Course analysis management system based on deep learning

Abstract

The invention relates to the technical field of course analysis management, in particular to a course analysis management system based on deep learning. The method has the advantages that multi-modal data deep analysis realizes full-dimensional analysis of unstructured data by integrating a 3D-CNN model, a Transform architecture and a BERT model and synchronously extracting an attention hot area, a voice emotional state and a text knowledge point association network of a classroom video; nonlinear behavior modeling adopts an LSTM network and time convolutional network fusion model, a knowledge internalization path and forgetting curve prediction are dynamically generated, parameters are optimized in combination with incremental learning, and a transition rule across knowledge points is captured; a teaching scene-evaluation threshold mapping table is constructed based on a reinforcement learning algorithm through dynamic decision and resource collaboration, collaborative optimization under multi-campus data privacy protection is achieved in combination with a federated learning framework, GPU computing nodes are dynamically allocated through a heterogeneous resource scheduling engine, and the analysis efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of course analysis management, and particularly relates to a course analysis management system based on deep learning. BACKGROUND

[0002] The field is a course analysis management system in educational informatization, and a core goal thereof is to realize course resource optimization, teaching process monitoring and teaching decision support through technical means. Current mainstream technologies are mostly based on traditional database architecture and shallow data analysis models, but as the complexity of educational scenarios increases, such as hybrid teaching and multi-modal data fusion, the existing systems have significant defects in the following aspects.

[0003] 1. Weak multi-modal data analysis capability

[0004] Existing systems rely on structured data, such as score and attendance analysis, but lack effective processing means for unstructured data such as classroom videos, voice discussions and text assignments, resulting in more than 60% of teaching behavior characteristics such as attention fluctuation and interaction quality not being quantified and utilized. Although some systems have introduced video analysis modules, they can only achieve basic behavior statistics such as attendance numbers, and cannot extract high-level semantic features such as emotional state and knowledge point association through deep learning.

[0005] 2. Single dimension of learning behavior modeling

[0006] Traditional algorithms such as decision trees and linear regression can only analyze explicit indicators and are difficult to capture non-linear relationships in knowledge internalization paths, such as knowledge point transition rules and long-term forgetting curves. They do not integrate time series data, such as the coordinated influence of stage test fluctuations and cognitive characteristics.

[0007] 3. Lack of dynamic adaptability mechanism

[0008] More than 80% of course management systems use fixed thresholds, such as score segmentation statistics and predefined rules such as attendance rate weight, which cannot dynamically adjust evaluation standards according to teaching progress. In a hybrid teaching system of a certain university, 42% of personalized recommendation schemes were ineffective due to the lack of real-time synchronization of online learning behavior and offline performance. SUMMARY

[0009] The present application aims to overcome the shortcomings of the prior art and provide a course analysis management system based on deep learning, which effectively solves the deficiencies of the prior art.

[0010] The purpose of the present application is achieved by the following technical solutions: a course analysis management system based on deep learning, comprising a multi-modal acquisition module, a semantic analysis engine, a behavior modeling module, a dynamic decision module, a real-time feedback module and a self-adaptive evaluation module, the multi-modal acquisition module is configured to synchronously acquire classroom video stream, voice discussion data, text homework files and structured attendance data and perform local preprocessing through an edge computing node, the semantic analysis engine is in communication connection with the multi-modal acquisition module, the semantic analysis engine comprises a video analysis unit, a voice processing unit and a text mining unit, the video analysis unit extracts student body movements and attention hot areas based on a 3D-CNN model, the voice processing unit identifies emotional states and knowledge point associations using a Transformer architecture, and the text mining unit constructs a knowledge point topology network through a BERT model, the behavior modeling module receives output data of the semantic analysis engine, the behavior modeling module integrates an LSTM network and an attention mechanism to generate a nonlinear learning model containing knowledge internalization paths, forgetting curves and cross-knowledge point transitions, the dynamic decision module interacts with the behavior modeling module in both directions, the dynamic decision module dynamically adjusts teaching evaluation thresholds based on a reinforcement learning algorithm and outputs personalized teaching strategies, the real-time feedback module receives strategy instructions from the dynamic decision module, the real-time feedback module pushes teaching intervention suggestions to a teacher end and outputs a self-adaptive learning path to a student end, and the self-adaptive evaluation module forms a closed loop with the multi-modal acquisition module and the dynamic decision module, and the self-adaptive evaluation module is configured to update model parameters through incremental learning and evaluate the effectiveness of teaching strategies.

