Method and system for dynamically evaluating online and offline mixed teaching performance based on PDCA (Packet Data Convergence Architecture) circulation

Abstract

The invention relates to the field of education data mining and teaching process evaluation, in particular to an online and offline mixed teaching performance dynamic evaluation method and system based on a PDCA (Packet Data Convergence Architecture) circulation, and the method comprises the steps: generating initial Rubric configuration and formulating a multi-source data collection scheme through the analysis of course types and learning condition information around the PDCA circulation; then, online and offline multi-modal information is collected and preprocessed, and data with labels and basic statistics are obtained; and in combination with causal graph construction and quantum multi-group fuzzy weight optimization, carrying out comprehensive evaluation on the data, outputting adaptive Rubric configuration and student scores, and generating a readable report and an early warning through a neural fuzzy abstract. And according to the score and early warning, implementing difference enhancement and holographic AR feedback, collecting intervention process data again, and dynamically adjusting Rubric to complete iteration. According to the invention, the teaching performance can be continuously optimized, and accurate early warning and personalized intervention are realized.

Description

Technical Field

[0001] The present invention relates to the field of educational data mining and teaching process evaluation, and particularly to a dynamic evaluation method and system for the performance of online-offline hybrid teaching based on the PDCA cycle. Background Art

[0002] In teaching practice, teachers can continuously revise teaching plans and accurately guide students' learning processes through the cycle of "Plan - Do - Check - Act". However, most of the existing technologies focus on static evaluation based on results (Chinese invention patent, publication number: CN116934543A, title: Online-offline hybrid teaching effect evaluation and learning warning method), or only collect a small amount of data at specific stages; the traditional multi-dimensional scoring, clustering, and single-test warning methods used are prone to ignoring real-time multi-modal information during the teaching process and the dynamic performance of students in aspects such as emotion and practice, resulting in difficulties in deep, personalized, and cyclic intervention. Summary of the Invention

[0003] In view of the many problems existing in the above-mentioned prior art, the present invention provides a dynamic evaluation method and system for the performance of online-offline hybrid teaching based on the PDCA cycle. The present invention first generates an initial Rubric and formulates a data collection plan to collect students' behavior, emotion, and achievement information in the online-offline multi-modal environment. Subsequently, the Rubric weights are constructed through a causal graph and optimized by quantum multi-group search to generate student scores, and a readable warning report is generated using a neuro-fuzzy summary. After performing difference reinforcement and AR feedback on key students, new data is collected and the Rubric is adjusted again to achieve cyclic improvement, ultimately improving teaching performance.

[0004] A dynamic evaluation method for the performance of online-offline hybrid teaching based on the PDCA cycle includes the following steps:

[0005] Analyze the type, teaching syllabus, and learning situation information of the target course, clarify the evaluation dimensions and the initial weight range and threshold, generate the initial Rubric configuration data, and formulate a multi-source data collection plan to obtain the collection configuration data; based on the data collection plan, obtain multi-modal online-offline information and perform preprocessing to generate labeled multi-modal data and basic statistical data;

[0006] Use causal graph construction and quantum multi-group fuzzy weight tuning to comprehensively evaluate the labeled multi-modal data and basic statistical data, obtain the adaptive Rubric configuration data and the phased student score data, and generate readable summary information and warning report data through a neuro-fuzzy summary;

[0007] Implement differential reinforcement training and holographic augmented reality feedback based on phased student scoring data and early warning report data, collect new teaching intervention process data, and adjust the adaptive Rubric configuration data to obtain the iterated Rubric configuration data, so as to achieve dynamic improvement of the online-offline hybrid teaching performance.

[0008] Preferably, after analyzing the type of the target course, preliminary evaluation dimensions are respectively set for the knowledge objectives and practice links in the teaching syllabus, and the preliminary weight intervals and thresholds of each dimension are determined after associating the preliminary evaluation dimensions with the learning situation information.

[0009] Preferably, when formulating the multi-source data collection scheme, the feasibility of the collection equipment and privacy policy is evaluated, and the acquisition frequencies and time alignment rules of camera images, heart rate sensors, and discussion text information are respectively determined for the online and offline teaching environments.

