A campus bullying prevention intelligent early warning and evidence collection system

Abstract

The application relates to the technical field of campus bullying prevention, and discloses a campus anti-bullying intelligent early warning and evidence collection system, which comprises a data acquisition module, a data preprocessing module, a bullying risk grading module, an evidence encryption storage module and a graded early warning processing module. Through multi-source heterogeneous coupling layered architecture and split type multi-parameter wearable sensing, various bullying scenes can be covered, not only the explicit physical fighting can be identified, but also the implicit bullying such as whispering, cold violence and hidden invasion under the table can be captured, the collection blind area is avoided, the early warning and evidence collection result is matched with the real bullying scene, and objective judicial evidence is provided for the campus bullying dispute mediation.

Description

Technical Field

[0001] This invention relates to the field of school bullying prevention technology, and more specifically discloses a smart early warning and evidence collection system for school bullying prevention. Background Technology

[0002] With the deepening of educational informatization and campus safety governance, the problem of school bullying has evolved from the traditional perception of "minor squabbles between classmates" to complex behaviors including physical violence, verbal abuse, social isolation, and cyberbullying. It is characterized by its high degree of concealment, long duration, and difficulty in obtaining evidence. Implementing intelligent early warning and standardized evidence collection for school bullying aims to detect potential bullying trends as early as possible and intervene in a timely manner to prevent escalation. Simultaneously, by reconstructing the truth through a complete and tamper-proof chain of evidence, it provides reliable evidence for school handling, parent-school communication, and even legal accountability, thereby effectively protecting the rights of victims, holding perpetrators accountable, and fundamentally creating a safe and healthy campus environment.

[0003] In the prior art, document CN119559769B discloses "A Voice Early Warning System and Method for Addressing School Bullying", which includes: a data acquisition module for collecting audio data and point cloud data from school toilets; a data processing module for calculating decibel values ​​of the audio data and determining whether the decibel values ​​exceed a preset decibel threshold; and obtaining a 3D point cloud model and cubicle point cloud slices based on the point cloud data; and a bullying analysis module for performing bullying analysis based on the noise-reduced audio data, the 3D point cloud model, and the cubicle point cloud slices.

[0004] The document CN121121656A discloses "A Campus Intelligent Monitoring Method for Preventing Bullying on Campus". The method includes: acquiring audio and video data from campus surveillance, preprocessing and filtering suspected bullying segments at the edge; calculating an interaction anomaly index reflecting the asymmetry of physical conflict by analyzing the spatiotemporal changes of key human points in the video; extracting and analyzing words from the audio to obtain a semantic bullying index that assesses verbal threats; and dynamically weighting the above two indices according to the signal-to-noise ratio of the on-site environment to merge them into a more accurate comprehensive bullying index.

[0005] While existing technologies can provide bullying warnings in private areas such as restrooms through audio and point cloud analysis, or achieve behavior recognition and tiered push notifications by fusing audio and video features, achieving some success in reducing false alarms and protecting privacy, they rely solely on fixed-point audio and video and local environmental audio for information collection. They lack methods for collecting implicit data such as human physiology and plantar pressure. There are still blind spots in scenarios such as covert violence under tables and close-range whispered emotional abuse. Furthermore, it is difficult to achieve microsecond-level temporal alignment of various data, making it difficult to form a complete closed loop of evidence. Finally, the fixed recognition threshold is prone to imbalance and false alarms or missed detections. There is also a lack of tiered alarm and processing feedback archiving mechanisms. In addition, all raw terminal data is uploaded, resulting in high bandwidth and deployment costs. Summary of the Invention

[0006] The main technical problem solved by this invention is to provide an intelligent early warning and evidence collection system for preventing bullying on campus, which can solve the problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a campus anti-bullying intelligent early warning and evidence collection system, comprising: a data acquisition module, a data preprocessing module, a bullying risk classification module, an evidence encryption and storage module, and a graded early warning processing module; The data acquisition module collects bone conduction speech, human physiological signs, plantar pressure, and on-site audio and video information; The data preprocessing module performs time-series calibration and environmental noise removal on all types of heterogeneous collected data received. The bullying risk classification module receives the pre-processed data, performs fusion processing on multi-source collected features, dynamically adjusts the judgment criteria, and completes the bullying risk classification. The evidence encryption and storage module adopts a multi-sub-chain binding storage structure, classifies and archives different types of data, and synchronously sets up offline backup storage units. The tiered early warning processing module is equipped with multi-level alarm push links, collects on-site handling records for archiving and evidence preservation, and configures tiered data access permission control rules.

