Construction method of behavior prediction model based on ideological and political lesson learning

Abstract

The invention belongs to the technical field of student education management systems, and particularly relates to a behavior prediction model construction method based on ideological and political lesson learning, and the method comprises the steps: 1, installing an intelligent sensing device on a student desk, and enabling the intelligent sensing device to obtain a posture signal, an expression signal and a physiological signal of a student; 2, setting distributed edge computing nodes, locally operating a lightweight Transform model, and extracting a posture feature vector, an expression feature vector and a physiological feature vector of a student; 3, aggregating multi-classroom edge node features, and constructing a global behavior analysis model; weights are dynamically distributed according to class features, and federal learning aggregation frequency and model compression ratio are automatically selected; step 4, generating a behavior baseline based on student historical data, and identifying an abnormal behavior deviating from a normal state; and outputting a three-dimensional visualization report. The ideological and political lesson learning behaviors of students can be accurately monitored.

Description

Technical Field

[0001] The present invention belongs to the technical field of student education management systems, and particularly relates to a construction method of a behavior prediction model based on ideological and political education course learning. Background Art

[0002] In the ideological and political education courses for college students, by constructing a behavior prediction model, the classroom behaviors of students can be understood in real time, so as to further improve the effect of ideological and political education courses. For this reason, the present invention proposes a construction method of a behavior prediction model based on ideological and political education course learning. Summary of the Invention

[0003] The purpose of the present invention is to provide a construction method of a behavior prediction model based on ideological and political education course learning, which can accurately monitor the behaviors of students in ideological and political education courses.

[0004] The technical solution adopted by the present invention is specifically as follows:

[0005] A construction method of a behavior prediction model based on ideological and political education course learning, comprising the following steps:

[0006] Step 1: Install an intelligent sensing device on the student's desk, and the intelligent sensing device is used to obtain the posture signal, expression signal and physiological signal of the student;

[0007] Step 2: Set up a distributed edge computing node, locally run a lightweight Transformer model, extract the posture feature vector, expression feature vector and physiological feature vector of the student according to the posture signal, expression signal and physiological signal of the student, and use the differential privacy algorithm to add Gaussian noise to the feature vector and then complete the perturbation gradient parameter setting;

[0008] Step 3: Aggregate the edge node features of multiple classrooms to construct a global behavior analysis model; dynamically allocate weights according to class features, and automatically select the federated learning aggregation frequency and model compression rate;

[0009] Step 4: Generate a behavior baseline based on the historical data of the student, identify abnormal behaviors deviating from the norm; and output a three-dimensional visualization report.

[0010] Preferably, in step 1, a rectangular slot is opened at the middle position of each seat of the desk, and the intelligent sensing device is installed in the rectangular slot and fixed with bolts. The intelligent sensing device includes a shell, and the shell is provided with an inclined side panel, and a millimeter-wave radar and a collection camera are installed on the side panel. An edge computer is installed inside the shell, and the millimeter-wave radar and the collection camera are both communicatively connected to the edge computer; several edge computers in the classroom are communicatively connected to a main computer, step 2 is run in the edge computer, and step 3 and step 4 are run in the main computer; the operating frequency band of the millimeter-wave radar is 60-64GHz, and the spatial resolution is 0.5cm. It can detect the frequency of finger micro-movements (0.1-5Hz) and associate it with the attention level.

[0011] Preferably, in step 1, the millimeter-wave radar obtains physiological signals by non-invasively monitoring the heart rate and breathing rate of the students, and the acquisition camera is used to monitor the students' posture movements and facial expressions to obtain posture signals and expression signals; the acquisition camera adopts a detachable edge camera; and is equipped with an adaptive light compensation chip to correct low-light or backlight scenes in real time.

