A teaching behavior pattern recognition method based on big data technology

Abstract

The present application relates to a teaching behavior pattern recognition method based on big data technology, which includes obtaining a high-concurrency data stream, performing sharding processing on the high-concurrency data stream through distributed processing technology to obtain a preliminary sharded data stream; for the preliminary sharded data stream, using real-time judgment technology to determine the data stream state to obtain an optimized sharded data stream; extracting features from the optimized sharded data stream, training a behavior recognition model through a machine learning algorithm to obtain a set of feature vectors; according to the set of feature vectors, using a pattern recognition algorithm to generate a behavior recognition result to obtain a prediction result set; processing the prediction result set through real-time control technology and an event-driven mechanism to generate an adjustment instruction set; using a closed-loop control system to monitor the adjustment instruction set to obtain a synchronous adjustment parameter set; optimizing the behavior recognition model through a reinforcement learning algorithm to generate a feedback loop data stream; extracting performance indicators from the feedback loop data stream to determine the system optimization result.

Description

Technical Field

[0001] This application relates to the technical field of electrical data processing, and particularly relates to a teaching behavior pattern recognition method based on big data technology. Background Art

[0002] Currently, many solutions still remain at the level of static data analysis, relying on post-event statistics or offline processing, and are unable to respond to the dynamic needs of learners in a timely manner. This lag causes the teaching adjustment to be out of touch with the actual state of students, limiting the potential effectiveness of intelligent teaching systems. The limitations of existing methods are mainly reflected in the lack of real-time performance and low accuracy of behavior recognition. Although many systems can collect a large amount of data, it is difficult to quickly analyze and convert it into effective intervention during the teaching process, often missing key teaching opportunities. In addition, the complexity of teaching behaviors makes the pattern recognition algorithm insufficient in adaptability in diverse scenarios, and it is prone to misjudgment or missed judgment. These defects make it difficult for the intelligent teaching environment to truly achieve dynamic adjustment and instant feedback.

[0003] The core challenge focuses on how to use electrical digital data and big data technology to solve the problems of real-time performance and accuracy in teaching behavior pattern recognition. First, real-time analysis of the learning behavior data stream requires high-efficient data processing capabilities, and current technologies often face latency bottlenecks when dealing with high-concurrency data. Second, the accuracy of pattern recognition depends on a deep understanding of complex teaching scenarios, including subtle differences in attention dispersion or learning breakthroughs. Existing algorithms are difficult to comprehensively capture these dynamic features. Without solving these two technical factors, the system cannot accurately intervene at critical moments, thus affecting the adaptive adjustment ability of the teaching process.

[0004] Therefore, how to seamlessly combine teaching behavior recognition with an instant feedback mechanism through real-time control technology to achieve the adjustment of teaching content and resources based on the dynamic state of learners has become a key issue in constructing an intelligent teaching environment. This issue not only involves the fast processing of data streams but also requires improving the ability of pattern recognition to capture key teaching moments to ensure the efficient implementation of the closed-loop control of the teaching process. Summary of the Invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the purpose of this application is to provide a teaching behavior pattern recognition method based on big data technology.

[0006] A teaching behavior pattern recognition method based on big data technology according to this application includes the following steps:

[0007] S101. Obtain a high-concurrency data stream from learning behavior data, perform sharding processing on the data through distributed stream processing technology, and use preset thresholds based on latency and throughput to determine whether the data stream meets the real-time requirement, obtaining a preliminarily processed data stream shard;

[0008] S102. For the sharded data stream after preliminary processing, use the incremental calculation method to update the data processing ability. When high-concurrency data streams arrive, allocate computing resources through a parallel computing framework to determine the data stream state after the real-time bottleneck is alleviated.

[0009] S103. From the data stream state after the real-time bottleneck is alleviated, obtain the action and time feature sets related to the behavior recognition accuracy, and train the multi-dimensional features in the complex teaching scenario through a deep learning model to obtain the feature vectors at the key teaching moments.

[0010] S104. Extract the identification information from the feature vectors at the key teaching moments, and use the attention mechanism to enhance the adaptability of the pattern recognition algorithm to the complex teaching scenario to obtain the prediction result set after the behavior recognition accuracy is improved.

[0011] S105. From the prediction result set after the behavior recognition accuracy is improved, obtain the identification information at the key teaching moments, trigger the dynamic adjustment ability by combining real-time control technology with an event-driven architecture, and generate an instruction set for teaching content adjustment.

[0012] S106. Extract the specific parameters from the instruction set for teaching content adjustment, monitor the change trend of the learning behavior data through a closed-loop control system, and obtain the adjustment parameter set synchronized with the high-concurrency data stream.

[0013] S107. Extract the execution status of the dynamic adjustment ability from the adjustment parameter set synchronized with the high-concurrency data stream, optimize the accuracy of the pattern recognition algorithm using a reinforcement learning algorithm, and generate a feedback loop data stream optimized by the closed-loop control system.

[0014] S108. Extract the delay and accuracy indicators from the feedback loop data stream optimized by the closed-loop control system, analyze and verify the execution effect of the dynamic adjustment ability, and obtain the verified data stream after the system performance is improved.

[0015] S109. Obtain the operating state of the closed-loop control system from the verified data stream after the system performance is improved, and seamlessly combine the teaching content adjustment with the instant feedback mechanism through an event-driven architecture to generate an optimized teaching parameter set for precise intervention at the key teaching moments.

[0016] Preferably, in step S101, the sharding process of the high-concurrency data stream by using the distributed processing technology includes: obtaining the sensor data stream, and sharding the sensor data stream by using the distributed stream processing technology to obtain the initial sharded data stream;

[0017] For the initial sharded data stream, judge by using a preset delay threshold. If the delay exceeds the delay threshold, mark it as a non-real-time data stream.

[0018] According to the judgment result, obtain the throughput data of the initial sharded data stream, and use a preset throughput threshold for judgment. If the throughput is lower than the throughput threshold, adjust the sharding strategy to obtain an optimized sharded data stream;

[0019] Reallocate the optimized sharded data stream through the consistent hashing algorithm to obtain a balanced sharded data stream.