[0010] Preferably, when obtaining multi-modal online-offline information based on the data collection scheme, coordinate correction is performed on the body key point trajectories generated by the motion capture device, and the trajectories after coordinate correction and the heart rate sensor information are synchronized according to the timestamps.

[0011] Preferably, when performing preprocessing, the key indicators in the video viewing logs and discussion texts are mapped to the preliminary evaluation dimensions through noise filtering and label annotation to generate labeled multi-modal data and basic statistical data.

[0012] Preferably, when constructing the causal graph, the causal strength between the behavior nodes, emotion nodes, and teaching result nodes is extracted from the labeled multi-modal data based on statistical independence tests, and a confidence threshold is assigned to each causal edge.

[0013] Preferably, quantum multi-population fuzzy weight tuning screens out the optimized adaptive Rubric configuration data and phased student scoring data through multiple iterations of particles in the Rubric parameter search space and in combination with the causal strength evaluation index.

[0014] Preferably, the neuro-fuzzy summary uses the phased student scoring data in the adaptive Rubric configuration data and the causal relationships constructed by the causal graph to generate readable summary information and early warning report data and label the key behavior defects or skill deficiency links.

[0015] Preferably, during the implementation of differential reinforcement training and holographic augmented reality feedback, time alignment is performed on the new teaching intervention process data and compared with the adaptive Rubric configuration data, and the parameters of each dimension are adjusted according to the comparison results to generate the iterated Rubric configuration data.

[0016] A dynamic evaluation system for online and offline hybrid teaching performance based on the PDCA cycle is used to implement the dynamic evaluation method for online and offline hybrid teaching performance based on the PDCA cycle. The system includes:

[0017] The course analysis and rubric generation module is used to analyze the type, syllabus and learning information of the target course, clarify the evaluation dimensions and preliminary weight range and threshold, generate initial rubric configuration data, and formulate a multi-source data collection plan to obtain collection configuration data;

[0018] A multimodal acquisition and preprocessing module, which is used to acquire multimodal online and offline information based on the data acquisition scheme and perform time alignment, noise filtering and labeling on the information to generate labeled multimodal data and basic statistical data;

[0019] The causal graph and quantum tuning module is used to comprehensively evaluate the labeled multimodal data and basic statistical data using causal graph construction and quantum multi-group fuzzy weight tuning to obtain adaptive rubric configuration data and staged student scoring data, and generate readable summary information and early warning report data through neuro-fuzzy summarization;

[0020] The differential reinforcement and feedback module is used to implement differential reinforcement training and holographic augmented reality feedback based on periodic student scoring data and early warning report data, and to adjust the adaptive rubric configuration data after collecting new teaching intervention process data to obtain iterative rubric configuration data and achieve dynamic improvement of online and offline hybrid teaching performance.

[0021] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0022] The present invention realizes the deep integration of online and offline multi-source data through causal graph construction and quantum multi-group fuzzy weight tuning technology, which can accurately identify key influencing factors and dynamically adjust Rubric weights.

[0023] The present invention uses neuro-fuzzy summary technology to achieve a readable summary of evaluation results and causal chains, and automatically generates early warning report data to timely discover students' weaknesses.

[0024] The present invention realizes targeted intervention on problem students in the PDCA cycle through differentiated reinforcement training and holographic augmented reality feedback technology, and continuously collects new data to adjust Rubric, significantly improving the performance and flexibility of online and offline hybrid teaching. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the process of the present invention;

[0026] Figure 2 Schematic diagram of the relationship between causal graph construction and quantum tuning in the present invention;

[0027] Figure 3 Schematic diagram of difference degree enhancement and holographic augmented reality feedback in the present invention;

[0028] Figure 4 Schematic structural diagram of the system of the present invention. Detailed implementation manners

[0029] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for the sake of explanation, numerous specific details are set forth in order to provide a thorough understanding of the embodiments of the present disclosure. However, obviously, one or more embodiments can be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily obscuring the concepts of the present disclosure.