[0012] Preferably, step 2 includes the following specific steps:

[0013] Step 201: Distributed edge computing environment settings:

[0014] Each edge computing node is installed with a corresponding computing framework, so that a lightweight Transformer model can be run on the edge computing node. A distributed system architecture based on message queues is used to coordinate tasks and data flows between nodes. The gesture signals, expression signals, and physiological signals collected by each edge computing node need to be processed in real time through the edge computing node to complete data collection.

[0015] Step 202: Build a lightweight Transformer model:

[0016] Reduce the model size by reducing the number of layers, hidden units, and attention heads; use sparse attention or low-rank attention to reduce computational complexity; reduce the computational complexity of small models through knowledge distillation; quantize model weights and activations to low precision through model quantization and pruning, and prune unimportant neurons and connections; use lightweight architecture design to adopt lightweight models designed for resource-constrained devices;

[0017] Step 203: extracting feature vectors;

[0018] First, gesture signal processing is performed to obtain gesture signals, expression signals, and physiological signals, and gesture feature vectors are extracted using a Transformer-based gesture recognition model.

[0019] Then, perform facial expression signal processing through a facial expression recognition model; analyze the facial expressions of students; represent the emotional and mood changes of students;

[0020] Then, perform physiological signal processing, directly read the physiological signals monitored by the millimeter-wave radar, and extract physiological feature vectors to reflect the physical condition of students.

[0021] Preferably, step 2 further includes the following steps: Step 203: Differential privacy algorithm setting, adding Gaussian noise to each feature vector; when training the model each time, noise needs to be added during gradient update; setting the privacy budget (ε) to control the intensity of the noise and the degree of data privacy protection; a smaller ε value will provide stronger privacy protection; for each feature vector, such as the feature vectors of posture, expression, and physiological signals, add noise according to the following formula where is the feature vector after adding noise, x is the original feature vector, and N(0,σ 2 ) is Gaussian noise with a mean of 0 and a variance of σ 2 .

[0022] Preferably, step 3 includes the following steps:

[0023] Step 301: Global behavior analysis model:

[0024] Use deep learning or other machine learning models to model the global behavior data, analyze the collective behavior patterns, classroom participation, and mood fluctuations of students; and aggregate the data of each edge node to the central server or cloud platform for training the global model; use federated learning to protect data privacy and avoid directly uploading data to the cloud;

[0025] Step 302: Dynamically allocate weights

[0026] According to the characteristics of each class, dynamically adjust the weights of each node in the global model; based on the class characteristics, use the weighted average method to allocate the weights of each class;

[0027] According to the class characteristics; dynamically adjust the aggregation frequency; use reinforcement learning or adaptive algorithms to determine the optimal aggregation period;

[0028] Each edge node regularly uploads its updated model weights, and the central server aggregates these updates to generate a global model; automatically select an appropriate aggregation frequency to balance performance and efficiency.

[0029] Preferably, step 3 further includes the following steps:

[0030] Step 304: Automatically select the model compression ratio

[0031] Model compression technology: Use techniques such as quantization, pruning, and low-rank decomposition to reduce the computational overhead and storage requirements of the model; select the compression rate dynamically; automatically adjust the model compression rate based on the computing power and network conditions of the edge nodes.

[0032] Preferably, step 3 further includes the following steps:

[0033] Step 305:

[0034] Real-time optimization: Continuously optimize the aggregation frequency and compression rate of the model according to real-time data and feedback;

[0035] End-to-end testing: Conduct tests in different classroom scenarios, evaluate the accuracy, real-time performance, and resource consumption of the model; and perform integrated optimization.

