An AI-based campus safety hazard monitoring system

Abstract

This invention discloses an artificial intelligence-based campus safety hazard monitoring system, belonging to the field of campus security technology. Specifically, it includes: collecting campus environmental data through a multimodal sensor network and converting it into a low-level physical event stream; using causal rules in a pre-set campus safety knowledge base for matching and reasoning to generate safety hazard hypotheses containing expected consequence event sequences and initial confidence levels; identifying discrepancy nodes and selecting key verification evidence types through comparative analysis; calculating the temporal matching degree based on the enhanced collection of verification event sequences to generate risk assessment values ​​corresponding to each safety hazard hypothesis; and automatically triggering a standardized handling protocol when the risk assessment value exceeds a threshold. This invention, through a multi-level causal reasoning and verification mechanism, effectively solves the gap between physical perception signals and behavioral semantics in existing technologies, achieving accurate identification and rapid handling of campus safety hazards.

Description

Technical Field

[0001] This invention relates to the field of campus security technology, specifically to an artificial intelligence-based campus security hazard monitoring system. Background Technology

[0002] With the continuous advancement of smart campus construction, artificial intelligence technology has provided new solutions for campus security management. Currently, mainstream intelligent monitoring methods mainly rely on two technological paths: behavior recognition systems based on video analytics and anomaly detection systems based on audio analytics. Video analytics systems use deep learning algorithms to identify specific behavioral patterns such as running and gathering in monitored footage; audio analytics systems detect abnormal sound events such as screams and impacts through acoustic feature analysis. These technologies achieve automated identification of overt behaviors through feature extraction and pattern matching.

[0003] However, these existing technologies all face a fundamental technical challenge: the semantic ambiguity of physical perception and the unobservability of behavioral intent. There is a huge semantic gap between the actual received sensor signals (such as movement trajectories, sound frequencies, and other underlying physical characteristics) and real security threats (such as bullying, theft, and other behaviors with clear intent).

[0004] The same physical behavior can have completely different safety implications in different contexts—running in a corridor could be an emergency escape or a chase and fight; whispers in a corner could be an academic discussion or a bullying threat. However, failing to break through the simple mapping paradigm of "feature-behavior" makes it impossible to understand the causal relationships between events, let alone construct a complete behavioral narrative from discrete perceptual signals, ignoring the intention dimension behind the behavior, resulting in a large number of false alarms and false negatives in practical applications. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based campus safety hazard monitoring system to solve the following technical problems:

[0006] How can we overcome the limitations of the simple "feature-behavior" mapping in existing technologies, establish the ability to understand discrete events into complete behavioral narratives, and achieve in-depth identification of security risks?

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] An artificial intelligence-based campus safety hazard monitoring system includes:

[0009] The data acquisition module is used to continuously collect raw data through a multimodal sensor network deployed in the campus environment and convert it into a low-level physical event stream, which includes object displacement events, acoustic events and human posture events.

[0010] The event reasoning module is used to match the underlying physical event stream with causal rules in the preset campus safety knowledge base. The causal rules include a sequence of premise events, a sequence of expected consequences events, and corresponding initial confidence levels, and generate several safety hazard hypotheses including the sequence of expected consequences events and the initial confidence levels.

[0011] The data analysis module is used to identify discrepancy nodes based on the expected consequence event sequence contained in each security hazard hypothesis, and select key verification evidence types from the event types contained in the underlying physical event stream based on the discrepancy nodes.

[0012] The hypothesis verification module is used to control the corresponding sensing unit to increase the acquisition frequency based on the key verification evidence type to obtain a high-density verification event sequence; by comparing the temporal consistency between the high-density verification event sequence and the expected consequence event sequence of each safety hazard hypothesis, and combining the initial confidence level of each safety hazard hypothesis, a risk assessment value corresponding to each safety hazard hypothesis is generated.

[0013] The result generation module is used to automatically trigger the standardized handling protocol bound to the safety hazard category when the risk assessment value corresponding to any safety hazard assumption exceeds the preset risk threshold.

[0014] As a further aspect of the present invention: the specific process of converting the data acquisition module into an underlying physical event stream is as follows:

[0015] The video stream data is subjected to continuous inter-frame difference calculation. The motion vector field of the pixel is analyzed by optical flow method. When the continuous motion vector forms a stable motion trajectory, an object movement event containing motion direction, velocity and acceleration features is generated.

[0016] The audio stream data is subjected to short-time Fourier transform to obtain a spectrogram. The harmonic structure and energy distribution features in the spectrogram are extracted. By matching the predefined acoustic feature template, an acoustic event containing sound source type, loudness and pitch features is generated.

[0017] Feature point detection is performed on video stream data. By analyzing the spatial relationship changes of feature points between consecutive frames, dynamic pose events are generated when structural changes that conform to a preset pattern are identified, and the pose type and spatial distribution of feature points are recorded.

[0018] Establish a unified spatiotemporal coordinate system, transform the coordinates of various events to this coordinate system, and attach precise timestamps and spatial location information to each event;

[0019] A density-based spatial clustering algorithm is used to aggregate spatiotemporally adjacent events. Redundant events are eliminated by calculating the similarity of event feature vectors, forming a standardized underlying physical event stream.

[0020] As a further aspect of the present invention: the specific construction process of the preset campus safety knowledge base in the event reasoning module is as follows:

[0021] Typical security incident cases are extracted from historical campus surveillance records. Temporal correlation analysis is performed on the continuous event sequences contained in each case to identify event combination patterns with causal relationships.

[0022] The identified event combination patterns are abstracted into causal rules, where each causal rule contains a sequence of prerequisite events and a sequence of expected consequences events. The sequence of prerequisite events consists of an initial event chain that triggers a security hazard, and the sequence of expected consequences events consists of a subsequent event chain that may be triggered by the prerequisite events.

[0023] The initial confidence level of each causal rule is calculated based on historical event statistics. The initial confidence level is obtained by statistically analyzing the ratio of the frequency of the causal rule occurring in historical data to the total frequency of all similar event sequences.

[0024] Establish a hierarchical organizational structure for the causal rule base, classify and store causal rules according to the type of safety hazard, and create an index relationship for each rule;

[0025] The completed causal rule base and its associated initial confidence values ​​are stored as a preset campus safety knowledge base.