[0030] The terms used herein are merely for describing specific embodiments and are not intended to limit the present disclosure. The terms "including", "comprising", etc. used herein indicate the presence of the described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0031] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0032] As Figure 1 shown, an online-offline hybrid teaching performance dynamic evaluation method based on the PDCA cycle includes the following steps:

[0033] Analyze the type, syllabus, and learning situation information of the target course, clarify the evaluation dimensions and the initial weight range and threshold, generate initial Rubric configuration data, and formulate a multi-source data collection plan to obtain collection configuration data; Based on the data collection plan, obtain multi-modal online and offline information and perform preprocessing to generate labeled multi-modal data and basic statistical data;

[0034] Based on the "Plan" and "Do" links of the PDCA cycle, the present invention analyzes the type of target course, the syllabus and the learning information, determines the multiple dimensions that need to be measured in course teaching (such as theoretical mastery, practical application, collaborative communication, etc.), and sets a preliminary weight interval and threshold for each dimension to form an initial Rubric configuration data. Rubric here refers to an indicator system used for quantitative or hierarchical evaluation in the field of teaching. By decomposing and weighting the teaching objectives and learning characteristics, it lays a quantitative foundation for subsequent evaluation.

[0035] While generating the initial Rubric configuration data, the present invention has developed a multi-source data collection plan to guide various online and offline data sources, such as camera image information of offline classes, heart rate sensor values, online video logs, discussion texts, and test homework records, etc., which are uniformly acquired and integrated in subsequent steps. In order to allow subsequent processes to be associated with specific students or a certain teaching period, the present invention performs timestamp comparison and noise filtering on the collected multimodal information, and labels valuable behavioral features to form labeled multimodal data and basic statistical data. For example, for a student, his camera expression features, heart rate fluctuation values, and key speech records in online discussions can be associated at the same time, and these features can be marked under the same label as input for subsequent analysis.

[0036] In actual applications, the generation of the initial Rubric configuration includes:

[0037] Position the course into multiple evaluation dimensions such as knowledge dimension K, practice dimension P, and collaboration dimension C, and assign a preliminary weight w to each dimension K 、w P 、w C and threshold range; in mathematical form, the initial evaluation score can be set as:

[0038] S initial =w K ×K+w P ×P+w C ×C

[0039] Where S initial represents the overall evaluation of the student under the initial Rubric, K, P, and C represent the scores of the student in the three dimensions of knowledge, practice, and collaboration, respectively, and w K 、w P 、w C are the preliminary weights of these dimensions respectively; if the syllabus clearly states that a certain type of knowledge point or practical link is more important, the corresponding weight can be increased, and the threshold can be set based on the experience of past learning data.

[0040] The formulation of the multi-source data acquisition plan includes: considering the teaching site environment, listing the "offline classroom acquisition sources" (such as camera image data, motion capture device values) and "online platform acquisition sources" (such as video viewing logs, discussion speech texts); specifying the acquisition frequencies and time alignment methods of various devices or platforms (such as uniformly using the course start time as zero and accumulating seconds backward), and stipulating the mapping rules for identifying student IDs in the privacy policy to ensure that multi-modal information can be bound to specific students during subsequent analysis; setting noise filtering rules, for example, discarding a camera frame when its blur degree is higher than a certain threshold, and inserting the median or previous value if there is a breakpoint in the heart rate sensor.

[0041] Multi-modal data acquisition and preprocessing include: continuously obtaining various online and offline information from the actual teaching process according to the data acquisition plan; performing time alignment. Assume the timestamp of the camera image data is T video , and the timestamp of the heart rate sensor is T heart . It is necessary to use a fixed offset or matching algorithm to make ΔT ≤ δ (δ is the acceptable alignment deviation amount), so that the image frames and heart rate records at the same moment can be corresponding to the same behavior state; through label annotation, unify and summarize the aligned image frames, heart rate, online quiz scores and other information under the same student ID and label, and output the labeled multi-modal data; at the same time, generate statistical indicators such as the total viewing time of the video and the number of discussion speeches to form basic statistical data.

