School violence prevention model and its construction method

Abstract

This invention relates to the field of campus safety management technology, and discloses a campus violence prevention model and its construction method. The construction method includes the following steps: Step 1, acquiring environmental parameters, physiological indicators, and behavioral characteristics in real time as basic data through a multimodal data acquisition device. The environmental parameters include concentration; the physiological indicators include body temperature fluctuations; and the behavioral characteristics include limb movements and voiceprint data. Step 2, preprocessing the basic data and generating a fused feature vector. Step 3, training a bullying behavior prediction model based on a federated learning platform using a support vector machine and neural network fusion algorithm; and outputting a behavioral risk value from the bullying behavior prediction model. This invention can achieve timely and accurate detection and early warning of risky behaviors, with strong objectivity and real-time performance, contributing to the construction of a zero-violence, high-efficiency smart campus.

Description

Technical Field

[0001] This invention relates to the field of campus safety management technology, specifically to a campus violence prevention model and its construction method. Background Technology

[0002] In campus safety management, the issue of school bullying has always been a key focus of prevention and control. According to relevant research, school bullying often leads to psychological problems such as depression and anxiety in victims. The timeliness of prevention and detection of school bullying incidents, and the accuracy of handling and judgment, are directly related to the healthy growth of adolescents and the positive development of society.

[0003] Currently, the prevention and control of school bullying mainly relies on three traditional methods: education and publicity, collaboration between families and schools, and system optimization. However, due to limitations in technological development and inherent flaws in the mechanisms, existing prevention and control methods have several shortcomings: First, most existing methods depend on physical surveillance cameras installed on campus, which have many blind spots and cannot comprehensively cover every corner of the campus. Moreover, relying solely on capturing and judging from surveillance footage requires a significant amount of manpower and resources, has a high false alarm rate, cannot cover multi-dimensional scenarios, and the response is not timely or accurate enough. Second, most existing prevention and control methods are in a "passive response" mode, with a long delay in intervention, only allowing for accountability after a bullying incident occurs, and failing to provide second-level early warning. Summary of the Invention

[0004] The present invention aims to provide a campus violence prevention model and its construction method, which can realize timely and accurate risk behavior detection and early warning, with strong objectivity and real-time performance, and helps to build a zero-violence, high-efficiency smart campus.

[0005] To achieve the above objectives, the present invention provides the following basic solution.

[0006] Option 1

[0007] The method for constructing a school violence prevention model includes the following steps:

[0008] Step 1: Real-time acquisition of environmental parameters, physiological indicators, and behavioral characteristics as basic data using multimodal data acquisition equipment. The environmental parameters include CO2 concentration; the physiological indicators include body temperature fluctuations; and the behavioral characteristics include limb movements and voiceprint data. Step 2: Preprocess the basic data and generate a fused feature vector; Step 3: Based on the federated learning platform, train the bullying behavior prediction model using a fusion algorithm of support vector machine and neural network; the bullying behavior prediction model outputs the behavior risk value. Furthermore, the preprocessing includes: using wavelet transform denoising and moving average filtering to process CO2 concentration data, using Kalman filtering to process body temperature fluctuation data, using spectral subtraction denoising to process voiceprint data; and performing feature extraction. Furthermore, the features extracted include: environmental features, physiological features, and behavioral features; the environmental features include CO2 concentration change rate and spatial air mass retention index; the physiological features include body temperature fluctuations; the behavioral features include the frequency of target behavior and voiceprint keyword hit rate; the target behavior includes pushing, blocking, and pulling actions; and the voiceprint keywords include abusive and threatening words.

[0009] Furthermore, the generation of the fused feature vector includes: normalizing each feature and concatenating them into a vector, dynamically allocating weights through an attention mechanism, and then generating a fused feature vector with uniform dimensions for subsequent model input.

[0010] Furthermore, it also includes step 4, which dynamically adjusts the early warning threshold of the basic data based on the venue capacity, ventilation efficiency and real-time collected multimodal basic data of the analyzed area; when any data in the basic data exceeds the corresponding early warning threshold, or the behavioral risk value reaches the corresponding risk threshold, a multi-level response mechanism is triggered.

[0011] Furthermore, the behavioral risk value ranges from 0 to 1; the multi-level response mechanism includes: when the behavioral risk value is 0.6 to 0.8, a level 1 warning is triggered, linking the internal campus system to activate LCD screen alerts and directional broadcasts; when the behavioral risk value is greater than 0.8, a level 2 warning is triggered, linking the internal campus system to activate buzzers and 5G NR technology to push the warning information to the security terminal with low latency, and simultaneously adjusting the campus fresh air system to manage carbon concentration.