[0026] As a further aspect of the present invention: the specific process of generating several safety hazard hypotheses, including the expected consequence event sequence and initial confidence levels, in the event reasoning module is as follows:

[0027] Iterate through all causal rules in the campus safety knowledge base and perform time-series matching of the prerequisite event sequence of each rule with the real-time underlying physical event stream;

[0028] When the event stream contains all the event types defined in the premise event sequence, and the time interval between the events meets the temporal constraints defined in the rule, and the spatial distribution satisfies the proximity condition defined in the rule, the rule is considered to be successfully matched.

[0029] For each successfully matched causal rule, a corresponding security risk hypothesis is created. The security risk hypothesis inherits the expected consequence event sequence and initial confidence level of the rule, resulting in several security risk hypotheses.

[0030] As a further aspect of the present invention: the specific process for determining the dissimilar nodes in the data analysis module is as follows:

[0031] The expected consequences event sequences of each safety hazard hypothesis are aligned using a dynamic time warping algorithm, and the positions with different event types after sequence alignment are identified as candidate difference points.

[0032] One-hot encoding is performed on the event types corresponding to the candidate difference points to construct event type feature vectors, and the cosine similarity of the feature vectors of corresponding positions of adjacent sequences is calculated.

[0033] The Euclidean distance is calculated based on the spatial coordinates of the events corresponding to the candidate difference points, and the time difference is calculated based on the timestamp. The spatial distance and the time difference are weighted and fused into a spatiotemporal distribution distance.

[0034] The cosine similarity is converted into an event type difference, and the event type difference and the spatiotemporal distribution distance are combined according to a preset weight to generate a comprehensive difference index. The candidate difference point corresponding to the highest comprehensive difference index is selected and marked as a difference node.

[0035] As a further aspect of the present invention: the specific process for generating risk assessment values ​​corresponding to each safety hazard hypothesis in the hypothesis verification module is as follows:

[0036] Increase the acquisition frequency of the corresponding sensing unit in the control sensor network to obtain a high-density verification event sequence;

[0037] Extract the attribute parameters contained in each event in the high-density verification event sequence. For object displacement events, the attribute parameters include motion direction, velocity and acceleration. For acoustic events, the attribute parameters include sound source type, loudness and pitch features. For human posture events, the attribute parameters include posture type and spatial distribution of feature points.

[0038] The attribute parameters of the verified event are compared with the corresponding attribute parameters of the expected event at the corresponding difference node position for each security hazard hypothesis;

[0039] The matching degree is calculated based on the comparison results of the attribute parameters. The current confidence level of each safety hazard hypothesis is used as the prior probability, and the calculated matching degree is used as the likelihood. The risk assessment value corresponding to each safety hazard hypothesis is calculated by Bayes' formula.

[0040] As a further aspect of the present invention, if there are two or more candidate difference points with equal comprehensive difference indices and all of them being the maximum value, then all of the candidate difference points are marked as difference nodes.

[0041] As a further aspect of the present invention: the specific process by which the standardized handling protocol bound to the safety hazard category is automatically triggered in the result generation module is as follows:

[0042] Based on the safety hazard assumptions corresponding to the corrected confidence level exceeding the preset threshold, extract the hazard type identifiers bound to them;

[0043] Based on the hazard type identifier, query the predefined standardized handling protocol library to obtain the corresponding handling instruction sequence;

[0044] The sequence of processing instructions is compiled into an executable set of control instructions, which includes the target terminal address, execution time window, and action parameters.

[0045] The control instruction set is distributed to the corresponding execution terminal device through the campus communication network; the instruction confirmation signal returned by the execution terminal device is received, and the execution status record of the handling protocol is updated.

[0046] As a further aspect of the present invention, it also includes receiving an instruction confirmation signal and on-site handling record returned by the execution terminal device; and comparing and verifying the on-site handling record with the safety hazard hypothesis that triggered the handling.

[0047] If the comparison and verification are consistent, a positive optimization sample containing the verification event sequence is generated, and the initial confidence of the corresponding causal rule in the campus safety knowledge base is improved based on the positive optimization sample.

[0048] If the comparison and verification are inconsistent, the verification event sequence is extracted and compared with the expected consequence event sequence of the causal rule corresponding to the safety hazard hypothesis. The difference nodes are identified, and negative judgment conditions are established based on the identified difference nodes.

[0049] The established negative judgment conditions are added to the expected consequence event sequence of the corresponding causal rule and re-injected into the campus safety knowledge base.

[0050] The beneficial effects of this invention are:

[0051] 1) This invention effectively bridges the gap between physical sensing signals and behavioral semantics by establishing a causal rule base and a hypothesis verification mechanism. This invention generates safety hazard hypotheses containing expected consequence event sequences by matching the underlying physical event stream with the premise event sequence of causal rules. This reasoning method based on causality can understand the safety implications of different behaviors in specific contexts. It is understandable that the meaning of a single physical behavior is highly uncertain and ambiguous, while its temporal position and causal relationship within a series of events can provide decisive constraints for resolving this uncertainty. Therefore, after receiving sensor data, this invention performs behavior recognition without judging a single behavior in isolation. Instead, it places it within a specific event sequence and comprehensively analyzes its safety implications in conjunction with other events occurring before and after it. Through subsequent verification modules, these hypotheses are continuously verified and corrected, thereby significantly reducing the risk of misjudgment due to behavioral semantic ambiguity and achieving accurate identification of safety hazards.

[0052] 2) This invention establishes the ability to understand discrete events into complete behavioral narratives through a multi-level event synthesis and verification process. First, it converts raw sensor data into a standardized underlying physical event stream. Then, it generates multiple parallel security hazard hypotheses through an event reasoning module. Next, a data analysis module identifies key discrepancies, and finally, a hypothesis verification module performs focused verification on these nodes. This progressive processing mechanism can construct complete behavioral narratives from isolated physical events, achieving in-depth identification of security hazards and ensuring accurate understanding of behavioral narratives in complex security scenarios.

