A multi-source teaching behavior data collaborative evaluation method and system

Abstract

The application discloses a kind of multi-source teaching behavior data collaborative evaluation method and system, it is related to educational information technology field, including by obtaining multi-source teaching behavior data, multi-source evaluation data and objective reference data, generate unit evaluation dataset;Baseline deviation, group deviation, behavior pattern deviation and time sequence deviation data of each evaluation source are calculated, and evaluation source deviation matrix is constructed;Baseline correction, group correction, behavior pattern correction and time sequence correction are sequentially executed according to deviation matrix, and corrected evaluation data is obtained;When divergence is greater than threshold value, combine historical stability, correction residual and objective reference consistency degree to calculate abnormality degree, generate abnormal evaluation source marker data and adjust corresponding fusion contribution value;Unit collaborative evaluation data is generated and collaborative evaluation result is aggregated and output, and deviation matrix and fusion contribution value are updated.The application improves the accuracy, consistency and self-adaptability of teaching evaluation.

Description

Technical Field

[0001] This invention relates to the field of educational information technology, and in particular to a collaborative evaluation method and system for multi-source teaching behavior data. Background Technology

[0002] With the development of educational informatization, classroom teaching evaluation increasingly relies on the fusion and analysis of multi-source heterogeneous data. Multimodal data such as video, audio, and operation logs, together with subjective evaluation data from different subjects such as students, peers, and supervisors, constitute a complex data foundation. Existing evaluation methods typically treat various data sources in isolation or simply summarize them at the final scoring stage. They lack effective means to unify multi-source data to a common behavioral granularity for structured organization and correlation, resulting in difficulty in aligning different modal data in the time and behavioral semantic dimensions, and a serious lack of refinement and interpretability of evaluation information. Due to significant differences in scoring standards, rigor, personal preferences, and judgment methods among various evaluators in different teaching scenarios, existing methods generally treat each evaluation source as equivalent or synthesize them based solely on fixed weights. This fails to identify the inherent systematic biases of each evaluation source, nor can it distinguish the changing patterns of biases across different groups of students, different teaching behavior patterns, and different evaluation time stages. Consequently, the evaluation results are mixed with a large number of factors that do not reflect actual teaching performance, resulting in poor objectivity of the data and poor comparability between evaluation sources. When multiple evaluation sources show significant discrepancies in their assessments of the same teaching behavior unit, existing technologies lack a scientific anomaly detection mechanism. They often rely solely on simple statistical rules to eliminate so-called outliers or simply average the results. This approach fails to accurately distinguish between truly anomalous evaluation sources and valid information reflecting genuine teaching differences. Furthermore, it struggles to dynamically adjust the contribution weights of each data source during fusion based on the historical stability of the evaluation sources and their consistency with objective references. Consequently, the fusion process lacks resilience against anomalous evaluation sources, and the determination of fusion weights lacks interpretable justification. More significantly, most existing evaluation systems terminate the processing flow after outputting the final result. The evaluation source bias structure, fusion weights, and correction parameters cannot self-correct or continuously evolve based on subsequent feedback reference data. The evaluation model remains static and cannot be continuously optimized with teaching practice and data accumulation, making it difficult to adapt to dynamically changing classroom environments and the needs of teaching improvement. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a collaborative evaluation method and system for multi-source teaching behavior data.

[0004] In a first aspect, the present invention provides a collaborative evaluation method for multi-source teaching behavior data, comprising the following steps: Acquire multi-source teaching behavior data, multi-source evaluation data, and objective reference data for the same teaching object in the same teaching activity, and divide the multi-source teaching behavior data into multiple teaching behavior units according to the boundaries of behavior events, and map the multi-source evaluation data to the corresponding teaching behavior units to generate unit evaluation datasets; Based on the historical unit evaluation dataset and objective reference data, the baseline deviation data, group deviation data, behavioral pattern deviation data and time series deviation data corresponding to each evaluation source are calculated, and the evaluation source deviation matrix is ​​constructed. The original evaluation data in the unit evaluation dataset is decomposed into deviation components to obtain the true evaluation components, system deviation components, situational deviation components and disturbance components. Based on the evaluation source deviation matrix, baseline correction, group correction, behavioral pattern correction and time series correction are performed on the original evaluation data in sequence to obtain the corrected evaluation data. For the corrected evaluation data in the same teaching behavior unit, the evaluation divergence degree is calculated. When the evaluation divergence degree is greater than the preset threshold, the abnormality degree is calculated by combining the historical stability of the evaluation source, the correction residual and the consistency of the objective reference, generating abnormal evaluation source label data, and adjusting the fusion contribution value of the corresponding corrected evaluation data according to the abnormal evaluation source label data. Based on the fusion contribution value of each evaluation source, the correction evaluation data in the same teaching behavior unit is collaboratively fused to generate unit collaborative evaluation data. The collaborative evaluation data of multiple units is aggregated to output the collaborative evaluation results corresponding to the teaching object. The results are compared with the feedback reference data obtained later to update the evaluation source deviation matrix and fusion contribution value for collaborative evaluation processing of multi-source evaluation data.

[0005] Secondly, a collaborative evaluation system for multi-source teaching behavior data includes: The data acquisition module is used to acquire multi-source teaching behavior data, multi-source evaluation data, and objective reference data of the same teaching object in the same teaching activity. It also divides the multi-source teaching behavior data into multiple teaching behavior units according to the boundaries of the behavior events, maps the multi-source evaluation data to the corresponding teaching behavior units, and generates a unit evaluation dataset. The evaluation source deviation matrix generation module is used to calculate the baseline deviation data, group deviation data, behavioral pattern deviation data and time series deviation data corresponding to each evaluation source based on the historical unit evaluation dataset and objective reference data, and to construct the evaluation source deviation matrix. The multi-dimensional deviation correction module is used to decompose the original evaluation data in the unit evaluation dataset into deviation components to obtain the true evaluation components, system deviation components, situational deviation components and disturbance components. Based on the evaluation source deviation matrix, the module sequentially performs baseline correction, group correction, behavioral pattern correction and time series correction on the original evaluation data to obtain the corrected evaluation data. The tagging module is used to calculate the evaluation divergence degree for the corrected evaluation data in the same teaching behavior unit. When the evaluation divergence degree is greater than the preset threshold, the abnormality degree is calculated by combining the historical stability of the evaluation source, the correction residual and the consistency of the objective reference, generating abnormal evaluation source tagging data, and adjusting the fusion contribution value of the corresponding corrected evaluation data according to the abnormal evaluation source tagging data. The dynamic update module is used to collaboratively integrate the correction evaluation data in the same teaching behavior unit based on the fusion contribution value of each evaluation source, generate unit collaborative evaluation data, aggregate the collaborative evaluation data of multiple units, output the collaborative evaluation results corresponding to the teaching object, compare them with the feedback reference data obtained later, and update the evaluation source deviation matrix and fusion contribution value for collaborative evaluation processing of multi-source evaluation data.

[0006] In summary, the present invention has at least one of the following beneficial technical effects: 1. The collaborative evaluation method for multi-source teaching behavior data provided by this invention acquires multi-source teaching behavior data, multi-source evaluation data, and objective reference data for the same teaching object in the same teaching activity. The multi-source teaching behavior data is then segmented into multiple teaching behavior units according to behavioral event boundaries. Multi-source evaluation data is mapped to corresponding teaching behavior units, thus forming a structured unit evaluation dataset with teaching behavior units as the common granularity. Based on this, according to the historical unit evaluation dataset and objective reference data, baseline deviation data, group deviation data, behavioral pattern deviation data, and temporal deviation data corresponding to each evaluation source are calculated, constructing an evaluation source deviation matrix that comprehensively characterizes the multi-dimensional deviation features of each evaluation source. By decomposing the original evaluation data in the unit evaluation dataset into deviation components, true evaluation components, systematic deviation components, contextual deviation components, and disturbance components are obtained. Based on the evaluation source deviation matrix, baseline correction, group correction, behavioral pattern correction, and temporal correction are sequentially performed on the original evaluation data. This achieves the gradual removal of various systematic and contextual deviations mixed in the original evaluation data, making the corrected evaluation data more realistically reflect the actual performance of the teaching object. This method breaks away from the limitations of traditional evaluation methods that simply aggregate scores from various evaluation sources or rely solely on fixed weights for synthesis. By separating and correcting the long-term scoring habits of evaluation sources, their biased judgments on different group attributes, the scale differences under different teaching behavior patterns, and the drift that occurs over time, it significantly improves the comparability and consistency of multi-source evaluation data. This allows evaluation information from different subjects and backgrounds to be integrated within a unified and objective reference framework, providing a reliable data foundation for the precision and scientific nature of teaching evaluation.

[0007] 2. This invention further utilizes the difference distribution analysis of corrected evaluation data within the same teaching behavior unit to calculate the evaluation divergence degree. When the evaluation divergence degree exceeds a preset threshold, it combines the historical stability of the evaluation source, the corrected residual, and the consistency of the objective reference to calculate the anomaly degree, generating abnormal evaluation source marker data. Based on this, the fusion contribution value of the corresponding corrected evaluation data is dynamically adjusted. This allows for the scientific differentiation between truly abnormal evaluation sources and reasonable opinions reflecting the complexity of the teaching behavior itself when significant discrepancies occur in the evaluation results, avoiding the erroneous deletion of effective evaluation information by traditional extreme value elimination or simple averaging. Furthermore, based on the fusion contribution value of each evaluation source, the corrected evaluation data within the same teaching behavior unit is collaboratively fused to generate unit collaborative evaluation data. Multiple unit collaborative evaluation data are then aggregated to output the collaborative evaluation results corresponding to the teaching object. Simultaneously, this result is compared with subsequently acquired feedback reference data to update the evaluation source deviation matrix and fusion contribution value, forming a closed-loop evaluation mechanism that can continuously evolve with teaching practice. This design not only provides a quantitative basis for the historical stability and reference consistency of the weight allocation of each evaluation source during the fusion process, but also enables the evaluation source bias structure and fusion parameters to be adaptively corrected based on actual feedback results. The system's responsiveness to changes in different teaching scenarios and evaluation ecosystems is greatly enhanced, and the interpretability and credibility of evaluation conclusions are systematically improved. This provides a technical solution with practical application value for the continuous improvement of classroom teaching quality and the intelligent evolution of the teaching evaluation system. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart of a collaborative evaluation method based on multi-source teaching behavior data.

