A large model-based intelligent teaching supervision evaluation method and system

Abstract

The present application relates to the technical field of neural networks, in particular to a wisdom teaching supervision evaluation method and system based on a large model, comprising the following steps: collecting teacher multi-source data and labelizing and normalizing processing, constructing a time sequence behavior set and a label order sequence, clustering high-frequency concurrent behaviors and marking conflicts, extracting change trend adjustment boundaries, sorting and scoring jump dominant labels, connecting behaviors to establish a scoring linkage path, and outputting a teaching supervision evaluation scheme. In the present application, through multi-source data label construction and time sequence integration, the structural reorganization and dynamic expression of teaching behaviors are realized, the behavior classification accuracy is improved by combining behavior frequency clustering and conflict identification, key behavior characteristics are extracted by means of trend ratio and score jump alignment, the scoring influence factor ordering and behavior flow chain are established, the chain mapping of teaching behavior and score change is realized, the evaluation discrimination ability and data response efficiency are enhanced, and the accuracy of teaching behavior identification and the clarity of score mechanism analysis are improved.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to a smart teaching supervision and evaluation method and system based on a large model. Background Technology

[0002] The field of neural network technology involves artificial intelligence technologies that simulate the structure and information processing mechanisms of the human nervous system. Core aspects include the construction, training, and application of artificial neural networks, encompassing deep neural networks, convolutional neural networks, recurrent neural networks, graph neural networks, and their specific implementations in image recognition, natural language processing, predictive modeling, and intelligent control. Overall, it belongs to intelligent computing systems that simulate neurons and their connections to process complex data and abstract features. The development of this technology has driven the widespread application of large-scale pre-trained models, big data-driven intelligent systems, and adaptive learning mechanisms, gradually expanding into industries such as education, healthcare, and finance. Traditional intelligent teaching supervision and evaluation methods refer to addressing the subjectivity and limitations of teaching evaluation by combining indicator-based scoring with expert observation during the process of monitoring teaching quality and assessing teacher behavior. This is typically achieved through methods such as recording and collecting teaching process data, setting evaluation dimension indicators, and quantifying expert scores. By employing large-scale neural network models to extract features and perform semantic analysis on teaching behavior data, and combining this with data sources such as teacher classroom audio and video and teaching text content, intelligent judgment and structured expression of teaching supervision data can be completed, thereby achieving complex cognitive behavior modeling of the teaching process and analysis of multi-dimensional data correlations.

[0003] Existing technologies rely on indicator setting and human judgment. In behavior recognition, they depend on data collection and manual scoring for quantification. Since scoring is based on subjective standards, behavior labels do not have a clear structural sequence relationship. The evaluation content is difficult to cover the continuous changes and concurrent situations of teaching behavior. When faced with complex scenarios such as high-density behavior clusters or score jumps, it is difficult to effectively extract key behavioral features and time node correlations. In scenarios with large score fluctuations, evaluation is prone to ambiguity. Behavioral patterns lack clear direction and cannot support refined analysis and the establishment of causal chains. This leads to generalized evaluation conclusions, one-sided behavior analysis, and insufficiently targeted feedback suggestions. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a smart teaching supervision and evaluation method based on a large model.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a smart teaching supervision and evaluation method based on a large model, comprising the following steps:

[0006] S1: Collect teacher's voice, operation behavior and action trajectory data, label the behavior according to the start and end time of the event, call the voice duration, trajectory change range and number of instructions for normalization processing, and splice them into a behavior set according to the time series to construct a label sorting sequence;

[0007] S2: Based on the continuous time position of the label sorting sequence, the frequency of label occurrence in adjacent periods is counted, high-frequency concurrent behaviors within the period are clustered, conflict behavior combinations are extracted and marked according to the mutual exclusion rule, and conflict shielding label fragment groups are output.

[0008] S3: Using the temporal behavior content of the conflict masking tag fragment group, construct the ratio feature of tag change frequency to the number of pause behaviors, extract the behavior change trend and compare it with adjacent time periods, align it by the position of jump change nodes, adjust the boundary, and form a time period boundary structure.

[0009] S4: Extract the behavior and rating sequence from the time period boundary structure, calculate the fluctuation range of the rating curve and locate the jump point, extract all behavior labels within the range, sort them according to the difference between the frequency of occurrence in the segment and the overall frequency, and output the set of dominant behaviors of the jump segment.

[0010] As a further embodiment of the present invention, the label sorting sequence includes a behavior type label sequence, a standardized time node sequence, and a periodic behavior pattern; the conflict shielding label fragment group includes a structural conflict label group, a set of shielding identifier sites, and concurrent behavior clustering results; the time period boundary structure includes behavior density feature parameters, jump node alignment boundaries, and adjusted segment divisions; and the jump fragment dominant behavior set includes a score mutation area label set, a dominant behavior type list, and label frequency difference sorting results.

[0011] As a further aspect of the present invention, the step of obtaining the label sorting sequence is as follows:

[0012] S101: Based on the teacher's voice, operation behavior and action trajectory data in the classroom, each type of behavior is labeled according to the start and end time of the event, and the duration of the voice, the trajectory change range and the number of instruction events are extracted to generate a set of behavior parameters.

[0013] S102: Based on the set of behavioral parameters, normalize the speech duration, trajectory change range and number of command events to generate a periodic parameter normalization sequence.

[0014] S103: Call the normalized sequence of the periodic parameters, concatenate them in chronological order to construct a multidimensional vector set, and perform combined analysis based on the correlation between the sorting trend of behavioral parameters, trajectory changes and speech span to obtain the label sorting sequence.

[0015] As a further aspect of the present invention, the step of obtaining the conflict masking tag fragment group is as follows:

[0016] S201: Based on the continuous time positions in the label sorting sequence, calculate the frequency of occurrence of labels in adjacent periods, identify labels with changing frequency differences, and obtain a label frequency difference sequence.

