with manually labeled sentiments), the loss function forces the distance between the two feature vectors in the 128-dimensional Euclidean space to be less than 0.1. For negative sample pairs, the loss function forces their distance to be greater than the edge limit of 1.0. After 10,000 iterations of convergence, the originally heterogeneous visual and auditory features are forced to map to the same basis metric space. Based on the clear logical consistency and physical meaning of the cosine value of the angle between vectors, the training establishes the geometric metric isomorphism of the two heterogeneous vector spaces, ensuring the consistency of the direction cosine value. The objective representation of the overlap between physical behavior and subjective intention at the semantic level is described. In addition, this embodiment also describes in detail the pre-deployment calibration procedure to solve the system adaptability problem caused by differences in deployment environment. When the system first connects to the target campus network or when the monitoring area undergoes physical changes, the calibration procedure is automatically triggered to collect environmental background noise data under no-load conditions, including light intensity distribution, acoustic noise spectrum and network latency jitter parameters. Based on these baseline data, the gain coefficient and filtering threshold of the multi-source heterogeneous data acquisition module are initialized and configured. The system requires the execution of a set of preset standard test actions. By comparing the deviation between the system's measured output and the standard test template, the projection matrix parameters in the safety semantic space mapping module are automatically fine-tuned until the error of the regulatory logic divergence index in the standard test scenario converges to the preset accuracy range, thereby ensuring that the risk judgment logic of the system always maintains a high degree of stability and reproducibility under different physical environments.

[0038] Example 5: This example establishes and solidifies standardized engineering procedures for pre-deployment calibration and model building, ensuring the reproducibility and stability of the technical solution under different batches of hardware and heterogeneous network environments. It performs calibration and optimization of physical geometric parameters. For the camera deployment in the panoramic vision capture submodule, the procedure does not employ empirical random installation, but rather determines the optimal parameters based on the engineering principles of maximizing coverage and minimizing occlusion. The simulation environment constructs a 3D model including a standard classroom desk and chair layout, personnel flow trajectories, and lighting changes. In the simulation, the camera's field of view is set as a variable. The values ​​range from 90 degrees to 150 degrees, with a step size of 5 degrees. The installation pitch angle is also a variable, ranging from 15 degrees to 45 degrees, with a step size of 5 degrees. Through Monte Carlo simulation, the system calculates the effective coverage of key interactive areas in the classroom and the success rate of resolving overlapping occlusions under different parameter combinations. Simulation results show that when the field of view is set to 120 degrees and the installation pitch angle is 30 degrees, the system can ensure full coverage while minimizing the error rate in behavior recognition caused by personnel occlusion. Therefore, the regulations explicitly use these two parameters as the benchmark optimal values ​​for system deployment and require that they be used in actual installation with a laser rangefinder and level. Precise calibration was performed, with errors controlled within ±2 degrees. The procedure then moved to the stage of refining the algorithm model's implementation path and optimizing hyperparameters. For the action feature extraction algorithm, the procedure explicitly adopted a deep learning architecture based on a spatiotemporal graph convolutional network (ST-GCN). To ensure the model's generalization ability in real-world scenarios, the training dataset was not limited to publicly available standard behavior libraries, but was augmented with a dedicated dataset, C-Bully-2025, constructed by the applicant. This dataset contains over 5000 expert-annotated bullying and non-bullying interaction segments, covering complex scenes with varying lighting, perspectives, and crowding levels. This data was used for model training. During the process, the hyperparameter settings followed a Bayesian optimization strategy. Specifically, the learning rate, batch size, and number of convolutional kernels were used as optimization variables, and the objective function was set as the average accuracy (mAP) on the validation set. The objective function was fitted using a Gaussian process regression model, and the parameter combination with the greatest expected improvement was selected in each iteration. After 50 rounds of optimization, the optimal hyperparameter combination was determined: an initial learning rate of 0.01, decayed using a cosine annealing strategy; a batch size of 32; and 64 convolutional kernels in the first layer. This optimization process ensured that the model performance did not depend on random luck, but was based on scientific statistical optimization results.