[0010] Figure 2 This is a framework diagram of a collaborative evaluation system based on multi-source teaching behavior data. Detailed Implementation

[0011] The following description, in conjunction with the implementation of the present invention, is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention, and all such modifications and additions should fall within the protection scope of the present invention.

[0012] Example 1 This invention addresses the practical problems in classroom teaching evaluation, such as multiple data sources, inconsistent evaluation scales, significant evaluator bias, and difficulty in continuously correcting results. It is applicable to smart classrooms, online teaching platforms, blended learning management platforms, and teacher training assessment scenarios. Its core idea is to unify multi-source teaching behavior data, multi-source evaluation data, and objective reference data for the same learner in the same teaching activity into a common granularity of teaching behavior unit. First, it identifies the inherent bias structure of each evaluation source. Then, it decomposes and hierarchically corrects the original evaluation data. Combining evaluation divergence degree, it identifies abnormal evaluation sources and dynamically adjusts the fusion contribution value, ultimately forming a collaborative evaluation result that is interpretable and continuously updated with feedback reference data. This process not only reduces systematic drift between different evaluation sources but also retains effective evaluation information consistent with actual teaching performance, making the evaluation results for learners more stable and closer to subsequent feedback.

[0013] This application discloses a collaborative evaluation method for multi-source teaching behavior data.

[0014] Reference Figure 1 A collaborative evaluation method for multi-source teaching behavior data, specifically including: In this implementation, a 45-minute offline and platform-linked class at a higher education institution is used as the single teaching activity, with the instructor as the single learning target. The instructor's teaching behavior in this class is then collaboratively evaluated. The data sources for the evaluation include teaching video data, classroom audio data, teaching platform operation log data, and evaluation record data. The teaching video data is collected by two cameras deployed at the front and back of the classroom, with H.264 encoding, a frame rate of 25 frames per second, and millisecond-level timestamps. Classroom audio data is collected using a combination of overhead and lavalier microphones, with a sampling rate of 16 kHz, and millisecond-level timestamps. Teaching platform operation log data comes from the teaching management platform's log server, and the record fields include at least the learning target identifier, operation type, operation time, resource object identifier, and page dwell time. Evaluation record data comes from student mobile evaluations, peer observation terminal evaluations, and supervisor evaluation terminal evaluations, and the fields include the evaluation source identifier, evaluation time, evaluation target identifier, evaluation content identifier, and original evaluation data. Objective reference data can be obtained from objective measurement results such as changes in students' in-class quiz scores, completion rates of classroom tasks, response rates to classroom interactions, and short-term feedback scores after class. Its data structure includes the collection time, the identification of the teaching object, the name of the reference indicator, and the value of the reference indicator.

[0015] The entire collaborative evaluation process is executed by the server-side processor. The server can adopt an eight-core CPU and GPU architecture, with at least 64 GB of memory, and persistent storage using a combination of relational and time-series databases. The data acquisition module is responsible for data collection and initial alignment; the evaluation source deviation matrix generation module is responsible for historical deviation modeling; the multi-dimensional deviation correction module is responsible for component decomposition and hierarchical correction; the labeling module is responsible for evaluation divergence analysis and generation of abnormal evaluation source labeling data; and the dynamic update module is responsible for collaborative fusion, result aggregation, and continuous updates based on feedback reference data. Intermediate results are transmitted between modules through a shared data bus. The primary key of the data uniformly adopts a combination key of teaching object identifier + teaching activity identifier + teaching behavior unit identifier + evaluation source identifier + time identifier to ensure consistency in subsequent mapping and backtracking.

[0016] Acquire multi-source teaching behavior data, multi-source evaluation data, and objective reference data for the same teaching object in the same teaching activity, and divide the multi-source teaching behavior data into multiple teaching behavior units according to the boundaries of behavior events, and map the multi-source evaluation data to the corresponding teaching behavior units to generate unit evaluation datasets; Specifically, the processing flow begins by acquiring multi-source teaching behavior data, multi-source evaluation data, and objective reference data for the same teaching subject in the same teaching activity. The multi-source teaching behavior data is then segmented into multiple teaching behavior units based on behavioral event boundaries. Multi-source evaluation data is mapped to corresponding teaching behavior units to generate unit evaluation datasets. Specifically, the data acquisition module first extracts time stamp information from teaching video data, classroom audio data, teaching platform operation record data, and evaluation record data. For teaching video data, the time stamp information is the acquisition timestamp of each frame or keyframe; for classroom audio data, the time stamp information is the start and end times of each audio buffer segment; for teaching platform operation record data, the time stamp information is the log write time and business event submission time; for evaluation record data, the time stamp information is the evaluation terminal submission time. Because the internal clocks of various devices have offsets, the data is first aligned according to a unified time benchmark. The unified time benchmark can be the standard time of the campus network time server. During the synchronization phase, the original timestamps of various data types are subtracted from the clock offset of the corresponding device to obtain the calibration timestamp. The device clock offset can be measured by sending three consecutive heartbeat synchronization messages before the start of class. If the range of the three measurements does not exceed 20 ms, the average value is taken as the device's clock offset; if it exceeds 20 ms, synchronization continues until the range converges to below 20 ms. After the above processing, an aligned dataset is generated.

[0017] After obtaining the aligned dataset, the data acquisition module identifies behavioral event boundaries based on the behavior occurrence time, behavior duration, and behavior switching marker in the aligned dataset. It then segments the multi-source teaching behavior data into multiple teaching behavior units according to these boundaries, generating a sequence of teaching behavior units. In this implementation, the behavior occurrence time is defined as the first identifiable teaching behavior of a teaching object; the behavior duration is defined as the length of time from the occurrence to the end of the teaching behavior, measured in seconds; and the behavior switching marker is defined as the boundary marker on the timeline when a switch occurs between two adjacent types of teaching behaviors. The behavior switching marker can be determined by combining video action recognition results, audio semantic segment switching results, and platform operation type change results. For example, if the teaching video data identifies a teacher switching from writing on the blackboard to patrolling, the classroom audio data identifies a switch from explanation to questioning, and the teaching platform operation record data shows a voting operation, then the earliest common time point within the overlapping time interval of these three events is taken as the new behavioral event boundary. To avoid overly fragmented segmentation, this implementation specifies a minimum teaching behavior unit duration of 15 seconds. If the time interval between two adjacent behavioral event boundaries is less than 15 seconds, the interval is merged into the previous teaching behavior unit. The maximum teaching behavior unit duration can be set to 300 seconds. If the duration exceeds this and there is a significant semantic shift within it, secondary boundaries are used for further segmentation. After segmentation, a sequence of teaching behavior units arranged in chronological order is obtained. For example, unit U1 is the introduction and explanation, unit U2 is the question-and-answer interaction, unit U3 is the demonstration, unit U4 is the student practice and observation, and unit U5 is the summary and feedback.

[0018] The data acquisition module extracts the evaluation time, evaluation object identifier, and evaluation content identifier from the multi-source evaluation data, maps the multi-source evaluation data to the corresponding teaching behavior units, and generates unit-specific evaluation data. In this embodiment, the evaluation time is matched with the time interval of the teaching behavior unit. If the evaluation time falls between the start and end times of a certain teaching behavior unit, the evaluation data is mapped to that teaching behavior unit. If the evaluation time is within a fuzzy interval of no more than 3 seconds before and after the boundary of two adjacent teaching behavior units, the evaluation content identifier is used for auxiliary discrimination. For example, if the evaluation content identifier is "interactive organization," it is preferentially mapped to the question-and-answer interaction unit; if the evaluation content identifier is "class summary," it is preferentially mapped to the summary and feedback unit. The evaluation object identifier is used to confirm that the evaluation corresponds to the current instructor rather than an individual student or teaching assistant in the classroom. The evaluation content identifier is used to divide the evaluation dimensions within the same teaching behavior unit, such as clarity of expression, effectiveness of interaction, and pacing control. In this embodiment, the original evaluation data is uniformly mapped to a percentage range of 0 to 100. If an evaluation source originally used a 5-point scale, it is multiplied by 20 according to the linear conversion rule; if it originally used a grading system of A, B, C, and D, it is mapped to 95, 85, 75, and 65 respectively. After mapping, unit-specific evaluation data is generated for each teaching behavior unit.

[0019] The data acquisition module associates the collection time and teaching object identification in the objective reference data with the corresponding evaluation data of the unit to generate objective reference data for each teaching behavior unit. Since the collection rhythm of objective reference data may not be completely synchronized with the teaching behavior unit, this implementation uses a time window association method. For each teaching behavior unit, a 120-second association time window is established before and after the end time of that teaching behavior unit. Within this time window, objective reference data with matching collection times and consistent teaching object identification are retrieved. If multiple objective reference data are retrieved, they are synthesized according to preset weights of the reference indicators to form the objective reference data corresponding to that teaching behavior unit. For example, the weight of classroom interaction response rate is 0.35, the weight of instant score change in in-class quizzes is 0.40, and the weight of classroom task completion rate is 0.25. All items have been normalized to the range of 0 to 100, so the dimension of the synthesized value is still a score, and the dimensions are consistent. Based on the teaching behavior unit, the corresponding evaluation data of the unit, and the objective reference data corresponding to each teaching behavior unit, a unit evaluation dataset is generated. Each record in the unit evaluation dataset includes at least the following: teaching object identifier, teaching activity identifier, teaching behavior unit identifier, behavior pattern identifier, evaluation source identifier, original evaluation data, objective reference data, evaluation time, and group attribute data.