[0008] Furthermore, the data acquisition module includes: a voice acquisition module, a wrist physiological acquisition module, a foot pressure acquisition module, and a fixed-point audio acquisition module; Voice acquisition module: Collects mandibular bone conduction voice through piezoelectric bone conduction pickup components deployed on the neck strap, and collects close-range abusive and whispered verbal abuse voices; Wrist physiological data acquisition module: integrates skin conductance sensing element, pulse variability acquisition device, and micro-motion accelerometer to collect three types of physiological stress data: skin conductance fluctuation, short-term pulse change, and sudden limb tremor. Foot pressure acquisition module: It collects instantaneous impact pressure data of the sole of the foot caused by kicking, dragging and falling under the table through a thin film piezoelectric pressure array set in the heel; Fixed-point audio acquisition module: Acquires environmental audio and real-time surveillance video of public areas.

[0009] Furthermore, the data preprocessing module includes: a timing calibration module and an audio noise reduction decoupling module; Time series calibration module: performs microsecond-level timestamp anchoring and binding on multi-source heterogeneous data; Audio noise reduction and decoupling module: By collecting ambient noise data from fixed points across the entire domain, it removes ambient noise and extracts targeted speech features for verbal bullying from bone conduction speech.

[0010] Furthermore, the bullying risk classification module includes: a feature fusion module, a judgment criterion adjustment module, and a risk level classification module; Feature fusion module: performs multi-dimensional coupling mapping on three types of features: physiological data, bone conduction speech, and on-site video, and calculates bullying risk score based on feature overlap. Judgment Criterion Adjustment Module: Built-in physiological behavior benchmark library for primary and secondary school students, dynamically changes the weight of various features according to the school stage, and automatically updates the recognition threshold monthly based on the daily activity data of the class; Risk level classification module: It has a five-level classification standard, which distinguishes between normal activities, playful behavior during breaks, verbal and emotional abuse, minor physical bullying, and severe malicious bullying.

[0011] Furthermore, the evidence encryption and storage module includes: a categorized data archiving module, a data anti-tampering module, and a local offline storage module; The classification data archiving module: builds independent storage sub-chains for wearable sensor data, fixed-point audio and video data, and engine analysis data respectively; Data tamper prevention module: It adopts a multi-sub-chain cross-hash anchoring architecture. Multiple sub-chain blocks corresponding to a single event are bound to the global master hash. Sub-chain data modification will cause the master hash to become invalid. Local offline storage module: When the network is down or the cloud fails, it generates chained storage locally on the gateway and synchronizes it to the cloud main chain after the network is restored.

[0012] Furthermore, the tiered early warning processing module includes: an alarm push module, a handling record archiving module, and an access permission management module; Alarm push module: Sends warnings to homeroom teachers, moral education officers, campus security personnel, student guardians, and local education authorities according to risk level; The handling record archiving module receives the handling text and on-site evidence materials uploaded by staff. The receipt data is synchronously written to the corresponding event storage chain sub-chain. When the handling receipt is missing, the system will supervise the process level by level on a regular basis. Access control module: Viewing permissions are divided according to personnel identity. Parents can only view abnormal data related to their own children, and school administrators are assigned the scope of data viewing according to their positions.

[0013] Furthermore, the risk levels, from low to high, are as follows: Level 0: Normal daily activities; Level 1: Benign play during breaks; Level 2: Verbal isolation and emotional abuse; Level 3: Mild physical bullying; and Level 4: Severe malicious bullying.

[0014] The beneficial effects of this invention's intelligent early warning and evidence collection system for preventing school bullying are as follows: Through a multi-source heterogeneous coupling layered architecture and split-type multi-parameter wearable sensing, it can cover various bullying scenarios. It can not only identify overt physical fights, but also capture covert bullying such as whispered bullying and hidden attacks under desks, avoiding blind spots in data collection and ensuring that the early warning and evidence collection results are consistent with real bullying scenarios, providing objective judicial evidence for the mediation of school bullying disputes. In addition, through evidence storage and time-series calibration, it achieves unified solidification of the time sequence of multi-source evidence, breaking down the fragmented barriers of audio-visual and physiological data evidence, and improving the judicial acceptance validity of evidence materials and the accuracy of AI analysis. At the same time, through hierarchical linkage and closed-loop on-chain receipt, it provides hierarchical alarm and full-process traceability control methods, which can match the corresponding personnel as needed and quickly implement the handling after bullying incidents of different levels, greatly improving the efficiency of school bullying handling and the ability to trace and control responsibilities. Attached Figure Description