[0012] Furthermore, the bullying behavior prediction model includes an input layer, a feature extraction layer, and a classification output layer; the feature extraction layer includes a DNN branch, an LSTM branch, and a feature fusion module; the classification output layer includes an SVM classifier.

[0013] Furthermore, the pretreatment also includes CO2 concentration compensation, temperature compensation, and light compensation;

[0014] The CO2 concentration compensation includes: synchronously collecting ambient temperature, ambient humidity, and CO2 concentration per unit time, and correcting the CO2 concentration through polynomial regression; the polynomial includes: CO 2_校正 =CO 2_实时 ×(1+α×ΔT+β×ΔH), where α and β are calibration coefficients; CO 2_校正 The corrected CO2 concentration, CO 2_实时The concentration of CO2 is measured in real time; ΔT and ΔH are the changes in ambient temperature and ambient humidity per unit time, respectively. The temperature compensation includes: synchronously collecting ambient temperature and body temperature fluctuations within a unit of time, correcting body temperature using a dynamic compensation formula, and thus correcting body temperature fluctuations; the dynamic compensation formula includes: T _校正 =T _实时 -k×(T _环境 -25℃); among which, T _校正 T is the corrected body temperature. _实时 T represents the real-time measured body temperature. _环境 is the ambient temperature per unit time, and k is the compensation coefficient; The body movements are captured by a camera; the light compensation includes using the CLAHE algorithm to enhance the contrast of the images captured by the camera. Furthermore, step 4 also includes: when some data in the basic data exceeds the corresponding warning threshold, multimodal association verification is also performed; the multimodal association verification includes: verifying whether environmental features, physiological features and behavioral features exceed the corresponding warning threshold, and if so, triggering a multi-level response mechanism.

[0015] Option 2 The school violence prevention model, constructed using the method described in Scheme 1, includes: an input layer, a feature extraction layer, and a classification output layer; the feature extraction layer includes a DNN branch, an LSTM branch, and a feature fusion module; the classification output layer includes an SVM classifier. The input layer takes preprocessed basic data as input, and the classification output layer outputs behavioral risk values.

[0016] The working principle and advantages of this invention are as follows: This invention presents a campus violence prevention model and its construction method, which enables timely and accurate detection and early warning of risky behaviors. It possesses strong objectivity and real-time capability, contributing to the construction of a zero-violence, highly efficient smart campus. The key feature is that this solution involves the collection and fusion of multimodal data, organically combining environmental parameters (CO2 concentration), physiological indicators (body temperature fluctuations), and behavioral characteristics (body movements, voiceprints) to construct a comprehensive campus violence perception system. Compared to existing single-dimensional or human-judgment-dependent prevention methods, this solution offers stronger objectivity and comprehensiveness, enabling more accurate violence perception and prediction. By combining a bullying behavior prediction model with regional carbon concentration changes and students' body temperature and behavioral patterns, it predicts and identifies potential bullying behaviors (e.g., abnormal gatherings can be identified through carbon concentration changes, and emerging behavioral conflicts can be identified through body temperature and behavioral pattern analysis). This allows for early warning and intervention against bullying, thereby improving campus safety management and protecting students' physical and mental health. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of Embodiment 1 of the campus violence prevention model and its construction method of the present invention. Detailed Implementation

[0018] The following detailed explanation illustrates the specific implementation methods:

[0019] Example 1

[0020] The basic implementation examples are as follows: Figure 1 The method for constructing a school violence prevention model, as shown, includes the following steps:

[0021] Step 1: Real-time acquisition of environmental parameters, physiological indicators, and behavioral characteristics as basic data using multimodal data acquisition equipment. The environmental parameters include CO2 concentration, ambient temperature, and ambient humidity; the physiological indicators include body temperature fluctuations; and the behavioral characteristics include limb movements and voiceprint data.

[0022] Specifically, the multimodal data acquisition device includes a CO2 sensor for acquiring CO2 concentration data, a thermal imaging camera for acquiring body temperature fluctuation data, a LoRa temperature and humidity node for acquiring ambient temperature and humidity, an optical camera and millimeter-wave radar for acquiring limb movement information, and a directional microphone array for acquiring voiceprint data. In this embodiment, a high-precision CO2 sensor driven by an STM32 microcontroller is selected, with a detection range of 0–5000 ppm and an error of ±2%; a thermal imaging camera in the 8–14 μm far-infrared band is selected to accurately acquire environmental information; and a 2-megapixel optical camera is selected to clearly acquire movement information.