[0017] S202: Call the label frequency difference sequence, extract the label combinations that occur frequently and concurrently within the period, perform type clustering based on the label co-occurrence relationship, and generate label combination cluster groups;

[0018] S203: Based on the clustering group of the label combinations, call the mutual exclusion judgment rule of teaching behavior, identify the label combinations with structural conflict characteristics, mark the corresponding periodic positions, and obtain the conflict shielding label fragment group.

[0019] As a further aspect of the present invention, the step of obtaining the time period boundary structure is as follows:

[0020] S301: Based on the temporal behavior content in the conflict shielding tag segment group, the change frequency and the number of pause behaviors of each type of tag in the segment are statistically analyzed, the ratio between the two is calculated, and a tag pause ratio sequence is generated.

[0021] S302: Call the tag pause ratio sequence, extract the trend curve, compare it with the trend of the previous and next time periods, calculate the difference of the continuous trend, and obtain the difference value of the behavior change trend;

[0022] S303: Based on the difference value of the behavior change trend, call the sequence of jump change node positions, determine the alignment offset value of the trend jump position, and synchronously adjust the segment boundary position to obtain the segment change amplitude sequence and establish the time period boundary structure.

[0023] As a further aspect of the present invention, the step of obtaining the set of dominant behaviors of the transition segment is as follows:

[0024] S401: Based on the time period boundary structure, extract the behavior label sequence and the rating sequence within the segment, calculate the continuous fluctuation range between the node values ​​of the rating curve, determine the position of the jump point in the fluctuation range, and obtain the set of rating jump positions.

[0025] S402: Call the set of scoring jump locations, extract all teacher behavior type labels within the corresponding jump segment, count the frequency of the labels in the segment, and obtain the behavior frequency distribution result of the jump segment.

[0026] S403: Based on the frequency distribution results of the jump segment behavior, compare the difference between the tag frequency within the segment and the overall tag frequency, calculate the difference ranking index and perform weighted judgment in combination with the behavior distribution characteristics, calculate and obtain the frequency offset intensity value of the tag, sort and filter the dominant items to obtain the set of dominant behaviors of the jump segment.

[0027] As a further aspect of the present invention, the method further includes:

[0028] S5: Based on the time position of the dominant behavior set of the jump segment, arrange adjacent behaviors in the order of first appearance and connect them linearly to establish a teaching action flow path sequence on the time axis, perform chain linkage mapping between action content and score changes, and output a smart teaching supervision and evaluation scheme.

[0029] The intelligent teaching supervision and evaluation scheme includes a teaching action flow map, a scoring change linkage mapping path, and a time series event chain structure.

[0030] The steps for obtaining the intelligent teaching supervision and evaluation scheme are as follows:

[0031] S501: Based on the set of dominant behaviors of the jump segment, extract the first appearance position of each type of label in the time axis, arrange all label items in chronological order, construct the label time progression sequence, and obtain the behavior label time sorting;

[0032] S502: Invoke the behavior label time sorting, construct a linear connection path according to the time interval of adjacent labels, and combine them in sequence to form a continuous behavior flow path to generate a teaching action flow path sequence;

[0033] S503: Based on the teaching action flow path sequence, track the time mapping relationship between labels and scoring jump points in the path, analyze the linkage degree between labels and scoring responses, calculate and obtain the scoring linkage response value of the path, and construct a dynamic mapping mechanism between labels and scores based on the linkage degree to obtain a smart teaching supervision and evaluation scheme.

[0034] A smart teaching supervision and evaluation system based on a large model includes:

[0035] The behavior cycle splicing module obtains the teacher's voice length, number of operation commands and trajectory range. After segmenting the voice according to the event time, it extracts the voice according to energy, sorts the operations according to the cumulative number, calculates the cumulative offset value of the trajectory, and splices it according to time after labeling to obtain the label sorted sequence.

[0036] The concurrent conflict detection module calls the label sorting sequence to count the number of labels in the time interval, compares the periodic distribution and uses mutual exclusion rules to determine the label combination, filters out the ones that need to be blocked and marks the time, and obtains the conflict blocking label fragment group;

[0037] The trend jump alignment module extracts label changes based on the conflict shielding label fragment group, calculates the sequence of change counts and pause ratios, compares adjacent differences, and adjusts the positions according to jump nodes to obtain the time period boundary structure.

[0038] The scoring band extraction module calls the labels and scoring sequences in the time period boundary structure, measures the scoring fluctuation and locates the jump, extracts the jump range labels and sorts them according to the difference between the number of times and the total number, and obtains the set of dominant behaviors of the jump segment.

[0039] The flow link establishment module establishes a link by first sorting the dominant behaviors in the jump segment set and connecting the tag time series to form a link, thereby obtaining a smart teaching supervision and evaluation scheme.

[0040] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0041] In this invention, by constructing and integrating multi-source data labels over time, the structure of teaching behavior is reorganized and dynamically expressed. By combining behavior frequency clustering and conflict identification, the accuracy of behavior classification is improved. Key behavioral features are extracted by aligning trend ratios and score jumps. A ranking of score influencing factors and a behavior flow chain are established to realize a chain mapping between teaching behavior and score changes, thereby enhancing evaluation and judgment capabilities and data response efficiency, and improving the accuracy of teaching behavior identification and the clarity of score mechanism analysis. Attached Figure Description

[0042] Figure 1 This is a flowchart of the main steps of the present invention;

[0043] Figure 2 This is a flowchart of the present invention, S1.

[0044] Figure 3 This is a flowchart of the S2 process of the present invention;

[0045] Figure 4 This is a flowchart of the S3 process of the present invention;

[0046] Figure 5 This is a flowchart of the S4 process of the present invention;

[0047] Figure 6 This is a flowchart of the S5 process of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0049] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0050] Please see Figure 1 A smart teaching supervision and evaluation method based on a large model includes the following steps:

[0051] S1: Collect teachers' voice, operation behavior and action trajectory data in the classroom, label each type of behavior according to the start and end time of the event in each cycle, call the voice duration, action trajectory change range and instruction event number to perform normalization processing, splice the processing results according to the time series and uniformly construct the behavior set content, and construct the label sorting sequence.