[0011] Preferably, according to any of the above solutions, the video analysis unit comprises an eye movement tracking subunit and an interaction quality evaluation subunit, the eye movement tracking subunit identifies the coordinates of the line of sight focus based on the OpenPose algorithm and calculates the attention fluctuation index, and the interaction quality evaluation subunit uses DBSCAN clustering analysis to generate a classroom interaction heat map.

[0012] Preferably, according to any of the above solutions, the behavior modeling module is configured to fuse the knowledge point topology network and the forgetting curve prediction model to generate a dynamic knowledge graph, and the behavior modeling module captures the cooperative influence law of stage test scores through a time convolution network (TCN).

[0013] Preferably, according to any of the above solutions, the dynamic decision module comprises a threshold self-adaptive unit and a cross-modal fusion unit, the dynamic decision module constructs a teaching scene-evaluation threshold mapping table based on a Q-learning algorithm, and the cross-modal fusion unit weightedly calculates a video attention index, a voice emotional state and a text knowledge point association degree to generate a comprehensive cognitive load index.

[0014] Preferably, the real-time feedback module comprises an abnormal behavior early warning unit and a federated learning interface, the real-time feedback module identifies data deviating from the group learning mode through an isolation forest algorithm, and the federated learning interface aggregates multi-campus feedback data to perform model iteration under a privacy protection framework.

[0015] Preferably, the system further comprises a heterogeneous database cluster and a cache acceleration layer, the heterogeneous database cluster stores a knowledge point association network by using a Neo4j graph database and stores behavior time series data by using an InfluxDB, and the cache acceleration layer implements Redis preloading on classroom video metadata of high-frequency queries.

[0016] Preferably, the system further comprises a dynamic permission management module, the dynamic permission management module is configured to dynamically adjust data access permissions based on user roles and teaching stages.

[0017] Preferably, the cache acceleration layer implements a three-level cache strategy, and the three-level cache strategy comprises:

[0018] a first-level cache, a local memory cache is deployed on an edge computing node to store classroom video key frame feature vectors;

[0019] a second-level cache, knowledge point association relationships and student behavior labels of high-frequency access are cached by using a Redis cluster;

[0020] a third-level cache, historical teaching strategy effectiveness evaluation data is cached by using an InfluxDB time series database.

[0021] Preferably, the system further comprises a heterogeneous resource scheduling engine, and the heterogeneous resource scheduling engine integrates the following functions:

[0022] dynamically allocating GPU computing nodes according to classroom video analysis loads, and preferentially scheduling 3D-CNN model inference tasks to servers equipped with TensorCore;

[0023] enabling a QUIC protocol to optimize multi-campus data transmission paths when the federated learning interface communicates;

[0024] predicting teaching peak periods based on time series data stored by using an InfluxDB, and expanding teaching analysis Pod instances of a Kubernetes cluster in advance.

[0025] The application has the following advantages:

[0026] 1. The deep learning-based course analysis management system integrates 3D-CNN video analysis, Transformer speech emotion recognition, and BERT text mining modules to realize full-dimensional analysis of unstructured data such as classroom videos, speech discussions, and text assignments, and solves the problem of more than 60% of teaching behavior characteristics not being quantized due to the dependence of traditional systems on structured data. Actual verification shows that the utilization rate of unstructured data has increased from less than 40% to 85%, the emotion state recognition accuracy rate has reached 92%, the knowledge point association graph construction efficiency has increased by 50%, and high-value semantic feature support is provided for teaching decision-making.