[0053] 3) This invention receives on-site handling records returned by the execution terminal device and compares them with the safety hazard assumptions that triggered the handling. Based on the verification results, it generates positive optimization samples or reconstructs the triggering rule process, thereby continuously optimizing the causal rules and initial confidence levels in the campus safety knowledge base. When identification deviations occur, by analyzing the characteristic patterns of false alarm cases, it adds exclusionary conditions or modifies the discrimination logic in the causal rules, ensuring that it can adapt to the ever-changing security threats in the campus environment and continuously improve regulatory effectiveness. Attached Figure Description

[0054] The invention will now be further described with reference to the accompanying drawings.

[0055] Figure 1 This is a schematic diagram of the structure of a campus safety hazard monitoring system based on artificial intelligence according to the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Please see Figure 1 As shown, this invention is an artificial intelligence-based campus safety hazard monitoring system, comprising:

[0058] The data acquisition module is used to continuously collect raw data through a multimodal sensor network deployed in the campus environment and convert it into a low-level physical event stream, which includes object displacement events, acoustic events and human posture events.

[0059] Environmental data is continuously collected through a multimodal sensor network deployed throughout the campus. Continuous frame analysis is performed on the video stream; when continuous motion trajectories such as "students running quickly" or "multiple people surrounding someone" are detected, object displacement events are generated, recording features such as direction of movement (e.g., moving from the east side of the playground to the west) and speed (e.g., 5 meters per second). Spectral analysis is performed on the audio stream; when specific acoustic patterns such as "screams" or "insults" are identified, acoustic events are generated, recording parameters such as sound source type (e.g., human voice) and loudness (e.g., 90 decibels). Human skeletal keypoint detection is performed on the video data; when specific postures such as "raising hands to fight" or "bending over to avoid" are identified, human posture events are generated. All events are standardized using a unified campus coordinate system; for example, "east end of the third-floor corridor of the teaching building" is mapped to coordinates (3,15,1), forming a low-level physical event stream containing precise spatiotemporal information.

[0060] The event reasoning module is used to match the underlying physical event stream with causal rules in the preset campus safety knowledge base. The causal rules include a sequence of premise events, a sequence of expected consequences events, and corresponding initial confidence levels, and generate several safety hazard hypotheses including the sequence of expected consequences events and the initial confidence levels.

[0061] The system matches real-time event streams with causal rules in the campus safety knowledge base. Taking the "bullying incident" rule as an example, its premise sequence is: [multiple people surrounding and moving] → [threatening language] → [defensive posture], and the expected consequence sequence is: [pushing action] → [object snatching] → [crying sound], with an initial confidence level of 0.4. When a match between the actual event stream and the premise sequence is detected, a "bullying incident" hypothesis is generated, inheriting the complete consequence sequence and the initial confidence level. Simultaneously, competing hypotheses such as "playing around" may be generated, forming a set of multiple hypotheses to be verified in parallel.

[0062] The data analysis module is used to identify discrepancy nodes based on the expected consequence event sequence contained in each security hazard hypothesis, and select key verification evidence types from the event types contained in the underlying physical event stream based on the discrepancy nodes.

[0063] Compare the expected consequences of each hypothetical safety hazard event sequence. Taking the distinction between "bullying incidents" and "playful roughhousing" as an example, sequence alignment reveals the core difference lies in whether "violent physical contact" occurs. Calculate the event type difference (e.g., the semantic difference between "pushing" and "patting") and spatiotemporal distribution distance (e.g., the difference in the concealment of the incident location) of this node, and select the node with the highest overall difference as the key difference node. Based on this, "high-intensity contact incidents" are selected as the key type of verification evidence.

[0064] The hypothesis verification module is used to control the corresponding sensing unit to increase the acquisition frequency based on the key verification evidence type to obtain a high-density verification event sequence; by comparing the temporal consistency between the high-density verification event sequence and the expected consequence event sequence of each safety hazard hypothesis, and combining the initial confidence level of each safety hazard hypothesis, a risk assessment value corresponding to each safety hazard hypothesis is generated.

[0065] For selected key verification evidence types, such as "high-intensity contact events," the acquisition frequency of cameras in the corresponding areas was increased from 15 frames / second to 30 frames / second, and the microphone sampling rate was increased from 16kHz to 44.1kHz. Detailed video evidence of "obvious pushing actions" and high-definition audio evidence of "painful cries" were obtained through enhanced acquisition. The characteristics of the actual verification events were compared with the expected characteristics of each hypothesis, such as the degree of matching between measured pushing force, duration, and other parameters. Based on the comparison results, the confidence of the hypotheses was updated using Bayesian inference; for example, the confidence of "bullying events" was increased from 0.4 to 0.9.

[0066] The result generation module is used to automatically trigger the standardized handling protocol bound to the safety hazard category when the risk assessment value corresponding to any safety hazard assumption exceeds the preset risk threshold.

[0067] When the revised confidence level of any safety hazard hypothesis exceeds a preset threshold (e.g., 0.8), a standardized handling protocol is automatically triggered. Taking a "bullying incident" with a confidence level of 0.9 as an example, the system automatically executes the following: sends an instruction to the nearest security terminal in the incident area to "immediately proceed to the east corner of the playground to handle the bullying incident," sends a notification to the Moral Education Department terminal to "record and follow up on the bullying incident," and simultaneously activates all cameras in the area to continuously track and record. All instructions include specific location coordinates, incident level, and handling time limit parameters, and receive execution confirmation signals from each terminal in real time through the campus IoT.

[0068] In a preferred embodiment of the present invention, the specific process of converting the data acquisition module into an underlying physical event stream is as follows:

[0069] The video stream data is subjected to continuous inter-frame difference calculation. The motion vector field of the pixel is analyzed by optical flow method. When the continuous motion vector forms a stable motion trajectory, an object movement event containing motion direction, velocity and acceleration features is generated.

[0070] The audio stream data is subjected to short-time Fourier transform to obtain a spectrogram. The harmonic structure and energy distribution features in the spectrogram are extracted. By matching the predefined acoustic feature template, an acoustic event containing sound source type, loudness and pitch features is generated.

[0071] Feature point detection is performed on video stream data. By analyzing the spatial relationship changes of feature points between consecutive frames, dynamic pose events are generated when structural changes that conform to a preset pattern are identified, and the pose type and spatial distribution of feature points are recorded.

[0072] Establish a unified spatiotemporal coordinate system, transform the coordinates of various events to this coordinate system, and attach precise timestamps and spatial location information to each event;

[0073] A density-based spatial clustering algorithm is used to aggregate spatiotemporally adjacent events. Redundant events are eliminated by calculating the similarity of event feature vectors, forming a standardized underlying physical event stream.