[0039] Finally, the procedure defines the quantification of logical judgment conditions and anomaly handling strategies. For fuzzy concepts such as high risk and low confidence, it transforms them into quantized logic gates based on deterministic inputs, such as the magnitude of the behavioral intensity feature vector. The intensity is divided into three discrete intervals: [0, 0.3) for low intensity, [0.3, 0.7) for medium intensity, and [0.7, 1.0] for high intensity; similarly, the direction cosine value of the confidence intention feature vector... Mapped to continuous confidence levels within the [-1,1] interval, the system incorporates a self-checking, fault-tolerant state machine. Before each initiation of the logical mutual exclusion potential energy calculation, it checks the dimensional integrity and numerical range compliance of the input vector. If a zero vector or numerical overflow anomaly caused by a sensor disconnection is detected, the state machine jumps to a degraded operation mode, automatically shielding abnormal channel data and using only reliable data from a single dimension for conservative early warning, while simultaneously triggering maintenance alarm signals. This ensures that the system can maintain basic safety monitoring functions even when some components fail, avoiding complete paralysis due to a single point of failure; monitoring logical divergence index The weighted balance coefficients α and β in the calculation formula are determined by the field boundary approximation calibration method. During the deployment initialization phase, the signal processing unit is sequentially input with a negative sample set containing high-frequency physical contact but without malicious intent, and a positive sample set containing weak physical movements and high-frequency tremors in the voice. The calibration procedure fixes β as the initial benchmark and gradually increases the value of α in 0.05 steps until the false alarm rate of the negative sample set reaches the preset noise baseline, at which point the value of α is locked. The value of β is monotonically adjusted with a synchronization length until the recall rate of the positive sample set reaches the response threshold. The physical calibration process establishes the mapping relationship between the coefficient values ​​and the environmental background noise level in a specific monitoring area, and defines the linear working range of the logical mutual exclusion potential energy calculation procedure.

[0040] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A campus physical bullying monitoring and early warning system based on multi-source data fusion, characterized in that, include: The multi-source heterogeneous data acquisition module is configured to acquire in real time the objective behavior time-series data stream within the target monitoring area and simultaneously acquire the subjective interactive feedback signals of the monitored object; The security semantic space mapping module, connected to the multi-source heterogeneous data acquisition module, is configured to construct a unified-dimensional security management semantic space. Based on the preset action feature extraction algorithm, it transforms the objective behavior time-series data stream into behavior intensity feature vectors mapped in the security management semantic space. At the same time, based on the preset semantic tendency analysis algorithm, it transforms the subjective interaction feedback signals into confidence and willingness feature vectors mapped in the security management semantic space. The logical mutual exclusion potential energy verification module, connected to the security semantic space mapping module, is configured to execute a logical mutual exclusion potential energy calculation procedure based on the geometric positional relationship between the behavior intensity feature vector and the confidence intention feature vector in the security management semantic space, in order to generate a regulatory logic divergence index that characterizes the degree of logical contradiction between objective physical behavior and subjective feedback information; wherein, the magnitude of the regulatory logic divergence index not only responds to the magnitude of the behavior intensity feature vector, but is also independently controlled by the difference in the direction of the vector angle between the behavior intensity feature vector and the confidence intention feature vector; The graded early warning execution module, connected to the logical mutual exclusion potential energy verification module, is configured to compare the regulatory logic divergence index with the preset risk mutual exclusion threshold, and generate and output a physical intervention instruction indicating that the system is in a state of hidden risk when the regulatory logic divergence index is greater than the risk mutual exclusion threshold.

2. The campus physical bullying monitoring and early warning system based on multi-source data fusion according to claim 1, characterized in that, The multi-source heterogeneous data acquisition module includes: a panoramic vision capture submodule, configured to acquire high-resolution on-site environmental image sequences and extract human skeletal key point coordinate data as the spatial feature source of the objective behavior time-series data stream; a non-contact vital sign monitoring submodule, configured to simultaneously acquire the heart rate variability parameters and skin conductivity parameters of the monitored object and use the parameters as physiological auxiliary dimension data of the objective behavior time-series data stream; and an interactive feedback acquisition submodule, configured to generate a control signal to start an inquiry program after detecting a motion event with acceleration exceeding a preset threshold, in order to obtain subjective interactive feedback signals.