[0042] In one embodiment, in a course that emphasizes both theory and practical operation, the initial Rubric configuration data defines: the weight of the knowledge dimension w K = 0.4, the weight of the practice dimension w P = 0.4, and the weight of the collaboration dimension w C = 0.2. If within a teaching period, a certain student's knowledge dimension score K = 80, practice dimension score P = 70, and collaboration dimension score C = 60, then according to the above formula:

[0043] S initial = 0.4×80 + 0.4×70 + 0.2×60 = 74

[0044] This example means that his comprehensive initial score is 74. If camera shooting and wearable heart rate collection are enabled in the classroom, relevant timestamps and frame features can be obtained after 20 minutes; after noise filtering and alignment, the multi-modal information of the student's performance from the 10th minute to the 15th minute is labeled under the "phase research quiz" label, and the basic statistical data can also record his performance during this period, such as the correct rate, abnormal heart rate peak, etc. Subsequent steps can further evaluate the teaching performance based on these labeled multi-modal data, and perform differential reinforcement training or holographic augmented reality feedback to continuously improve the teaching.

[0045] Through this targeted analysis of target courses, teaching outlines and learning information, the priorities of different teaching objectives can be clearly distinguished in the initial stage (Plan) of the PDCA cycle, thereby establishing initial Rubric configuration data that fits actual teaching needs; with the help of multi-source data collection solutions, both offline and online teaching scenarios are covered, making the collected information more comprehensive; finally, the pre-processed synthesized labeled multimodal data and basic statistical data provide sufficient basis for the subsequent "Check" and "Act" links in the PDCA cycle, enabling teachers to accurately locate learning pain points, reasonably allocate teaching resources and dynamically adjust scoring weights and lesson plan planning, thereby improving the performance of online and offline hybrid teaching.

[0046] Preferably, after analyzing the type of target course, preliminary evaluation dimensions are set for the knowledge objectives and practical aspects in the syllabus respectively, and the preliminary weight range and threshold of each dimension are determined after the preliminary evaluation dimensions are correspondingly associated with the learning situation information.

[0047] In the early "planning" phase of the PDCA cycle, the present invention extracts knowledge objectives and practice aspects from the syllabus after analyzing the types of target courses, and regards them as two preliminary evaluation dimensions. In order to make the evaluation process more appropriate to the students' actual learning situation, it is necessary to introduce learning information (referring to factors such as class size, students' professional background, and course familiarity), and associate these factors with the evaluation dimensions, so as to clarify the preliminary weight intervals and thresholds of each dimension.

[0048] The main ideas of the above approach are:

[0049] The "knowledge goal dimension" and "practical aspect dimension" are established based on the course type and teaching syllabus. The former focuses on measuring the degree of theoretical mastery, while the latter focuses on measuring operational, experimental or practical application capabilities.

[0050] Collect learning information (such as students' historical grades in similar courses and cognitive ability performance) and map it to the above dimensions to adjust the initial weight and threshold range of each dimension.

[0051] When generating a preliminary rubric, the final score S can be preliminarily defined using the following formula:

[0052] S=w K ×K+w P ×P

[0053] Among them, K represents the evaluation score under the knowledge goal dimension, P represents the evaluation score under the practice link dimension, and w K With w PThe initial weights for the two major dimensions of knowledge and practice respectively (satisfying w K + w P = 1). Through the association of learning situation information, different initial thresholds and reference intervals can be set for K and P to cope with the diverse differences in students' theoretical levels and practical abilities.

[0054] In practical applications, the knowledge points listed in the teaching syllabus (such as theoretical concepts, formula derivations, analysis skills) are mapped to the knowledge objective dimension, and the practical links (such as experimental operations, case drills, project implementations) are mapped to the practical link dimension. In the learning situation information, if there are a large number of students with a high degree of investment in practical teaching, the upper limit of the interval of w P is initially increased (such as adjusted within [0.4, 0.6]), and the interval of w K is correspondingly narrowed; if most students have a weak theoretical foundation in previous courses, the threshold of the practical dimension can be lowered to prevent a large number of failures caused by overly high practical requirements.

[0055] Set the initial thresholds for each evaluation dimension based on the previously collected class or department data (such as the passing rate of students in practical course exams and the average score in theoretical tests). For example, in the knowledge dimension, it is determined that a student passes when the score ≥ 60, and the determination threshold for the practical dimension is set to ≥ 50 to determine passing. At this time, if the learning situation information indicates that nearly half of the students in this class have no relevant practical foundation, the threshold of the practical dimension is delimited within an acceptable interval (such as between 50 and 65). If it is observed that the practical ability has been significantly improved during the teaching process, the corresponding threshold can be further increased in the "Check" or "Act" link of the subsequent PDCA cycle.