[0020] Based on the historical unit evaluation dataset and objective reference data, the baseline deviation data, group deviation data, behavioral pattern deviation data and time series deviation data corresponding to each evaluation source are calculated, and the evaluation source deviation matrix is ​​constructed. Specifically, after forming the unit evaluation dataset corresponding to the current teaching activity, the evaluation source deviation matrix generation module calculates the baseline deviation data, group deviation data, behavioral pattern deviation data, and time-series deviation data for each evaluation source based on the historical unit evaluation dataset and objective reference data, thus constructing the evaluation source deviation matrix. In this embodiment, the historical unit evaluation dataset is taken from historical data of similar courses within the past semester, covering at least 200 teaching activities, with each evaluation source participating in the evaluation of at least 1000 teaching behavior units to ensure the stability of the deviation estimation. The original evaluation data of each evaluation source in different teaching behavior units is extracted from the historical unit evaluation dataset, and the deviation between the original evaluation data and the objective reference data is calculated based on the objective reference data to generate evaluation deviation data. The evaluation deviation data is denoted as Δ, defined as the original evaluation data minus the objective reference data, with the unit being points; therefore, the dimension of Δ is also points. If a student's evaluation source gives an original evaluation score of 88 for a certain teaching behavior unit, while the corresponding objective reference score for that teaching behavior unit is 80, then the evaluation deviation score is 8. If the original evaluation score is 72 and the objective reference score is 78, then the evaluation deviation score is -6. This definition directly reflects the degree to which the evaluation source overestimates or underestimates the objective reference score, and both the sign and magnitude have clear physical meaning.

[0021] Baseline deviation data is obtained by aggregating evaluation deviation data from the same evaluation source across all teaching behavior units. Baseline deviation data describes the overall deviation trend of an evaluation source without considering the group attributes, behavioral patterns, and time periods of the learners. In this embodiment, the baseline deviation data is calculated using a truncated mean. First, all evaluation deviation data from the same evaluation source are sorted numerically, and the top and bottom 5% of extreme values ​​are removed. Then, the remaining 90% of the evaluation deviation data is averaged. The truncated mean is used because there may be anomalies in the evaluation deviation data, and direct averaging is easily influenced by a small number of extreme values. The calculation results are still expressed in points. For example, a baseline deviation of +4.2 points from a peer evaluation source indicates that this evaluation source tends to give an evaluation 4.2 points higher than the objective reference data in the long term.

[0022] In calculating group deviation data, evaluation deviation data is grouped according to the group attributes of the teaching objects to obtain group deviation data. In this embodiment, group attribute data may include professional title level, teaching experience range, course category, department type, and class size level, etc. More specifically, group attribute data corresponding to the teaching object identifiers is first extracted from the historical unit evaluation dataset, and then the group attribute data is correlated with the evaluation deviation data output from the same evaluation source to form a group evaluation deviation sequence. For example, for the same student evaluation source, a group evaluation deviation sequence for junior-level teachers, a group evaluation deviation sequence for intermediate-level teachers, and a group evaluation deviation sequence for senior-level teachers can be formed.

[0023] For each group attribute in the group evaluation deviation sequence, the deviation center value and deviation dispersion value are statistically calculated to generate group deviation characteristic data. The deviation center value can be taken as the median, with the unit being points; the deviation dispersion value can be taken as the median absolute deviation, also with the unit being points. These two quantities are used because they are more robust to outliers and can more realistically characterize the central position and degree of fluctuation of the evaluation deviation data under a certain group attribute.

[0024] Group deviation data is determined based on the distribution differences of group deviation characteristic data among various group attributes, and this data is then written into the evaluation source deviation matrix. The distribution differences can be represented by the difference between the deviation center value of each group attribute and the overall deviation center value. For example, for a certain supervisory evaluation source, the overall deviation center value is +1.5 points, while the deviation center value under the large-class teaching group attribute is +3.0 points. Therefore, the group deviation data corresponding to this group attribute is +1.5 points. Thus, the meaning of group deviation data is the additional offset of the evaluation source on a specific group attribute, in addition to the baseline deviation data.

[0025] The calculation method for behavioral pattern deviation data corresponds to that of group deviation data, but the grouping dimension is changed to the behavioral pattern corresponding to the teaching behavior unit. Behavioral patterns can be simultaneously marked when the aforementioned teaching behavior units are segmented, such as introduction and explanation, questioning and interaction, demonstration, supervision and guidance, and summary and feedback. The evaluation source deviation matrix generation module groups the evaluation deviation data according to the behavioral patterns corresponding to the teaching behavior units, and statistically analyzes the deviation center in each behavioral pattern to obtain the behavioral pattern deviation data. If a student's evaluation source generally scores 2.8 points higher than the objective reference data in the questioning and interaction behavioral pattern, while their overall baseline deviation data is only 0.9 points, then the behavioral pattern deviation data can be recorded as +1.9 points, indicating that this evaluation source has a tendency to artificially inflate evaluations in the questioning and interaction scenario.

[0026] The time-series deviation data is obtained by segmenting the evaluation deviation data according to the evaluation time sequence. Segmentation can be done using a fixed-length time window or by teaching week. In this implementation, considering that classroom evaluation is easily affected by the semester's pace, a two-level segmentation method of teaching week + in-class time period is adopted. The teaching weeks are divided into four segments: weeks 1-4, weeks 5-8, weeks 9-12, and weeks 13 and beyond. The in-class time period is divided into four segments: 0-10 minutes after class starts, 10-25 minutes, 25-40 minutes, and beyond 40 minutes. The mean of the evaluation deviation data for each evaluation source is calculated within each segment, and the baseline deviation data for that evaluation source is subtracted to obtain the time-series deviation data. This time-series deviation data reflects the temporal drift of the evaluation source during both the semester's progression and the classroom's progress.

[0027] Based on baseline deviation data, group deviation data, behavioral pattern deviation data, and time-series deviation data, an evaluation source deviation matrix is ​​constructed. This matrix can be stored as a four-dimensional index structure, with the primary index being the evaluation source identifier, and the secondary indices being group attributes, behavioral patterns, and time segments, respectively. All matrix elements have numerical values ​​of points. Each element in the matrix corresponds to an estimated deviation value for a given evaluation source under a specific combination of conditions.

[0028] The original evaluation data in the unit evaluation dataset is decomposed into deviation components to obtain the true evaluation components, system deviation components, situational deviation components and disturbance components. Based on the evaluation source deviation matrix, baseline correction, group correction, behavioral pattern correction and time series correction are performed on the original evaluation data in sequence to obtain the corrected evaluation data. Specifically, after obtaining the evaluation source deviation matrix, the multi-dimensional deviation correction module decomposes the original evaluation data in the unit evaluation dataset into deviation components, obtaining the true evaluation component, system deviation component, contextual deviation component, and disturbance component. In this embodiment, the original evaluation data, evaluation source identifier, teaching behavior unit identifier, and objective reference data are first extracted from the unit evaluation dataset. Evaluation deviation data are determined based on the correspondence between the original evaluation data and the objective reference data. These evaluation deviation data are aggregated according to the evaluation source identifier to obtain the system deviation component. The physical meaning of the system deviation component is a stable offset caused by the long-term habits, scoring scale preferences, and evaluation styles of the evaluation source; therefore, it mainly corresponds to the baseline deviation data in the evaluation source deviation matrix, with the unit being points. The evaluation deviation data is cross-aggregated according to the group attributes of the teaching objects and the behavioral patterns of the teaching behavior units to obtain the contextual deviation component. The contextual deviation component reflects the additional offset generated by the evaluation source under a certain combination of teaching objects and behavioral patterns. For example, a more lenient approach to questioning and interaction with new teachers, or a more stringent approach to supervision and guidance in large classes. This offset is not purely a habitual behavior of the evaluation source, but is coupled with the evaluation context; therefore, it is treated as a separate contextual deviation component.

[0029] After identifying the systematic bias and situational bias components, the original evaluation data is subtracted from these components to obtain the true evaluation component and the perturbation component. This separation is achieved by comparing the direction of change with the objective reference data. If the direction of change of the remaining quantity after subtraction is consistent with the direction of change of the objective reference data between adjacent teaching behavior units, this consistent portion is recorded as the true evaluation component; the remaining portion that does not conform to the direction of change of the objective reference data and exhibits short-term random fluctuations is recorded as the perturbation component. For example, if the remaining quantities of an evaluation source after subtraction in units U2, U3, and U4 are 76, 82, and 79 points respectively, while the objective reference data are 74, 80, and 78 points respectively, and both increase from U2 to U3 and decrease from U3 to U4, then the main portion consistent with this direction can be considered the true evaluation component, and the remaining small fluctuations can be considered the perturbation component. The perturbation component usually originates from attention drift, terminal operation errors, or occasional emotional fluctuations during evaluation, and the unit is still points. After the above processing, the true evaluation component, system bias component, situational bias component, and disturbance component are used as input data for the correction process.

[0030] Based on this, the multi-dimensional deviation correction module sequentially performs baseline correction, group correction, behavioral pattern correction, and time-series correction on the original evaluation data according to the evaluation source deviation matrix, resulting in corrected evaluation data. This order is chosen because baseline deviation data corresponds to the most stable and prevalent long-term biases of the evaluation source and should be removed first; group deviation data and behavioral pattern deviation data represent conditional biases based on this and should be processed layer by layer after baseline correction; time-series deviation data reflects fine-grained drifts in the time dimension and is suitable for fine-tuning in the later stages to reduce interference from large adjustments in the earlier stages on the time correction.