[0015] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0016] Figure 1 This is a schematic diagram of the system module architecture. Detailed Implementation

[0017] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0018] According to one aspect of the invention, such as Figure 1 As shown, a campus anti-bullying intelligent early warning and evidence collection system is provided, including: a data acquisition module, which collects bone conduction speech, human physiological signs, plantar pressure, and on-site audio and video information. This module includes: Voice acquisition module: Collects mandibular bone conduction voice through piezoelectric bone conduction pickup components deployed on the neck strap, and collects close-range abusive and whispered verbal abuse voices; Specifically, the piezoelectric bone conduction pickup unit is fixedly mounted close to the user's chin skin. The hardware structure is independently made into a neck-mounted split component. This component and the two acquisition hardware at the wrist and heel are interconnected through low-power wired lines to form a complete acquisition terminal. The entire three-part hardware uses a custom-developed micro synchronous clock chip to uniformly control the sampling timing. The pickup unit relies on the physical isolation structure design of the hardware to block the ambient sound waves propagating in the air. It only receives the sound source signal transmitted through the medium of the human mandibular bone. From the hardware structure, it avoids the problem of environmental noise such as classroom noise and outdoor wind noise mixing into the effective speech, and achieves accurate pickup of close-range whispers and face-to-face insults. It also includes the addition of a miniature environmental noise auxiliary acquisition patch inside the neckband device. The patch and the bone conduction pickup unit are synchronously controlled by a unified clock chip to carry out sampling operations, and synchronously collect raw data of ambient background noise. This noise sampling data is packaged together with the raw bone conduction speech data packet and transmitted down to the back-end preprocessing stage. It provides a supporting reference data source for the subsequent speech feature decoupling and splitting work at the gateway end, and is synchronously included in the local pre-screened data pool of the entire terminal to participate in the initial data identification.

[0019] Wrist physiological data acquisition module: integrates skin conductance sensing element, pulse variability acquisition device, and micro-motion accelerometer to collect three types of physiological stress data: skin conductance fluctuation, short-term pulse change, and sudden limb tremor. Three types of sensing components are encapsulated in a sealed environment within the wrist module. All hardware wiring is connected to the built-in terminal MCU main control chip and is uniformly scheduled and controlled by a micro synchronous clock chip shared by the entire wearable terminal. The data acquisition work is started synchronously with the neckband voice module and the shoe heel pressure module according to the predetermined sampling rhythm of 125µs staggered time. The three types of physiological parameters can be captured synchronously at the same sampling time node, avoiding the problem of misaligned sampling time and disordered data sequence of multiple types of physiological data from the source of collection. During normal operation, the wrist-mounted MCU continuously records the student's normal physiological data when there are no abnormalities. This data is automatically aggregated to generate a user-specific physiological baseline reference value. The fluctuation range of various physiological indicators is quantified using the following formula:

[0020] In the formula, The standard deviation of the sampled physiological data. To accurately distinguish between the gradual physiological fluctuations caused by students' daily play and free running during breaks and the sudden and drastic changes in physiological parameters caused by verbal intimidation and physical bullying, the study relies on real-time dynamic comparison of the change range of the baseline data to make a preliminary judgment of stress abnormalities. In addition, the full amount of raw physiological data collected from the wrist is first pre-screened in the local MCU. The system automatically filters out invalid data caused by students running, laughing and playing during normal breaks based on preset physiological parameter smooth thresholds. Only suspected abnormal physiological data that exceeds the threshold range is encrypted and packaged. The remaining compliant and normal daily physiological data is directly overwritten and cleared in the local storage area of ​​the terminal without occupying uplink transmission bandwidth.