[0056] In the "Do" stage of the present invention, the system collects data (such as knowledge tests, hands-on experiment scores) according to the initial evaluation dimensions K and P, and conducts causal analysis and quantum multi-group fuzzy weight optimization on the collected data in the "Check" stage; at this time, these initially set weight intervals and thresholds will directly affect the results of the subsequent algorithm for evaluating students. In the "Act" link, if it is found that the distribution of theoretical and practical performances is too different from the original judgment, the weight intervals and thresholds are dynamically adjusted to continuously improve the performance of online and offline blended teaching.

[0057] By splitting the "knowledge goal" and "practical link" dimensions, teachers can clearly distinguish students' different weaknesses in theoretical understanding and practical operation; by combining learning information when setting the initial weight intervals and thresholds, they can make a reasonable match between course difficulty and student ability in the early stage of the PDCA cycle to avoid evaluation distortion caused by too high or too low requirements in the early stage; when the system enters subsequent links (causal diagram construction, quantum multi-group search, etc.), these preliminary dimensions and thresholds can be verified with the actual collected data, and further refined or adjusted in the next round of cycles to achieve continuous fit between the evaluation dimensions and the students' actual learning status.

[0058] In one embodiment, in a teaching class, the syllabus explicitly requires students to master theoretical analysis (knowledge goal dimension) and complete certain types of experimental operations (practical link dimension). The learning information shows that more than 70% of the students in the class have insufficient previous experimental experience. At this time, the teacher sets the knowledge dimension interval to [0.5, 0.7] and the practice dimension interval to [0.3, 0.5]. If a student scores K = 75 on the knowledge test and P = 65 on the practical test in the mid-term test, and takes the weight w K =0.6, w P =0.4, then the comprehensive preliminary evaluation S can be calculated as follows:

[0059] S = 0.6 × 75 + 0.4 × 65 = 71

[0060] Later, in the "check" stage of the PDCA cycle, the system will conduct causal mining and weight optimization on the overall performance of the class. If it is found that the practical test performance of most students is extremely deviated, the practical dimension threshold can be further lowered in the next round of cycles, and special intensive training can be carried out on related teaching links, so as to gradually optimize the overall performance of online and offline blended teaching.

[0061] Preferably, when formulating a multi-source data collection plan, a feasibility assessment is conducted on the collection equipment and privacy policy, and the acquisition frequency and time alignment rules of camera images, heart rate sensors, and discussion text information are determined for online and offline teaching environments respectively.

[0062] In the dynamic evaluation of online and offline hybrid teaching performance based on the PDCA cycle, formulating a multi-source data collection plan is a key operation in the "Plan" stage. Its purpose is to:

[0063] Determine from the hardware and system levels whether feasible collection equipment (such as cameras, heart rate sensors, etc.) can be used to implement the present invention; evaluate the compliance of the required equipment and collection links in view of personal privacy and data security requirements; configure data acquisition frequency and time alignment strategies for different teaching scenarios (offline classrooms and online platforms) respectively, so as to efficiently obtain fusible data streams in the subsequent "Execution (Do)" stage.

[0064] The present invention regards camera image data, heart rate sensor data, and online discussion text information as the most representative sources for recording teaching behaviors and states. By clarifying the acquisition frequency and time alignment rules, multi-modal information can be accurately mapped to the same time axis in subsequent steps (such as causal graph construction and quantum multi-group fuzzy weight optimization), so as to correlate features such as students' expressions, heart rate changes, and text speeches at a certain moment, providing sufficient multi-dimensional evidence for comprehensive evaluation.

[0065] In practical applications, it is necessary to basically detect the positions of the deployed cameras (front and side of the classroom) and their resolutions to ensure that the acquired image resolution ≥ 720p for extracting expression or action features; test the heart rate sensor to confirm that its wearing method and signal accuracy meet the requirements for continuous acquisition within the class time range (for example, 90 minutes), and protect students' privacy during use: only record in the form of encrypted IDs without storing personal identity information; for the online discussion text recording part, it is necessary to ensure that the teaching platform allows real-time speech content to be captured, and add basic information such as speech timestamps and user IDs to the log to ensure that specific teaching activities can be associated during subsequent processing.