[0052] S2: Based on the continuous time position in the label sorting sequence, the frequency of label occurrence in adjacent periods is statistically analyzed. The high-frequency concurrent behavior type combinations in each period are clustered and divided. Then, the behavioral label combinations with structural conflicts are extracted through the teaching behavior mutual exclusion judgment rule, and the positions are marked and assigned masking labels. The conflict masking label fragment group is output.

[0053] S3: Using the time behavior content in the conflict masking label fragment group, construct a ratio feature group according to the frequency of change of each type of label and the number of pause behaviors in the segment, extract the behavior change trend and compare it with the change trend in the previous and next time periods, call the jump change node position sequence to make front and back alignment judgment, perform adjustment operation on the segment boundary, and construct the time period boundary structure.

[0054] S4: Extract the behavior label sequence and score sequence from the time period boundary structure, calculate the continuous fluctuation range between the node values ​​of the score curve, locate the score jump point, extract all teacher behavior type labels within the range, sort the difference between the frequency of the label in the segment and the overall label frequency, and output the set of dominant behaviors of the jump segment.

[0055] S5: Use the time position corresponding to each type of label in the set of dominant behaviors of the jump segment, arrange the behavior labels according to the first appearance time of the label, and linearly connect the time progression order between adjacent behaviors to establish the teaching action flow path sequence in the time axis, complete the chain linkage mapping from action content to score change, and output the intelligent teaching supervision and evaluation scheme.

[0056] The label sorting sequence includes behavior type label sequence, standardized time node sequence, and periodic behavior pattern; the conflict shielding label segment group includes structural conflict label group, shielding marker site set, and concurrent behavior clustering results; the time period boundary structure includes behavior density feature parameters, jump node alignment boundary, and adjusted segment division; the jump segment dominant behavior set includes score mutation area label set, dominant behavior type list, and label frequency difference sorting results; the smart teaching supervision and evaluation scheme includes teaching action flow map, score change linkage mapping path, and time series event chain structure.

[0057] Please see Figure 2 The steps to obtain the label sorting sequence are as follows:

[0058] S101: Based on the teacher's voice, operation behavior and action trajectory data in the classroom, each type of behavior is labeled according to the start and end time of the event, and the duration of the voice, trajectory change range and number of instruction events are extracted to generate a set of behavior parameters.

[0059] Based on the teacher's voice, operational behavior, and motion trajectory data in the classroom, the recorded complete lecture audio is first divided into 10ms frames. Each frame is then judged as valid speech based on an energy threshold of 0.6W. If the frame energy is greater than 0.6W, it is marked as a speech frame. Speech segments are identified by the start and end frames of consecutive speech frames. The start and end times of each speech segment are calculated to obtain its duration. Operational behavior is collected from the teaching system's interaction log. The system records the timestamps and operation types of events such as clicking on courseware, playing videos, or switching PPTs. If the interval between operations is less than 5 seconds, it is considered a continuous trigger segment of the same behavior. The start and end times of each operation are combined and marked as the time segment of that behavior. For motion trajectories, three positioning beacons are deployed on the podium and in the classroom to collect three-dimensional positioning data from the positioning device worn by the teacher. The coordinate position is recorded once per second. If the sum of three consecutive sets of coordinate displacements exceeds 0.5 meters and the direction changes by more than 30 degrees, it is marked as a trajectory. The change range is defined by setting the displacement value as a trajectory change threshold of 0.5 meters and the direction threshold as 30 degrees, and recording the start and end times of the trajectory change. Command events are generated by converting speech content into text using speech recognition technology. The system pre-sets 15 standard command phrases, such as "Look here," "Read together," and "Turn to the next page." Text paragraphs are matched using keyword comparison. Each successful match of a standard phrase is recorded as a command event. The number of command events associated with each speech paragraph is recorded in the paragraph's label. For example, if a speech segment lasts 13 seconds from 09:02:10 to 09:02:23, matching 2 standard commands, and the teacher moves 1.1 meters, changes direction by 45 degrees, and performs two consecutive clicks during this time period, then the behavior label would be "13 seconds," the trajectory change range "1.1 meters," the number of command events "2," and the number of operation events "2." All of this information is mapped to the time period from 09:02:10 to 09:02:23, forming the set of behavior parameters for that period.

[0060] S102: Based on the set of behavioral parameters, normalize the speech duration, trajectory change range and number of command events to generate a normalized sequence of periodic parameters;

[0061] The acquired set of behavioral parameters is normalized. First, a normalization interval is set based on the historical maximum and minimum values ​​of each parameter type. The speech duration ranges from 1 second to 20 seconds in the collected samples, and the normalized length value ranges from [0, 1]. The trajectory displacement ranges from a minimum of 0.2 meters to a maximum of 2.0 meters, and the number of command events ranges from a minimum of 0 to a maximum of 5. Normalization is based on the minimum value, and the current value is linearly mapped to the interval to obtain the normalized value. For example, if the current speech duration is 13 seconds, the normalized value is (13-1) / (20-1) = 0.632; if the trajectory change is 1.1 meters, the normalized value is (1.1-0.2) / (2.0-0.1) = 0.632. .2) = 0.5, the number of instruction events is 2, and the corresponding normalized result is (2-0) / (5-0) = 0.4; to ensure the stability of the normalization process, the normalization error tolerance is set to no more than 0.01. For values ​​less than 0 or greater than 1 after normalization, they are processed according to the nearest boundary value; each normalized value is stored in the periodic vector dataset. Each row corresponds to a behavior record period, including three normalization results, and the corresponding time start is marked. For example, the normalized vector of the above example record is [0.632, 0.5, 0.4], and the corresponding time is 09:02:10. This process is continuously executed until all periodic behavior parameters are normalized, forming a time series vector set as input for subsequent analysis.