[0023] Furthermore, in this embodiment, the aforementioned multimodal data acquisition device and edge computing unit are integrated into an integrated module for easy deployment on campus.

[0024] Step 2: Preprocess the basic data and generate a fused feature vector.

[0025] The preprocessing includes: processing CO2 concentration data using wavelet transform denoising and moving average filtering (in this embodiment, the window size is set to 5 seconds to smooth instantaneous fluctuations), processing body temperature fluctuation data using Kalman filtering, processing voiceprint data using spectral subtraction denoising, removing environmental noise from the voiceprint data, and performing feature extraction.

[0026] The features extracted include: environmental features, physiological features, and behavioral features; the environmental features include CO2 concentration change rate and spatial air mass retention index; the physiological features include body temperature fluctuation; the behavioral features include the frequency of target behavior and voiceprint keyword hit rate; the target behavior includes pushing, blocking, and pulling actions; and the voiceprint keywords include abusive and threatening words.

[0027] The spatial air mass retention index AI is used to reflect the air flow within a confined space.

[0028] in, CO2 concentration gradient (ppm / m 3 V can be calculated by interpolating data from adjacent sensors; V refers to the spatial volume (m³). 3 Q can be obtained based on BIM models or measured data; Q refers to the ventilation rate (m³ / s). 3 / h). When AI is less than 0.8, it indicates good air circulation and no risk of air stagnation; when AI is greater than 0.8, it can be determined that there is a risk of air stagnation and there may be group conflict.

[0029] When extracting the frequency of target behaviors, the YOLOv7 algorithm is used, combined with skeletal key points, to extract action contours and identify target actions to statistically analyze frequency and avoid facial recognition. When extracting voiceprint keyword hit rates, the BERT model is used to extract voiceprint keywords based on voiceprint information to statistically analyze hit rates. Furthermore, when analyzing voiceprint information, the BERT model only extracts metadata such as intonation, speech rate, and keyword frequency to avoid the risks associated with storing raw audio. This setup effectively avoids the privacy risks of facial recognition and personal privacy leaks throughout the entire process, achieving contactless behavior recognition and privacy-preserving voiceprint analysis. The process of generating the fusion feature vector includes: normalizing each feature and concatenating them into a vector, dynamically allocating weights through an attention mechanism, and then generating a fusion feature vector with uniform dimensions for subsequent model input. The pretreatment also includes CO2 concentration compensation, temperature compensation, and light compensation; The CO2 concentration compensation includes: synchronously collecting ambient temperature, ambient humidity, and CO2 concentration per unit time, and correcting the CO2 concentration through polynomial regression; the polynomial includes: CO 2_校正 =CO 2_实时 ×(1+α×ΔT+β×ΔH), where α and β are calibration coefficients; CO 2_校正 The corrected CO2 concentration, CO 2_实时 The concentration of CO2 is measured in real time; ΔT and ΔH are the changes in ambient temperature and ambient humidity per unit time, respectively. The temperature compensation includes: synchronously collecting ambient temperature and body temperature fluctuations within a unit of time, correcting body temperature using a dynamic compensation formula, and thus correcting body temperature fluctuations; the dynamic compensation formula includes: T _校正 =T _实时 -k×(T _环境 -25℃); among which, T _校正 T is the corrected body temperature. _实时 T represents the real-time measured body temperature. _环境 is the ambient temperature per unit time, and k is the compensation coefficient; The body movements are captured using an optical camera; the light compensation includes using the CLAHE algorithm to enhance the contrast of the images captured by the camera. Step 3: Based on the federated learning platform, train the bullying behavior prediction model using a fusion algorithm of support vector machine and neural network; the bullying behavior prediction model outputs the behavior risk value.

[0030] The bullying behavior prediction model includes an input layer, a feature extraction layer, and a classification output layer; the feature extraction layer includes a DNN branch, an LSTM branch, and a feature fusion module; the classification output layer includes an SVM classifier.

[0031] When training the bullying behavior prediction model, a pre-labeled dataset (containing labeled historical conflict events and corresponding historical basic data) is used. Based on a federated learning platform, the model is trained using encrypted data streams and a transfer learning method, training the model in different scenarios. In this embodiment, the scenarios include classrooms, corridors, and restrooms. In transfer learning, the feature extraction layer is frozen, and only the classification output layer is updated. The model's performance is evaluated using cross-validation to ensure its accuracy and robustness across various scenarios—testing the model's generalization ability in different scenarios (classrooms, restrooms, corridors) to ensure a false negative rate ≤5% and a false positive rate ≤10%.