[0031] Specifically, based on the baseline deviation data corresponding to the evaluation source, the original evaluation data is baseline corrected to generate baseline-corrected data. If the original evaluation data is denoted as R and the baseline deviation data as B, then the baseline-corrected data is denoted as... Defined as =RB. Here R, B and... All scores are in points, and the dimensions are perfectly matched. For example, if the original evaluation data from a peer review source is 86 points and the baseline deviation data is +4.2 points, then the baseline-corrected data is 81.8 points. Based on the group attribute data corresponding to the teaching objects and the group deviation data in the evaluation source deviation matrix, the baseline-corrected data is adjusted for group bias, generating group-corrected data. If the group deviation data is denoted as G, then the group-corrected data... = -G. Based on the behavioral pattern data corresponding to the teaching behavior units and the behavioral pattern deviation data in the evaluation source deviation matrix, behavioral pattern correction is performed on the group correction data to generate behavioral pattern correction data. If the behavioral pattern deviation data is denoted as M, then the behavioral pattern correction data... = -M. Finally, based on the time-series deviation data corresponding to the evaluation time, the behavioral pattern correction data is time-series corrected to generate corrected evaluation data. If the time-series deviation data is denoted as T, then the corrected evaluation data... = -T. The entire chain of corrections can be expressed as =RBGMT, where each component is in points and represents the deviation to be deducted from the original evaluation data. Therefore, the physical meaning and dimensions of addition and subtraction are consistent. For extreme cases, this implementation sets a lower limit of 0 points and an upper limit of 100 points for the corrected evaluation data. If the calculated value exceeds this range, it is truncated at the boundary. After correction, the corrected evaluation data and the true evaluation components are stored accordingly for subsequent evaluation divergence calculation. During storage, the original evaluation data, correction results at each level, and the true evaluation components are simultaneously retained in the same record, facilitating subsequent tracing of the evaluation source behavior.

[0032] For the corrected evaluation data in the same teaching behavior unit, the evaluation divergence degree is calculated. When the evaluation divergence degree is greater than the preset threshold, the abnormality degree is calculated by combining the historical stability of the evaluation source, the correction residual and the consistency of the objective reference, generating abnormal evaluation source label data, and adjusting the fusion contribution value of the corresponding corrected evaluation data according to the abnormal evaluation source label data. Specifically, for multiple evaluation sources within the same teaching behavior unit, the labeling module then calculates the evaluation divergence degree. When the evaluation divergence degree exceeds a preset threshold, the anomaly degree is calculated by combining the historical stability of the evaluation source, the correction residual, and the consistency with objective references, generating abnormal evaluation source labeling data. In this embodiment, for the corrected evaluation data within the same teaching behavior unit, the difference distribution between the corrected evaluation data of each evaluation source is first calculated. If there are n evaluation sources within the teaching behavior unit, an n×n difference matrix is ​​constructed, with matrix elements... The evaluation divergence degree is defined as the absolute value of the difference between the corrected evaluation data of the i-th evaluation source and the j-th evaluation source, expressed in points. The evaluation divergence degree is generated based on the central tendency of the difference distribution. Here, the central tendency is calculated by subtracting the normalized dispersion index from 1. Let σc be the standard deviation of all corrected evaluation data in the same teaching behavior unit, and let the effective interval width of the corrected evaluation data be 100 points. Then, the normalized dispersion index is σc / 100, a dimensionless value; the evaluation divergence degree D is defined as D = σc / 100. The reason for directly using standard deviation normalization is that the standard deviation reflects the degree of dispersion of multiple evaluation sources around the mean; the larger the dispersion, the stronger the divergence. Furthermore, after normalization, the value is stable between 0 and 1, making it easy to set a uniform threshold. In this embodiment, the preset threshold is set to 0.12, meaning that when the standard deviation of the corrected evaluation data exceeds 12 points, the teaching behavior unit is considered to have significant evaluation divergence.

[0033] Once the evaluation discrepancy exceeds a preset threshold, the labeling module extracts the stability data of the corresponding evaluation source in the historical unit evaluation dataset, the correction residual data in the current teaching behavior unit, and the objective reference consistency data. It then jointly sorts the stability data, correction residual data, and objective reference consistency data to generate anomaly. Historical stability reflects the long-term stability of an evaluation source's output and can be defined as a normalized index in the reciprocal form of the fluctuation of the correction evaluation data in continuous historical teaching behavior units. The value ranges from 0 to 1, with a larger value indicating greater stability. To ensure clear physical meaning, this implementation denotes the standard deviation of the historical correction evaluation data of the evaluation source as σh, in points. A reference fluctuation upper limit of Sref = 20 points is taken, so historical stability is denoted as Sh = 1 - σh / Sref; when σh exceeds 20 points, Sh is truncated to 0. Thus, Sh is a dimensionless quantity and positively correlated with stability. The correction residual reflects the deviation between the current correction evaluation data and the true evaluation component or objective reference data. In this implementation, the correction residual is denoted as Er and defined as | -C|, where C is the objective reference data for the current teaching behavior unit, in points. Objective reference consistency reflects the degree of consistency between the evaluation source and the objective reference data in terms of direction of change and relative level. It can be obtained by linearly mapping the Pearson correlation coefficient between the evaluation source's corrected evaluation data sequence and the objective reference data sequence over the most recent k teaching behavior units to the interval of 0 to 1. If the correlation coefficient is denoted as ρ, and its value ranges from -1 to 1, then the objective reference consistency... =(ρ+1) / 2, a dimensionless quantity. During joint ranking, historical stability and objective reference consistency are sorted in descending order, while corrected residuals are sorted in ascending order. The ranking scores are then synthesized using preset weights. In this embodiment, the anomaly degree A is defined as A=0.4×(1- )+0.35× / 20+0.25×(1- In the formula, / 20 is used to normalize the corrected residuals to around 0 to 1. 20 points is an empirical reference scale with significant anomalous meaning in classroom evaluation differences, therefore A is a dimensionless quantity. The larger A is, the more likely the evaluation source is to be an anomalous evaluation source that leads to the current evaluation discrepancy.

[0034] After obtaining the anomaly scores of all evaluation sources, anomaly evaluation source labeling data is generated based on the predetermined ranking of anomalies among all evaluation sources, and this anomaly evaluation source labeling data is sent to the fusion contribution value adjustment step. The predetermined ranking can be set as the top 20% of evaluation sources in terms of anomaly score within the same teaching behavior unit, and the anomaly score is not lower than 0.55. In this embodiment, if there are 10 evaluation sources in the current teaching behavior unit, the top 2 evaluation sources with anomalies that meet the anomaly score threshold are assigned the value of anomaly evaluation source labeling data 1, and the remaining evaluation sources are assigned the value of 0; if there are fewer than 5 evaluation sources, only the evaluation source with the highest anomaly score that exceeds the threshold is labeled as anomaly. The anomaly evaluation source labeling data itself is a binary or hierarchical label, or it can be saved as a three-value structure of high anomaly, medium anomaly, and non-anomaly to adapt to different fusion strategies.

[0035] Based on the aforementioned labeling results, the labeling module adjusts the fusion contribution value of the corresponding corrected evaluation data according to the abnormal evaluation source labeling data. Specifically, it first extracts the historical stability, correction residual, objective reference consistency, and abnormal evaluation source labeling data corresponding to each evaluation source. Then, it uses historical stability and objective reference consistency as positive contribution data, and correction residual and abnormal evaluation source labeling data as suppression contribution data to generate the fusion contribution value corresponding to each evaluation source. In this embodiment, an initial contribution score is first calculated for each evaluation source. .

[0036] in Specifically, =0.35× +0.30× +0.20×(1- / 20)+0.15×(1- In the formula, For anomaly assessment, the data is labeled with 0 for non-anomalies and 1 for high anomalies; if a three-value structure is used, 0.5 can be used for medium anomalies. Because... , , / 20 and All of them are dimensionless quantities, therefore It is a dimensionless quantity, with values ​​generally between 0 and 1. Subsequently, for all evaluation sources within the same teaching behavior unit... Normalization is performed to obtain the fusion contribution value. Defined as = / Σ Therefore, the sum of the fusion contribution values ​​of all evaluation sources is 1, which facilitates subsequent weighted aggregation. When the abnormal evaluation source labeling data indicates that the evaluation source is located in a predetermined position with a high degree of abnormality, the fusion contribution value of the corresponding corrected evaluation data of that evaluation source is reduced. In this embodiment, the evaluation source can be normalized before normalization. Multiply by a suppression coefficient of 0.4; for moderately abnormal evaluation sources, multiply by 0.7. Conversely, when the data marking an abnormal evaluation source does not indicate that the evaluation source is in a predetermined position with a high degree of abnormality, the fusion contribution value of the corresponding corrected evaluation data for that evaluation source is maintained. The processed fusion contribution value is written back to the current teaching behavior unit data record and used for collaborative fusion. The direct result of this is that although the data of abnormal evaluation sources are not simply removed, their impact on the final result is weakened in a controlled manner, thereby avoiding a single extreme evaluation source from influencing the overall conclusion.

[0037] Based on the fusion contribution value of each evaluation source, the correction evaluation data in the same teaching behavior unit is fused collaboratively to generate unit collaborative evaluation data. The collaborative evaluation data of multiple units is aggregated, and the collaborative evaluation results corresponding to the teaching object are output. The results are compared with the feedback reference data obtained later, and the evaluation source deviation matrix and fusion contribution value are updated for collaborative evaluation processing of multi-source evaluation data. Specifically, in the fusion phase, the dynamic update module collaboratively merges the correction evaluation data within the same teaching behavior unit based on the fusion contribution values ​​of each evaluation source, generating unit-based collaborative evaluation data. Specifically, it extracts all correction evaluation data and their corresponding fusion contribution values ​​from the same teaching behavior unit, and then weights and aggregates the correction evaluation data according to the fusion contribution values. If the correction evaluation data from the i-th evaluation source is... The fusion contribution value Then the unit collaborative evaluation data Defined as =Σ × .because For dimensionless weights, In units of cents, therefore The units are still points, and the dimensions remain consistent. Taking unit U3 as an example, if there are four evaluation sources with corrected evaluation scores of 78, 82, 80, and 69 points respectively, and fusion contribution values ​​of 0.30, 0.28, 0.27, and 0.15 respectively, then the unit collaborative evaluation score is 78.87 points. This result better reflects the role of high-reliability evaluation sources compared to simple averaging.