[0062] S103: Call the normalized sequence of periodic parameters, concatenate them in chronological order to construct a multidimensional vector set, and perform combined analysis based on the correlation between the sorting trend of behavioral parameters, trajectory changes and speech span to obtain the label sorting sequence;

[0063] The normalized periodic behavior vectors are concatenated in chronological order to form a multi-dimensional vector set. Each periodic vector corresponds to three dimensions, representing the normalized value of speech continuity, the normalized value of trajectory change, and the normalized value of the number of commands, respectively. The system compares the direction and magnitude of change in each dimension between adjacent vectors cycle by cycle, extracting the increase or decrease relationship of each dimension. If the absolute value of the change in a certain dimension is less than 0.05, it is considered "stable"; if it is greater than 0.05, it is considered "increasing" or "decreasing". For example, if the speech dimension is 0.543 in the previous cycle and 0.632 in the next cycle, the change value is +0.089, which is classified as "increasing". In each cycle, the three-dimensional change trend is combined into a set of labels, such as "increasing-stable-decreasing", and a priority reference table is set for the labels corresponding to each change combination. If the priority of "speech rising, trajectory stable, instruction falling" is higher than that of "speech falling, trajectory falling, instruction rising", the system builds a label sorting dictionary based on this. After obtaining the changing combinations, the system assigns sorting label numbers according to the dictionary. For example, in the current period, the combination "rising-stable-falling" corresponds to the label sequence {speech, trajectory, instruction} in the dictionary, so the output sorting sequence is this label order. To build a reasonable dictionary, the system initially assigns 1 point to each combination. Then, it collects data from 300 lessons and counts the frequency of each label sorting. The higher the frequency, the higher the score. Finally, the label sorting table is updated through statistics. The final label sorting result is stored in the corresponding period record as a numbered sequence, completing the generation of the label sorting sequence for this segment.

[0064] Please see Figure 3 The steps for obtaining conflict masking tag fragment groups are as follows:

[0065] S201: Based on the continuous time position in the label sorting sequence, calculate the frequency of the label in adjacent periods, identify the labels with frequency differences, and obtain the label frequency difference sequence.

[0066] Based on the continuous time positions in the label sorting sequence, the frequency of occurrence of each type of label in the label sequence is extracted periodically. The time window length is set to 5 periods. The label types "voice", "trajectory" and "instruction" are used as monitoring dimensions. The frequency of each type of label is counted in each time window. For example, in periods 1 to 5, voice appears 3 times, trajectory appears 2 times and instruction appears 4 times, with frequencies of 0.6, 0.4 and 0.8 respectively. In periods 6 to 10, voice appears 5 times, trajectory appears 3 times and instruction appears 1 time, with frequencies of 1.0, 0.6 and 0.2 respectively. Then the difference between adjacent windows is calculated, and the frequency difference of voice is 0.4, trajectory is 0.2 and instruction is -0. 6. Classify all label calculation results. If the absolute value of the frequency difference is greater than 0.5, it is defined as a "significantly changing label"; if it is less than or equal to 0.5, it is a "stationary label". Here, 0.5 is set as the frequency difference judgment threshold. It comes from the fact that 90% of the frequency fluctuation range in the previous sample analysis is concentrated within ±0.5. Therefore, using 0.5 as the judgment standard is repeatable and reasonable. For example, the instruction label change is -0.6, which meets the "significant change" judgment rule. This label is identified as a label with frequency difference change. This identification process is repeated once for the label frequency in each time window. Finally, the frequency difference value is recorded in chronological order to form a label frequency difference value sequence.

[0067] S202: Call the label frequency difference sequence, extract the label combinations that occur frequently and concurrently within the period, perform type clustering based on the label co-occurrence relationship, and generate label combination cluster groups;

[0068] The label frequency difference sequence is called, and the labels in the periodic segments with small frequency fluctuations are analyzed to determine whether two or more types of labels appear synchronously more than 3 times in consecutive periods. If so, they are recorded as high-frequency concurrent combinations. For example, in periods 12 to 16, the labels "voice" and "trajectory" appear together in each period, with corresponding occurrence counts of [1, 1, 1, 1, 1], which meets the concurrent occurrence condition. This combination is extracted and recorded as a high-frequency label combination. All combination samples are sorted by concurrency intensity, and the co-occurrence judgment threshold is set to "average concurrent occurrence greater than 3 times". This threshold is derived from the average co-occurrence count of highly correlated label pairs in the sample training set, which is 3.12, and rounded to 3 after statistical analysis. In the selected high-frequency combinations, the start and end positions, label composition, and concurrent count of each combination in the periodic sequence are extracted. Then, the string composed of label names is used for mapping and classification, such as "voice-track" and "track-instruction" as the basis for clustering. During the clustering process, if there are identical labels among the label combinations or the absolute value of the frequency difference is less than 0.2, they are defined as belonging to the same class and grouped into a cluster group. The difference threshold of 0.2 in the clustering judgment condition is set according to the distribution range of the average difference of frequency value changes after label standardization. Finally, combinations with label homogeneity and similar frequency characteristics are clustered into one group. For example, if the frequency difference between "voice-track" and "voice-track-instruction" is small, they are grouped into the same label combination cluster group.

[0069] S203: Based on the label combination cluster group, call the mutual exclusion judgment rule of teaching behavior, identify the label combination with structural conflict characteristics, mark the corresponding period position, and obtain the conflict shielding label fragment group;

[0070] Based on the label combination clusters, the mutual exclusion judgment rules for teaching behaviors are called in each group to check the behavioral logic between the labels in each group and determine whether there are structural conflict behaviors. The mutual exclusion rule set contains 12 prohibited label combinations. Sample mutual exclusion combinations include "explaining-stillness", "asking questions-large movements", and "instruction-pause operation". If any of these combinations appears in any cluster, it is judged as a structural conflict combination, and the start and end positions of the cluster in the original periodic sequence are recorded. For example, the combination "voice-trajectory-instruction" is identified in periods 22 to 24. The “voice” indicates the lecturing state, the “trajectory” displacement exceeds 1.5 meters, and the “instruction” appears simultaneously, constituting a logical conflict of “lecturing-large movement-command overlap”. The system uses this rule to determine that the segment is a conflict segment, setting the minimum displacement threshold for action conflict to 1.2 meters. That is, if the periodic displacement corresponding to the trajectory label is greater than 1.2 meters, it is considered a large movement. This threshold is derived by referring to the median value of the average movement behavior of teachers in classroom recordings, which is 1.1 meters, plus 0.1 meters, and has the basis of repeated measurement. The system uniformly labels all identified conflict segments to form a conflict masking label segment group.