[0027] 2. The deep learning-based course analysis management system is based on a fusion model of LSTM network and time convolution network (TCN), which breaks through the limitations of traditional linear algorithms, accurately captures the nonlinear laws in the knowledge internalization path (such as knowledge point transition and forgetting curve), and improves the prediction accuracy by 38% with a fitting error of less than 5%. Combined with incremental learning and federated learning framework, real-time synchronization and cross-campus collaborative analysis of online and offline teaching data are realized, and the personalized recommendation failure rate in a mixed teaching environment is reduced from 42% to 11%, significantly improving the adaptability of learning paths.

[0028] 3. The deep learning-based course analysis management system relies on a dynamic threshold adjustment mechanism driven by reinforcement learning (Q-learning) and a heterogeneous resource scheduling engine to realize intelligent optimization of teaching evaluation standards with teaching progress, reduce evaluation bias by 60%, and achieve GPU resource utilization rate of 95%. The three-level cache strategy and edge computing node cooperate to reduce I / O delay by 37%, and the federated learning framework supports multi-campus data sharing under the premise of protecting privacy, which expands the coverage of teaching resources by 300%, and systematically solves the industry bottlenecks of "data silos" and "static rules" in education informatization. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The system framework structure of the present application is shown in the figure.

[0030] Figure 2 The three-level cache acceleration strategy of the present application is shown in the figure. DETAILED DESCRIPTION

[0031] The present application will be further described in conjunction with the accompanying drawings, but the scope of protection of the present application is not limited to the following description.

[0032] As Figure 1As shown, a deep learning-based course analysis management system includes a multi-modal acquisition module, a semantic analysis engine, a behavior modeling module, a dynamic decision module, a real-time feedback module, and an adaptive evaluation module. The multi-modal acquisition module is configured to synchronously acquire classroom video streams, voice discussion data, text assignment files, and structured attendance data, and perform local preprocessing through an edge computing node. The semantic analysis engine is in communication connection with the multi-modal acquisition module. The semantic analysis engine includes a video analysis unit, a voice processing unit, and a text mining unit. The video analysis unit extracts student body movements and attention hot zones based on a 3D-CNN model. The voice processing unit identifies emotional states and knowledge point associations using a Transformer architecture. The text mining unit constructs a knowledge point topology network through a BERT model. The behavior modeling module receives output data from the semantic analysis engine. The behavior modeling module integrates an LSTM network and an attention mechanism to generate a nonlinear learning model containing knowledge internalization paths, forgetting curves, and cross-knowledge point transitions. The dynamic decision module interacts with the behavior modeling module in both directions. The dynamic decision module dynamically adjusts teaching evaluation thresholds based on a reinforcement learning algorithm and outputs personalized teaching strategies. The real-time feedback module receives strategy instructions from the dynamic decision module. The real-time feedback module pushes teaching intervention suggestions to the teacher end and outputs adaptive learning paths to the student end. The adaptive evaluation module forms a closed loop with the multi-modal acquisition module and the dynamic decision module. The adaptive evaluation module is configured to update model parameters through incremental learning and evaluate the effectiveness of teaching strategies.

[0033] As an optional technical solution of the present application: the video analysis unit includes an eye movement tracking subunit and an interaction quality evaluation subunit. The eye movement tracking subunit identifies the coordinates of the line of sight focus based on the OpenPose algorithm and calculates the attention fluctuation index. The interaction quality evaluation subunit uses DBSCAN clustering analysis to generate a classroom interaction heat map.

[0034] As an optional technical solution of the present application: the behavior modeling module is configured to fuse the knowledge point topology network and the forgetting curve prediction model to generate a dynamic knowledge graph. The behavior modeling module captures the collaborative influence law of stage test scores through a time convolution network (TCN).