[0066] Camera image data: In the present invention, the interval time Δt camera (unit: second) is usually set, for example, a key frame is captured every 10 seconds for expression recognition. If higher requirements are needed, continuous video recording can also be performed first and then segmented sampling;

[0067] Heart rate sensor data: Generally, a sampling frequency of 1 second or 2 seconds is used to obtain the heart rate value HR(t). If worried about excessive data volume, a certain degree of moving average processing can be performed on the device side.

[0068] Discussion text information: It mostly belongs to event-driven, that is, timestamps are automatically generated when students post or reply on the platform, and can be included in the background log as T discussion . If finer granularity is required, the entry time marks of each sentence of text can be added to the chat area or forum system.

[0069] In order to combine and analyze camera frames, heart rate sensor records, and discussion texts in subsequent steps, a unified time reference is needed. The start time of the class can be set as T = 0 and continuously recorded backward;

[0070] Attach the timestamp T camera to each frame or sampling moment of the camera data, heart attach the timestamp T discussion to each sampling of the heart rate sensor, and camera attach the timestamp T to the text events generated in the discussion area. When subsequent causal or correlation analysis of students' expressions and heart rates within the same time period is required, it can be judged by |T camera -Theart If |≤δ, the records within δ are regarded as information in the same time period.

[0071] Example calculation: If the timestamp of a frame of an image captured by a camera is T camera = 450s (450 seconds after the class starts), and the timestamp of a piece of discussion text is T discussion = 452s, then if δ is set to 2 seconds, the two can be considered to be in the same time period and can be merged under the same label event.

[0072] Through the feasibility assessment of the acquisition device, combinations of cameras and heart rate sensors with low image quality or difficult deployment can be excluded first to avoid the risk of subsequent data gaps or illegal collection of personal information; ensure the privacy of students by means such as encrypted IDs and encrypted transmission, so that the information collected in the "Do" stage can legally and completely enter the analysis in the "Check" stage. When using camera images and heart rate sensor data, a unified time alignment strategy enables the system to recognize the mental and behavioral states of a certain student at the same moment. For example, when T camera = 600s, it is detected that the student has a confused expression, and when T heart = 601s, a heart rate peak appears, which can form a "confused - nervous" multimodal association, providing strong data support for subsequent causal analysis. By using the event timestamp method for discussion text information, the specific content of the online speech can be associated with the heart rate or offline expression simultaneously to judge the changes in the dimensions of "emotional investment" or "communication quality".

[0073] In one embodiment, during a class, the teacher decides to use a camera with a frame rate of 1 frame per 10 seconds to collect expression data and let the heart rate sensor collect the students' heart rates at a sampling frequency of 2 seconds per time. The system marks each captured image as T camera ∈{0, 10, 20,...}s, and the heart rate data is T heart ∈{0, 2, 4,...}s. If a student participates in online discussion and speaks 4 times from the 300th second (5 minutes) to the 360th second (6 minutes) of the class, then the corresponding log records T discussion are distributed between 300 and 360. At this time, subsequent analysis will integrate the data in the interval under the label of "active discussion period" and perform causal or correlation calculations between expression - heart rate - speech quality to provide a quantitative basis for "Check - Measure" in the PDCA cycle.

[0074] Through such acquisition frequencies and time alignment rules, the behavioral characteristics and emotional responses of students at different time periods can be accurately located in the subsequent analysis link, which not only ensures the operability of the acquisition process but also enables the effective integration of online and offline teaching behavior data, improving the fineness of the entire dynamic evaluation process of blended teaching performance.

[0075] Preferably, when obtaining multi-modal online and offline information based on the data acquisition scheme, coordinate correction is performed on the body key point trajectories generated by the motion capture device, and the trajectories after coordinate correction are synchronized with the heart rate sensor information according to the time stamps.

[0076] In the process of applying the PDCA cycle to the dynamic evaluation of the performance of online and offline hybrid teaching, the body key point trajectories generated by the motion capture device are of great significance for subsequent causal analysis and practical dimension evaluation. In the stage of obtaining multi-modal online and offline information based on the data acquisition scheme of the present invention, coordinate correction is performed on the original body key point trajectories output by the motion capture device, and the corrected trajectories are synchronized with the heart rate sensor information through time stamps to form multi-dimensional data that can be fused. This enables the teaching system to match the student's body movement trajectories with physiological data in the "check" and "measure" links, so as to discover potential practical ability characteristics and signs of emotional investment.