[0071] Please see Figure 4 The steps for obtaining the time-boundary structure are as follows:

[0072] S301: Based on the temporal behavior content in the conflict-masking tag segment group, statistically analyze the change frequency and the number of pause behaviors of each type of tag in the segment, calculate the ratio between the two, and generate a tag pause ratio sequence.

[0073] Based on the temporal behavior content in the conflict masking tag segment group, the start and end cycle numbers of each conflicting segment are read one by one. Within each segment, the system statistically analyzes three types of tags: voice, trajectory, and command. The statistical method is as follows: if the value of a certain type of tag changes between two adjacent cycles, it is recorded as one tag change. The total number of changes for that type of tag is accumulated and then divided by the number of cycles covered by the segment to obtain the unit cycle change frequency of the tag. For example, if segment 1 spans 5 cycles, with voice state changing 3 times, trajectory state changing 1 time, and command state changing 2 times, for a total of 6 changes, then the total tag change frequency of this segment is 6 ÷ 5 = 1.2 times / cycle. At the same time, the pause behavior within the segment is also statistically analyzed. A pause is defined as a certain... The total number of changes in the three types of labels during a period is either 0 or 1. The system sets "less than 1 time" as the pause judgment threshold. This threshold is derived from the analysis of the lower limit of label activity fluctuation in the initial sample set. 90% of the pause periods have no more than 1 label change, so 1 is taken as the standard threshold. For example, in the above segment, there is only 1 label change in periods 1 and 3, so the total pause behavior is 2 times. Finally, the system divides the total label change frequency by the number of pauses to obtain the label pause ratio. This ratio reflects the density of label activity in a unit pause behavior. If the ratio is too low, it indicates that the label changes are sparse. For example, if the ratio is less than 1.0, it means that there is only one change behavior before each pause. The system executes the ratio calculation process one by one in the segment number and forms a label pause ratio sequence for subsequent use.

[0074] S302: Call the tag pause ratio sequence, extract the trend curve, compare it with the trend of the previous and next time periods, calculate the difference of the continuous trend, and obtain the difference value of the behavior change trend.

[0075] The system invokes the tag pause ratio sequence and establishes a time-varying curve for the tag pause ratio based on the segment number order. The numerical difference between the ratios of adjacent segments is used as the trend indicator. The trend change between each pair of adjacent segments is the absolute value of the difference between their ratios. For example, if the ratio of segment 2 is 6.0 and the ratio of segment 1 is 3.0, then the trend is |6.0−3.0|=3.0. The system recursively calculates the ratio difference between all adjacent segments in this way and summarizes these differences to construct a trend change sequence. After constructing the trend change curve, the system further averages the trend differences between the current segment and its preceding and following segments to obtain a continuous trend change. The overall magnitude, for example, the differences before and after segment 2 are 3.0 and 3.75 respectively, with a mean of 3.375; to determine whether this mean reaches the set threshold of "significant trend change", the system sets this judgment threshold to 2.5. This value comes from the statistical results of the trend change mean of 1000 sample segments, 90% of which are concentrated between 2.0 and 2.5. The system takes the upper limit of 2.5 as the standard judgment threshold. If the mean trend difference of the current segment is higher than this threshold, it is marked as a "trend change segment", otherwise it is a stable trend segment. After all segments have undergone the above comparison process, the system outputs the trend change identification result and trend difference value of each segment, forming the final behavioral change trend difference value.

[0076] S303: Based on the difference value of behavioral change trend, call the sequence of jump change node positions, determine the alignment offset value of the trend jump position, and synchronously adjust the segment boundary position to obtain the segment change amplitude sequence and establish the time period boundary structure.

[0077] Based on the difference value of behavioral change trends, the system compares the corresponding positions of each trend change segment in the jump change node sequence. The node sequence is a list of previously identified cycles of drastic behavioral changes. For example, if a segment is numbered 2, corresponding to cycles 20 to 24, and the system finds relevant nodes in the node sequence for cycles 23 and 25, then it compares this node with the segment boundary cycle. If the offset between the jump node cycle and the boundary cycle does not exceed 2 cycles, it is recorded as a "boundary-alignable interval"; otherwise, it is recorded as a "boundary-offset interval." The system sets the offset judgment threshold to 2 cycles. This threshold is set based on the fact that the vast majority of teachers' teaching behavior changes occur within 3 cycles, and statistics show that 82% of the boundary deviations are actually within 2 cycles. Since the shift is within 2 periods, the value is repeatable and verifiable. For example, if the current segment boundary is between periods 21 and 24, and the nodes are at 23 and 25, then the front boundary shifts by 2 and the back boundary shifts by 1. The system synchronously adjusts the boundaries based on the node positions, correcting the starting boundary from 21 to 23 and the ending boundary from 24 to 25. After correction, the segment span is 3 periods, while the original span was 4 periods. The system records the change magnitude as 1 and stores this change magnitude in the corresponding segment structure. Following this rule, the system synchronously adjusts the boundaries of all trend-jumping segments and outputs the corrected segment period range and change magnitude for each segment, ultimately establishing a complete time period boundary structure for subsequent structural analysis of behavioral segments.

[0078] Please see Figure 5 The steps to obtain the set of dominant behaviors for jump fragments are as follows:

[0079] S401: Based on the time period boundary structure, extract the behavioral label sequence and rating sequence within the segment, calculate the continuous fluctuation range between the node values ​​of the rating curve, determine the position of the jump point in the fluctuation range, and obtain the set of rating jump positions.