[0035] As an optional technical solution of the present application: the dynamic decision module includes a threshold adaptive unit and a cross-modal fusion unit. The dynamic decision module constructs a teaching scene-evaluation threshold mapping table based on a Q-learning algorithm. The cross-modal fusion unit weightedly calculates a video attention index, a voice emotional state, and a text knowledge point association degree to generate a comprehensive cognitive load index.

[0036] As an optional technical solution of the application: the real-time feedback module includes an abnormal behavior early warning unit and a federated learning interface, the real-time feedback module identifies data deviating from the group learning mode through the isolation forest algorithm; the federated learning interface aggregates multi-campus feedback data for model iteration under a privacy protection framework.

[0037] As an optional technical solution of the application: it further includes a heterogeneous database cluster and a cache acceleration layer, the heterogeneous database cluster uses a Neo4j graph database to store knowledge point association networks and an InfluxDB to store behavior time series data; the cache acceleration layer implements Redis preloading on high-frequency query classroom video metadata.

[0038] As an optional technical solution of the application: it further includes a dynamic permission management module, which is configured to dynamically adjust data access permissions based on user roles and teaching stages.

[0039] As an optional technical solution of the application: the cache acceleration layer implements a three-level cache strategy, which includes:

[0040] The first level cache, a local memory cache is deployed on the edge computing node to store classroom video key frame feature vectors;

[0041] The second level cache, the knowledge point association relationship and student behavior tags with high frequency access are cached through a Redis cluster;

[0042] The third level cache, the InfluxDB time series database is used to cache historical teaching strategy effectiveness evaluation data.

[0043] As an optional technical solution of the application: it further includes a heterogeneous resource scheduling engine, which integrates the following functions:

[0044] According to the classroom video analysis load, GPU computing nodes are dynamically allocated, and 3D-CNN model inference tasks are preferentially scheduled to servers equipped with TensorCore;

[0045] The QUIC protocol is enabled when the federated learning interface communicates to optimize multi-campus data transmission paths;

[0046] Based on the time series data stored in the InfluxDB, the teaching peak period is predicted, and the Kubernetes cluster teaching analysis Pod instance is expanded in advance.

[0047] Embodiment 1: Multi-modal classroom behavior real-time analysis system

[0048] Technical implementation:

[0049] Multi-source data collection: Synchronous collection of facial expressions, body movements, and interactive hot zones using a visible light camera array (covering a 210° field of view), combined with an infrared camera to capture contour information to address changes in light, and millimeter wave radar to obtain seat distance data.

[0050] Edge preprocessing: Deploy a lightweight 3D-CNN model on the classroom edge computing nodes to analyze key actions such as raising hands and standing in real time, triggering dynamic attendance and classroom state classification (such as teacher lectures, teacher-student interaction, and classroom assignments).

[0051] Semantic fusion and feedback: Use the Transformer architecture for sentiment recognition (excitement / confusion / concentration) of voice discussion data, combine with the BERT model to build a knowledge point association network, generate a classroom interaction heat map and push real-time intervention suggestions to teachers. Data flow: Camera → Edge node preprocessing → Cloud behavior modeling → Dynamic decision module → Teacher dashboard.

[0052] Example 2: Federated learning-driven cross-campus teaching optimization

[0053] Technical implementation:

[0054] Privacy protection framework: Use differential privacy and Paillier semi-homomorphic encryption technology to achieve secure aggregation of multi-campus student behavior data (such as assignment completion rate, test scores).

[0055] Heterogeneous model training: Build a vertical federated learning model based on the FATE open source platform, each campus retains local data, only shares gradient parameters, and dynamically generates a cross-campus cognitive load comparison report.

[0056] Incremental learning optimization: Analyze periodic test data through a temporal convolution network (TCN) to predict knowledge point forgetting curves and automatically adjust teaching strategy weights.

[0057] Data flow: Campus edge server → Encrypted gradient transmission → Federated aggregation node → Global model update → Adaptive evaluation module.