[0020] Preferably, in step S102, the determining the data stream state by using the real-time judgment technology includes: obtaining an initial sharded data stream, and updating the initial sharded data stream by using an incremental calculation method to obtain a data stream with improved processing capacity;

[0021] When a high-concurrency data stream arrives, allocate computing resources through a parallel computing framework to obtain an optimized data stream;

[0022] Judge whether the optimized data stream exceeds a preset concurrency threshold. If it exceeds the concurrency threshold, adjust the resource allocation strategy;

[0023] According to the adjustment result, use stream processing technology to detect the bottleneck mitigation state to obtain a data stream after bottleneck mitigation;

[0024] Analyze the data stream after bottleneck mitigation through a state monitoring tool to obtain the final data stream state.

[0025] Preferably, in step S103, the training the behavior recognition model by using a machine learning algorithm to obtain a feature vector set includes: obtaining action features and time features to obtain an initial feature set;

[0026] Use a convolutional neural network to perform multi-dimensional feature extraction on the initial feature set to obtain a multi-dimensional feature matrix;

[0027] If the matching degree between the action feature and the time feature in the multi-dimensional feature matrix is higher than a preset threshold, train the multi-dimensional feature matrix through a long short-term memory network to obtain a behavior recognition model;

[0028] Group the critical moment candidate set through a clustering algorithm to obtain a feature vector set;

[0029] If there are outliers in the feature vector set, remove the outliers through a filtering tool to obtain an optimized feature vector set.

[0030] Preferably, in step S104, the generating the behavior recognition result by using a pattern recognition algorithm includes: obtaining teaching moment data, and extracting identification information through a preset rule to obtain a preliminary feature set;

[0031] For the preliminary feature set, the attention mechanism is used to assign weights to obtain an enhanced feature representation;

[0032] The enhanced feature representation is input into a pattern recognition algorithm to obtain behavioral features;

[0033] Determine whether the behavioral features match a preset category. If they match, a behavioral recognition result is generated through a classifier to obtain a behavioral classification set;

[0034] According to the behavioral classification set, a statistical tool is used to calculate the recognition accuracy to obtain accuracy improvement data;

[0035] Fuse the behavioral classification set and the accuracy improvement data to generate a prediction result set.

[0036] Preferably, in step S105, the processing of the prediction result set through real-time control technology and event-driven mechanism includes: obtaining the prediction result set, using behavioral recognition technology to extract identification information at critical moments to obtain preliminary classification data;

[0037] According to the preliminary classification data, use real-time control technology to match a preset event-driven architecture, and determine whether the trigger condition is met. If the trigger condition is met, determine a dynamic adjustment signal;

[0038] According to the dynamic adjustment signal, generate an adjustment ability parameter through an event-driven mechanism to obtain an adjustment requirement;

[0039] For the adjustment requirement, combine the adjustment ability parameter to generate an instruction set to obtain an adjustment plan;

[0040] Judge the integrity of the instruction set through a preset threshold to determine the final output sequence.

[0041] Preferably, in step S106, the monitoring of the adjustment instruction set by using a closed-loop control system includes: obtaining an initial parameter set, and using a closed-loop control method to obtain the real-time change trend;

[0042] Judge whether the real-time change trend exceeds a preset threshold. If it exceeds, determine the updated value of the adjustment parameter through a parameter extraction algorithm;

[0043] According to the state of the high-concurrency data stream, synchronously process the adjustment parameter and the monitoring data to obtain the synchronously processed data;

[0044] Use a support vector machine algorithm to classify the synchronously processed data to obtain an optimized parameter set;

[0045] According to the optimized parameter set, update the adjustment instruction set to obtain an updated instruction set;

[0046] The updated instruction set is monitored through the closed-loop control system to obtain a final set of adjustment parameters.

[0047] Preferably, in step S107, optimizing the behavior recognition model through the reinforcement learning algorithm includes: obtaining high-concurrency data, and through synchronous processing of the high-concurrency data, obtaining a set of adjustment parameters and an initial execution state;

[0048] Extracting dynamic adjustment ability features from the set of adjustment parameters to judge the change trend of the execution state;

[0049] Adopting a reinforcement learning algorithm to optimize the pattern recognition algorithm according to the change trend to obtain parameters with improved accuracy;

[0050] Adjusting the closed-loop control system according to the parameters with improved accuracy to generate an optimized feedback data stream;

[0051] Determining the system adjustment direction according to the optimized feedback data stream to obtain an adjusted execution state;

[0052] If the adjusted execution state does not reach the preset threshold, re-optimize the feedback data stream to obtain a stable state.

[0053] A method for recognizing teaching behavior patterns based on big data technology according to the present application has the advantages that this method performs real-time analysis on learning behavior data through distributed stream processing technology, uses a deep learning model to identify critical moments in complex teaching scenarios, and combines an attention mechanism to improve the accuracy of behavior recognition. The present invention utilizes real-time control technology and an event-driven architecture to dynamically adjust teaching content according to the recognition result, and continuously monitors the changes in learning behavior through a closed-loop control system. At the same time, the present invention applies a reinforcement learning algorithm to optimize the pattern recognition accuracy, realizing the seamless combination of teaching content adjustment and instant feedback. This method can achieve accurate recognition and timely intervention in teaching scenarios in a high-concurrency environment, effectively improving teaching quality and learning effects. Description of the Drawings

[0054] Figure 1 is the flow chart of a method for recognizing teaching behavior patterns based on big data technology according to the present application Figure 1 ;

[0055] Figure 2 is the flow chart of a method for recognizing teaching behavior patterns based on big data technology according to the present application Figure 2 。 Detailed Embodiments

[0056] As Figure 1 - Figure 2 shown, a method for recognizing teaching behavior patterns based on big data technology according to the present application.

[0057] A teaching behavior pattern recognition method based on big data technology described in this application includes the following steps:

[0058] S101. Obtain a high-concurrency data stream from learning behavior data, perform sharding processing on the data through distributed stream processing technology, and use preset thresholds based on latency and throughput to determine whether the data stream meets the real-time requirement, obtaining a preliminarily processed data stream shard;

[0059] S102. For the preliminarily processed data stream shard, adopt an incremental calculation method to update the data processing ability. When the high-concurrency data stream arrives, allocate computing resources through a parallel computing framework to determine the data stream state after the real-time bottleneck is alleviated;

[0060] S103. From the data stream state after the real-time bottleneck is alleviated, obtain the action and time feature sets related to the behavior recognition accuracy, and train the multi-dimensional features in complex teaching scenarios through a deep learning model to obtain the feature vectors at key teaching moments;

[0061] S104. Extract identification information from the feature vectors at key teaching moments, and adopt an attention mechanism to enhance the adaptability of the pattern recognition algorithm to complex teaching scenarios, obtaining a prediction result set with improved behavior recognition accuracy;