[0080] Based on the time-boundary structure, the system extracts the behavior label sequence and scoring sequence within the defined boundaries segment by segment. The label sequence is the teacher's behavior category record in each cycle. For example, a segment consists of cycles 10 to 15, and the label order is "voice, voice, trajectory, instruction, voice, instruction". At the same time, the scoring sequence is extracted as "8.3, 8.0, 7.5, 6.9, 7.1, 8.2". The scoring sequence is composed of classroom observation scores, with one score recorded for each cycle. The system calculates the difference between any two adjacent scoring nodes in the scoring sequence, and the continuously fluctuating difference sequence is "0". The sequence is 3, 0.5, 0.6, 0.2, 1.1. A threshold of 1.0 is set in this difference sequence, meaning that if a difference is greater than or equal to 1.0, it is considered a scoring jump point. Analysis of standard course scoring fluctuation data shows that 90% of this value falls within ±0.9. Therefore, setting 1.0 as the boundary is more universally applicable. The fifth pair of scoring differences in the current sequence, 1.1, meets the jump condition, corresponding to period 14 to 15. The system records the jump position as period 15. After all segments are calculated in this way, the jump node positions are output one by one, forming a set of scoring jump positions for behavior frequency statistics.

[0081] S402: Call the set of rating jump locations, extract all teacher behavior type labels within the corresponding jump segment, count the frequency of the labels within the segment, and obtain the behavior frequency distribution results of the jump segment.

[0082] The system calls the set of scoring jump locations and sequentially locates the boundary structure segment where each scoring jump point is located. It extracts the teacher behavior tags that appear in all cycles covered by the segment and counts the cumulative occurrence of each type of tag. For example, if a scoring jump location is in cycle 15, the cycle is located in segment number 3, the corresponding cycle of the segment is from 10 to 15, and the tag sequence is "voice, voice, trajectory, instruction, voice, instruction". The system counts voice as 3 times, trajectory as 1 time, and instruction as 2 times, for a total of 6 tags. The system builds a record dictionary by tag name, with voice = 3, trajectory = 1, and instruction = 2. It also records the segment number and the total number of tags, which is 6. The above operation is repeated for all scoring jump locations. Each record is bound to three pieces of information: "segment number - tag name - occurrence count". The system further marks the frequency sequence for subsequent offset intensity analysis, and generates a complete tag frequency distribution structure for each segment.

[0083] S403: Based on the frequency distribution results of the transition segment behavior, compare the difference between the tag frequency within the segment and the overall tag frequency, calculate the difference ranking index, and combine it with the behavior distribution characteristics for weighted judgment, using the formula:

[0084] ;

[0085] The frequency offset intensity value of the tag is obtained by calculation, sorted and the dominant items are filtered to obtain the set of dominant behaviors of the jump segment;

[0086] The parameters are defined as follows:

[0087] : No. Frequency offset strength value of class behavior label;

[0088] : The frequency of this tag's occurrence within the transition section;

[0089] Total number of tags within the transition segment;

[0090] The frequency of this tag throughout the entire time period;

[0091] Total number of tags throughout the entire time period;

[0092] The coefficient of variation (dimensionless, defined as) of the rating fluctuation of this label within the jump range. );

[0093] The influence index of the rating change within the jump range of this label (dimensionless after normalization).

[0094] The normalized value (dimensionless, usually defined as the mean of the behavior intervals within the transition zone) of this label. ).

[0095] Calculation Example 1 (labeled "speech")

[0096] The known parameters are as follows:

[0097] ;

[0098] ;

[0099] Step 1: Calculate the absolute value of the frequency difference

[0100] ;

[0101] Step 2: Multiply by the coefficient of variation

[0102] ;

[0103] Step 3: Calculate the factors affecting the score

[0104] ;

[0105] Step 4: Summing to obtain the offset intensity value

[0106] ;

[0107] Calculation Example 2 (labeled "Trajectory")

[0108] The known parameters are as follows:

[0109] ;

[0110] ;

[0111] Absolute value of frequency difference:

[0112] ;

[0113] Multiply by the coefficient of variation:

[0114] ;

[0115] Factors affecting the rating:

[0116] ;

[0117] Total offset intensity value:

[0118] ;

[0119] Filtering rules

[0120] Calculate for all tags After assigning values, sort them from highest to lowest, and take the label with the highest value as the dominant behavior label for the transition segment. For example, if the offset intensity value of the "speech" label is the highest among all labels, then the dominant behavior of the transition segment is "speech".

[0121] Please see Figure 6 The steps to obtain the smart teaching supervision and evaluation scheme are as follows:

[0122] S501: Based on the set of dominant behaviors of jump segments, extract the first appearance position of each type of label in the timeline, arrange all label items in chronological order, construct the label time progression sequence, and obtain the behavior label time sorting;

[0123] Based on the set of dominant behaviors in the jump segments, the system extracts the dominant label type one by one and records the cycle number in which it is first identified as the dominant behavior on the time axis. For example, the "voice" label first appears in cycle 5, "instruction" first appears in cycle 9, and "trajectory" first appears in cycle 12. The system establishes a mapping dictionary with the label as the key and the cycle number of the first appearance as the value. Then, all labels are sorted in chronological order to obtain a time-progressing label arrangement sequence, namely "voice-instruction-trajectory". This sequence represents the actual guiding order of the labels. This order is not based on the label frequency, but strictly judged according to the first appearance time. If two labels first appear in the same cycle, the order is judged according to the priority rule. The system priority rule is "voice > instruction > trajectory". This rule is derived from the average value of the label's influence coefficient on the score in the sample statistics. The first appearance of all labels is processed and sorted using this rule. Finally, a complete behavior label time sorting sequence is generated for subsequent behavior flow path construction.