[0036] The technical effects achieved by the present invention are:

[0037] In the present invention, it is possible to accurately monitor students' behaviors in ideological and political education courses: By collecting students' postures, expressions, and physiological signals through intelligent perception devices, it is possible to comprehensively monitor students' behaviors and emotional states, thereby providing accurate behavior prediction. The differential privacy algorithm is used to perturb the feature vectors of students, effectively protecting students' personal privacy and ensuring the security of data. Through distributed edge computing nodes and lightweight Transformer models, data is processed locally, reducing latency, improving real-time performance and efficiency; weights are dynamically assigned according to class characteristics, and the aggregation frequency and model compression rate of federated learning are automatically adjusted, enabling the system to flexibly adapt to different scenarios, improving personalized performance; aggregating the feature vectors of edge nodes in multiple classrooms to construct a global behavior analysis model, enhancing the generalization ability and accuracy of the system; generating a behavior baseline through historical data, capable of accurately identifying students' abnormal behaviors, helping to intervene and improve the learning process in a timely manner; generating a three-dimensional visualization report to intuitively display the results of students' behavior analysis, facilitating teachers or administrators to analyze and make decisions; the characteristics of the distributed architecture and federated learning enhance the fault tolerance of the system, enabling it to operate stably under multiple nodes and improving the system's fault tolerance. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flowchart of a method for constructing a behavior prediction model based on ideological and political course learning of the present invention;

[0039] Figure 2 is an overall structural schematic diagram of installing an intelligent perception device in the present invention;

[0040] Figure 3 is an overall structural schematic diagram of another perspective of installing an intelligent perception device in the present invention.

[0041] In the drawings, the list of components represented by each reference numeral is as follows:

[0042] 1. Housing; 2. Side plate; 3. Millimeter-wave radar; 4. Acquisition camera; 5. Edge computer; 6. General computer. Specific embodiments

[0043] In order to make the purpose and advantages of the present invention clearer, the present invention will be specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific implementation manners of the present invention, and does not strictly limit the scope of protection specifically claimed by the present invention.

[0044] As Figures 1 - 3 shown, a method for constructing a behavior prediction model based on ideological and political course learning includes the following steps:

[0045] Step 1: Install an intelligent perception device on the student's desk, and the intelligent perception device is used to obtain the posture signal, expression signal and physiological signal of the student;

[0046] Step 2: Set up a distributed edge computing node, locally run a lightweight Transformer model, extract the posture feature vector, expression feature vector and physiological feature vector of the student according to the posture signal, expression signal and physiological signal of the student, and use the differential privacy algorithm to add Gaussian noise to the feature vector and then complete the perturbation gradient parameter setting;

[0047] Step 3: Aggregate the features of multiple classroom edge nodes to construct a global behavior analysis model; dynamically allocate weights according to class features, and automatically select the federated learning aggregation frequency and model compression rate;

[0048] Step 4: Generate a behavior baseline based on the student's historical data, identify abnormal behaviors that deviate from the norm, such as frequent leg shaking caused by anxiety; and output a three-dimensional visualization report, such as a time axis + behavior heat map + physiological index correlation curve.

[0049] In the present invention, non-contact physiological monitoring by the millimeter-wave radar 3: Traditional classroom monitoring systems usually rely on cameras or wearable devices. The application of the millimeter-wave radar 3 in the educational scenario is a first; its 60-64 GHz frequency band selection; it penetrates books but avoids the risk of human radiation; it solves the contradiction between privacy and security; Hardware-level data desensitization: The original radar signal is directly converted into a heart rate value at the front-end chip; the original waveform is not stored; the risk of privacy leakage is avoided from the physical level, which is different from the traditional software encryption scheme; Experimental data shows that for the concentration model that fuses the physiological signal heart rate variability and sitting posture data, the false alarm rate is reduced by 12.7% compared with the single vision scheme.

[0050] In step 1, a rectangular notch is opened at the middle position of each seat on the desk, and the intelligent sensing device is installed in the rectangular notch and fixed with bolts. The intelligent sensing device includes a housing 1, an inclined side plate 2 is provided on the housing 1, a millimeter-wave radar 3 and a collection camera 4 are installed on the side plate 2, an edge computer 5 is installed inside the housing 1, and both the millimeter-wave radar 3 and the collection camera 4 are communicatively connected to the edge computer 5; several edge computers 5 in the classroom are communicatively connected to a general computer 6. Step 2 runs in the edge computer 5, and steps 3 and 4 run in the general computer 6; the working frequency band of the millimeter-wave radar 3 is 60 - 64 GHz, the spatial resolution is 0.5 cm, and it can detect the finger micro-movement frequency of 0.1 - 5 Hz and associate it with the attention level.