[0021] Foot pressure acquisition module: It collects instantaneous impact pressure data of the sole of the foot caused by kicking, dragging and falling under the table through a thin film piezoelectric pressure array set in the heel; First, the thin-film piezoelectric pressure array is completely embedded in the reserved cavity in the heel of the student shoe using a fully sealed embedded packaging process. The signal wiring led out from the array is connected to the wired interconnection line of the wrist module through the concealed wiring inside the shoe body. Finally, it is connected to the synchronous clock control system shared by the entire terminal. It relies on unified clock commands to achieve synchronous start and stop sampling with the other two separate modules, ensuring that the sampling timing of pressure data, voice, and physiological data is completely aligned. Then, the pressure array continuously captures the normal pressure values ​​of the soles of the feet when students are standing, walking slowly, or jumping normally. The terminal MCU continuously summarizes and generates the reference pressure range of the user's daily soles. Once there are abnormal force actions such as hidden kicks under the table, being dragged to the ground by external force, or sudden tripping, the pressure array immediately captures the instantaneous peak impact pressure and generates exclusive abnormal load raw data. Finally, the generated raw pressure data is synchronously imported into the terminal's local pre-screening process. It is then matched and verified locally with the physiological and voice data collected at the same time within the MCU. Only pressure data that is accompanied by abnormal physiological stress or abnormal close-range abusive voice will be judged as suspicious bullying-related data and uploaded separately to the backend gateway in encrypted form. Foot pressure data generated by normal walking or jumping in place will be cyclically erased and retained on-site and will not be transmitted upwards.

[0022] Fixed-point audio acquisition module: Acquires environmental audio and real-time surveillance video of public areas; Specifically, the entire campus is pre-divided into grid areas according to the spatial boundaries of classrooms, corridors, restrooms, playgrounds, and stairwells. Within each independent grid area, binocular panoramic vision cameras are deployed at fixed points, along with spatial array environmental acoustic acquisition boards to form fixed-point acquisition units. Both types of acquisition hardware are connected to the dedicated edge coupling gateway hardware of this grid through wired lines. All fixed-point acquisition points follow the grid division to achieve comprehensive control of key areas of the campus. Next, the fixed-point unit continuously collects ambient background noise data and real-time video footage within its grid area around the clock. The collected ambient noise data is stored separately and reserved for subsequent gateway-side removal of residual noise mixed in bone conduction speech. The video data stream follows the gateway's self-developed timing calibration logic and is then time-anchored and bound to the multi-source sensor data returned by the wearable terminal before entering the preprocessing link.

[0023] The data preprocessing module performs time-series calibration and environmental noise removal on all received heterogeneous data. This module includes: Time series calibration module: performs microsecond-level timestamp anchoring and binding on multi-source heterogeneous data; Firstly, the edge coupling gateways configured in each grid partition rely on the self-developed dedicated timing calibration chip built into the device to carry out hardware-level clock synchronization, avoiding the NTP general network time synchronization method used by traditional security NVR devices. All physiological, pressure, and bone conduction voice data packets, real-time video streams from grid binocular cameras, and environmental noise data collected by spatial array pickup devices uploaded by wearable terminals are connected to the gateway's dedicated data receiving channel. The hardware chip uniformly captures the native sampling time information of each data source, using the sampling benchmark of the wearable terminal's self-developed synchronization clock as a unified time scale for the entire domain. Then, the chip, referring to the inherent working rhythm of the wearable terminal's 125µs staggered time-sharing sampling, corrects the time difference caused by the clock drift of different hardware components and the delay of wired transmission lines, and smooths out the timing misalignment between the wearable device and the fixed-point acquisition device one by one. Finally, it achieves ±1µs accuracy timestamp anchoring and binding of heterogeneous data collected from all categories. All bound data are uniformly attached with the same global time identifier to ensure that the time dimension of multi-source evidence is completely consistent. After the final timing binding is completed, all data is classified and stored according to the gateway's three-level tiered caching rules. Abnormal data marked as suspected bullying is directly written to the instantaneous high-speed latch cache to achieve hardware locking and cannot be automatically overwritten by the system. Normally collected data is stored in a 15-minute cyclic temporary cache. In the event of a sudden power outage, all data in the cache is automatically written to the offline Flash backup block. After power is restored, the gateway actively reads the backup data and transmits it upwards according to the established link.

[0024] Audio noise reduction and decoupling module: By collecting ambient noise data from fixed points across the entire domain, it removes ambient noise and extracts targeted speech features for verbal bullying from bone conduction speech. First, the ambient background noise data collected in real time by the acoustic acquisition board of the same grid fixed-point array is retrieved. At the same time, the local ambient noise material collected synchronously by the noise auxiliary patch built into the neck bone conduction module is matched. The two sets of noise data and the original bone conduction speech data after time-series calibration are aggregated and collected locally on the gateway. The gateway only performs feature decomposition calculation and does not perform full encoding and forwarding of the original audio and video. Then, targeted feature decoupling processing logic is run locally on the gateway. Based on environmental noise reference data, irrelevant background noise mixed in with bone conduction audio, such as group chatter during breaks, outdoor traffic noise, and large-scale group play, is separated layer by layer using the following formula:

[0025] In the formula, The effective human voice power for bone conduction pickup, This represents the total noise power in the ambient space. The skeletal sound transmission coupling correction coefficient is used to extract effective speech features from the mixed audio, such as students verbally abusing each other at close range and isolating each other through whispered conversations. Finally, the effective voice features after noise reduction, decoupling and purification are integrated and packaged with the time-series unified physiological data, plantar pressure data and fixed-point monitoring screen data into a pre-processed data packet. The data packet is sent to the bullying risk classification module through the system's self-developed private encryption protocol. The invalid environmental noise data after disassembly and separation is directly overwritten and cleared in the local cache of the gateway, and no longer transmitted upwards to waste cloud computing power and communication bandwidth.

[0026] The bullying risk grading module receives pre-processed data, fuses features collected from multiple sources, dynamically adjusts the judgment criteria, and completes the bullying risk grading. This module includes: Feature fusion module: performs multi-dimensional coupling mapping on three types of features: physiological data, bone conduction speech, and on-site video, and calculates bullying risk score based on feature overlap. First, the system receives the integrated data packet transmitted from the edge gateway. The system then decomposes and classifies the feature fields of the data packet, including the physiological stress features after time alignment and noise reduction, the bullying voice features after removing noise, and the human body image features extracted by the camera. The system then transforms the feature parameters from the three different sources into standardized feature vectors that can be matched and calculated. A three-dimensional feature mapping space is built inside the engine, and the three types of feature parameters are loaded into the three-dimensional coordinates of the space respectively. Then, the system searches for overlapping points in the three-dimensional space where the three types of features are abnormal at the same time. When a single dimension has an abnormal value or two or more dimensions deviate from the normal baseline range at the same time, the system accumulates the risk score accordingly. Even if there are only abnormal physiological data and abusive voices, and the monitoring screen has no effective violent scenes, the system can still identify hidden bullying behavior without images by relying on the feature overlap results, filling the gap that pure video recognition cannot determine hidden bullying. After the final feature overlap calculation is completed, the generated real-time risk score is temporarily stored in the dedicated runtime cache of the judgment engine. The score data is synchronously passed down to the lower-level judgment standard adjustment module. The original three types of feature source data are archived and cached for later use, and are not directly pushed out to avoid the risk of leakage of original privacy data.

[0027] Judgment Criterion Adjustment Module: Built-in physiological behavior benchmark library for primary and secondary school students, dynamically changes the weight of various features according to the school stage, and automatically updates the recognition threshold monthly based on the daily activity data of the class; The module pre-stores three independent sets of student physiological behavior benchmark sample data for primary school, junior high school, and senior high school. When the system enters student registration information, it simultaneously binds the corresponding grade level tag. The engine automatically retrieves the corresponding grade level benchmark library content based on the tag and synchronously and adaptively adjusts the weight ratio of the three types of features (physiological, voice, and video) in the risk score calculation process. For primary school students, the weight of physiological data is automatically reduced and the weight of body image features is increased. For junior and senior high school students, the weight of voice and skin conductance physiological features is gradually increased and the weight of body movement judgment is reduced. The system automatically aggregates pre-processed daily data from all classes across the entire grade every month. The aggregated data includes the normal physiological fluctuation range of all students in the class, the characteristics of their voices during breaks, and the overall frequency of physical activities in the class. Based on the aggregated data, a unique activity baseline is generated for each corresponding class. For lively classes that are active and frequently playful, the threshold for recognizing playful behavior is automatically raised, while for classes that are usually quiet and reserved and have little physical interaction, the threshold for judging verbal abuse is automatically lowered. No manual input or modification of parameters is required throughout the entire process.

[0028] Risk level classification module: It has a five-level classification standard, which distinguishes between normal activities, playful behavior during breaks, verbal and emotional abuse, minor physical bullying, and severe malicious bullying. Specifically, the module pre-fixes the five-level risk classification judgment logic from 0 to 4. After receiving the final risk score output from the upstream, the system compares and matches the real-time score with the score range of each level one by one to automatically complete the risk classification. Level 0 corresponds to students having no abnormal daily activities and Level 1 is defined as benign playful behavior during breaks. The above two levels do not trigger any immediate warning actions. Level 2 corresponds to verbal isolation and whispered emotional abuse without physical contact; Level 3 corresponds to mild physical bullying such as pushing and tearing; and Level 4 corresponds to severe malicious bullying such as surrounding and beating. The medium-to-high risk classification results of Levels 2, 3, and 4 will generate exclusive event identifiers simultaneously. After binding with triple time anchor information, they will be pushed to the backend evidence encryption and storage module to start the evidence storage process. The classification results will be simultaneously transmitted to the classification early warning processing module to start the corresponding level alarm push.