[0124] S502: Sort the behavior labels by time, construct a linear connection path according to the time interval of adjacent labels, combine them in sequence to form a continuous behavior flow path, and generate a sequence of teaching action flow paths;

[0125] The system calls the behavior label time sorting sequence and calculates the difference between the first occurrence period of two adjacent labels in the sort to construct the time path of the label flow path. For example, if the label sequence is "voice-instruction-trajectory", its corresponding period is 5, 9, 12. The system calculates that the time interval from voice to instruction is 4 periods and from instruction to trajectory is 3 periods, forming the connection path "voice → instruction (Δ4) → trajectory (Δ3)". The system splices each connection according to the structure to generate a complete teaching action flow path sequence. If there are multiple dominant labels that appear repeatedly in the label path, such as "voice" being the dominant label again in period 18, the system adds a backlink connection "trajectory → voice (Δ6)" in the path to record the backtracking behavior path. The system performs this connection construction on the dominant labels in the entire teaching process and marks and numbers all connection segments in the path to finally generate a teaching action flow path sequence. This sequence can be used to track the time relationship between behavior response and score.

[0126] S503: Based on the teaching action flow path sequence, trace the time mapping relationship between labels and scoring transition points in the path, analyze the degree of linkage between labels and scoring responses, and use the formula:

[0127] ;

[0128] The scoring linkage response value of the obtained path is calculated, and a dynamic mapping mechanism between tags and scores is constructed based on the degree of linkage to obtain a smart teaching supervision and evaluation scheme.

[0129] in, This represents the score-linked response value for the r-th teaching behavior path. This represents the set of all behavior labels contained in path r. This represents the rating jump value corresponding to label m. This indicates the time span corresponding to the label. This represents the difference in rating trends before and after label m. This indicates the breadth of rating fluctuation in the segment containing label m. This indicates the repetition interval of label m within the path. This represents the average time interval between labels within the path.

[0130] Based on the teaching action flow path sequence, the system performs a scoring linkage analysis on the tag set contained in each path, using the scoring linkage response value formula:

[0131] ;

[0132] Taking the path "voice → command → trajectory" as an example, its tag set is: The parameters for each tag are set as follows:

[0133] Voice tag: rating jump value Time span (Periods 5 to 9), rating trend difference Rating fluctuation breadth Repeat time interval within the path Average time interval ;

[0134] Instruction tags: , , , , , ;

[0135] Track tags: , , , , , .

[0136] The calculation steps for substituting the phonetic terms are as follows:

[0137] Part One ;

[0138] Part Two ;

[0139] Part Three ;

[0140] The total is .

[0141] The calculation for the instruction item is as follows:

[0142] ;

[0143] ;

[0144] ;

[0145] The total is .

[0146] The trajectory term is calculated as follows:

[0147] ;

[0148] ;

[0149] ;

[0150] The total is .

[0151] The sum of the three items yields the score-linked response value for this path:

[0152] ;

[0153] This value represents the strength of the linkage between behavioral changes and score changes within the teaching path. A smaller value indicates a weaker influence of the behavior within that path on the score. The system determines the score based on the behavior within each path. The values ​​are sorted, and a list of path-linked response strengths is output, establishing a dynamic mapping structure between path behavior and scoring. The results indicate that the overall scoring response for the path "voice → command → trajectory" is weak.

[0154] A smart teaching supervision and evaluation system based on a large model includes:

[0155] The behavior cycle splicing module obtains the teacher's voice length, number of operation commands and trajectory range. After segmenting the voice according to the event time, it extracts the voice according to energy, sorts the operations according to the cumulative number, calculates the cumulative offset value of the trajectory, and splices it according to time after labeling to obtain the label sorted sequence.

[0156] The concurrent conflict detection module calls the label sorting sequence to count the number of labels in the time interval, compares the periodic distribution and uses mutual exclusion rules to determine the label combination, filters out the ones that need to be blocked and marks the time, and obtains the conflict blocking label fragment group;

[0157] The trend jump alignment module extracts label changes based on conflict-masking label fragment groups, calculates the sequence of changes and pause ratios, compares adjacent differences, and adjusts the positions according to jump nodes to obtain the time period boundary structure.

[0158] The scoring band extraction module calls the labels and scoring sequences in the time period boundary structure, measures the scoring fluctuations and locates the jumps, extracts the jump range labels and sorts them by the difference between the number of times and the total number, and obtains the set of dominant behaviors of the jump segments.

[0159] The flow link establishment module establishes a link by first sorting the dominant behaviors in the jump segment set and connecting the label time series to form a link, thereby obtaining a smart teaching supervision and evaluation scheme.

[0160] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A smart teaching supervision and evaluation method based on a large model, characterized in that, Includes the following steps: S1: Collect teacher's voice, operation behavior and action trajectory data, label the behavior according to the start and end time of the event, call the voice duration, trajectory change range and number of instructions for normalization processing, and splice them into a behavior set according to the time series to construct a label sorting sequence; S2: Based on the continuous time position of the label sorting sequence, the frequency of label occurrence in adjacent periods is counted, high-frequency concurrent behaviors within the period are clustered, conflict behavior combinations are extracted and marked according to the mutual exclusion rule, and conflict shielding label fragment groups are output. S3: Using the temporal behavior content of the conflict masking tag fragment group, construct the ratio feature of tag change frequency to the number of pause behaviors, extract the behavior change trend and compare it with adjacent time periods, align it by the position of jump change nodes, adjust the boundary, and form a time period boundary structure. S4: Extract the behavior and rating sequence from the time period boundary structure, calculate the fluctuation range of the rating curve and locate the jump point, extract all behavior labels within the range, sort them according to the difference between the frequency of occurrence in the segment and the overall frequency, and output the set of dominant behaviors of the jump segment. The steps for obtaining the time period boundary structure are as follows: S301: Based on the temporal behavior content in the conflict-blocking tag segment group, the change frequency and the number of pause behaviors of each type of tag in the segment are statistically analyzed, the ratio between the two is calculated, and a tag pause ratio sequence is generated. S302: Call the tag pause ratio sequence, extract the trend curve, compare it with the trend of the previous and next time periods, calculate the difference of the continuous trend, and obtain the difference value of the behavior change trend; S303: Based on the difference value of the behavior change trend, call the sequence of jump change node positions, determine the alignment offset value of the trend jump position, and synchronously adjust the segment boundary position to obtain the segment change amplitude sequence and establish the time period boundary structure.