[0038] After obtaining collaborative evaluation data for multiple teaching behavior units, the dynamic update module arranges the collaborative evaluation data according to the chronological order of the teaching activities and aggregates it based on the duration and importance of each teaching behavior unit to generate collaborative evaluation results corresponding to the teaching object. In this embodiment, the unit aggregation weight is determined by both duration and behavior importance. Let the duration of the u-th teaching behavior unit be denoted as . The unit is seconds, and the importance of the behavior is... The value ranges from 0 to 1 and is pre-configured according to the instructional design. For example, 0.25 can be used for interactive questioning, 0.20 for demonstration, and 0.15 for summary and feedback. To ensure that the duration item is dimensionally compatible with the behavior importance item, the duration is first normalized to... = / Σ After obtaining the dimensionless value, the unit aggregation weight is defined. = × +(1- )× / Σ ,in The value is set to 0.5. The collaborative evaluation result Rall corresponding to the teaching object is then defined as Rall = Σ × Pu is dimensionless. Since the score is zero, Rall remains zero. In this implementation, the output includes not only the collaborative evaluation result but also the corresponding evaluation source contribution distribution data and unit correction trajectory data, used for result interpretation and feedback updates. The evaluation source contribution distribution data records the cumulative percentage of each evaluation source's actual contribution in all teaching behavior units, while the unit correction trajectory data records the path of change from the original evaluation data to the corrected evaluation data and then to the unit collaborative evaluation data in each teaching behavior unit, facilitating administrators to review how a certain conclusion was formed.

[0039] In this embodiment, the collaborative evaluation results do not stop after generation but enter a continuous updating phase. The dynamic update module acquires subsequent feedback reference data corresponding to the same teaching object and associates the subsequent feedback reference data with the collaborative evaluation results according to the teaching object identifier and time identifier to generate feedback deviation data. The subsequent feedback reference data can come from the quality of homework, the results of stage tests, student satisfaction follow-ups, and subsequent classroom performance indicators. The collection time is generally between 24 hours and 14 days after the original teaching activity ends. The feedback deviation data is denoted as Δf, defined as the collaborative evaluation result minus the subsequent feedback reference data, in points. If the collaborative evaluation result is 83 points and the subsequent feedback reference data is 79 points, then the feedback deviation data is 4 points, indicating that the current collaborative evaluation is slightly high. Subsequently, the baseline deviation data, group deviation data, behavioral pattern deviation data, and time series deviation data in the evaluation source deviation matrix are corrected based on the feedback deviation data to generate an updated evaluation source deviation matrix. This embodiment adopts an exponential sliding update method. If an old deviation value is denoted as Pold and the new feedback-driven correction amount is denoted as Δadj, then the updated deviation value is... =(1- )×Pold+ ×Δadj, where The value is taken as 0.1 to 0.2, and 0.15 is preferred in this embodiment. Since both Pold and Δadj are in units of minutes, therefore... The unit is still points. The construction of Δadj can be determined based on the apportioned contribution of feedback deviation data across relevant evaluation sources, group attributes, behavioral patterns, and time segments. For example, it can be fed back according to the proportion of each evaluation source's fusion contribution value in this fusion. Simultaneously, the historical stability, correction residuals, and objective reference consistency of each evaluation source are corrected based on the feedback deviation data to generate updated fusion contribution values. If an evaluation source consistently exhibits small feedback deviation data across multiple feedbacks, its historical stability and objective reference consistency will gradually increase, and its fusion contribution value will rise accordingly; conversely, it will decrease. The updated evaluation source deviation matrix and the updated fusion contribution values ​​are written into the collaborative evaluation process and directly invoked when processing new multi-source evaluation data in the next iteration, thus forming an adaptive closed-loop evaluation mechanism.

[0040] Furthermore, multi-source teaching behavior data, multi-source evaluation data, and objective reference data are obtained for the same teaching subject in the same teaching activity. The multi-source teaching behavior data is then segmented into multiple teaching behavior units according to behavioral event boundaries. Multi-source evaluation data is mapped to corresponding teaching behavior units to generate unit evaluation datasets, including: Extract time stamp information from teaching video data, classroom audio data, teaching platform operation record data, and evaluation record data; align the various types of data according to a unified time benchmark to generate an aligned dataset. Based on the behavior occurrence time, behavior duration, and behavior switching marker in the aligned dataset, the behavior event boundaries are identified, and the multi-source teaching behavior data is divided into multiple teaching behavior units according to the behavior event boundaries to generate a sequence of teaching behavior units. Extract the evaluation time, evaluation object identifier, and evaluation content identifier from the multi-source evaluation data, map the multi-source evaluation data to the corresponding teaching behavior unit, and generate unit-specific evaluation data. By associating the collection time and teaching object identification in the objective reference data with the corresponding evaluation data of the unit, objective reference data corresponding to each teaching behavior unit is generated. A unit evaluation dataset is generated based on the teaching behavior units, the corresponding evaluation data for each unit, and the objective reference data corresponding to each teaching behavior unit.

[0041] In a specific embodiment, the execution method of data acquisition and teaching behavior unit generation can be further refined in practical applications. After extracting the time stamp information from teaching video data, classroom audio data, teaching platform operation record data, and evaluation record data, the data acquisition module can first establish a unified time-series buffer and write all records into the same timeline index table according to the calibrated timestamps. The time resolution of this timeline index table can be set to 100 ms. If the original sampling granularity of a certain data type is higher than 100 ms, it is written in the nearest neighbor mapping method; if it is lower than 100 ms, the original granularity is retained and backfilled to the 100 ms grid during querying. After this processing, the identification of behavior occurrence time, behavior duration, and behavior switching marker can be completed under the same index framework. For behavior switching markers, this embodiment can set at least two of the following three conditions to be determined as valid boundaries: the video behavior category confidence change exceeds 0.35, the audio semantic tag change duration exceeds 2 s, and the platform operation type changes from browsing to publishing tasks or from displaying resources to initiating quizzes. Using dual-condition triggering can reduce the erroneous segmentation caused by noise from a single data source. When mapping multi-source evaluation data to teaching behavior units, if an evaluation content identifier spans two adjacent teaching behavior units—for example, if an evaluator submits an impression evaluation of the previous unit within 5 seconds of switching units—the assignment is determined by the semantic matching degree between the evaluation content identifier and the behavioral pattern of the previous unit. If the semantic matching degree is higher than 0.6, it is still mapped to the previous teaching behavior unit; otherwise, it is mapped to the current teaching behavior unit. This makes the unit evaluation dataset more consistent with the actual psychological timing of the evaluation.

[0042] Furthermore, based on the historical unit evaluation dataset and objective reference data, baseline deviation data, group deviation data, behavioral pattern deviation data, and time series deviation data corresponding to each evaluation source are calculated to construct an evaluation source deviation matrix, including: The original evaluation data of each evaluation source in different teaching behavior units are extracted from the historical unit evaluation dataset, and the deviation between the original evaluation data and the objective reference data is calculated based on the objective reference data to generate evaluation deviation data. By aggregating evaluation deviation data from the same evaluation source across all teaching behavior units, baseline deviation data is obtained. The evaluation deviation data were grouped according to the group attributes of the teaching subjects to obtain group deviation data. The evaluation deviation data were grouped according to the behavioral patterns corresponding to the teaching behavior units to obtain behavioral pattern deviation data; the evaluation deviation data were also segmented and statistically analyzed according to the evaluation time sequence to obtain time sequence deviation data. An evaluation source deviation matrix is ​​constructed based on baseline deviation data, group deviation data, behavioral pattern deviation data, and time series deviation data.

[0043] In a specific embodiment, based on the previously generated unit evaluation dataset, the evaluation source deviation matrix generation module further retrieves historical unit evaluation datasets and objective reference data to perform a unified analysis of the original evaluation data of each evaluation source in different teaching behavior units: first, it calculates the deviation between the original evaluation data and the objective reference data to obtain evaluation deviation data that directly reflects the degree of overestimation or underestimation; then, it summarizes the evaluation deviation data of the same evaluation source in all teaching behavior units, extracts its long-term stable overall deviation, and forms baseline deviation data; finally, it groups the evaluation deviation data according to the group attributes of the teaching objects to identify the evaluation source's performance when facing different groups. Are there any additional biases, forming group bias data? Based on the behavioral patterns corresponding to the teaching behavior units, further group them to obtain behavioral pattern bias data in different teaching scenarios, and statistically analyze them in segments according to the evaluation time sequence to capture the time drift of evaluation sources as the classroom stage or semester progresses, forming time-series bias data. Write the baseline bias data, group bias data, behavioral pattern bias data and time-series bias data into the evaluation source bias matrix, so that subsequent multi-dimensional bias correction can directly call the corresponding bias items layer by layer according to the evaluation source identifier, group attribute, behavioral pattern and time segment, thereby providing a stable basis for the hierarchical correction and collaborative integration of the original evaluation data.

[0044] Further, group bias data specifically include: Extract group attribute data corresponding to teaching object identifiers from historical unit evaluation datasets, and associate the group attribute data with the evaluation deviation data output from the same evaluation source to form a group evaluation deviation sequence; For each group attribute in the group evaluation deviation sequence, the deviation center value and deviation discrete value are calculated to generate group deviation characteristic data; The group deviation data is determined based on the distribution differences of the group deviation characteristic data among the group attributes, and the group deviation data is written into the evaluation source deviation matrix.