[0062] S105. From the prediction result set with improved behavior recognition accuracy, obtain the identification information at key teaching moments, trigger the dynamic adjustment ability by combining real-time control technology with an event-driven architecture, and generate an instruction set for teaching content adjustment;

[0063] S106. Extract specific parameters from the instruction set for teaching content adjustment, monitor the change trend of learning behavior data through a closed-loop control system, and obtain an adjustment parameter set synchronized with the high-concurrency data stream;

[0064] S107. Extract the execution status of the dynamic adjustment ability from the adjustment parameter set synchronized with the high-concurrency data stream, optimize the accuracy of the pattern recognition algorithm using a reinforcement learning algorithm, and generate a feedback loop data stream optimized by the closed-loop control system;

[0065] S108. Extract the latency and accuracy metrics from the feedback loop data stream optimized by the closed-loop control system, analyze and verify the execution effect of the dynamic adjustment ability, and obtain a verified data stream with improved system performance;

[0066] S109. Obtain the operating state of the closed-loop control system from the verified data stream with improved system performance, seamlessly combine teaching content adjustment with an instant feedback mechanism through an event-driven architecture, and generate an optimized teaching parameter set for precise intervention at key teaching moments.

[0067] As shown Figure 1 - Figure 2 In step S101, a high-concurrency data stream is obtained from the learning behavior data, the data is fragmented through distributed stream processing technology, and a preset threshold based on latency and throughput is used to determine whether the data stream meets the real-time requirement, obtaining a preliminarily processed data stream fragment.

[0068] Further, in step S101, a high-concurrency data stream is obtained from the learning behavior data, and the data is fragmented through distributed stream processing technology to obtain an initial fragmented data stream;

[0069] For the initial fragmented data stream, a preset latency threshold is used for judgment. If the latency exceeds the latency threshold, it is marked as a non-real-time data stream, obtaining a preliminary real-time judgment result;

[0070] According to the preliminary real-time judgment result, the throughput data of the initial fragmented data stream is obtained, and a preset throughput threshold is used for judgment to obtain a judgment result on whether the throughput meets the requirements;

[0071] If the throughput data is lower than the throughput threshold, the fragmentation strategy is adjusted, and the adjusted strategy is executed for fragmentation operation through a distributed system to obtain an optimized data stream fragment;

[0072] The stream processing operation is executed on the optimized data stream fragment through a distributed system, and the processed data is redistributed using the consistent hashing algorithm to obtain a balanced fragmented data stream;

[0073] Key features are extracted from the balanced fragmented data stream, and the key features are analyzed using the support vector machine algorithm to determine whether the real-time performance meets the requirements, obtaining a final real-time judgment result;

[0074] According to the final real-time judgment result, the fragmentation result is stored through a log recording tool, and the data processing capacity is updated using the incremental calculation method to obtain a persistent fragmented data stream;

[0075] For the persistent fragmented data stream, computing resources are allocated through a parallel computing framework. If the high-concurrency data stream arrives and exceeds the preset concurrency threshold, the resource allocation strategy is adjusted to obtain an optimized data stream;

[0076] The bottleneck mitigation state of the optimized data stream is detected through stream processing technology, and the data state is analyzed using a state monitoring tool to obtain the final data stream state.

[0077] Specifically, in step S101, a high-concurrency data stream of 100,000 per second is collected from the learning behavior data, and the KeyBy operator of Apache Flink is used to fragment the data according to the user ID, generating an initial fragmented data stream;

[0078] Use a sliding window for the initial sharded data stream, including a window size of 1 second and a sliding step of 500 milliseconds to calculate the processing delay. If the 95th percentile delay exceeds the 200 - millisecond threshold, mark this shard as a non - real - time data stream;

[0079] Based on the real - time marking results, count the throughput metrics of each shard. When the throughput of a certain shard is lower than 5000 records per second, trigger a dynamic sharding strategy adjustment, and split the original shard into 2 sub - shards by the hash modulo method;

[0080] Use the Rebalance operator of Flink to perform streaming aggregation calculations on the optimized sharded data, and then re - allocate data partitions through the consistent hashing algorithm, where the number of virtual nodes is set to 1000 to ensure that the load difference between nodes does not exceed 15%;

[0081] Extract 12 - dimensional features such as timestamp interval and data volume volatility from the balanced sharded data, input them into an SVM classifier, including using the RBF kernel function and a penalty factor C = 1.0 for real - time determination, and set the classification confidence threshold to 0.85;

[0082] Write the determination results into the Elasticsearch log cluster, and at the same time update the system processing capacity metrics based on the exponential weighted moving average method, where the smoothing coefficient α = 0.3;

[0083] When it is detected that the instantaneous concurrency breaks through 80,000 records per second, automatically trigger the YARN resource scheduler to expand the computing nodes to 1.5 times the original;

[0084] Detect the network IO utilization rate through the backpressure monitoring mechanism of Flink. When it exceeds the 75% threshold for 3 consecutive cycles, start the data localization optimization strategy, and finally output a steady - state data stream.

[0085] As Figure 1 - Figure 2 shown, in step S102, for the sharded data stream after preliminary processing, use the incremental calculation method to update the data processing capacity. When a high - concurrency data stream arrives, allocate computing resources through a parallel computing framework to determine the data stream state after the real - time bottleneck is relieved.

[0086] Furthermore, in step S102, obtain the sharded data stream after preliminary processing, and use the incremental calculation method to update the data processing capacity;

[0087] Allocate computing resources through a parallel computing framework to obtain a data stream with improved processing capacity;

[0088] Judge whether the data stream with improved processing capacity exceeds the preset concurrency threshold. If it exceeds, adjust the resource allocation strategy;

[0089] Obtain the optimized data stream according to the adjustment result of the resource allocation strategy;

[0090] Detect the bottleneck mitigation status of the optimized data stream through stream processing technology to obtain the data stream after bottleneck mitigation;

[0091] Analyze the data stream after bottleneck mitigation using a status monitoring tool to obtain the final data stream status;

[0092] Extract key metrics from the final data stream status to determine whether the system is running stably;

[0093] According to the judgment result of the stable system operation, redistribute the data stream through the consistent hashing algorithm to obtain the evenly sharded data stream;

[0094] Extract the real-time characteristics from the evenly sharded data stream and use the support vector machine algorithm to judge the status of the data stream after real-time bottleneck mitigation.