2. The intelligent teaching supervision and evaluation method based on a large model according to claim 1, characterized in that, The label sorting sequence includes a behavior type label sequence, a standardized time node sequence, and a periodic behavior pattern. The conflict shielding label fragment group includes a structural conflict label group, a set of shielding marker sites, and concurrent behavior clustering results. The time period boundary structure includes behavior density feature parameters, jump node alignment boundaries, and adjusted segment divisions. The jump fragment dominant behavior set includes a score mutation area label set, a list of dominant behavior types, and label frequency difference sorting results.

3. The intelligent teaching supervision and evaluation method based on a large model according to claim 2, characterized in that, The steps for obtaining the label sorting sequence are as follows: S101: Based on the teacher's voice, operation behavior and action trajectory data in the classroom, each type of behavior is labeled according to the start and end time of the event, and the duration of the voice, trajectory change range and number of instruction events are extracted to generate a set of behavior parameters. S102: Based on the set of behavioral parameters, normalize the speech duration, trajectory change range and number of command events to generate a periodic parameter normalization sequence. S103: Call the normalized sequence of the periodic parameters, concatenate them in chronological order to construct a multidimensional vector set, and perform combined analysis based on the correlation between the sorting trend of behavioral parameters, trajectory changes and speech span to obtain the label sorting sequence.

4. The intelligent teaching supervision and evaluation method based on a large model according to claim 3, characterized in that, The steps for obtaining the conflict-blocking tag fragment group are as follows: S201: Based on the continuous time positions in the label sorting sequence, calculate the frequency of occurrence of labels in adjacent periods, identify labels with changing frequency differences, and obtain a label frequency difference sequence. S202: Call the label frequency difference sequence, extract the label combinations that occur frequently and concurrently within the period, perform type clustering based on the label co-occurrence relationship, and generate label combination cluster groups; S203: Based on the clustering group of the label combinations, call the mutual exclusion judgment rule of teaching behavior, identify the label combinations with structural conflict characteristics, mark the corresponding periodic positions, and obtain the conflict shielding label fragment group.

5. The intelligent teaching supervision and evaluation method based on a large model according to claim 1, characterized in that, The steps for obtaining the set of dominant behaviors of the transition fragment are as follows: S401: Based on the time period boundary structure, extract the behavior label sequence and the rating sequence within the segment, calculate the continuous fluctuation range between the node values ​​of the rating curve, determine the position of the jump point in the fluctuation range, and obtain the set of rating jump positions. S402: Call the set of scoring jump locations, extract all teacher behavior type labels within the corresponding jump segment, count the frequency of the labels in the segment, and obtain the behavior frequency distribution result of the jump segment. S403: Based on the frequency distribution results of the jump segment behavior, compare the difference between the tag frequency within the segment and the overall tag frequency, calculate the difference ranking index and perform weighted judgment in combination with the behavior distribution characteristics, calculate and obtain the frequency offset intensity value of the tag, sort and filter the dominant items to obtain the set of dominant behaviors of the jump segment.

6. The intelligent teaching supervision and evaluation method based on a large model according to claim 1, characterized in that, The method further includes: S5: Based on the time position of the dominant behavior set of the jump segment, arrange adjacent behaviors in the order of first appearance and connect them linearly to establish a teaching action flow path sequence on the time axis, perform chain linkage mapping between action content and score changes, and output a smart teaching supervision and evaluation scheme. The intelligent teaching supervision and evaluation scheme includes a teaching action flow map, a scoring change linkage mapping path, and a time series event chain structure. The steps for obtaining the intelligent teaching supervision and evaluation scheme are as follows: S501: Based on the set of dominant behaviors of the jump segment, extract the first appearance position of each type of label in the time axis, arrange all label items in chronological order, construct the label time progression sequence, and obtain the behavior label time sorting; S502: Invoke the behavior label time sorting, construct a linear connection path according to the time interval of adjacent labels, and combine them in sequence to form a continuous behavior flow path to generate a teaching action flow path sequence; S503: Based on the teaching action flow path sequence, track the time mapping relationship between labels and scoring jump points in the path, analyze the linkage degree between labels and scoring responses, calculate and obtain the scoring linkage response value of the path, and construct a dynamic mapping mechanism between labels and scores based on the linkage degree to obtain a smart teaching supervision and evaluation scheme.

7. A smart teaching supervision and evaluation system based on a large model, characterized in that, The system is used to execute the intelligent teaching supervision and evaluation method based on a large model as described in any one of claims 1-6, including: The behavior cycle splicing module obtains the teacher's voice length, number of operation commands and trajectory range. After segmenting the voice according to the event time, it extracts the voice according to energy, sorts the operations according to the cumulative number, calculates the cumulative offset value of the trajectory, and splices it according to time after labeling to obtain the label sorted sequence. The concurrent conflict detection module calls the label sorting sequence to count the number of labels in the time interval, compares the periodic distribution and uses mutual exclusion rules to determine the label combination, filters out the ones that need to be blocked and marks the time, and obtains the conflict blocking label fragment group; The trend jump alignment module extracts label changes based on the conflict shielding label fragment group, calculates the sequence of change counts and pause ratios, compares adjacent differences, and adjusts the positions according to jump nodes to obtain the time period boundary structure. The scoring band extraction module calls the labels and scoring sequences in the time period boundary structure, measures the scoring fluctuation and locates the jump, extracts the jump range labels and sorts them according to the difference between the number of times and the total number, and obtains the set of dominant behaviors of the jump segment. The flow link establishment module establishes a link by first sorting the dominant behaviors in the jump segment set and connecting the tag time series to form a link, thereby obtaining a smart teaching supervision and evaluation scheme.

Images