[0051] In step 1, the millimeter-wave radar 3 non-invasively monitors the student's heart rate and breathing frequency to obtain physiological signals, and the collection camera 4 is used to monitor the student's posture movements and facial expressions to obtain posture signals and expression signals; the collection camera 4 uses a detachable edge camera; and is equipped with a chip for adaptive light compensation, such as HDR-X3, to correct low-light or backlight scenes in real time.

[0052] Preferably, step 2 includes the following specific steps:

[0053] Step 201: Distributed edge computing environment setup:

[0054] Each edge computing node installs a corresponding computing framework, such as TensorFlow Lite, PyTorch Mobile, etc., so that a lightweight Transformer model can run on the edge computing node; in order to effectively process data and distributed inference of the model, a distributed system architecture based on message queues can be used to coordinate tasks and data streams between nodes. Posture signals, expression signals, and physiological signals collected by each edge computing node need to be processed in real time through the edge computing node to complete data collection;

[0055] Step 202: Construct a lightweight Transformer model:

[0056] Reduce the model size by reducing the number of layers, hidden units, and attention heads; use sparse attention or low-rank attention to reduce the computational complexity; reduce the computational amount of the small model through knowledge distillation; quantize the model weights and activations to low precision and prune unimportant neurons and connections through model quantization and pruning; use a lightweight architecture design to adopt a lightweight model designed for resource-constrained devices;

[0057] Step 203: Extract feature vectors;

[0058] First, perform attitude signal processing to obtain attitude signals, facial expression signals, and physiological signals, and use a Transformer-based attitude recognition model to extract attitude feature vectors; these features can represent the behavior characteristics of students, such as sitting postures, standing postures, walking, etc.

[0059] Then, perform facial expression signal processing through a facial expression recognition model; such as facial expression analysis based on convolutional neural networks or Transformers; analyze the facial expressions of students; extract the facial expression feature vectors of students, such as happy, sad, surprised, confused, etc., to represent the emotional and mood changes of students.

[0060] Next, perform physiological signal processing, directly read the physiological signals monitored by the millimeter-wave radar 3, and extract physiological feature vectors, such as heart rate, respiratory rate, body temperature, etc., and these data can reflect the physical condition of students.

[0061] Step 203: Set the differential privacy algorithm, add Gaussian noise to each feature vector; the standard deviation of the Gaussian noise determines the degree of privacy protection. The larger the standard deviation, the stronger the privacy protection; during each model training, noise needs to be added during gradient updates; especially when using optimization algorithms such as gradient descent; by adding perturbations, the gradients of the real data can be prevented from being exposed during each training process; set the privacy budget ε to control the intensity of the noise and the degree of data privacy protection; a smaller ε value will provide stronger privacy protection; for each feature vector, such as the feature vectors of attitude, facial expression, and physiological signals, add noise according to the following formula where is the feature vector after adding noise, x is the original feature vector, and N(0,σ 2 ) is Gaussian noise with a mean of 0 and a variance of σ 2 .