[0074] For video stream data, continuous inter-frame difference calculation is performed. Moving regions are detected by comparing changes in pixel intensity values ​​between adjacent frames. Optical flow analysis is then used to analyze the displacement vectors of pixels within the moving regions. The consistency of these vectors across multiple frames is used to distinguish between real object motion and environmental noise. When a set of continuous motion vectors forms a smooth and continuous motion trajectory, it is determined to be a meaningful object movement event. Simultaneously, the direction, velocity, and acceleration characteristics are calculated based on the changes in the direction and magnitude of the motion vectors. For audio stream data, frame segmentation is first performed. A short-time Fourier transform is performed on each frame signal to convert it from the time domain to the frequency domain, obtaining a spectrogram. Analysis of the harmonic structure in the spectrogram identifies the physical source of the sound. Calculation of the energy distribution in each frequency band distinguishes the sound type. The extracted spectral features are then matched with predefined acoustic feature templates for similarity matching, accurately identifying the sound source type and quantifying loudness and pitch characteristics. For video... The streaming data also performs human keypoint detection. By establishing a topological model of human joints, it analyzes the changing patterns of joint connection angles and relative positions between consecutive frames. When the spatial relationship change of a specific joint combination is detected to conform to a preset posture pattern, a corresponding human posture event can be generated. To ensure the spatiotemporal consistency of multi-source data, a unified coordinate system based on the campus geographic information system is established. The spatial coordinates of all events are mapped to this unified coordinate system through coordinate transformation, and a timestamp generated by a high-precision time synchronization mechanism is attached to each event. Finally, a density-based spatial clustering algorithm is used to aggregate spatiotemporally adjacent and similar events into more representative composite events by calculating the distribution density of events in the spatiotemporal dimension. At the same time, redundant events caused by sensor noise or repeated detection are identified and eliminated by calculating the similarity of event feature vectors, ultimately forming a standardized underlying physical event stream with a clear structure and rich information.

[0075] In another preferred embodiment of the present invention, the specific construction process of the preset campus safety knowledge base in the event reasoning module is as follows:

[0076] Typical security incident cases are extracted from historical campus surveillance records. Temporal correlation analysis is performed on the continuous event sequences contained in each case to identify event combination patterns with causal relationships.

[0077] The identified event combination patterns are abstracted into causal rules, where each causal rule contains a sequence of prerequisite events and a sequence of expected consequences events. The sequence of prerequisite events consists of an initial event chain that triggers a security hazard, and the sequence of expected consequences events consists of a subsequent event chain that may be triggered by the prerequisite events.

[0078] The initial confidence level of each causal rule is calculated based on historical event statistics. The initial confidence level is obtained by statistically analyzing the ratio of the frequency of the causal rule occurring in historical data to the total frequency of all similar event sequences.

[0079] Establish a hierarchical organizational structure for the causal rule base, classify and store causal rules according to the type of safety hazard, and create an index relationship for each rule;

[0080] The completed causal rule base and its associated initial confidence values ​​are stored as a preset campus safety knowledge base.

[0081] When extracting typical security incident cases from historical campus surveillance records, the first step is to screen for security incidents with complete process records. For each case, a temporal correlation analysis is performed on the continuous event sequences. By analyzing the chronological order and statistical dependencies of events, causal event combination patterns are identified. For example, observing that the event sequence "crowd gathering → sudden increase in sound decibels → physical contact" repeatedly appears in historical data and has statistical significance, it can be determined that it constitutes a causal event combination pattern. When abstracting the identified event combination patterns into causal rules, based on the temporal logic and conditional probability relationships of event occurrences, frequently co-occurring event sequences with causal potential are solidified into rules. The prerequisite event sequence consists of event chains that precede in time and can trigger subsequent chain reactions, such as the sequence "abnormal electromagnetic signal → equipment offline alarm." The expected consequence event sequence consists of subsequent events that the prerequisite events may trigger. The event chain is structured as follows: "Network connection interruption → data transmission anomaly". When calculating the initial confidence of each causal rule based on historical event statistics, the relative frequency of occurrence is calculated by statistically analyzing the actual occurrence frequency of the event sequence described by the causal rule in historical data and comparing it with the total occurrence frequency of all similar event sequences. This frequency-based confidence calculation method can objectively reflect the reliability of the rule in history. When establishing the hierarchical organizational structure of the causal rule base, it is classified according to the severity of safety hazards and the occurrence scenario. For example, rules for fire safety, personal safety, and property safety are stored separately, and a multi-dimensional index relationship based on event type, occurrence scenario, and severity is established for each rule. This organizational method facilitates rapid retrieval and matching in the future. Finally, the completed causal rule base and its associated initial confidence values ​​are stored in a structured form to form a preset campus safety knowledge base that can be used by the event reasoning module.

[0082] By deeply mining the causal patterns hidden in historical monitoring data, scattered security incident cases are transformed into a predictive causal rule system. Extracting event sequences from historical cases and performing temporal correlation analysis reveals statistically significant event combination patterns with causal potential, avoiding biases that may arise from rule formulation based on subjective experience. Abstracting these event combination patterns into causal rules comprising precondition event sequences and expected consequence event sequences enables the system not only to identify existing security risks but also to predict potential subsequent risks. A mechanism for calculating initial confidence based on historical frequency assigns a quantitative reliability index to each rule, ensuring that the system prioritizes historically validated, highly reliable rules during inference. A hierarchical organizational structure and multi-dimensional indexing mechanism enable rapid identification of applicable causal rules in complex and ever-changing campus security scenarios, improving system response efficiency.

[0083] In another preferred embodiment of the present invention, the specific process of generating several safety hazard hypotheses, including a sequence of expected consequences and an initial confidence level, in the event reasoning module is as follows:

[0084] Iterate through all causal rules in the campus safety knowledge base and perform time-series matching of the prerequisite event sequence of each rule with the real-time underlying physical event stream;

[0085] When the event stream contains all the event types defined in the premise event sequence, and the time interval between the events meets the temporal constraints defined in the rule, and the spatial distribution satisfies the proximity condition defined in the rule, the rule is considered to be successfully matched.