[0095] Specifically, in step S102, obtain the shards of the preliminarily processed data stream, and update the data processing capacity using the incremental calculation method, including dynamically adjusting the load of computing nodes based on the sliding window mechanism, setting the window size to 5 seconds, and the incremental factor to 0.2;

[0096] Allocate computing resources through a parallel computing framework, adopt the Spark dynamic resource scheduling strategy, set the minimum number of Executors to 4, the maximum to 16, and the memory allocation to 8GB / Executor to obtain the data stream with improved processing capacity;

[0097] Judge whether the data stream with improved processing capacity exceeds the preset concurrency threshold, including QPS≥5000. If it exceeds, adjust the resource allocation strategy and adopt the Kubernetes auto-scaling mechanism to increase the number of Pod replicas from 4 to 8;

[0098] According to the adjustment result of the resource allocation strategy, obtain the optimized data stream, with the latency reduced from 200ms to 80ms;

[0099] Detect the bottleneck mitigation status of the optimized data stream through stream processing technology, adopt the backpressure monitoring mechanism of Flink, and determine that the bottleneck is mitigated when the buffer utilization rate is lower than 70% to obtain the data stream after bottleneck mitigation;

[0100] Analyze the data stream after bottleneck mitigation using the status monitoring tool Prometheus, collect metrics such as CPU utilization ≤60% and network IO ≤1Gbps to obtain the final data stream status;

[0101] Extract key metrics from the final data stream status, including throughput volatility ≤5%, to determine whether the system is running stably;

[0102] According to the judgment result of the stable operation of the system, redistribute the data stream through the consistent hashing algorithm, set the number of virtual nodes to 200, ensure that the shard offset rate ≤ 3%, and obtain a balanced shard data stream;

[0103] Extract real-time features from the balanced shard data stream, including end-to-end delay and processing rate. Through the support vector machine algorithm, including the kernel function RBF and the penalty factor C = 1.0, train a classification model with an accuracy rate ≥ 95% to judge the data stream state after the real-time bottleneck is alleviated.

[0104] As Figure 1 - Figure 2 shown, in step S103, obtain the action and time feature sets related to the behavior recognition accuracy from the data stream state after the real-time bottleneck is alleviated, and train the multi-dimensional features in the complex teaching scenario through a deep learning model to obtain the feature vectors of the key teaching moments.

[0105] Furthermore, in step S103, extract action features and time features from the data stream state after the real-time bottleneck is alleviated to obtain an initial feature set;

[0106] Use a convolutional neural network to perform multi-dimensional feature extraction on the initial feature set to obtain a multi-dimensional feature matrix;

[0107] If the matching degree of the action features and time features in the multi-dimensional feature matrix is higher than the preset threshold, train the multi-dimensional feature matrix through a long short-term memory network to obtain a behavior recognition model for the teaching scenario;

[0108] Analyze the change trend of the action features in the multi-dimensional feature matrix according to the behavior recognition model to obtain a candidate set for critical moments;

[0109] Group the candidate set for critical moments through a clustering algorithm to obtain a set of feature vectors for critical moments;

[0110] If there are outliers inconsistent with the time features in the set of feature vectors, eliminate the outliers through a filtering tool to obtain an optimized set of feature vectors;

[0111] Verify the behavior recognition accuracy for the optimized set of feature vectors to obtain the final feature vectors;

[0112] Extract identification information from the final feature vectors, and use the attention mechanism to enhance the pattern recognition algorithm to obtain a set of prediction results with improved behavior recognition accuracy;

[0113] Analyze the multi-dimensional feature changes in the complex teaching scenario through the set of prediction results to determine the feature vectors of the key teaching moments.

[0114] Specifically, in step S103, action features and time features are extracted from the data stream state after real-time bottleneck mitigation. The OpenPose algorithm is used to detect the coordinates of human key points, and the initial feature set is constructed by combining timestamp information. The feature dimension is 25×3, including 25 key points, and each point contains x, y coordinates and confidence;

[0115] A convolutional neural network is used to perform multi-dimensional feature extraction on the initial feature set. The ResNet-50 architecture is used to extract spatial features, and a 128-dimensional feature vector is output to form a multi-dimensional feature matrix;

[0116] If the matching degree between the action features and time features in the multi-dimensional feature matrix is higher than the preset threshold, including cosine similarity > 0.85, the multi-dimensional feature matrix is trained through a bidirectional LSTM network with 64 hidden layer units, and the behavior recognition model of the teaching scenario is output;

[0117] According to the behavior recognition model, the change trend of action features in the multi-dimensional feature matrix is analyzed, and the mean square error of the feature difference between adjacent frames is calculated. The frames with MSE > 0.1 are marked as the candidate set of critical moments;

[0118] The DBSCAN clustering algorithm is used to group the candidate set of critical moments, with the neighborhood radius ε = 0.5 and the minimum number of samples min_samples = 3, to obtain the set of feature vectors at critical moments;

[0119] If there are outliers inconsistent with the time features in the set of feature vectors, including isolated points with a time interval exceeding 5 seconds, the Kalman filter is used to eliminate the outliers to obtain an optimized set of feature vectors;

[0120] The behavior recognition accuracy of the optimized set of feature vectors is verified. The feature vectors with F1-score > 0.9 calculated by cross-validation (k = 5) are used as the final output;

[0121] The identification information is extracted from the final feature vectors, and the multi-head attention mechanism (number of heads = 8) is used to enhance the pattern recognition algorithm, with an output dimension of 256, to obtain the prediction result set with improved behavior recognition accuracy;

[0122] The multi-dimensional feature changes in the complex teaching scenario are analyzed through the prediction result set, and the dynamic time warping (DTW) distance of the feature vectors is calculated to determine the feature vectors at critical teaching moments.

[0123] As Figure 1 - Figure 2 shown, in step S104, the identification information is extracted from the feature vectors at critical teaching moments, and the attention mechanism is used to enhance the adaptability of the pattern recognition algorithm to the complex teaching scenario, to obtain the prediction result set with improved behavior recognition accuracy.