[0062] In the present invention, the differential privacy algorithm ensures that personal data will not be leaked during the calculation process by adding noise to the feature vector. In this way, it can effectively prevent the leakage of students' sensitive information. Reducing latency: Running the model locally on the edge computing node instead of uploading the data to the central server for processing can significantly reduce latency and improve the response speed of the system. Lightweight model: Using a lightweight Transformer model can ensure low computational resource consumption while ensuring high model performance, making it suitable for deployment on edge devices. Real-time processing: Since the calculation occurs at the local edge node, the system can process students' physiological signals, expressions, and posture features in real time, providing support for real-time monitoring and interaction. High efficiency: Distributed computing can distribute tasks to multiple nodes, improving the overall computational power and efficiency of the system, thereby accelerating the data processing process. Scalability: The distributed architecture allows nodes to be expanded as needed, improving the scalability of the system and enabling it to handle more user data and tasks. Personalized service: By combining multiple signals such as posture, expression, and physiological signals, it is possible to more accurately capture the state of students and provide personalized feedback and support. Enhancing model security: By adding Gaussian noise and differential privacy, it can effectively prevent the model from being attacked or reverse-engineered to obtain private data, enhancing the security of the overall system. These advantages make this solution more advantageous when dealing with sensitive data, while improving the efficiency and response speed of the system.

[0063] Preferably, step 3 includes the following steps:

[0064] Step 301: Global behavior analysis model:

[0065] Use deep learning or other machine learning models, such as LSTM, GRU, Transformer, etc., to model the global behavior data, analyze students' collective behavior patterns, classroom participation, and emotional fluctuations; and aggregate the data of each edge node to the central server or cloud platform for training the global model; use federated learning to protect data privacy and avoid directly uploading data to the cloud.

[0066] Step 302: Dynamically allocate weights

[0067] According to the characteristics of each class, such as the number of students, learning level, classroom environment, teaching method, etc., dynamically adjust the weights of each node in the global model; for example, the behavior data of a certain class may contribute more to the global model than other classes; based on the class characteristics, use the weighted average method to allocate the weights of each class.

[0068] Step 303: Automatic selection of federated learning aggregation frequency

[0069] Aggregation Frequency Optimization: In federated learning, frequent model aggregation increases computational and communication overheads; Dynamically adjust the aggregation frequency according to class characteristics, such as the number of students, device performance, etc.; Use reinforcement learning or adaptive algorithms to determine the optimal aggregation period;

[0070] Model Synchronization and Update: Each edge node regularly uploads its updated model weights, and the central server aggregates these updates to generate a global model; Automatically select an appropriate aggregation frequency to balance performance and efficiency.

[0071] Step 304: Automatic Selection of Model Compression Ratio

[0072] Model Compression Techniques: Utilize techniques such as quantization, pruning, and low-rank decomposition to reduce the computational overhead and storage requirements of the model, especially on resource-constrained devices; such as edge nodes; By dynamically selecting the compression ratio, ensure that the compressed model can still maintain good performance; Compression Ratio Adaptation: Automatically adjust the model compression ratio based on the computational power and network conditions of the edge node; Reinforcement learning or algorithms based on performance evaluation can be used to optimize the compression strategy;

[0073] Step 305:

[0074] Real-time Optimization: Continuously optimize the aggregation frequency and compression ratio of the model based on real-time data and feedback to ensure that the entire system can operate efficiently.

[0075] End-to-End Testing: Conduct tests in different classroom scenarios to evaluate the accuracy, real-time performance, and resource consumption of the model to ensure that the model can adapt to different teaching environments; And perform integrated optimization.

[0076] In the present invention, through this method, efficient aggregation of multi-classroom edge node features, global behavior analysis, dynamic weight allocation, and intelligent optimization of the aggregation frequency and model compression ratio of federated learning can be achieved, thereby realizing efficient and accurate teaching analysis and personalized recommendation systems.