[0086] For each successfully matched causal rule, a corresponding security risk hypothesis is created. The security risk hypothesis inherits the expected consequence event sequence and initial confidence level of the rule, resulting in several security risk hypotheses.

[0087] When traversing all causal rules in the campus safety knowledge base, a parallel processing approach is used to simultaneously compare the premise event sequences of multiple rules with the real-time underlying physical event stream. For example, in a playground activity scenario, the system will simultaneously match the premise event sequences of the "chasing and playing" rule (including events such as running fast, shouting loudly, and physical contact) and the premise event sequences of the "accidental fall" rule (including events such as sudden fall, cries of pain, and crowd gathering). When creating corresponding safety hazard hypotheses for each successfully matched causal rule, a complete attribute inheritance mechanism is used to completely copy the expected consequence event sequence defined in the rule into the newly generated hypothesis. At the same time, the initial confidence level of the rule obtained through historical data statistics is used as the initial confidence level of the hypothesis. For example, when the "playground conflict" rule is successfully matched, the generated hypothesis will completely include the expected consequence event sequence defined by the rule (including subsequent events such as conflict escalation, personnel injury, and crowd gathering) and inherit the initial confidence level of the rule based on historical statistics. Multiple hypotheses for different safety hazards are generated to form a complete hypothesis set, providing a foundation for subsequent analysis and verification.

[0088] This hypothesis generation mechanism, based on multi-rule parallel matching, fully utilizes the causal rule system in the knowledge base to achieve comprehensive coverage and accurate identification of typical campus safety scenarios. By employing parallel traversal of all causal rules, it ensures that no potential safety hazards are overlooked in complex campus environments. For example, in a playground scenario where there might be simultaneous risks of chasing and roughhousing and accidental falls, the system can generate corresponding safety hazard hypotheses for each. Strict temporal matching requirements guarantee that the causal relationships between events conform to actual development patterns, avoiding incorrect association of accidental and unrelated events. For instance, it ensures that the "chasing" event actually occurs before the "blocking" event, rather than being two independent events. Spatial proximity verification ensures that these events do indeed occur within a reasonable range of the same scenario, eliminating interference from unrelated events occurring simultaneously in different locations. By fully inheriting the expected consequence event sequence of the causal rules, each hypothesis contains a complete prediction of the potential development of the safety hazard, providing a clear monitoring target for subsequent verification. The mechanism of inheriting initial confidence levels allows the system to distinguish the reliability of different hypotheses, prioritizing high-confidence hypotheses that have been validated by numerous historical cases. This hypothesis generation method enables the system to accurately identify multiple potential security threats from complex campus activity scenarios and provides detailed development predictions and reliability assessments for each threat. This provides crucial decision-making support for accurate campus security early warnings and rapid, effective emergency response. By generating multiple parallel hypotheses, the system avoids the risk of misjudgment that may result from making a single judgment on complex campus security scenarios, greatly improving the comprehensiveness and accuracy of security supervision.

[0089] In another preferred embodiment of the present invention, the specific process for determining the dissimilar nodes in the data analysis module is as follows:

[0090] The expected consequences event sequences of each safety hazard hypothesis are aligned using a dynamic time warping algorithm, and the positions with different event types after sequence alignment are identified as candidate difference points.

[0091] One-hot encoding is performed on the event types corresponding to the candidate difference points to construct event type feature vectors, and the cosine similarity of the feature vectors of corresponding positions of adjacent sequences is calculated.

[0092] The Euclidean distance is calculated based on the spatial coordinates of the events corresponding to the candidate difference points, and the time difference is calculated based on the timestamp. The spatial distance and the time difference are weighted and fused into a spatiotemporal distribution distance.

[0093] The cosine similarity is converted into an event type difference, and the event type difference and the spatiotemporal distribution distance are combined according to a preset weight to generate a comprehensive difference index. The candidate difference point corresponding to the highest comprehensive difference index is selected and marked as a difference node.

[0094] First, a dynamic time warping algorithm is used to align the expected consequences of each safety hazard hypothesis with the event sequences. This method can handle sequences of different lengths and find the optimal alignment path. For example, when comparing the expected sequence of the "playground conflict" hypothesis (including pushing, arguing, and crowd gathering) with the expected sequence of the "accidental fall" hypothesis (including falling, calling for help, and crowd watching), the time axis is automatically adjusted to align similar events in time, thereby accurately identifying positions in the sequences where the event types are significantly different after alignment as candidate difference points. Next, one-hot encoding is performed on the event types corresponding to the candidate difference points, converting each event type into a binary vector with only one bit set to 1 and the rest to 0. This encoding method can eliminate the numerical differences between different event types, ensuring that each type is on an equal footing in the feature space. Finally, the cosine similarity of the feature vectors at corresponding positions in adjacent sequences is calculated. The degree of difference in event types is quantified by measuring the similarity of two vectors in direction. Simultaneously, Euclidean distance is calculated based on the spatial coordinates of the events corresponding to candidate difference points. This distance calculation accurately reflects the actual proximity of the two events. Time difference is calculated based on timestamps to assess the temporal proximity of the events. Then, the spatial distance and time difference are weighted and fused using preset weighting coefficients to form a comprehensive spatiotemporal distribution distance. This fusion method considers both spatial proximity and temporal continuity. Finally, cosine similarity is mathematically transformed to obtain the event type difference degree, which is then combined with the spatiotemporal distribution distance according to preset importance weights to generate a comprehensive difference degree index that fully reflects the degree of difference. By comparing the index values ​​of all candidate difference points, the candidate difference point with the largest value is selected as the most critical difference node.

[0095] In another preferred embodiment of the present invention, the specific process of generating the risk assessment value corresponding to each safety hazard hypothesis in the hypothesis verification module is as follows:

[0096] Increase the acquisition frequency of the corresponding sensing unit in the control sensor network to obtain a high-density verification event sequence;

[0097] Extract the attribute parameters contained in each event in the high-density verification event sequence. For object displacement events, the attribute parameters include motion direction, velocity and acceleration. For acoustic events, the attribute parameters include sound source type, loudness and pitch features. For human posture events, the attribute parameters include posture type and spatial distribution of feature points.