[0124] Further, in step S104, identification information is extracted from the feature vectors of critical teaching moments, and an attention mechanism is used to enhance the adaptability of the pattern recognition algorithm to complex teaching scenarios, obtaining a prediction result set with improved behavior recognition accuracy;

[0125] According to the identification information in the prediction result set, an action and time feature set related to behavior recognition accuracy is obtained, and preliminary classification data of multi-dimensional features is determined;

[0126] The action and time feature set is trained through a deep learning model to obtain the feature vectors of critical teaching moments in complex teaching scenarios;

[0127] Identification information of critical moments is extracted from the feature vectors, and the improved degree of behavior recognition accuracy is judged using the data stream state after real-time bottleneck mitigation;

[0128] If it is judged that the behavior recognition accuracy meets the preset threshold, the start signal for dynamic adjustment is determined by matching the event-driven architecture through real-time control technology;

[0129] According to the start signal, an adjustment ability parameter is generated using the event-driven mechanism, and an instruction set corresponding to the adjustment requirement of teaching content is obtained;

[0130] The integrity of the instruction set is judged using the preset threshold through the core fields of the instruction set, obtaining the final output sequence;

[0131] After obtaining the final output sequence, the accuracy of the behavior recognition technology is optimized using machine learning algorithms, obtaining an updated prediction result set;

[0132] According to the updated prediction result set, the identification information of critical moments is processed in a loop to obtain the next round of teaching content adjustment instructions.

[0133] Specifically, in step S104, identification information is extracted from the feature vectors of critical teaching moments, and a Transformer model based on the multi-head attention mechanism is used to enhance the features of the teaching video frame sequence. The number of attention heads is set to 8, the hidden layer dimension is 512, and the weight distribution is calculated through Softmax normalization, obtaining a prediction result set with the behavior recognition accuracy improved to 92.3%;

[0134] According to the timestamp identification in the prediction result set, action features are extracted at 0.5-second intervals using the sliding window algorithm, combined with the time series features output by the LSTM network, forming a 128-dimensional feature vector containing joint angles and movement speeds, and determining the preliminary classification data;

[0135] Train the feature set using the ResNet-50 deep learning model, set the batch size to 32, the learning rate to 0.001, use the cross-entropy loss function to iterate 100 rounds, and output a 2048-dimensional feature vector for the key teaching moments;

[0136] Extract the TOP-10 significant region identifiers from the feature vectors, use the Kalman filter algorithm to process the data stream delay. When the recognition frame rate stabilizes at 25 FPS, determine that the improvement amplitude of the behavior recognition accuracy exceeds the preset threshold of 15%;

[0137] If the accuracy meets the standard, adjust the event trigger frequency through a PID controller to match the event-driven architecture based on the Kafka message queue. When the message throughput reaches 5000 messages per second, generate a start signal;

[0138] According to the signal strength parameter, use the decision tree model to generate adjustment weights, set the learning concentration threshold to 70%, and output 12 types of instruction templates including knowledge point reorganization and speech rate adjustment;

[0139] By parsing the keyword frequencies in the instructions, use the TF-IDF algorithm to calculate the field weights. When the integrity score exceeds 0.85, generate the final JSON format output sequence;

[0140] Input the sequence into the XGBoost model for incremental training, adjust the feature selection parameter to the optimal split point to increase the behavior recognition F1 value to 94.6%;

[0141] Based on the updated confusion matrix results, re-extract the key frame identifiers with a timestamp error less than 0.2 seconds, and trigger the generation of a new round of dynamic adjustment instructions.

[0142] As Figure 1 - Figure 2 shown, in step S105, obtain the identification information of the key teaching moments from the prediction result set after the improvement of the behavior recognition accuracy, trigger the dynamic adjustment ability through the combination of real-time control technology and the event-driven architecture, and generate an instruction set for teaching content adjustment.

[0143] Furthermore, in step S105, obtain the prediction result set, use the behavior recognition technology to process the prediction result set, extract the identification information of the critical moments, and obtain the preliminary classification data;

[0144] According to the preliminary classification data, use the real-time control technology to match the preset event-driven architecture, judge whether the preset trigger conditions are met, and determine the start signal for dynamic adjustment;

[0145] If the start signal for dynamic adjustment is determined, generate the adjustment ability parameters through the event-driven mechanism, and obtain the adjustment requirements for the teaching content;

[0146] Generate a corresponding instruction set based on the adjustment requirements and the adjustment ability parameters to obtain an adjustment plan for the teaching content;

[0147] Extract the core fields of the instruction set from the adjustment plan of the teaching content, judge the integrity of the instruction set through a preset threshold, and determine the final output sequence;

[0148] After obtaining the final output sequence, use a machine learning algorithm to optimize the accuracy of the behavior recognition technology to obtain an updated prediction result set;

[0149] According to the updated prediction result set, use the attention mechanism to enhance the pattern recognition algorithm to process complex teaching scenarios, extract the feature vectors of key teaching moments, and obtain a prediction result set with improved behavior recognition accuracy;

[0150] Extract the identification information of critical moments from the prediction result set with improved behavior recognition accuracy, and generate adjustment ability parameters by combining real-time control technology with an event-driven architecture;

[0151] For the adjustment ability parameters, generate a teaching optimization parameter set by combining the operating state of the closed-loop control system to obtain an instruction set with precise intervention at key teaching moments.

[0152] Specifically, in step S105, after obtaining the prediction result set, use a behavior recognition model based on YOLOv5 to analyze the teaching behaviors in the video stream, set a confidence threshold of 0.85 to extract key frames, and divide the behavior data into 5 types of teaching scenarios through a time series clustering algorithm to obtain preliminary classified data with timestamps;

[0153] According to the teacher-student interaction frequency parameter in the classified data, use the PID control algorithm to match the rule engine in the event-driven architecture, and when it is detected that the interaction frequency is lower than the preset threshold of 0.6 for 3 consecutive times, trigger a dynamic adjustment signal;

[0154] Publish an adjustment instruction through the event bus, use a weighted moving average algorithm to calculate the change rate of behavior characteristics in the last 10 seconds, and generate an ability parameter matrix including the adjustment range (±15%) and direction (content difficulty / presentation speed);

[0155] Combine the metadata index of the teaching resource library, use the Dijkstra algorithm to search for the optimal adjustment path in the knowledge graph, and generate an instruction set including 3 alternative solutions;

[0156] Extract the core fields in the instruction set, including the knowledge point ID and the presentation duration, calculate the semantic integrity score through the BERT model, and output the final sequence when the score exceeds 0.92;

[0157] The incremental random forest algorithm is adopted to update the recognition model with 2,000 newly collected behavior data, and the key behavior recognition accuracy is improved from 86% to 91%.

[0158] The multi-head attention mechanism (8 heads) is applied to the updated prediction set to compress the dimension of the feature vector (from 512 dimensions to 128 dimensions), enhancing the recognition ability for complex scenarios such as group discussions.