[0077] In step 3, the features of edge nodes in multiple classrooms are aggregated to construct a global behavior analysis model, dynamically allocate weights, and automatically select the federated learning aggregation frequency and model compression rate, which have the following advantages: Global optimization: By aggregating the features of edge nodes in multiple classrooms, a global behavior analysis model can be established to obtain more comprehensive and accurate behavior recognition and prediction results. This global perspective improves the generalization ability and accuracy of the model. Personalization and customization: Dynamically allocating weights according to class characteristics can adjust the importance weights of the model according to the needs and characteristics of different classes, thereby improving the personalization and customization capabilities of the model to meet the needs in different scenarios. Improving efficiency: Federated learning reduces the burden of data transmission and network latency by performing local calculations instead of centralizing all data on the server, and can protect privacy. Dynamically selecting the aggregation frequency and model compression rate can optimize communication and computational efficiency. Privacy protection: Federated learning enables data to be processed at local edge nodes, avoiding centralized storage of sensitive data, reducing the risk of data leakage, and ensuring data privacy. Adaptability: Dynamically selecting the aggregation frequency and model compression rate can automatically adjust according to the actual situation, so as to achieve optimal performance under different network conditions, computing capabilities, and data distributions. This flexibility enables the system to adapt to various different environments. Reducing bandwidth consumption: Through model compression, the amount of data transmitted during each aggregation is reduced, especially in edge computing scenarios, which can effectively save bandwidth resources and reduce network burden. Enhancing fault tolerance: The distributed characteristics of multi-node aggregation and federated learning enhance the fault tolerance of the system. When some nodes have problems or go offline, other nodes can still continue to participate in the aggregation to ensure the stability of the system. Lower computational burden: By offloading the computational pressure through edge computing and distributing the computational tasks to local nodes, the computational burden on the central server is reduced, and the response speed and computational efficiency of the entire system are improved. Scalability: As the number of classrooms and edge nodes increases, the system can be flexibly scaled to continue to maintain good performance and response speed, supporting large-scale applications. Through these advantages, this solution can provide a more efficient, flexible, and secure global behavior analysis system to adapt to the dynamically changing teaching environment and requirements.

[0078] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. The structures, devices, and operation methods not specifically described and explained in the present invention are implemented according to the conventional means in the art without special description and limitation.

Claims

1. A method for constructing a behavior prediction model based on ideological and political course learning, characterized in that: Including the following steps: Step 1: Install intelligent sensing devices on students' desks. The intelligent sensing devices are used to obtain students' posture signals, expression signals, and physiological signals; Step 2: Set up distributed edge computing nodes, locally run a lightweight Transformer model, extract students' posture feature vectors, expression feature vectors, and physiological feature vectors according to students' posture signals, expression signals, and physiological signals, and complete the perturbation gradient parameter setting after adding Gaussian noise to the feature vectors using the differential privacy algorithm; Step 3: Aggregate the features of multiple classroom edge nodes to construct a global behavior analysis model; dynamically allocate weights according to class features, and automatically select the federated learning aggregation frequency and model compression ratio; Step 4: Generate a behavior baseline based on students' historical data, identify abnormal behaviors deviating from the norm; and output a three-dimensional visualization report.

2. The construction method of a behavior prediction model based on ideological and political course learning according to claim 1, characterized in that: In the said Step 1, a rectangular notch is opened at the middle position of each seat of the desk, and the intelligent sensing device is installed in the rectangular notch and fixed with bolts. The intelligent sensing device includes a housing (1), and an inclined side plate (2) is provided on the housing (1). A millimeter-wave radar (3) and a collection camera (4) are installed on the side plate (2). An edge computer (5) is installed inside the housing (1). Both the millimeter-wave radar (3) and the collection camera (4) are communicatively connected to the edge computer (5); several edge computers (5) in the classroom are communicatively connected to a main computer (6). The Step 2 runs in the edge computer (5), and the Step 3 and the Step 4 run in the main computer (6); the working frequency band of the millimeter-wave radar (3) is 60 - 64 GHz, the spatial resolution is 0.5 cm, and it can detect the finger micro-movement frequency (0.1 - 5 Hz) and correlate it to the attention level.

3. The construction method of a behavior prediction model based on ideological and political course learning according to claim 2, characterized in that: In the said Step 1, the millimeter-wave radar (3) obtains physiological signals by non-invasively monitoring students' heart rates and breathing frequencies. The collection camera (4) is used to monitor students' posture movements and facial expressions to obtain posture signals and expression signals; the collection camera (4) adopts a detachable edge camera; and is equipped with an adaptive light compensation chip to correct low-light or backlight scenarios in real time.