[0077] In practical applications, the initial coordinates output by the motion capture device are often offset or distorted due to factors such as the installation position and lens distortion. The present invention combines the method of calibration plane or marker points to correct each body key point trajectory, including:

[0078] Collect the ideal coordinates of each key point in several reference postures;

[0079] Let the key point coordinates output by the original device be denoted as p raw , and the reference posture comparison coordinates be denoted as p ref ;

[0080] According to the linear or non-linear transformation model f(·), map p raw to the corrected coordinates p corr = f(p raw ); the parameters of f here are obtained by fitting the reference postures;

[0081] This correction process can reduce the error caused by uneven installation of the device, and make the trajectory more accurately reflect the actual movement amplitude and posture of the student.

[0082] In the previously formulated multi-source data acquisition scheme, the heart rate sensor also periodically outputs physiological information, recording the heart rate value at time h(t).

[0083] To synchronize the motion trajectory with the heart rate data, time stamps T corr and T heart need to be recorded in each motion capture output frame and heart rate output. If the difference between the two satisfies |T corr - T heartIf |≤Δ, the actions and heart rate information can be regarded as those in the same time period; by indexing the corrected trajectory points p corr (t) and heart rate h(t) within the same time period, it can be observed whether significant physiological fluctuations occur in the corresponding action nodes of students during subsequent causal analysis or difference degree enhancement.

[0084] When the body key point trajectory is successfully synchronized with the heart rate sensor information, the system can judge the participation degree or tension degree of students in the practical dimension according to the changes of the trajectory amplitude and heart rate at the same moment; in the "Check" stage of PDCA, the construction of the causal diagram may discover the potential relationship of "excessive action amplitude - abnormal increase in heart rate - decrease in task completion degree", providing a basis for the improvement design in the next "Action" stage; in the subsequent "Action" stage, such as difference degree enhancement training or AR feedback, it is necessary to use the corrected trajectory and physiological data to determine whether a certain operation drill triggers the unstable emotions of students, and then incorporate this situation into the improvement strategy to improve the teaching performance.

[0085] In one embodiment, in a certain hybrid experimental course, to monitor the action stability of students during the equipment assembly link, the present invention uses a motion capture device and a set of heart rate sensors.

[0086] First, collect the reference posture to obtain the reference coordinates p ref ; Subsequently, compare the device output p raw with p ref and fit the transformation model f(·) to obtain the corrected p corr .

[0087] By uniformly recording the start time of the class as zero, the upper limit of the time stamp difference between the action frames output per second and the heart rate sampling is set to Δ = 1s. If an action mutation p corr (50s) occurs at t = 50s and the heart rate h(50s) also rises significantly, then the action trajectory and tension emotion in this time period can be associated in subsequent analysis to infer whether the operation difficulty is too high.

[0088] This embodiment shows that by correcting the motion capture coordinates and synchronizing with the heart rate sensor information, the teaching system can comprehensively identify the key problems in the learning process of students through body movements and physiological states, and further conduct targeted intervention in the next round of improvement of the PDCA cycle.

[0089] Preferably, during preprocessing, the key indicators in the viewing video log and discussion text are mapped to the preliminary evaluation dimension through noise filtering and label annotation to generate labeled multi-modal data and basic statistical data.

[0090] In the process of applying the PDCA cycle to the dynamic evaluation of the performance of online-offline hybrid teaching, the preprocessing stage is used to map the key information in the collected video viewing logs and discussion texts to the previously established preliminary evaluation dimensions, so that subsequent analysis can accurately evaluate students' performance in aspects such as knowledge, practice, or collaboration. Through noise filtering and label annotation, this process establishes connections between the text and viewing behavior indicators and the previously defined dimensions, generating labeled multimodal data and basic statistical data, providing detailed basis for the next round of "Check" or "Act".