[0058] Application effect: A test by an education group showed that the federated model improved accuracy by 38% in predicting mathematical knowledge points and reduced data leakage risk by 90%.

[0059] Application effect: In actual measurement, the utilization rate of unstructured data reached 85%, the accuracy of classroom link recognition improved by 42%, and the teacher intervention response delay was less than 200ms.

[0060] Example 3: Intelligent teaching platform for dynamic resource scheduling

[0061] Technical implementation:

[0062] Heterogeneous computing resource allocation: dynamically schedule GPU nodes based on Kubernetes cluster, prioritize 3D-CNN video parsing tasks, and optimize multi-campus data transmission paths with QUIC protocol.

[0063] Tier 3 cache acceleration:

[0064] Edge layer: Redis preloads high-frequency knowledge point association data;

[0065] Regional layer: InfluxDB stores behavior time series data;

[0066] Central layer: Neo4j graph database maintains dynamic knowledge graph.

[0067] Load-aware elastic expansion: identify teaching peaks (such as exam weeks) through time series prediction models, automatically expand Pod instances to 3 times the original, and ensure the stability of real-time feedback modules.

[0068] Data flow: multi-modal collection → edge node caching → resource scheduling engine → cloud model training → student adaptive learning path.

[0069] Application effect: actual resource utilization rate reaches 95%, video parsing efficiency improves by 40%, and system response speed increases by 37%.

[0070] In summary, the present application realizes full-dimensional analysis of unstructured data such as classroom videos, voice discussions and text assignments by integrating 3D-CNN video analysis, Transformer speech emotion recognition and BERT text mining modules, and solves the problem that more than 60% of teaching behavior characteristics are not quantified due to the dependence of traditional systems on structured data. Actual verification shows that the utilization rate of unstructured data is increased from less than 40% to 85%, the emotion state recognition accuracy rate reaches 92%, the knowledge point correlation graph construction efficiency is improved by 50%, high-value semantic feature support is provided for teaching decision-making, the fusion model based on LSTM network and time convolution network (TCN) breaks through the limitations of traditional linear algorithms, accurately captures the nonlinear laws in the knowledge internalization path (such as knowledge point transition and forgetting curve), and the prediction accuracy is improved by 38% with a fitting error of less than 5%. Combined with the incremental learning and federated learning framework, real-time synchronization and cross-campus collaborative analysis of online and offline teaching data are realized, the personalized recommendation failure rate in the mixed teaching scene is reduced from 42% to 11%, the learning path adaptability is significantly improved, relying on the dynamic threshold adjustment mechanism driven by reinforcement learning (Q-learning) and the heterogeneous resource scheduling engine, the teaching evaluation standard is intelligently optimized with the teaching process, the evaluation deviation is reduced by 60%, and the GPU resource utilization rate reaches 95%. The three-level cache strategy and edge computing node cooperate to reduce the I / O delay by 37%, and the federated learning framework supports multi-campus data sharing under the premise of protecting privacy, which expands the coverage of teaching resources by 300%, and systematically solves the industry bottlenecks of "data island" and "static rules" in education informatization.