[0045] In a specific embodiment, based on the deviation matrix construction scheme, the formation of group deviation data can also adopt a more refined statistical caliber. In addition to professional title level and teaching experience range, the group attribute data corresponding to the teaching object identifier can also include the type of teaching language, the strength of the course's theoretical content, and the class's academic foundation level. After associating these group attribute data with the evaluation deviation data output from the same evaluation source, the group evaluation deviation sequence can be established either based on a single group attribute or a combination of group attributes. For example, senior professional title + large class teaching constitutes a composite group attribute. When calculating the deviation center value and deviation dispersion value for each group attribute in the group evaluation deviation sequence, this implementation method requires each group attribute to contain at least 50 evaluation deviation data points. If the number is less than this, it falls back to the single attribute level to avoid estimation distortion caused by insufficient samples. If the deviation center value in the group deviation feature data uses the median, and the deviation dispersion value uses the median absolute deviation, both are retained to two decimal places. When determining group bias data based on the distribution differences of group bias characteristic data among various group attributes, group attributes that deviate from the center value by more than 2 points and deviate from the discrete value by more than 1.5 points can be preferentially included in the evaluation source bias matrix. Group attributes that do not reach the threshold are recorded as 0 bias items. This preserves significant differences while controlling the matrix size.

[0046] Furthermore, the original evaluation data in the unit evaluation dataset is decomposed into bias components to obtain the true evaluation components, system bias components, context bias components, and disturbance components, including: Extract the original evaluation data, evaluation source identifiers, teaching behavior unit identifiers, and objective reference data from the unit evaluation dataset; The evaluation deviation data is determined based on the correspondence between the original evaluation data and the objective reference data; the evaluation deviation data is aggregated according to the evaluation source identifier to obtain the systematic deviation components; The evaluation deviation data were cross-aggregated according to the group attributes of the teaching objects and the behavioral patterns of the teaching behavior units to obtain the situational deviation component; After subtracting the system bias component and the situation bias component from the original evaluation data, we obtain the true evaluation component and the disturbance component. The part that changes in the same direction as the objective reference data is recorded as the true evaluation component, and the remaining fluctuating part is recorded as the disturbance component. The true evaluation component, system bias component, situational bias component, and disturbance component are used as input data for the correction process.

[0047] In a specific embodiment, based on the deviation component decomposition scheme, if it is desired to make the division between the true evaluation component and the disturbance component more stable, a cross-unit smoothing constraint can be introduced. Specifically, the first-order difference is performed on the remaining sequence after deducting the system deviation component and the situational deviation component within three consecutive teaching behavior units for the same evaluation source, resulting in a change sequence. This change sequence is compared with the first-order difference sequence of the corresponding objective reference data. If the signs are consistent and the change amplitude ratio is between 0.5 and 1.5, the change segment is recorded as the true evaluation component. If the signs are inconsistent, or the amplitude ratio exceeds 1.5, the portion exceeding the threshold is recorded as the disturbance component. Here, the amplitude ratio is the ratio of the absolute values ​​of the two changes, a dimensionless quantity. Setting the range from 0.5 to 1.5 is to tolerate differences in the sensitivity of different evaluation sources to objective changes, without excessively treating reasonable deviations as noise. After this decomposition, the true evaluation component is closer to the evaluation source's response to real teaching changes, while the disturbance component more centrally carries random errors.

[0048] Furthermore, based on the evaluation source deviation matrix, baseline correction, group correction, behavioral pattern correction, and time series correction are sequentially performed on the original evaluation data to obtain corrected evaluation data, including: Based on the baseline deviation data of the evaluation source corresponding to the original evaluation data, the original evaluation data is baseline corrected to generate baseline corrected data. Based on the group attribute data corresponding to the teaching objects and the group deviation data in the evaluation source deviation matrix, the baseline correction data is corrected for the group to generate the group correction data. Based on the behavioral pattern data corresponding to the teaching behavior unit and the behavioral pattern deviation data in the evaluation source deviation matrix, the group correction data is corrected to generate behavioral pattern correction data. Based on the time-series deviation data corresponding to the evaluation time, the behavioral pattern correction data is corrected in time series to generate correction evaluation data. The corrected evaluation data is stored in correspondence with the actual evaluation components for the calculation of evaluation discrepancy.

[0049] In a specific embodiment, based on the multi-dimensional bias correction stage, when the original evaluation data undergoes baseline correction, group correction, behavioral pattern correction, and time-series correction sequentially, a step-by-step verification mechanism can be set. After each level of correction is completed, the absolute value of the residual between the current correction result and the objective reference data is calculated. If the absolute value of the residual does not decrease but increases by more than 3 points compared to the previous stage, the correction magnitude reduction for that level is triggered. Taking group correction as an example, if the residual deteriorates after applying group bias data, the group bias data is multiplied by a reduction factor of 0.5 and corrected again; if it still deteriorates, that level of correction is skipped, and behavioral pattern correction is directly entered. This step-by-step verification mechanism can avoid overcorrection caused by the historical bias matrix when local samples drift. When the corrected evaluation data is stored corresponding to the real evaluation components, the change amount of each level of correction can also be stored in addition, which facilitates the visualization of subsequent unit correction trajectory data.

[0050] Furthermore, for the corrected evaluation data within the same teaching behavior unit, the evaluation discrepancy degree is calculated. When the evaluation discrepancy degree exceeds a preset threshold, the anomaly degree is calculated by combining the historical stability of the evaluation source, the corrected residual, and the consistency of the objective reference, generating abnormal evaluation source marker data, including: For the corrective evaluation data in the same teaching behavior unit, calculate the difference distribution between the corrective evaluation data of each evaluation source, and generate the evaluation divergence degree based on the concentration of the difference distribution; When the evaluation discrepancy exceeds a preset threshold, the stability data of the corresponding evaluation source in the historical unit evaluation dataset, the correction residual data in the current teaching behavior unit, and the objective reference consistency data are extracted. The stability data, correction residual data, and objective reference consistency data are then jointly sorted to generate an anomaly score. Based on the predetermined ranking of the anomaly degree among all evaluation sources, anomaly evaluation source labeling data is generated and sent to the fusion contribution value adjustment step.

[0051] In one specific embodiment, the threshold for evaluating the degree of divergence is not fixed, based on the calculation scheme for evaluation divergence and anomaly. When the number of evaluation sources is small, even a few differences may lead to a large standard deviation. Therefore, in this embodiment, the preset threshold can be adaptively adjusted according to the number of evaluation sources, n. If the base threshold is 0.12, where n is not less than 3, the threshold remains at 0.12 when n=6; when n=3, the threshold is increased to approximately 0.17 to reduce false alarms when the number of evaluation sources is too small; when n=12, the threshold is decreased to approximately 0.085 to more sensitively identify anomalies in a large set of evaluation sources. When jointly ranking stability data, corrected residual data, and objective reference consistency data, a ranking method can be used instead of direct numerical weighting to reduce the impact of normalization errors of different dimensions. The higher the historical stability ranking, the better; the lower the corrected residual ranking, the better; and the higher the objective reference consistency ranking, the better. Then, the anomaly ranking is synthesized by 0.4, 0.35, and 0.25. The anomaly evaluation source labeling data is still determined by the predetermined position of the anomaly among all evaluation sources.

[0052] Furthermore, the fusion contribution value of the corresponding corrected evaluation data is adjusted based on the anomaly evaluation source marker data, including: Extract historical stability, corrected residuals, objective reference consistency, and abnormal evaluation source labeling data for each evaluation source; Historical stability and consistency with objective references are used as positive contribution data, while corrected residuals and abnormal evaluation source labeling data are used as suppression contribution data to generate fusion contribution values ​​for each evaluation source. When the abnormal evaluation source labeling data indicates that the evaluation source is located in a predetermined position with a high degree of abnormality, the fusion contribution value of the corresponding corrected evaluation data of the evaluation source is reduced. When the abnormal evaluation source labeling data does not indicate that the evaluation source is located in a predetermined position with a high degree of abnormality, the fusion contribution value of the corresponding corrected evaluation data of the evaluation source is maintained. The adjusted fusion contribution value will be used for collaborative fusion.

[0053] In a specific embodiment, based on the fusion contribution value adjustment scheme, if the abnormal evaluation source marker data indicates that the evaluation source is located in a predetermined position with a high degree of abnormality, in addition to reducing the fusion contribution value of the corresponding corrected evaluation data of the evaluation source, a minimum retention contribution value can also be set to prevent a certain evaluation source from being completely marginalized. In this embodiment, the minimum retention contribution value is set to 0.03. That is, even if the normalized fusion contribution value is significantly reduced due to anomaly suppression, it will not be lower than 0.03. The reason for this arrangement is that abnormal evaluation sources are not the same as invalid evaluation sources; their data may still contain local information that other evaluation sources have failed to capture. Retaining a small proportion of contribution is more in line with the robustness principle of multi-source fusion. Correspondingly, when the abnormal evaluation source marker data does not indicate that the evaluation source is located in a predetermined position with a high degree of abnormality, the fusion contribution value of the corresponding corrected evaluation data of the evaluation source is maintained and not further amplified, avoiding the system from becoming overly dependent on individual evaluation sources.