[0159] Key segments with a time window of 5 seconds are extracted from the optimized feature vectors, and the behavior trajectory prediction is corrected in real time through a Kalman filter to generate adjustment parameters including confidence intervals.

[0160] Finally, 30 operation indicators in the closed-loop system are fused, including latency < 50ms and throughput > 200 QPS, and fuzzy logic inference is used to generate a teaching optimization parameter set including priority weights (0 - 1).

[0161] As Figure 1 - Figure 2 shown, in step S106, specific parameters are extracted from the instruction set for teaching content adjustment, and the change trend of learning behavior data is monitored through a closed-loop control system to obtain an adjusted parameter set synchronized with the high-concurrency data stream.

[0162] Furthermore, in step S106, specific parameters are extracted from the instruction set for teaching content adjustment to obtain an initial parameter set;

[0163] The learning behavior data is collected in real time through a closed-loop control system to obtain the real-time change trend;

[0164] It is judged whether the real-time change trend exceeds the preset threshold to determine whether to trigger parameter adjustment;

[0165] If parameter adjustment is triggered, a parameter extraction algorithm is used to calculate the updated value to obtain the adjustment parameters;

[0166] The adjustment parameters and system monitoring data are synchronously processed according to the state of the high-concurrency data stream to obtain the synchronously processed data;

[0167] The support vector machine algorithm is used to classify the synchronously processed data to obtain an optimized parameter set;

[0168] The instruction set of teaching content is updated with the optimized parameter set to obtain the updated instruction set;

[0169] The updated instruction set is monitored through a closed-loop control system to obtain the final adjusted parameter set synchronized with the data stream;

[0170] According to the final adjusted parameter set, combined with behavior recognition technology, key moment identification information is extracted to generate an instruction set for the next round of adjustment.

[0171] Specifically, in step S106, specific parameters are extracted from the instruction set for teaching content adjustment, including parameters such as learning duration, knowledge point difficulty coefficient, and interaction frequency parsed from the instruction set to obtain an initial parameter set;

[0172] The learning behavior data is collected in real time through a closed-loop control system, including data such as the correct rate of students' answers, learning duration distribution, and number of interactions, to obtain a real-time change trend;

[0173] It is determined whether the real-time change trend exceeds a preset threshold, including setting that the correct rate of answering questions is lower than 60% or the learning duration exceeds 120 minutes as abnormal to determine whether to trigger parameter adjustment;

[0174] If parameter adjustment is triggered, a parameter extraction algorithm is used to calculate the updated value, including optimizing the knowledge point difficulty coefficient through the gradient descent method to obtain the adjusted parameters;

[0175] The adjusted parameters and system monitoring data are synchronously processed according to the state of the high-concurrency data stream, including matching the adjusted parameters with the real-time monitoring data in the case of processing 1000 data streams per second to obtain the synchronously processed data;

[0176] The support vector machine algorithm is used to classify the synchronously processed data, including classifying the data into three categories of high, medium, and low learning effects to obtain an optimized parameter set;

[0177] The instruction set of the teaching content is updated through the optimized parameter set, including adjusting the knowledge point difficulty coefficient from 0.8 to 0.6 to obtain the updated instruction set;

[0178] The updated instruction set is monitored through a closed-loop control system, including real-time monitoring whether the correct rate of students' answers has increased to 70% to obtain the final adjusted parameter set synchronized with the data stream;

[0179] According to the final adjusted parameter set, key moment identification information is extracted in combination with behavior recognition technology, including identifying that the learning efficiency peak appears at the 15th minute and the 45th minute of the students to generate an instruction set for the next round of adjustment.

[0180] As Figure 1 - Figure 2 shown, in step S107, the execution state of the dynamic adjustment ability is extracted from the adjusted parameter set synchronized with the high-concurrency data stream, and the accuracy of the pattern recognition algorithm is optimized using the reinforcement learning algorithm to generate an optimized feedback loop data stream of the closed-loop control system.

[0181] Furthermore, in step S107, high-concurrency data is obtained, and through synchronous processing of the high-concurrency data, an adjusted parameter set and an initial execution state are obtained;

[0182] Extract the dynamic adjustment ability features from the set of adjustment parameters, and judge the change trend of the execution state;

[0183] Adopt a reinforcement learning algorithm to optimize the pattern recognition algorithm according to the change trend, and obtain the parameters with improved accuracy;

[0184] Adjust the closed-loop control system according to the parameters with improved accuracy, and generate an optimized feedback data stream;

[0185] Determine the system adjustment direction according to the optimized feedback data stream, and obtain the adjusted execution state;

[0186] If the adjusted execution state does not reach the preset threshold, re-optimize the feedback data stream according to the set of adjustment parameters to obtain an updated feedback data stream;

[0187] Judge whether the system reaches a stable state according to the updated feedback data stream, and obtain a stable state determination result;

[0188] Update the synchronous processing process of the high-concurrency data according to the stable state determination result, and generate a new data stream of the closed-loop control system;

[0189] Extract the adjusted dynamic adjustment ability features from the new data stream, and judge the optimized execution state of the closed-loop control system.

[0190] Specifically, in step S107, when obtaining high-concurrency data, use the distributed message queue Kafka to process 100,000 real-time data streams per second, align the data timestamps through the time window synchronization algorithm, and generate a set of adjustment parameters including delay thresholds and throughput metrics and the initial execution state;

[0191] When extracting the dynamic adjustment ability features from the set of adjustment parameters, calculate whether the standard deviation of data delay within the sliding window exceeds 50ms, and judge the change trend in combination with the fluctuation range of CPU utilization of the execution state;

[0192] Adopt a reinforcement learning algorithm based on DQN, set the reward function as the percentage increase in pattern recognition accuracy, and iteratively optimize the number of convolutional kernels and learning rate of the convolutional neural network through Q-learning, so that the classification accuracy is improved from 92% to 96%;

[0193] Adjust the proportional coefficient of the PID controller of the closed-loop control system according to the optimized parameters, and generate a feedback data stream including error correction values;

[0194] Analyze the slope of the error convergence curve in the feedback data stream, determine that the system adjustment direction is to increase the weight of the integral link, and obtain the adjusted execution state;

[0195] If the steady-state error of the execution state is still higher than the preset threshold of 0.5%, the genetic algorithm is used to perform multi-objective optimization on the feedback data stream. The crossover probability is set to 0.8 and the mutation probability is 0.1 to generate an updated feedback data stream;

[0196] Calculate the variance value of the updated data stream through Kalman filtering. When the variance is less than 0.01 for 5 consecutive iterations, it is determined that the system reaches a stable state;

[0197] Based on the stable state determination result, reconstruct the data synchronization processing thread pool, dynamically adjust the number of threads from 200 to 150, and generate a data stream containing new sampling period parameters;

[0198] When extracting the adjusted dynamic adjustment ability characteristics from the new data stream, use the LSTM network to predict the execution state in the next 3 cycles and output the response time distribution histogram after system optimization.