[0098] The attribute parameters of the verified event are compared with the corresponding attribute parameters of the expected event at the corresponding difference node position for each security hazard hypothesis;

[0099] The matching degree is calculated based on the comparison results of the attribute parameters. The current confidence level of each safety hazard hypothesis is used as the prior probability, and the calculated matching degree is used as the likelihood. The risk assessment value corresponding to each safety hazard hypothesis is calculated by Bayes' formula.

[0100] First, for the key difference nodes identified by the data analysis module, the corresponding sensing units in the sensor network are controlled to increase their acquisition frequency. For example, the frame rate of the surveillance camera is increased to a level that can clearly capture continuous motion, and the sampling rate of the microphone array is adjusted to a level that can distinguish speech details. This enhanced acquisition mode obtains a high-density verification event sequence containing richer temporal information. Then, detailed attribute parameters for each event are extracted from the verification event sequence. For object displacement events, the direction of motion is obtained by calculating the tangent direction of the trajectory, which reflects the object's intention to move. The speed of motion is obtained by calculating the displacement per unit time, which... To differentiate between normal walking and running, acceleration is analyzed by calculating the rate of change of velocity, which can identify abnormal behaviors such as sudden starts or stops. For acoustic events, sound source identification involves matching the spectral characteristics of the sound with typical sound source templates, distinguishing between different categories such as human voices and impact sounds. Loudness is measured by calculating the amplitude intensity of the sound wave, reflecting the magnitude of the sound. Pitch features are extracted by analyzing the fundamental frequency of the sound and its variation patterns, which can identify the emotional characteristics of speech. For human posture events, posture type identification involves matching the spatial configuration formed by connecting key points of the human body with typical posture patterns, which can distinguish between standing and standing postures. Analyzing the spatial distribution of feature points in different postures such as standing, running, and falling involves calculating the relative positional relationships of each joint, which reflects the specific details of limb movements. Then, the actual collected verification event attribute parameters are compared one by one with the expected event attribute parameters at the corresponding difference nodes for each safety hazard hypothesis. For example, when distinguishing between the hypotheses of "playground conflict" and "chasing and playing," the actual collected personnel movement speed and acceleration are compared with the expected movement characteristics of the two hypotheses for the difference node "personnel interaction behavior." Conflict scenarios are expected to exhibit movement patterns such as rapid changes of direction and sudden stops, while playful scenarios are expected to exhibit regular back-and-forth running. The dynamic model calculates the matching degree based on the comparison results of attribute parameters. It evaluates the degree of conformity between the actual observed value and the expected value of each attribute parameter, and performs weighted fusion by comprehensively considering the importance of each parameter to obtain an overall matching degree score. Finally, the current confidence of each safety hazard hypothesis is used as the prior probability of the hypothesis being true, and the calculated matching degree is used as the likelihood of observing the current verification evidence under the condition that the hypothesis is true. The prior probability and likelihood are combined by Bayesian probability inference formula to calculate the revised confidence of each hypothesis after incorporating new evidence. This process is essentially an objective update of the credibility of the original hypothesis based on the latest collected verification evidence.

[0101] By introducing actual observation data to dynamically evaluate and update initial hypotheses, the accuracy and reliability of safety hazard identification are effectively improved. Enhanced data collection targeting key discrepancies allows for the acquisition of richer and more detailed event characteristic information, providing sufficient evidence to accurately distinguish between different hypotheses. Extracting and analyzing event attribute parameters from multiple dimensions comprehensively captures the key characteristics of events, avoiding biases arising from single-feature judgments. Detailed comparison of actual observed attributes with hypothesized expected attributes accurately quantifies the difference between actual conditions and theoretical predictions, providing a reliable basis for probability calculations. For example, in the event of a large gathering on a playground, analyzing the movement characteristics of the participants can distinguish between conflict-related pushing and playful chasing; analyzing sound characteristics can differentiate between angry shouts and joyful laughter. This multi-feature fusion verification effectively improves identification accuracy. Through this rigorous verification and correction process, the system can dynamically adjust the confidence level of each security risk hypothesis based on actual evidence, gradually strengthening hypotheses that conform to the actual situation and weakening hypotheses that contradict the evidence. This ensures that the system can accurately identify real security threats, minimize false alarms and missed alarms, and improve the efficiency and intelligence level of campus security supervision.

[0102] In another preferred embodiment of the present invention, if there are two or more candidate difference points with equal comprehensive difference indices and all of them being the maximum value, then all of the candidate difference points are marked as difference nodes.

[0103] In another preferred embodiment of the present invention, the specific process of automatically triggering the standardized handling protocol bound to the safety hazard category in the result generation module is as follows:

[0104] Based on the safety hazard assumptions corresponding to the corrected confidence level exceeding the preset threshold, extract the hazard type identifiers bound to them;

[0105] Based on the hazard type identifier, query the predefined standardized handling protocol library to obtain the corresponding handling instruction sequence;

[0106] The sequence of processing instructions is compiled into an executable set of control instructions, which includes the target terminal address, execution time window, and action parameters.

[0107] The control instruction set is distributed to the corresponding execution terminal device through the campus communication network; the instruction confirmation signal returned by the execution terminal device is received, and the execution status record of the handling protocol is updated.

[0108] When the revised confidence level of a certain safety hazard hypothesis exceeds a preset threshold, the hazard type identifier bound to the hypothesis is first extracted. This identifier has been associated with each causal rule during the construction of the campus safety knowledge base, such as identifiers with clear semantics like "bullying incident" and "chasing danger." Based on this identifier, a predefined standardized handling protocol library is queried. This protocol library stores detailed handling procedures hierarchically according to hazard type. For example, the protocol for "bullying incident" includes a multi-step instruction sequence such as notifying nearby security personnel, starting on-site audio and video recording, and notifying the homeroom teacher. When the obtained handling instruction sequence is compiled into an executable control instruction set, it is then processed according to the campus Internet of Things (IoT). The device's communication protocols and interface specifications convert natural language-described instructions into control signals that the device can recognize. The target terminal address is determined through a campus device location database, the execution time window is set according to the urgency of the potential hazard, and the action parameters are configured according to specific handling requirements. The compiled control instruction set is distributed to the corresponding execution terminal devices via the campus communication network. Multicast technology from the campus IoT is used to send instructions to multiple relevant terminals simultaneously, and data encryption and authentication mechanisms ensure secure transmission. Upon receiving confirmation signals from the execution terminal devices, the system updates the execution progress of each instruction in the handling status record, establishing a complete handling tracking file. For example, when the system detects a bullying incident in a corner of the playground with a confidence level reaching a threshold, it automatically sends an "Immediately proceed to verify" instruction to the security patrol terminal in that area, an "Adjust angle to track and film" instruction to nearby cameras, and a "Pay attention to the student involved" notification to the relevant homeroom teacher's terminal, forming a multi-party collaborative handling network.