[0054] Furthermore, based on the fusion contribution values ​​of each evaluation source, the corrective evaluation data in the same teaching behavior unit is collaboratively fused to generate unit collaborative evaluation data. Multiple unit collaborative evaluation data are then aggregated to output the collaborative evaluation results corresponding to the teaching object. These results are compared with subsequently acquired feedback reference data to update the evaluation source deviation matrix and fusion contribution values. This is used for collaborative evaluation processing of multi-source evaluation data, including: Extract all correction evaluation data and their corresponding fusion contribution values ​​from the same teaching behavior unit, and weight and aggregate the correction evaluation data according to the fusion contribution values ​​to generate unit collaborative evaluation data; The collaborative evaluation data of multiple teaching behavior units are arranged in chronological order according to the teaching activities, and aggregated according to the duration and importance of each teaching behavior unit to generate the collaborative evaluation results corresponding to the teaching object. Output the evaluation source contribution distribution data and unit correction trajectory data corresponding to the collaborative evaluation results for result interpretation and feedback updates; Obtain subsequent feedback reference data corresponding to the same teaching object, and associate the subsequent feedback reference data with the collaborative evaluation results according to the teaching object identifier and time identifier to generate feedback deviation data; Based on the feedback deviation data, the baseline deviation data, group deviation data, behavioral pattern deviation data, and time series deviation data in the evaluation source deviation matrix are corrected to generate an updated evaluation source deviation matrix. Simultaneously, based on the feedback deviation data, the historical stability, correction residuals, and objective reference consistency of each evaluation source are corrected to generate an updated fusion contribution value. The updated evaluation source deviation matrix and the updated fusion contribution value are written into the collaborative evaluation process.

[0055] In a specific embodiment, based on the collaborative integration and aggregation scheme, the collaborative evaluation data of multiple teaching behavior units, arranged chronologically according to the teaching activities, can be further aggregated in segments according to the teaching activity structure. For example, a lesson can be divided into three stages: introduction, development, and conclusion. The collaborative evaluation results for each stage can be calculated separately, and then aggregated into the final collaborative evaluation results with stage weights of 0.2, 0.6, and 0.2. This approach is particularly suitable for classrooms with relatively stable teaching structures, helping to prevent the development stage from completely overshadowing the important information of the introduction and conclusion stages due to the large number of units. When outputting the evaluation source contribution distribution data corresponding to the collaborative evaluation results, it can be output simultaneously at three levels: evaluation source type, individual evaluation source, and teaching behavior unit. This allows administrators to see the total contribution of student evaluation, peer evaluation, and supervisor evaluation, and also to trace back to which specific type of evaluation source played a major role in a particular unit. The unit correction trajectory data can be recorded sequentially in the order of original evaluation data—baseline correction data—group correction data—behavioral pattern correction data—corrected evaluation data—unit collaborative evaluation data. This ensures that the generation logic of the evaluation conclusions is complete and traceable during result interpretation.

[0056] The correction of the evaluation source deviation matrix and fusion contribution value based on feedback deviation data can be divided into two paths: fast update and slow update. For baseline deviation data and time-series deviation data, which are more susceptible to recent classroom pace and evaluation environment, a larger update coefficient, such as 0.20, can be used. For group deviation data and behavioral pattern deviation data, which reflect longer-term and more structured deviation characteristics, a smaller update coefficient, such as 0.08, can be used. The updates of historical stability, correction residuals, and objective reference consistency can be set at different speeds according to the evaluation source type. For example, for student evaluation sources, which are numerous and fluctuate greatly, a sliding window length of 50 teaching behavior units can be used; for supervisory evaluation sources, which are few and highly specialized, a window length of 100 teaching behavior units can be used. With this layered update, the system can respond to recent changes in the evaluation ecosystem without disrupting the long-term deviation structure due to short-term noise.

[0057] Example 2 This application discloses a collaborative evaluation system for multi-source teaching behavior data, including: The data acquisition module is used to acquire multi-source teaching behavior data, multi-source evaluation data, and objective reference data of the same teaching object in the same teaching activity. It also divides the multi-source teaching behavior data into multiple teaching behavior units according to the boundaries of the behavior events, maps the multi-source evaluation data to the corresponding teaching behavior units, and generates a unit evaluation dataset. The evaluation source deviation matrix generation module is used to calculate the baseline deviation data, group deviation data, behavioral pattern deviation data and time series deviation data corresponding to each evaluation source based on the historical unit evaluation dataset and objective reference data, and to construct the evaluation source deviation matrix. The multi-dimensional deviation correction module is used to decompose the original evaluation data in the unit evaluation dataset into deviation components to obtain the true evaluation components, system deviation components, situational deviation components and disturbance components. Based on the evaluation source deviation matrix, the module sequentially performs baseline correction, group correction, behavioral pattern correction and time series correction on the original evaluation data to obtain the corrected evaluation data. The tagging module is used to calculate the evaluation divergence degree for the corrected evaluation data in the same teaching behavior unit. When the evaluation divergence degree is greater than the preset threshold, the abnormality degree is calculated by combining the historical stability of the evaluation source, the correction residual and the consistency of the objective reference, generating abnormal evaluation source tagging data, and adjusting the fusion contribution value of the corresponding corrected evaluation data according to the abnormal evaluation source tagging data. The dynamic update module is used to collaboratively integrate the correction evaluation data in the same teaching behavior unit based on the fusion contribution value of each evaluation source, generate unit collaborative evaluation data, aggregate the collaborative evaluation data of multiple units, output the collaborative evaluation results corresponding to the teaching object, compare them with the feedback reference data obtained later, and update the evaluation source deviation matrix and fusion contribution value for collaborative evaluation processing of multi-source evaluation data.

[0058] Specifically, this embodiment also provides a system implementation structure combined with the method. The data acquisition module can be composed of a data access unit, a time-series alignment unit, a behavior segmentation unit, and a mapping and association unit. The data access unit receives teaching video data, classroom audio data, teaching platform operation record data, and evaluation record data through a streaming interface; the time-series alignment unit performs clock offset correction and alignment dataset generation under a unified time reference; the behavior segmentation unit is responsible for identifying behavioral event boundaries and outputting teaching behavior unit sequences; the mapping and association unit is responsible for mapping multi-source evaluation data to corresponding teaching behavior units and associating objective reference data with the corresponding evaluation data of the unit to form a unit evaluation dataset. The evaluation source deviation matrix generation module can be composed of a historical data extraction unit, a deviation calculation unit, a group statistics unit, and a matrix writing unit, wherein the group statistics unit calculates baseline deviation data, group deviation data, behavior pattern deviation data, and time-series deviation data respectively. The multi-dimensional deviation correction module can be composed of a component decomposition unit and a hierarchical correction unit. The former outputs true evaluation components, system deviation components, context deviation components, and disturbance components, while the latter completes baseline correction, group correction, behavior pattern correction, and time-series correction in a predetermined order. The labeling module can consist of a divergence analysis unit, an anomaly calculation unit, and a contribution adjustment unit, which respectively process the evaluation divergence degree, anomaly degree, and fusion contribution value. The dynamic update module can consist of a fusion unit, an aggregation unit, a feedback correlation unit, and a parameter update unit, which is responsible for generating collaborative evaluation data, outputting collaborative evaluation results, and writing back and updating the evaluation source deviation matrix and fusion contribution value. Each module can be deployed on the same server or in a distributed microservice architecture; as long as its processing logic and data flow relationship remain consistent, the collaborative evaluation processing effect of this invention can be achieved.

[0059] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention, they should all fall within the protection scope of the present invention.

[0060] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0061] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A collaborative evaluation method for multi-source teaching behavior data, characterized in that, Includes the following steps: Acquire multi-source teaching behavior data, multi-source evaluation data, and objective reference data for the same teaching object in the same teaching activity, and divide the multi-source teaching behavior data into multiple teaching behavior units according to the boundaries of behavior events, and map the multi-source evaluation data to the corresponding teaching behavior units to generate unit evaluation datasets; Based on the historical unit evaluation dataset and objective reference data, the baseline deviation data, group deviation data, behavioral pattern deviation data and time series deviation data corresponding to each evaluation source are calculated, and the evaluation source deviation matrix is ​​constructed. The original evaluation data in the unit evaluation dataset is decomposed into deviation components to obtain the true evaluation components, system deviation components, situational deviation components and disturbance components. Based on the evaluation source deviation matrix, baseline correction, group correction, behavioral pattern correction and time series correction are performed on the original evaluation data in sequence to obtain the corrected evaluation data. For the corrected evaluation data in the same teaching behavior unit, the evaluation divergence degree is calculated. When the evaluation divergence degree is greater than the preset threshold, the abnormality degree is calculated by combining the historical stability of the evaluation source, the correction residual and the consistency of the objective reference, generating abnormal evaluation source label data, and adjusting the fusion contribution value of the corresponding corrected evaluation data according to the abnormal evaluation source label data. Based on the fusion contribution value of each evaluation source, the correction evaluation data in the same teaching behavior unit is collaboratively fused to generate unit collaborative evaluation data. The collaborative evaluation data of multiple units is aggregated to output the collaborative evaluation results corresponding to the teaching object. The results are compared with the feedback reference data obtained later to update the evaluation source deviation matrix and fusion contribution value for collaborative evaluation processing of multi-source evaluation data.

2. The collaborative evaluation method for multi-source teaching behavior data according to claim 1, characterized in that, Acquire multi-source teaching behavior data, multi-source evaluation data, and objective reference data for the same learner in the same teaching activity. Divide the multi-source teaching behavior data into multiple teaching behavior units based on behavioral event boundaries. Map the multi-source evaluation data to the corresponding teaching behavior units to generate unit evaluation datasets, including: Extract time stamp information from teaching video data, classroom audio data, teaching platform operation record data, and evaluation record data; align the various types of data according to a unified time benchmark to generate an aligned dataset. Based on the behavior occurrence time, behavior duration, and behavior switching marker in the aligned dataset, the behavior event boundaries are identified, and the multi-source teaching behavior data is divided into multiple teaching behavior units according to the behavior event boundaries to generate a sequence of teaching behavior units. Extract the evaluation time, evaluation object identifier, and evaluation content identifier from the multi-source evaluation data, map the multi-source evaluation data to the corresponding teaching behavior unit, and generate unit-specific evaluation data. By associating the collection time and teaching object identification in the objective reference data with the corresponding evaluation data of the unit, objective reference data corresponding to each teaching behavior unit is generated. A unit evaluation dataset is generated based on the teaching behavior units, the corresponding evaluation data for each unit, and the objective reference data corresponding to each teaching behavior unit.