4. The construction method of a behavior prediction model based on ideological and political course learning according to claim 1, characterized in that: The said Step 2 includes the following specific steps: Step 201: Distributed edge computing environment setup: Install corresponding computing frameworks on each edge computing node so that a lightweight Transformer model runs on the edge computing node; use a distributed system architecture based on message queues to coordinate tasks and data streams between nodes. Signals such as posture signals, expression signals, and physiological signals collected by each edge computing node need to be processed in real time through the edge computing node to complete data collection; Step 202: Construct a lightweight Transformer model: Reduce the model size by decreasing the number of layers, hidden units, and attention heads; use sparse attention or low-rank attention to reduce computational complexity; reduce the computational load of the small model through knowledge distillation; quantize the model weights and activations to low precision and prune unimportant neurons and connections through model quantization and pruning; use lightweight architecture design to adopt lightweight models designed for resource-constrained devices; Step 203: Extract feature vectors; First, perform pose signal processing to obtain pose signals, facial expression signals, and physiological signals, and use a Transformer-based pose recognition model to extract pose feature vectors; Then, perform facial expression signal processing through a facial expression recognition model; analyze the facial expressions of students; represent the emotional and mood changes of students; Then, perform physiological signal processing, directly read the physiological signals monitored by the millimeter-wave radar (3), and extract physiological feature vectors to reflect the physical condition of students.

5. The construction method of a behavior prediction model based on ideological and political course learning according to claim 1, characterized in that: Step 2 further includes the following steps: Step 203: Differential privacy algorithm setting, adding Gaussian noise to each feature vector; at each model training, noise needs to be added during gradient update; setting the privacy budget (ε) to control the intensity of the noise and the degree of data privacy protection; a smaller ε value provides stronger privacy protection; for each feature vector, such as the feature vectors of posture, expression, and physiological signals, add noise according to the following formula where is the feature vector after adding noise, x is the original feature vector, and N(0,σ 2 ) is Gaussian noise with a mean of 0 and a variance of σ 2 .

6. The construction method of a behavior prediction model based on ideological and political course learning according to claim 1, characterized in that: Step 3 includes the following steps: Step 301: Global behavior analysis model: Use deep learning or other machine learning models to model global behavior data, analyze the collective behavior patterns, class participation, and emotional fluctuations of students; and aggregate the data of each edge node to a central server or cloud platform for training of the global model; use federated learning to protect data privacy and avoid directly uploading data to the cloud; Step 302: Dynamically assign weights According to the characteristics of each class, dynamically adjust the weights of each node in the global model; based on class characteristics, use the weighted average method to assign weights to each class; According to class characteristics; dynamically adjust the aggregation frequency; use reinforcement learning or adaptive algorithms to determine the optimal aggregation period; Each edge node regularly uploads its updated model weights, and the central server aggregates these updates to generate a global model; automatically select an appropriate aggregation frequency to balance performance and efficiency.

7. The construction method of a behavior prediction model based on ideological and political course learning according to claim 1, characterized in that: Step 3 also includes the following steps: Step 304: Automatically select the model compression ratio Model compression technology: Use techniques such as quantization, pruning, and low-rank decomposition to reduce the computational overhead and storage requirements of the model; automatically adjust the model compression ratio based on the computational power and network conditions of the edge nodes by dynamically selecting the compression ratio.

8. The construction method of a behavior prediction model based on ideological and political course learning according to claim 1, characterized in that: Step 3 also includes the following steps: Step 305: Real-time optimization: Continuously optimize the aggregation frequency and compression ratio of the model based on real-time data and feedback; End-to-end testing: Conduct tests in different classroom scenarios to evaluate the accuracy, real-time performance, and resource consumption of the model; and perform integrated optimization.

Images