[0109] In another preferred embodiment of the present invention, the method further includes receiving an instruction confirmation signal and an on-site handling record returned by the execution terminal device; and comparing and verifying the on-site handling record with the safety hazard hypothesis that triggered the handling.

[0110] If the comparison and verification are consistent, a positive optimization sample containing the verification event sequence is generated, and the initial confidence of the corresponding causal rule in the campus safety knowledge base is improved based on the positive optimization sample.

[0111] If the comparison and verification are inconsistent, the verification event sequence is extracted and compared with the expected consequence event sequence of the causal rule corresponding to the safety hazard hypothesis. The difference nodes are identified, and negative judgment conditions are established based on the identified difference nodes.

[0112] The established negative judgment conditions are added to the expected consequence event sequence of the corresponding causal rule and re-injected into the campus safety knowledge base.

[0113] The system continuously receives instruction confirmation signals and on-site handling records returned by the execution terminal device. The instruction confirmation signals verify whether the handling instructions were correctly received and executed, while the on-site handling records contain a description of the actual situation confirmed by security personnel upon arrival. When comparing and verifying the on-site handling records with the safety hazard hypotheses that triggered the handling, natural language processing technology is used to parse key information in the handling records and perform a matching degree analysis with the hazard characteristics described in the hypotheses. For example, when the system triggers the "bullying incident" hypothesis due to the detection of an event sequence such as "running fast, shouting loudly, and physical contact," the on-site confirmation result by security personnel might be "students playing around" or "bullying behavior actually exists." The system will compare these two results with the original hypothesis for consistency. If the comparison verification is consistent, a positive optimization sample containing a complete verification event sequence is generated. This sample records the entire process data from the initial event occurrence to the handling confirmation. Based on these correctly verified cases, the initial confidence of the corresponding causal rules in the campus safety knowledge base is appropriately increased, with the increase based on... The rule is dynamically adjusted based on its historical accuracy. If the comparison and verification are inconsistent, the actual verification event sequence is extracted from the handling records and compared in detail with the expected consequence event sequence of the causal rule corresponding to the safety hazard assumption that triggered the handling. By analyzing the differences in multiple dimensions such as event type, spatiotemporal distribution, and intensity characteristics, key difference nodes that lead to misjudgment are identified. For example, the system expects "bullying incidents" to produce consequences such as "continuous crying and avoidance behavior," but what is actually observed are events such as "laughter and chasing, and normal social interaction." These differences are the difference nodes that need attention. Based on the identified difference nodes, negative judgment conditions are established in the expected consequence event sequence of the corresponding causal rule. These conditions serve as exclusion rules to help the system avoid repeated misjudgments when encountering similar scenarios in the future. For example, a negative condition is added to the "bullying incident" rule: "If laughter is detected and the movement trajectory is regularly cyclical, then the confidence of this rule is reduced." Finally, the optimized causal rule is re-injected into the campus safety knowledge base to complete the iterative update of the knowledge base.

[0114] This feedback optimization mechanism, based on practical verification, continuously compares actual handling results with system judgments, enabling the system to learn and continuously improve. Receiving confirmation signals and handling records from the execution terminal ensures the system obtains reliable feedback information, providing factual basis for system optimization. By comparing and verifying on-site handling records with safety hazard hypotheses, the system can objectively assess its own identification accuracy and promptly identify misjudgments. Generating positive optimization samples and increasing the confidence of corresponding rules allows correctly verified rules to play a greater role in similar future scenarios, improving the system's identification efficiency. When misjudgments occur, feature comparison between the actual verification event sequence and the expected consequence sequence accurately pinpoints the source of identification deviation, providing a clear direction for subsequent rule optimization. Establishing negative judgment conditions based on difference nodes essentially adds exclusionary clauses to the original rules, making them more rigorous and precise, effectively preventing the recurrence of similar misjudgments. Re-injecting the optimized rules into the knowledge base completes a closed loop from practice to theory and back to practice, enabling the system to continuously accumulate experience and improve its knowledge system. For example, when the system misclassifies playful roughhousing among students as bullying, analysis reveals that such roughhousing is usually accompanied by regular back-and-forth running and laughter, while genuine bullying often involves one-sided chasing and crying. The system then adds a corresponding negative condition to its bullying incident judgment rules, thereby improving the accuracy of future identifications. This significantly reduces the false alarm rate, making campus safety management more precise and efficient.

[0115] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An artificial intelligence-based campus safety hazard supervision system, characterized in that, The method comprises the following steps: a data acquisition module for continuously collecting raw data through a multi-modal sensor network deployed in a campus environment and converting the raw data into a bottom layer physical event stream, the bottom layer physical event stream comprising object displacement events, acoustic events and human posture events; an event reasoning module for matching the bottom layer physical event stream with preset campus safety knowledge base, the preset campus safety knowledge base comprising a premise event sequence, an expected consequence event sequence and an initial confidence degree, generating a plurality of safety hazard hypotheses comprising the expected consequence event sequence and the initial confidence degree; a data analysis module for determining a difference node according to the expected consequence event sequence contained in each safety hazard hypothesis, and selecting a key verification evidence type from the event types contained in the bottom layer physical event stream based on the difference node; The specific process of determining the difference node is as follows: aligning the expected consequence event sequences of each safety hazard hypothesis through a dynamic time warping algorithm, identifying the positions with different event types after sequence alignment as candidate difference points; one-hot encoding the event types corresponding to the candidate difference points, constructing an event type feature vector, and calculating the cosine similarity of the event type feature vectors of the positions corresponding to adjacent expected consequence event sequences; calculating the Euclidean distance based on the spatial coordinates of the events corresponding to the candidate difference points, calculating the time difference value based on the time stamps, and weighting and fusing the spatial distance and the time difference value into a spatio-temporal distribution distance; converting the cosine similarity into an event type difference degree, combining the event type difference degree and the spatio-temporal distribution distance according to a preset weight to generate a comprehensive difference degree index, and selecting the candidate difference point corresponding to the highest comprehensive difference degree index as the difference node; a hypothesis verification module for controlling the corresponding sensing unit to increase the collection frequency based on the key verification evidence type, obtaining a high-density verification event sequence, and generating a risk assessment value corresponding to each safety hazard hypothesis by comparing the time sequence consistency between the high-density verification event sequence and the expected consequence event sequence of each safety hazard hypothesis and combining the initial confidence degree of each safety hazard hypothesis; a result generation module for extracting the hazard type identifier bound to the safety hazard hypothesis corresponding to the corrected confidence degree exceeding the preset threshold when the risk assessment value corresponding to any safety hazard hypothesis exceeds the preset risk threshold; querying a predefined standardized disposal protocol library based on the hazard type identifier to obtain a corresponding disposal instruction sequence; compiling the disposal instruction sequence into an executable control instruction set, the control instruction set comprising a target terminal address, an execution time window and an action parameter; distributing the control instruction set to the corresponding execution terminal device through a campus communication network; receiving an instruction confirmation signal returned by the execution terminal device, and updating a disposal protocol execution state record.