3. The collaborative evaluation method for multi-source teaching behavior data according to claim 2, characterized in that, Based on historical unit evaluation datasets and objective reference data, baseline deviation data, group deviation data, behavioral pattern deviation data, and time series deviation data corresponding to each evaluation source are calculated to construct an evaluation source deviation matrix, including: The original evaluation data of each evaluation source in different teaching behavior units are extracted from the historical unit evaluation dataset, and the deviation between the original evaluation data and the objective reference data is calculated based on the objective reference data to generate evaluation deviation data. By aggregating evaluation deviation data from the same evaluation source across all teaching behavior units, baseline deviation data is obtained. The evaluation deviation data were grouped according to the group attributes of the teaching subjects to obtain group deviation data. The evaluation deviation data were grouped according to the behavioral patterns corresponding to the teaching behavior units to obtain behavioral pattern deviation data; the evaluation deviation data were also segmented and statistically analyzed according to the evaluation time sequence to obtain time sequence deviation data. An evaluation source deviation matrix is ​​constructed based on baseline deviation data, group deviation data, behavioral pattern deviation data, and time series deviation data.

4. The collaborative evaluation method for multi-source teaching behavior data according to claim 3, characterized in that, Group bias data, specifically including: Extract group attribute data corresponding to teaching object identifiers from historical unit evaluation datasets, and associate the group attribute data with the evaluation deviation data output from the same evaluation source to form a group evaluation deviation sequence; For each group attribute in the group evaluation deviation sequence, the deviation center value and deviation discrete value are calculated to generate group deviation characteristic data; The group deviation data is determined based on the distribution differences of the group deviation characteristic data among the group attributes, and the group deviation data is written into the evaluation source deviation matrix.

5. The collaborative evaluation method for multi-source teaching behavior data according to claim 1, characterized in that, The original evaluation data in the unit evaluation dataset is decomposed into bias components to obtain the true evaluation components, system bias components, context bias components, and disturbance components, including: Extract the original evaluation data, evaluation source identifiers, teaching behavior unit identifiers, and objective reference data from the unit evaluation dataset; The evaluation deviation data is determined based on the correspondence between the original evaluation data and the objective reference data; the evaluation deviation data is aggregated according to the evaluation source identifier to obtain the systematic deviation components; The evaluation deviation data were cross-aggregated according to the group attributes of the teaching objects and the behavioral patterns of the teaching behavior units to obtain the situational deviation component; After subtracting the system bias component and the situation bias component from the original evaluation data, we obtain the true evaluation component and the disturbance component. The part that changes in the same direction as the objective reference data is recorded as the true evaluation component, and the remaining fluctuating part is recorded as the disturbance component. The true evaluation component, system bias component, situational bias component, and disturbance component are used as input data for the correction process.

6. The collaborative evaluation method for multi-source teaching behavior data according to claim 5, characterized in that, Based on the evaluation source deviation matrix, baseline correction, group correction, behavioral pattern correction, and time series correction are sequentially performed on the original evaluation data to obtain corrected evaluation data, including: Based on the baseline deviation data of the evaluation source corresponding to the original evaluation data, the original evaluation data is baseline corrected to generate baseline corrected data. Based on the group attribute data corresponding to the teaching objects and the group deviation data in the evaluation source deviation matrix, the baseline correction data is corrected for the group to generate the group correction data. Based on the behavioral pattern data corresponding to the teaching behavior unit and the behavioral pattern deviation data in the evaluation source deviation matrix, the group correction data is corrected to generate behavioral pattern correction data. Based on the time-series deviation data corresponding to the evaluation time, the behavioral pattern correction data is corrected in time series to generate correction evaluation data. The corrected evaluation data is stored in correspondence with the actual evaluation components for the calculation of evaluation discrepancy.

7. The collaborative evaluation method and system for multi-source teaching behavior data according to claim 1, characterized in that, For corrected evaluation data within the same teaching behavior unit, the evaluation discrepancy degree is calculated. When the evaluation discrepancy degree exceeds a preset threshold, the anomaly degree is calculated by combining the historical stability of the evaluation source, the corrected residual, and the consistency of objective references, generating anomaly evaluation source marker data, including: For the corrective evaluation data in the same teaching behavior unit, calculate the difference distribution between the corrective evaluation data of each evaluation source, and generate the evaluation divergence degree based on the concentration of the difference distribution; When the evaluation discrepancy exceeds a preset threshold, the stability data of the corresponding evaluation source in the historical unit evaluation dataset, the correction residual data in the current teaching behavior unit, and the objective reference consistency data are extracted. The stability data, correction residual data, and objective reference consistency data are then jointly sorted to generate an anomaly score. Based on the predetermined ranking of the anomaly degree among all evaluation sources, anomaly evaluation source labeling data is generated and sent to the fusion contribution value adjustment step.

8. The collaborative evaluation method for multi-source teaching behavior data according to claim 7, characterized in that, Adjust the fusion contribution value of the corresponding corrected evaluation data based on the anomaly evaluation source marker data, including: Extract historical stability, corrected residuals, objective reference consistency, and abnormal evaluation source labeling data for each evaluation source; Historical stability and consistency with objective references are used as positive contribution data, while corrected residuals and abnormal evaluation source labeling data are used as suppression contribution data to generate fusion contribution values ​​for each evaluation source. When the abnormal evaluation source labeling data indicates that the evaluation source is located in a predetermined position with a high degree of abnormality, the fusion contribution value of the corresponding corrected evaluation data of the evaluation source is reduced. When the abnormal evaluation source labeling data does not indicate that the evaluation source is located in a predetermined position with a high degree of abnormality, the fusion contribution value of the corresponding corrected evaluation data of the evaluation source is maintained. The adjusted fusion contribution value will be used for collaborative fusion.

9. The collaborative evaluation method for multi-source teaching behavior data according to claim 1, characterized in that, Based on the fusion contribution values ​​of each evaluation source, the corrective evaluation data in the same teaching behavior unit is collaboratively fused to generate unit collaborative evaluation data. Multiple unit collaborative evaluation data are then aggregated to output the collaborative evaluation results corresponding to the teaching object. These results are compared with subsequently acquired feedback reference data to update the evaluation source deviation matrix and fusion contribution values. This process is used for collaborative evaluation processing of multi-source evaluation data, including: Extract all correction evaluation data and their corresponding fusion contribution values ​​from the same teaching behavior unit, and weight and aggregate the correction evaluation data according to the fusion contribution values ​​to generate unit collaborative evaluation data; The collaborative evaluation data of multiple teaching behavior units are arranged in chronological order according to the teaching activities, and aggregated according to the duration and importance of each teaching behavior unit to generate the collaborative evaluation results corresponding to the teaching object. Output the evaluation source contribution distribution data and unit correction trajectory data corresponding to the collaborative evaluation results for result interpretation and feedback updates; Obtain subsequent feedback reference data corresponding to the same teaching object, and associate the subsequent feedback reference data with the collaborative evaluation results according to the teaching object identifier and time identifier to generate feedback deviation data; Based on the feedback deviation data, the baseline deviation data, group deviation data, behavioral pattern deviation data, and time series deviation data in the evaluation source deviation matrix are corrected to generate an updated evaluation source deviation matrix. Simultaneously, based on the feedback deviation data, the historical stability, correction residuals, and objective reference consistency of each evaluation source are corrected to generate an updated fusion contribution value. The updated evaluation source deviation matrix and the updated fusion contribution value are written into the collaborative evaluation process.

10. A collaborative evaluation system for multi-source teaching behavior data, applied to the collaborative evaluation method for multi-source teaching behavior data as described in any one of claims 1-9, comprising: The data acquisition module is used to acquire multi-source teaching behavior data, multi-source evaluation data, and objective reference data of the same teaching object in the same teaching activity. It also divides the multi-source teaching behavior data into multiple teaching behavior units according to the boundaries of the behavior events, maps the multi-source evaluation data to the corresponding teaching behavior units, and generates a unit evaluation dataset. The evaluation source deviation matrix generation module is used to calculate the baseline deviation data, group deviation data, behavioral pattern deviation data and time series deviation data corresponding to each evaluation source based on the historical unit evaluation dataset and objective reference data, and to construct the evaluation source deviation matrix. The multi-dimensional deviation correction module is used to decompose the original evaluation data in the unit evaluation dataset into deviation components to obtain the true evaluation components, system deviation components, situational deviation components and disturbance components. Based on the evaluation source deviation matrix, the module sequentially performs baseline correction, group correction, behavioral pattern correction and time series correction on the original evaluation data to obtain the corrected evaluation data. The tagging module is used to calculate the evaluation divergence degree for the corrected evaluation data in the same teaching behavior unit. When the evaluation divergence degree is greater than the preset threshold, the abnormality degree is calculated by combining the historical stability of the evaluation source, the correction residual and the consistency of the objective reference, generating abnormal evaluation source tagging data, and adjusting the fusion contribution value of the corresponding corrected evaluation data according to the abnormal evaluation source tagging data. The dynamic update module is used to collaboratively integrate the correction evaluation data in the same teaching behavior unit based on the fusion contribution value of each evaluation source, generate unit collaborative evaluation data, aggregate the collaborative evaluation data of multiple units, output the collaborative evaluation results corresponding to the teaching object, compare them with the feedback reference data obtained later, and update the evaluation source deviation matrix and fusion contribution value for collaborative evaluation processing of multi-source evaluation data.

Images