[0199] As Figure 1 - Figure 2 shown, in step S108, extract the delay and accuracy metrics from the optimized feedback loop data stream of the closed-loop control system, and verify the execution effect of the dynamic adjustment ability through analysis to obtain the verified data stream after system performance improvement.

[0200] Further, in step S108, obtain the data stream in the optimized feedback loop of the closed-loop system, extract the delay metric and accuracy metric from the data stream to obtain a preliminary quantification result;

[0201] Use statistical methods to calculate the distribution characteristics of the delay metric and accuracy metric in the preliminary quantification result to determine the execution effect of the dynamic adjustment;

[0202] If the delay metric exceeds the preset threshold, adjust the parameters of the feedback loop to obtain an optimized data stream;

[0203] Extract the new delay metric and accuracy metric from the optimized data stream, and use machine learning algorithms to judge the change trend of the system performance;

[0204] According to the change trend, obtain the verification basis for system performance improvement by comparing the differences between the new and old accuracy metrics;

[0205] After obtaining the verification basis, use a preset classification algorithm to judge the stability of the performance improvement to obtain the final quantified data stream;

[0206] By analyzing the final quantified data stream, determine the optimized control ability of the closed-loop system and obtain the verification result of the execution effect;

[0207] Extract the operating state of the closed-loop control system from the verification result and generate a teaching optimization parameter set through an event-driven architecture;

[0208] According to the teaching optimization parameter set, adjust the teaching content through an instant feedback mechanism to obtain the data stream after precise intervention at key teaching moments.

[0209] Specifically, in step S108, obtain the data stream in the optimized closed-loop system feedback loop, and use a sliding window algorithm to extract the delay metrics in the data packet at intervals of 10 ms, including an average delay of 15 ms and the accuracy metric, including a classification accuracy rate of 95%, to obtain a preliminary quantization result including timestamps and values.

[0210] Use statistical methods to calculate the distribution characteristics of the delay and accuracy metrics, including using a normal distribution to test the mean of the delay metric (μ = 18 ms, σ = 3 ms) and the confidence interval of the accuracy metric (95% ± 2%) to determine whether the execution effect of the dynamic adjustment meets the expectations.

[0211] If it is detected that the delay metric exceeds the preset threshold, including 20 ms, then adjust the proportional gain parameter of the feedback loop (Kp decreases from 1.2 to 0.8) through a PID controller to generate an optimized data stream.

[0212] Extract the updated delay metric (12 ms) and accuracy metric (96%) from the new data stream, and use an LSTM model to predict the performance change trend within the next 5 cycles (error rate < 1%).

[0213] According to the trend analysis results, compare the difference between the new and old accuracy metrics (+1%), and use a t-test (p < 0.05) to verify the significance of the performance improvement.

[0214] Input the verification basis into a random forest classifier, and based on the feature importance ranking, including a delay weight of 0.6 and an accuracy weight of 0.4, determine the stability level (Level A) and output the final quantified data stream.

[0215] Perform dimensionality reduction processing on the final data stream through principal component analysis (PCA), extract the first two principal components (cumulative contribution rate of 85%) to determine the optimization control ability boundary of the closed-loop system (delay < 15 ms, accuracy rate > 94%).

[0216] Analyze the system operation status from the verification results, including a load rate of 70%, trigger the teaching parameter generation module through an event-driven architecture, and output a teaching optimization parameter set including the adjustment range (±5%) and response time (200 ms).

[0217] Based on the parameter set, call a reinforcement learning strategy (ε-greedy algorithm, ε = 0.1) to dynamically adjust the teaching content push frequency (increase from 1 time / minute to 1.2 times / minute) to generate an intervened data stream for the next round of closed-loop optimization.

[0218] As Figure 1 - Figure 2 shown, in step S109, the operating state of the closed-loop control system is obtained from the verified data stream after the system performance is improved, and the teaching content adjustment and the instant feedback mechanism are seamlessly combined through the event-driven architecture to generate an optimized teaching parameter set after precise intervention at the critical teaching moment.

[0219] Further, in step S109, the operating state data of the closed-loop control system is extracted from the verified data stream after the system performance is improved to obtain the system performance indicators after data stream parsing;

[0220] According to the dynamic change characteristics in the operating state data, it is judged whether the critical teaching moment is triggered;

[0221] If the critical teaching moment is triggered, the dynamic change characteristics are processed through the event-driven architecture to determine the adjustment direction of the teaching content;

[0222] An intervention strategy set is generated by integrating the instant feedback mechanism according to the adjustment direction;

[0223] The execution effect at the critical moment is analyzed for the intervention strategy set to obtain a preliminary set of optimized parameters;

[0224] The preliminary set is classified through the random forest algorithm to obtain an optimized teaching parameter set;

[0225] It is judged whether the matching degree of the optimized teaching parameter set is lower than the preset threshold to obtain the matching degree analysis result;

[0226] If the matching degree is lower than the preset threshold, the optimized teaching parameter set is iteratively adjusted through the gradient descent algorithm to obtain an adjusted parameter set;

[0227] The adjusted parameter set is monitored through the closed-loop control system to obtain the final optimized teaching parameter set synchronized with the data stream.