[0029] The evidence encryption and storage module employs a multi-sub-chain binding storage structure to classify and archive different types of data, and simultaneously sets up offline backup storage units. This module includes: The classification data archiving module: builds independent storage sub-chains for wearable sensor data, fixed-point audio and video data, and engine analysis data respectively; The system receives the risk rating results from the bullying risk rating module and the full amount of evidence data preprocessed by the edge gateway. It isolates and splits the data according to the data source attributes. The skin conduction, pulse, bone conduction voice, and plantar pressure uploaded by the wearable terminal are all included in the wearable sensor data sub-chain and built into a separate block. The grid binocular camera video and the environmental audio collected by the spatial array are uniformly stored in the fixed-point audio-visual data sub-chain. The five-level risk scores, rating judgment conclusions and other judgment information are independently built into the engine judgment conclusion sub-chain. The three types of data of the same bullying event are marked with the same event code and are stored in the chain respectively. Level 0 and Level 1 data without abnormalities are only compressed and silently archived without generating judicial evidence blocks. Level 2 to 4 medium and high risk data are automatically based on the wearable sampling time, gateway time synchronization time and judgment landing time to form a triple time anchor, and are then encapsulated into judicial evidence data packets before being written into the corresponding sub-chain block. The preliminary preparation of sub-chain blocks is completed locally at the corresponding grid edge gateway. The gateway performs data classification and data packet encapsulation locally before uploading the data in segments to the cloud to aggregate it into a global chain. The cloud only performs the aggregation and integration of sub-chains across the entire campus, avoiding the instantaneous centralized uploading of massive amounts of original evidence that would cause a surge in cloud storage resources. Each sub-chain adds blocks in strict accordance with the order of events, and the block generation order cannot be manually inserted, deleted, or changed.

[0030] Data tamper prevention module: It adopts a multi-sub-chain cross-hash anchoring architecture. Multiple sub-chain blocks corresponding to a single event are bound to the global master hash. Sub-chain data modification will cause the master hash to become invalid. First, for the three sub-chains corresponding to the blocks generated synchronously in a single bullying event, a unique hash identifier for each sub-chain block is generated during the block encapsulation and finalization stage using the following formula:

[0031] In the formula, For standard hash compression permutation functions, The hash value of the previous block. The original plaintext data of the current block is used as the basis for the unified calculation of the hash values ​​of the three sub-chains to generate a unique global master hash. The global master hash is synchronously written back to the corresponding event blocks of the three sub-chains, so that the three independent sub-chains can be cross-locked and bound through a single master hash. Once the content of any block in a sub-chain is modified or fields are deleted, the original hash value of that block will change immediately, and the global master hash synchronization verification will fail. The system will automatically retain hash anomaly records and lock the location of the abnormal block. It is impossible to destroy the integrity of the entire evidence chain by tampering with a single type of audio or video or physiological data, thus meeting the hard requirements of judicial evidence collection to prevent tampering. Finally, after all the cross-anchored blocks are uploaded to the chain, an independent evidence storage operation log is generated synchronously. The log is stored separately from the three types of business sub-chains. When judicial authorities retrieve case files, they can use the log and the main hash to double verify the originality of the evidence across the entire chain.

[0032] Local offline storage module: When the network is down or the cloud fails, it generates chained storage on the gateway and synchronizes it to the cloud main chain after the network is restored; Specifically, each mesh-coupled gateway hardware pre-defines a dedicated Flash offline storage partition. When communication failures occur, such as fiber optic cable interruption in the park, cloud server crash, or temporary power outage in the area, the gateway automatically disconnects the cloud interaction link and switches to local evidence storage operation mode. The entire evidence storage logic, including categorized chain building and three-chain hash cross-anchoring, runs completely and independently in the gateway's local hardware. All evidence generated during the failure period is still generated into local temporary three-chain sub-chains as usual, and there is no evidence loss due to network outage. After the network communication failure is resolved and the link is restored, the gateway automatically triggers the breakpoint resume transmission program, and uploads the local offline sub-chain data to the global chain in the cloud in batches, strictly following the order of block generation. After synchronization, the system automatically compares the local sub-chain hash with the cloud-uploaded hash. If the verification is consistent, the archiving is completed. Local offline backup data can be automatically cleaned up in a preset cycle, and there is no need for maintenance personnel to manually export backup data throughout the process.