[0032] Step 4: Based on the site capacity, ventilation efficiency, and real-time collected multimodal basic data of the analyzed area, dynamically adjust the early warning threshold of the basic data; when any data in the basic data exceeds the corresponding early warning threshold, or the behavioral risk value reaches the corresponding risk threshold, trigger a multi-level response mechanism.

[0033] The LSTM model is used to dynamically adjust the warning threshold of the basic data. The LSTM model is trained using a dataset consisting of the site capacity, ventilation efficiency, historical multimodal basic data, and manually labeled historical warning thresholds of the analyzed area. After training, the warning thresholds corresponding to the site capacity, ventilation efficiency, and real-time multimodal basic data of the analyzed area are used as inputs, and the warning thresholds corresponding to CO2 concentration change rate, spatial air mass retention index, body temperature fluctuation, target behavior frequency, and voiceprint keyword hit rate are used as outputs.

[0034] Furthermore, when any data in the basic data exceeds the corresponding warning threshold, multimodal association verification is also performed; the multimodal association verification includes: verifying whether environmental features, physiological features and behavioral features exceed the corresponding warning threshold, and if so, triggering a multi-level response mechanism.

[0035] The behavioral risk value ranges from 0 to 1. The multi-level response mechanism includes: when the behavioral risk value is between 0.6 and 0.8, a level 1 warning is triggered, which links the internal campus system, activates the LCD screens and broadcasts deployed on campus, and provides graphic warnings and targeted broadcasts; when the behavioral risk value is greater than 0.8, a level 2 warning is triggered, which links the internal campus system, activates the buzzers and 5G NR technology deployed on campus, pushes the warning information to the security terminal with low latency, and simultaneously adjusts the campus fresh air system to manage carbon concentration.

[0036] This embodiment also provides a school violence prevention model, which is constructed using the school violence prevention model construction method described in Scheme 1, and includes: an input layer, a feature extraction layer, and a classification output layer; the feature extraction layer includes a DNN branch, an LSTM branch, and a feature fusion module; the classification output layer includes an SVM classifier; The input layer takes preprocessed basic data as input, and the classification output layer outputs behavioral risk values.

[0037] This embodiment provides a campus violence prevention model and its construction method, which can achieve timely and accurate detection and early warning of risky behaviors. It has strong objectivity and real-time performance, and helps to build a zero-violence, high-efficiency smart campus.

[0038] Example 2

[0039] A method for constructing a school violence prevention model has been improved based on Example 1, as follows.

[0040] In step 3, a quantization model is also constructed based on the federated learning platform.

[0041] Specifically, the quantization model is composed of the following formula:

[0042] Behavioral risk value = α·AI + β·ΔCO2 + γ·f 动作 +δ·P 情感 . Wherein, α, β, γ, and δ are weight coefficients. In this embodiment, supervised learning (labeling historical conflict events and corresponding historical basic data) is used to train the quantization model, and gradient descent is used to optimize the weight coefficients. AI is the space air mass retention index; ΔCO2 is the real-time CO2 concentration fluctuation value; f 动作 The frequency of the target behavior reflects the level of activity in physical conflict; P 情感 This represents the voiceprint keyword hit rate; a higher value indicates a greater risk of conflict.

[0043] The behavioral risk value ranges from 0 to 1. The multi-level response mechanism includes: when the behavioral risk value is between 0.6 and 0.8, a level 1 warning is triggered, which links the internal campus system, activates the LCD screens and broadcasts deployed on campus, and provides graphic warnings and targeted broadcasts; when the behavioral risk value is greater than 0.8, a level 2 warning is triggered, which links the internal campus system, activates the buzzers and 5G NR technology deployed on campus, pushes the warning information to the security terminal with low latency, and simultaneously adjusts the campus fresh air system to manage carbon concentration.

[0044] Optionally, when the behavior risk value output by the bullying behavior prediction model is greater than 0.6, the behavior risk value is also calculated simultaneously by the quantitative model. If the behavior risk values ​​output by both models are greater than 0.6, a multi-level response mechanism is triggered, and a first-level warning or a second-level warning is triggered according to the average of the behavior risk values ​​output by the two models, so as to reduce the false alarm rate.

[0045] This embodiment provides a school violence prevention model and its construction method, which has a higher early warning accuracy.