[0228] Specifically, in step S109, the operating state data of the closed-loop control system is extracted from the verified data stream after the system performance is improved, and the data stream parsing algorithm is adopted, including sliding window analysis, with a window size of 5 seconds and a sampling frequency of 100 Hz to calculate the system performance indicators, including response delay, average delay ≤ 50 ms, and throughput ≥ 1000 pieces / second;

[0229] According to the dynamic change characteristics in the operating state data, including that the fluctuation range of the learning behavior data exceeds ±15%, the threshold judgment model with a set threshold of 0.7 is used to detect whether the critical teaching moment is triggered;

[0230] If a critical teaching moment is triggered, through an event-driven architecture, including Kafka message queues to handle dynamic change characteristics, combined with LSTM time series analysis to determine the direction of teaching content adjustment, including reducing the difficulty coefficient by 0.2;

[0231] According to the adjustment direction, integrate an instant feedback mechanism, including real-time collection of students' answer correct rates with a sampling interval of 2 seconds, and generate an intervention strategy set containing 3 strategies, including adjusting the explanation speed, increasing the number of example questions, and switching media forms;

[0232] For the intervention strategy set, adopt an effect evaluation model, including weighted scoring based on accuracy, time consumption, and interaction rate, with weights of 0.5, 0.3, and 0.2 respectively, to analyze the execution effect at critical moments and output a preliminary set of optimization parameters, including a parameter range of [0.1, 0.9];

[0233] Set 100 decision trees with a maximum depth of 10 layers through the random forest algorithm to classify the preliminary set, and screen out a teaching optimization parameter set, including an optimal parameter combination of 0.6 / 0.8 / 0.4;

[0234] Use cosine similarity calculation to judge the matching degree of the teaching optimization parameter set, with a threshold of 0.85. If the matching degree is lower than the threshold, apply the gradient descent algorithm with a learning rate of 0.01 and iterate 50 times to adjust the parameter set until the error rate ≤ 5%;

[0235] Finally, through the PID controller of the closed-loop control system, with a proportional coefficient of 1.2 and an integral time of 0.5 seconds, monitor the adjusted parameter set to achieve the final teaching optimization parameter set synchronized with the data stream, including a dynamic adjustment frequency of 1Hz.

[0236] For those skilled in the art, various corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all such changes and deformations should fall within the protection scope of the claims of this application.

Claims

1. A teaching behavior pattern recognition method based on big data technology, characterized in that, It includes the following steps: Perform sharding processing on the high-concurrency data stream of learning behavior data through distributed stream processing technology, calculate the latency quantile value in real time based on the dynamic sliding window mechanism. When the 95% quantile latency exceeds the preset threshold of 200ms, mark the non-real-time data stream, and dynamically adjust the sharding strategy according to the throughput threshold to generate an optimized sharded data stream; Adopt the incremental calculation method combined with the smoothing coefficient α = 0.3 of the exponential weighted moving average method to dynamically adjust the processing capacity of the sharded data stream, and trigger the dynamic resource scheduling strategy through the parallel computing framework. When the instantaneous concurrency exceeds 80,000 records per second, automatically expand the computing nodes to 1.5 times the original resources to obtain a balanced sharded data stream; Extract action features and time features from the balanced sharded data stream, use the OpenPose algorithm to detect the coordinates of human key points, construct an initial feature set in combination with timestamp information, and use a convolutional neural network to perform multi-dimensional feature extraction on the initial feature set to form a multi-dimensional feature matrix; If the matching degree between the action features and time features in the multi-dimensional feature matrix is higher than the preset threshold, train through a bidirectional LSTM network with 64 hidden layer units to output a behavior recognition model for the teaching scenario; analyze the change trend of the action features in the multi-dimensional feature matrix according to the behavior recognition model, calculate the mean square error of the feature difference between adjacent frames, and mark the frames with MSE > 0.1 as the candidate set for critical moments; use the DBSCAN clustering algorithm to group the candidate set for critical moments and adopt Kalman filtering to eliminate outliers to generate a set of feature vectors for critical teaching moments; Perform weight allocation on the set of feature vectors based on the 8 attention heads of the multi-head Transformer model, generate an enhanced feature representation analysis through Softmax normalization to improve the behavior recognition accuracy, and output a prediction result set; Combine real-time control technology and event-driven architecture, trigger dynamic adjustment instructions according to the prediction result set, and generate an instruction set for teaching content adjustment, including obtaining the prediction result set, using behavior recognition technology to extract identification information for critical moments to obtain preliminary classification data; According to the teacher-student interaction frequency parameter in the preliminary classification data, use the PID control algorithm to match the rule engine in the event-driven architecture. When it is detected that the interaction frequency is lower than the preset threshold of 0.6 for 3 consecutive times, calculate the adjustment weight through a decision tree model and generate a JSON format instruction set including knowledge point reorganization and speech rate adjustment to trigger dynamic adjustment of teaching content.

2. The teaching behavior pattern recognition method based on big data technology according to claim 1, characterized in that Monitor the execution status of the instruction set through a closed-loop control system to synchronously adjust parameters to match the change trend of the high-concurrency data stream; Adopt a reinforcement learning algorithm to optimize the accuracy of the behavior recognition model to obtain a feedback loop data stream; Extract the latency index and accuracy index from the feedback loop data stream to verify the improvement effect of system performance; Generate a teaching optimization parameter set according to the verification result and use it for intervention at critical teaching moments through the event-driven architecture.

3. The teaching behavior pattern recognition method based on big data technology according to claim 1, characterized in that, The initial sharded data stream is redistributed through the consistent hashing algorithm. If the latency exceeds the preset threshold or the throughput is lower than the threshold, the sharding strategy is dynamically adjusted. The sharding strategy adjustment adopts the consistent hashing redistribution algorithm with 1000 virtual nodes. The support vector machine algorithm is used to perform real-time classification on the sharded data stream, and the classification confidence threshold is set to 0.

85. The parallel computing framework adopts a dynamic resource scheduling strategy, including: When it is detected that the instantaneous concurrency exceeds 80,000 requests per second, the computing nodes are triggered to expand to 1.5 times the original resources. The data localization strategy is adjusted through the stream processing backpressure monitoring mechanism to make the network IO utilization rate lower than 75%.

4. The teaching behavior pattern recognition method based on big data technology according to claim 2, characterized in that The reinforcement learning algorithm is the deep Q network DQN, including: The percentage increase in pattern recognition accuracy is used as the reward function to optimize the number of convolutional kernels and the learning rate of the convolutional neural network. When the steady-state error exceeds 0.5%, the genetic algorithm is used to perform multi-objective optimization on the feedback data stream.

5. The teaching behavior pattern recognition method based on big data technology according to claim 2, wherein The system performance verification includes: The LSTM model is used to predict the trends of latency and accuracy metrics within the next 5 cycles. The t-test with p < 0.05 is used to verify the significance of performance improvement, and a quantitative data stream with a stability level of A is output.

6. The teaching behavior pattern recognition method based on big data technology according to claim 2, wherein The generation of the teaching optimization parameter set includes: The optimal parameter combination is screened through the random forest algorithm, and the matching degree threshold is set to 0.

85. If the matching degree does not reach the threshold, the gradient descent algorithm is used to iteratively adjust the parameters until the error rate ≤ 5%.

Images