[0033] The tiered early warning and processing module sets up multi-level alarm push links, collects on-site handling records for archiving and evidence preservation, and configures tiered data access permission control rules. This module includes: Alarm push module: Sends warnings to homeroom teachers, moral education officers, campus security personnel, student guardians, and local education authorities according to risk level; First, after receiving the risk level classification results (0-4) from the risk level classification module, the system automatically matches the preset graded push strategy. Level 0 daily activities and Level 1 benign playfulness do not trigger instant message pushes. Level 0 data is directly archived and saved silently. Level 1 data is only summarized into a class hazard ledger at the end of the day and sent to the corresponding class teacher's system account. Level 2 verbal bullying triggers a real-time alert via an app pop-up for the class teacher and the grade's moral education officer. Level 3 mild bullying simultaneously activates three push channels: a 5-minute time limit reminder is sent to the campus security terminal, and SMS is sent to the class teacher and the guardian of the student involved. Level 4 severe bullying simultaneously executes a multi-channel notification mechanism: emergency dispatch instructions from the security terminal, direct phone calls from the school leader in charge, voice calls from the guardian, and automatic filing and uploading by the local education bureau's supervision backend. Each level of warning information is accompanied by an event number and a triple time anchor evidence index, making it easy for recipients to quickly retrieve the corresponding evidence data. Then, different levels of warnings adopt differentiated transmission channels. Low-level messages rely on internal system applications for push, while medium and high-risk warnings reuse dual-channel transmission of operator SMS and voice dedicated lines to avoid the problem of warning omissions caused by single communication link failure. All push records generate push logs in real time and are temporarily cached. The logs are subsequently entered into the evidence storage link along with the processing data. Finally, the push link and the self-developed private protocol of the five-layer architecture are connected. The warning key information is sent through private encrypted messages, while ordinary ledger messages can be transmitted using general TCP, which takes into account both data security and transmission efficiency.

[0034] The handling record archiving module receives the handling text and on-site evidence materials uploaded by staff. The receipt data is synchronously written to the corresponding event storage chain sub-chain. When the handling receipt is missing, the system will supervise the process level by level on a regular basis. Specifically, after receiving the corresponding level of warning, the relevant personnel will complete the on-site handling work and need to enter the handling process, on-site photos and other supporting materials into the corresponding event entry in the system to form a handling receipt data packet. The receipt data packet carries the original bullying event's unique code, and the system will automatically match the corresponding three independent sub-chains in the evidence encryption and storage module based on the code. Next, the compliant and complete receipt data is encrypted and directly written into the sub-chain block corresponding to the event, and simultaneously participates in the global master hash anchoring and binding, so that the handling process information and the original bullying evidence form a complete closed-loop evidence chain. If no receipt is entered after the set time limit after the corresponding warning is issued, the system will automatically send a supervision reminder to the next higher level of management every 2 hours. If the class teacher fails to handle the matter within the time limit, it will be pushed to the moral education department. If the moral education department fails to supervise and handle the matter, it will continue to report to the school-level person in charge. The supervision at each level is recorded in the whole process.

[0035] Access control module: Viewing permissions are divided according to personnel identity. Parents can only view abnormal data related to their own children, and school administrators are assigned the scope of data viewing according to their positions. Specifically, all platform users are pre-assigned access tags based on six identity categories: parents, homeroom teachers, moral education officers, security personnel, school leaders, and education bureau administrators. A data access isolation mechanism is then implemented in the backend based on these tags. After parents log in, the system automatically filters out abnormal evidence fragments from the database that are only linked to their own child's ID number, and automatically blocks the physiological and audio / video privacy data of other students in the same class. School administrators at various levels define the scope of grade and class data they can access according to their job responsibilities. Security personnel can only view information related to their own incidents, while school administrators have access to data across their entire jurisdiction. All data access operations generate access logs and are backed up in the corresponding system to prevent unauthorized access to or leakage of student privacy information.

[0036] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention are also within the protection scope of the present invention.