[0046] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for constructing a school violence prevention model, characterized in that, Includes the following steps: Step 1: Acquire environmental parameters, physiological indicators, and behavioral characteristics in real time using multimodal data acquisition equipment as basic data. The environmental parameters include... Concentration; the physiological indicators include body temperature fluctuations; the behavioral characteristics include limb movements and voiceprint data; Step 2: Preprocess the basic data and generate a fused feature vector; Step 3: Based on the federated learning platform, train the bullying behavior prediction model using a fusion algorithm of support vector machine and neural network; the bullying behavior prediction model outputs the behavior risk value.

2. The method for constructing a school violence prevention model according to claim 1, characterized in that, The preprocessing includes: wavelet transform denoising and moving average filtering. Concentration data, body temperature fluctuation data processed by Kalman filtering, and voiceprint data processed by spectral subtraction noise reduction method; and feature extraction was performed.

3. The method for constructing a school violence prevention model according to claim 2, characterized in that, The features extracted during feature extraction include: environmental features, physiological features, and behavioral features; the environmental features include... Concentration change rate, spatial air mass retention index; the physiological characteristics include body temperature fluctuation; the behavioral characteristics include the frequency of occurrence of target behavior, voiceprint keyword hit rate; the target behavior includes pushing, blocking, and pulling actions; the voiceprint keywords include abusive and threatening words.

4. The method for constructing a school violence prevention model according to claim 3, characterized in that, The process of generating the fusion feature vector includes: normalizing each feature and concatenating them into a vector, dynamically allocating weights through an attention mechanism, and then generating a fusion feature vector with uniform dimensions for subsequent model input.

5. The method for constructing a school violence prevention model according to claim 1, characterized in that, It also includes step 4, which dynamically adjusts the early warning threshold of the basic data based on the venue capacity, ventilation efficiency and real-time collected multimodal basic data of the analyzed area; when any data in the basic data exceeds the corresponding early warning threshold, or the behavioral risk value reaches the corresponding risk threshold, a multi-level response mechanism is triggered.

6. The method for constructing a school violence prevention model according to claim 5, characterized in that, The behavioral risk value ranges from 0 to 1; the multi-level response mechanism includes: when the behavioral risk value is 0.6 to 0.8, a level 1 warning is triggered, linking the internal campus system, activating LCD screen alerts and directional broadcasts; when the behavioral risk value is greater than 0.8, a level 2 warning is triggered, linking the internal campus system, activating buzzers and 5G NR technology, pushing the warning information to the security terminal with low latency, and simultaneously adjusting the campus fresh air system to manage carbon concentration.

7. The method for constructing a school violence prevention model according to claim 1, characterized in that, The bullying behavior prediction model includes an input layer, a feature extraction layer, and a classification output layer; the feature extraction layer includes a DNN branch, an LSTM branch, and a feature fusion module; the classification output layer includes an SVM classifier.

8. The method for constructing a school violence prevention model according to claim 2, characterized in that, The preprocessing also includes Concentration compensation, temperature compensation, light compensation; The Concentration compensation includes: synchronously collecting ambient temperature, ambient humidity, and [other parameters] per unit time. Concentration, corrected by multinomial regression Concentration; the polynomial includes: ,in, , For calibration coefficients; CO 2_校正 For the corrected concentration, For real-time measurement concentration; and These represent the changes in ambient temperature and ambient humidity per unit time, respectively. The temperature compensation includes: synchronously collecting ambient temperature and body temperature fluctuations within a unit of time, correcting the body temperature using a dynamic compensation formula, and thus correcting the body temperature fluctuations; the dynamic compensation formula includes: ;of which, CO 2_校正 This is the corrected body temperature. This refers to the body temperature measured in real time. is the ambient temperature per unit time, and k is the compensation coefficient; The body movements are captured by a camera; the light compensation includes using the CLAHE algorithm to enhance the contrast of the images captured by the camera.

9. The method for constructing a school violence prevention model according to claim 5, characterized in that, Step 4 also includes: when some data in the basic data exceeds the corresponding warning threshold, multimodal association verification is also performed; the multimodal association verification includes: verifying whether environmental features, physiological features and behavioral features exceed the corresponding warning threshold, and if so, triggering a multi-level response mechanism.

10. A school violence prevention model, characterized in that, The model is constructed using the method described in any one of claims 1-9 for preventing school violence, and includes: an input layer, a feature extraction layer, and a classification output layer; the feature extraction layer includes a DNN branch, an LSTM branch, and a feature fusion module; the classification output layer includes an SVM classifier. The input layer takes preprocessed basic data as input, and the classification output layer outputs behavioral risk values.

Images