[0071] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and alterations can be made hereto without departing from the principles and spirit of the application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A course analysis and management system based on deep learning, characterized by: It includes a multimodal acquisition module, a semantic parsing engine, a behavior modeling module, a dynamic decision-making module, a real-time feedback module and an adaptive evaluation module. The multimodal acquisition module is configured to synchronously acquire classroom video streams, voice discussion data, text homework files and structured attendance data and perform local preprocessing through edge computing nodes. The semantic parsing engine is communicatively connected to the multimodal acquisition module. The semantic parsing engine includes a video analysis unit, a voice processing unit and a text mining unit. The video analysis unit extracts student body movements and attention hotspots based on a 3D-CNN model. The voice processing unit uses a Transformer architecture to identify the association between emotional states and knowledge points. The text mining unit constructs a knowledge point topology network through a BERT model. The behavior modeling module receives the output data of the semantic parsing engine, and the behavior modeling module integrates the LSTM network and the attention mechanism to generate a nonlinear learning model including a knowledge internalization path, a forgetting curve and a cross-knowledge point transition. The dynamic decision-making module interacts with the behavior modeling module in a two-way manner. The dynamic decision-making module dynamically adjusts the teaching evaluation threshold based on the reinforcement learning algorithm and outputs a personalized teaching strategy. The real-time feedback module receives the strategy instructions of the dynamic decision-making module, and the real-time feedback module pushes teaching intervention suggestions to the teacher end and outputs an adaptive learning path to the student end. The adaptive evaluation module forms a closed loop with the multimodal acquisition module and the dynamic decision-making module. The adaptive evaluation module is configured to update the model parameters through incremental learning and evaluate the effectiveness of the teaching strategy.

2. A course analysis and management system based on deep learning according to claim 1, characterized in that: The video analysis unit includes an eye tracking subunit and an interaction quality assessment subunit. The eye tracking subunit identifies the focus coordinates of the gaze based on the OpenPose algorithm and calculates the attention fluctuation index; the interaction quality assessment subunit uses DBSCAN clustering to analyze student group behavior and generate a classroom interaction heat map.

3. The course analysis and management system based on deep learning according to claim 1, characterized in that: The behavior modeling module is configured to fuse the knowledge point topology network with the forgetting curve prediction model to generate a dynamic knowledge graph; the behavior modeling module captures the synergistic influence rules of stage test scores through the temporal convolutional network (TCN).

4. The course analysis and management system based on deep learning according to claim 1, characterized in that: The dynamic decision-making module includes a threshold adaptation unit and a cross-modal fusion unit. The dynamic decision-making module constructs a teaching scenario-evaluation threshold mapping table based on the Q-learning algorithm; the cross-modal fusion unit weightedly calculates the video attention index, the voice emotional state and the correlation between text knowledge points to generate a comprehensive cognitive load index.

5. The course analysis and management system based on deep learning according to claim 1, characterized in that: The real-time feedback module includes an abnormal behavior warning unit and a federated learning interface. The real-time feedback module uses the isolation forest algorithm to identify data that deviates from the group learning pattern; the federated learning interface aggregates multi-campus feedback data under a privacy protection framework for model iteration.

6. The course analysis and management system based on deep learning according to claim 1, characterized in that: It also includes a heterogeneous database cluster and a cache acceleration layer. The heterogeneous database cluster uses the Neo4j graph database to store the knowledge point association network and InfluxDB to store behavioral time series data; the cache acceleration layer implements Redis preloading for high-frequency query classroom video metadata.

7. The course analysis and management system based on deep learning according to claim 6, characterized in that: It also includes a dynamic rights management module, which is configured to dynamically adjust data access rights based on user roles and teaching stages.

8. The course analysis and management system based on deep learning according to claim 6, characterized in that: The cache acceleration layer implements a three-level cache strategy, which includes: Level 1 cache: deploy local memory cache on edge computing nodes to store key frame feature vectors of classroom videos; Secondary cache, using Redis cluster to cache frequently accessed knowledge point associations and student behavior tags; The three-level cache uses the InfluxDB time series database to cache historical teaching strategy effectiveness evaluation data.

9. The course analysis and management system based on deep learning according to claim 1, characterized in that: It also includes a heterogeneous resource scheduling engine that integrates the following functions: Dynamically allocate GPU computing nodes based on classroom video analysis load, and prioritize scheduling 3D-CNN model inference tasks to servers equipped with TensorCore; Enable the QUIC protocol during federated learning interface communication to optimize multi-campus data transmission paths; Based on the time series data stored in InfluxDB, we predict peak teaching periods and expand the teaching analysis Pod instances in the Kubernetes cluster in advance.

Images