Claims

1. A campus anti-bullying intelligent early warning and evidence collection system, characterized in that, include: Data acquisition module, data preprocessing module, bullying risk classification module, evidence encryption and storage module, and graded early warning processing module; The data acquisition module collects bone conduction speech, human physiological signs, plantar pressure, and on-site audio and video information; The data preprocessing module performs time-series calibration and environmental noise removal on all types of heterogeneous collected data received. The bullying risk classification module receives the pre-processed data, performs fusion processing on multi-source collected features, dynamically adjusts the judgment criteria, and completes the bullying risk classification. The evidence encryption and storage module adopts a multi-sub-chain binding storage structure, classifies and archives different types of data, and synchronously sets up offline backup storage units. The tiered early warning processing module is equipped with multi-level alarm push links, collects on-site handling records for archiving and evidence preservation, and configures tiered data access permission control rules.

2. The campus anti-bullying intelligent early warning and evidence collection system according to claim 1, characterized in that: The data acquisition module includes: a voice acquisition module, a wrist physiological acquisition module, a foot pressure acquisition module, and a fixed-point audio acquisition module; Voice acquisition module: Collects mandibular bone conduction voice through piezoelectric bone conduction pickup components deployed on the neck strap, and collects close-range abusive and whispered verbal abuse voices; Wrist physiological data acquisition module: integrates skin conductance sensing element, pulse variability acquisition device, and micro-motion accelerometer to collect three types of physiological stress data: skin conductance fluctuation, short-term pulse change, and sudden limb tremor. Foot pressure acquisition module: It collects instantaneous impact pressure data of the sole of the foot caused by kicking, dragging and falling under the table through a thin film piezoelectric pressure array set in the heel; Fixed-point audio acquisition module: Acquires environmental audio and real-time surveillance video of public areas.

3. The intelligent early warning and evidence collection system for campus anti-bullying according to claim 1, characterized in that: The data preprocessing module includes: a timing calibration module and an audio noise reduction and decoupling module; Time series calibration module: performs microsecond-level timestamp anchoring and binding on multi-source heterogeneous data; Audio noise reduction and decoupling module: By collecting ambient noise data from fixed points across the entire domain, it removes ambient noise and extracts targeted speech features for verbal bullying from bone conduction speech.

4. The intelligent early warning and evidence collection system for campus anti-bullying according to claim 1, characterized in that: The bullying risk classification module includes: a feature fusion module, a judgment criterion adjustment module, and a risk level classification module; Feature fusion module: performs multi-dimensional coupling mapping on three types of features: physiological data, bone conduction speech, and on-site video, and calculates bullying risk score based on feature overlap. Judgment Criterion Adjustment Module: Built-in physiological behavior benchmark library for primary and secondary school students, dynamically changes the weight of various features according to the school stage, and automatically updates the recognition threshold monthly based on the daily activity data of the class; Risk level classification module: It has a five-level classification standard, which distinguishes between normal activities, playful behavior during breaks, verbal and emotional abuse, minor physical bullying, and severe malicious bullying.

5. The intelligent early warning and evidence collection system for campus anti-bullying according to claim 1, characterized in that: The evidence encryption and storage module includes: a categorized data archiving module, a data anti-tampering module, and a local offline storage module; The classification data archiving module: builds independent storage sub-chains for wearable sensor data, fixed-point audio and video data, and engine analysis data respectively; Data tamper prevention module: It adopts a multi-sub-chain cross-hash anchoring architecture. Multiple sub-chain blocks corresponding to a single event are bound to the global master hash. Sub-chain data modification will cause the master hash to become invalid. Local offline storage module: When the network is down or the cloud fails, a chain of storage is generated locally on the gateway, and synchronized to the cloud main chain after the network is restored.

6. The intelligent early warning and evidence collection system for campus anti-bullying according to claim 1, characterized in that: The hierarchical early warning processing module includes: an alarm push module, a handling record archiving module, and an access control module; Alarm push module: Sends warnings to homeroom teachers, moral education officers, campus security personnel, student guardians, and local education authorities according to risk level; The handling record archiving module receives the handling text and on-site evidence materials uploaded by staff. The receipt data is synchronously written to the corresponding event storage chain sub-chain. When the handling receipt is missing, the system will supervise the process level by level on a regular basis. Access control module: Viewing permissions are divided according to personnel identity. Parents can only view abnormal data related to their own children, and school administrators are assigned the scope of data viewing according to their positions.

7. The intelligent early warning and evidence collection system for campus anti-bullying according to claim 6, characterized in that: The risk levels, from low to high, are as follows: Level 0: Normal daily activities; Level 1: Benign play during breaks; Level 2: Verbal isolation and emotional abuse; Level 3: Mild physical bullying; Level 4: Severe malicious bullying.

Images