2. The campus safety hazard monitoring system based on artificial intelligence according to claim 1, wherein, In the data acquisition module, the specific process of converting the raw data into the bottom layer physical event stream is as follows: performing continuous frame difference calculation on video stream data, analyzing the motion vector field of pixel points through an optical flow method, and generating an object movement event comprising motion direction, speed and acceleration characteristics when detecting that continuous motion vectors form a stable motion trajectory; The audio stream data is subjected to short-time Fourier transform to obtain a frequency spectrum, harmonic structure and energy distribution features in the frequency spectrum are extracted, and an acoustic event containing a sound source type, loudness and tone features is generated by matching a predefined acoustic feature template; Feature point detection is performed on the video stream data, and a dynamic posture event is generated when a structure change conforming to a preset mode is recognized by analyzing spatial relationship changes of feature points between consecutive frames, and a posture type and spatial distribution of feature points are recorded; A unified space-time coordinate system is established, coordinates of various events are converted to the coordinate system, and accurate time stamps and spatial position information are attached to each event; A density-based spatial clustering algorithm is used to aggregate events adjacent in space-time, redundant events are eliminated by calculating similarity of event feature vectors, and a standardized bottom-layer physical event stream is formed.

3. The campus safety hazard monitoring system based on artificial intelligence according to claim 1, wherein, In the event reasoning module, the specific construction process of the preset campus safety knowledge base is as follows: Typical safety event cases are extracted from historical campus monitoring records, time sequence correlation analysis is performed on continuous event sequences contained in each case, and event combination modes with causal relationships are identified; The identified event combination modes are abstracted into causal rule forms, each causal rule contains a premise event sequence and an expected consequence event sequence, the premise event sequence is composed of an initial event chain triggering a safety hazard, and the expected consequence event sequence is composed of a subsequent event chain triggered by the premise event; The initial confidence of each causal rule is calculated based on historical event statistical data, the initial confidence is obtained by calculating the proportion of the frequency of the causal rule in the historical data to the total frequency of all similar event sequences; A hierarchical organization structure of the causal rule base is established, the causal rules are classified and stored according to safety hazard types, and an index relationship is established for each rule; The constructed causal rule base and its associated initial confidence values are stored as the preset campus safety knowledge base.

4. The campus safety hazard monitoring system based on artificial intelligence according to claim 1, wherein, In the event reasoning module, the specific process of generating a plurality of safety hazard hypotheses containing expected consequence event sequences and initial confidences is as follows: All causal rules in the campus safety knowledge base are traversed, and the premise event sequence of each rule is time-matched with the real-time bottom-layer physical event stream; When it is detected that the event stream contains all event types defined by the premise event sequence, the time interval of event occurrence meets the time sequence constraints defined by the rule, and the spatial distribution meets the proximity conditions defined by the rule, it is determined that the rule is matched successfully; A corresponding safety hazard hypothesis is created for each matched causal rule, the safety hazard hypothesis inherits the expected consequence event sequence and the initial confidence of the rule, and a plurality of safety hazard hypotheses are obtained.

5. The campus safety hazard monitoring system based on artificial intelligence according to claim 2, characterized in that, In the hypothesis verification module, the specific process of generating risk assessment values corresponding to each safety hazard hypothesis is as follows: The acquisition frequency of the corresponding sensing unit in the sensor network is increased to obtain a high-density verification event sequence; extracting attribute parameters contained in each event in the high-density verification event sequence, the attribute parameters including motion direction and speed and acceleration for object displacement events, sound source type, loudness and tone characteristics for acoustic events, and posture type and feature point spatial distribution for human posture events; comparing the attribute parameters of the verification events with the corresponding attribute parameters of the expected events of each security risk hypothesis at the corresponding difference node positions; calculating a matching degree according to the comparison results of the attribute parameters, taking the current confidence of each security risk hypothesis as a prior probability and the calculated matching degree as a likelihood, and calculating a risk assessment value corresponding to each security risk hypothesis through a Bayesian formula.

6. The campus safety hazard monitoring system based on artificial intelligence according to claim 1, wherein, It also includes that if the comprehensive difference degree indexes corresponding to two or more candidate difference points are equal and are all maximum, then all the candidate difference points are marked as difference nodes.

7. The campus safety hazard monitoring system based on artificial intelligence according to claim 1, wherein, It also includes receiving an instruction confirmation signal and a field disposal record returned by the execution terminal device; comparing and verifying the field disposal record with the security risk hypothesis triggering the disposal; if the comparison and verification are consistent, generating a positive optimization sample containing the verification event sequence, and improving the initial confidence of the corresponding causal rule in the campus safety knowledge base based on the positive optimization sample; if the comparison and verification are inconsistent, extracting the verification event sequence and performing feature comparison with the expected consequence event sequence of the corresponding causal rule of the security risk hypothesis, identifying difference nodes, and establishing a negative judgment condition based on the identified difference nodes; adding the established negative judgment condition to the expected consequence event sequence of the corresponding causal rule and re-injecting the campus safety knowledge base.

Images