3. The campus physical bullying monitoring and early warning system based on multi-source data fusion according to claim 1, characterized in that, The safety semantic space mapping module includes: a spatiotemporal action convolution calculation submodule, configured to encode the spatiotemporal features of the objective behavior time-series data stream, extract physical feature parameters including limb contact frequency, center of gravity displacement amplitude, and action intensity, and normalize the physical feature parameters to generate a behavior intensity feature vector; and a semantic tendency parsing submodule, configured to perform natural language processing on subjective interaction feedback signals, extract semantic feature values ​​representing the denial, concealment, or fear state of the monitored object, and map the semantic feature values ​​into directional components to generate a credibility intention feature vector.

4. The campus physical bullying monitoring and early warning system based on multi-source data fusion according to claim 1, characterized in that, The logic mutual exclusion potential energy calculation procedure executed in the logic mutual exclusion potential energy verification module determines the regulatory logic divergence index based on the following mathematical relationship: ,in, As a divergence indicator of regulatory logic, For behavior intensity feature vectors, This is the feature vector of confidence and willingness. The intensity modulus characterizing objective physical behavior. The cosine value representing the directional consistency between objective physical behavior and subjective feedback information in the semantic space of safety management is α and β, which are preset weighted balance coefficients.

5. A campus physical bullying monitoring and early warning system based on multi-source data fusion according to claim 4, characterized in that, The logical mutual exclusion potential energy verification module further includes a dynamic weighted calibration submodule, configured to adjust the weighted balance coefficients in the relation in real time based on the confidence feature components contained in the confidence intention feature vector. The numerical value; specifically configured as follows: when the reliability intention feature vector indicates that the monitored object is in a high stress characteristic state, the weighted balance coefficient is automatically increased. The value is used to improve the sensitivity of the regulatory logic divergence indicator to logical contradictions.

6. A campus physical bullying monitoring and early warning system based on multi-source data fusion according to claim 1, characterized in that, The system also includes: a digital rule base storage module, which stores and manages digital entries of rules and maps each entry to a rule constraint boundary vector in the security management semantic space; and a logical mutual exclusion potential energy verification module, which is also configured to calculate the Euclidean distance between the behavior intensity feature vector and the rule constraint boundary vector, and to superimpose the Euclidean distance as a correction factor into the regulatory logic divergence index.

7. A campus physical bullying monitoring and early warning system based on multi-source data fusion according to claim 1, characterized in that, The graded early warning execution module includes: a hidden risk locking submodule, configured to generate a hidden risk event marker when the regulatory logic divergence index is greater than the risk mutual exclusion threshold and the magnitude of the behavior intensity feature vector is less than the preset violence threshold; and an evidence chain closed-loop generation submodule, configured to automatically associate and generate the original image sequence, subjective interaction feedback signal and divergence calculation parameters corresponding to the time period of the regulatory logic divergence index in response to physical intervention instructions, and digitally sign them to generate a digital evidence package.

8. A campus physical bullying monitoring and early warning system based on multi-source data fusion according to claim 1, characterized in that, The multi-source heterogeneous data acquisition module also includes an environmental interference filtering submodule, which is configured to monitor the ambient background noise decibel value and the flow density of irrelevant targets in real time, and perform adaptive noise reduction processing on the data stream based on the ambient background noise decibel value only before the objective behavior time-series data stream is input into the security semantic space mapping module, so as to eliminate the interference features generated by non-target events.

9. A campus physical bullying monitoring and early warning system based on multi-source data fusion according to claim 1, characterized in that, The system also includes: a historical behavior baseline construction module, which is used to statistically analyze the average behavior intensity feature value of the group in the target monitoring area based on the historical records within the sliding time window, and construct a normalized behavior baseline model; and a logical mutual exclusion potential energy verification module, which is configured to calculate the statistical deviation of the behavior intensity feature vector relative to the normalized behavior baseline model, and use the statistical deviation exceeding the preset benchmark value as a prerequisite trigger condition for starting the logical mutual exclusion potential energy calculation procedure.

10. A campus physical bullying monitoring and early warning system based on multi-source data fusion according to claim 1, characterized in that, The graded early warning execution module also includes a multi-level response communication submodule, which is configured to selectively activate different physical response strategies based on the preset value range into which the regulatory logic divergence index falls; specifically, when the index is in the first preset range, an activation signal is generated to drive the on-site audible and visual warning equipment; when the index is in the second preset range greater than the first preset range, an emergency distress signal containing location coordinate data and on-site summary data is simultaneously sent to the remote management center through an encrypted wireless channel.

Images