[0091] In practical applications, there are situations such as jumping playback and long-term stagnation in the video viewing logs, and these invalid or incomplete records need to be removed during preprocessing. If the continuous interruption duration exceeds a certain threshold δ T , then this segment is regarded as noise to avoid forming biases in the statistical indicators. The discussion texts may contain content unrelated to the course. Keyword screening or short text clustering can be used to filter out the speeches containing irrelevant topics and retain the speeches that are truly related to the course theme or a certain knowledge point. The "noise filtering" here aims to ensure that the data input for analysis has a strong correlation with the teaching objectives and does not affect subsequent evaluations due to messy information.

[0092] The filtered video viewing logs and discussion texts are labeled according to the preliminary evaluation dimensions, such as "related to knowledge understanding" or "related to practical skills". For the video viewing logs, the number of times students repeatedly watch or pause the video segments in the knowledge section can be counted and labeled as the "degree of participation in the knowledge objective dimension"; for the discussion texts, topic recognition is carried out through a pre-set topic dictionary or text classification model (such as theoretical concept elaboration, experimental operation experience), and corresponding labels corresponding to the preliminary evaluation dimensions are added on this basis. Each label, in addition to the belonging dimension, also comes with a timestamp and the student ID, which are used for subsequent association with other modal data (such as heart rate, motion capture, etc.).

[0093] Labeled multimodal data: After the text and viewing behavior information are labeled and stored together with other previously obtained modal features (such as camera images or heart rate), the corresponding dimension labels can be directly called for clustering or causal analysis in subsequent analysis.

[0094] Basic statistical data: Refers to some key indicators obtained by quantitatively integrating the video viewing logs and discussion texts, such as the average viewing duration of each student for the theoretical video segments, the playback frequency of the experimental teaching video segments, the number of speeches on the experimental topics in the discussion area, etc.

[0095] Subsequently, in the "Check" stage of the PDCA cycle, these data will be used in conjunction with algorithms to evaluate the students' participation in the knowledge dimension and the practice dimension, and targeted interventions will be made in the "Act" stage.

[0096] After excluding invalid records through noise filtering, the formed multimodal data is more accurate and authentic; based on label annotation, mapping the watched videos and discussion texts to the preliminary evaluation dimensions respectively facilitates quick retrieval and summarization in subsequent steps. For example, when analyzing "knowledge mastery", all the watched and spoken data under the labels of this dimension can be immediately extracted; the basic statistical data provides a concise overview for the "Check" step of PDCA, thereby quickly locating the weak areas of students in terms of knowledge or practice directions and shortening the positioning time of teaching intervention.

[0097] In one embodiment, in a certain course, the teacher defines the knowledge objective dimension as "mastery of theoretical concepts" and the practice session dimension as "experimental operation skills". Combining video chapter division and text keyword filtering, the total staying time of students watching theoretical concept videos and the number of key speeches around experimental operation skills in the discussion area are counted. If a certain student repeatedly views a theoretical video paragraph many times and at the same time has very few speeches related to experimental operation skills in the discussion text, then in the labeled multimodal data, it can be seen that the corresponding "knowledge understanding" label appears frequently and the "practical skills" label is significantly low. This pattern may indicate that the student pays more attention to theory but lacks in operation. Subsequently, when analyzing the causal chain or carrying out differential reinforcement training, the performance of its practical dimension can be improved targeted, realizing more effective teaching intervention in the entire PDCA cycle.

[0098] As Figure 2 shown, using causal graph construction and quantum multi-group fuzzy weight optimization to comprehensively evaluate the labeled multimodal data and basic statistical data, obtaining adaptive Rubric configuration data and stage student scoring data, and generating readable summary information and early warning report data through neuro-fuzzy summarization;

[0099] This process belongs to the "Check" link in the PDCA cycle. By combining causal graph construction and quantum multi-group fuzzy weight optimization with labeled multimodal data and basic statistical data for comprehensive evaluation, adaptive Rubric configuration data and stage student scoring data are obtained. Subsequently, using neuro-fuzzy summarization to generate readable summary information and early warning report data provides a basis for implementing differential reinforcement training and other teaching interventions in the subsequent "Act" link.

[0100] The labeled multimodal data contains information in different dimensions such as learning behaviors, emotional expressions, and evaluation results. Through statistical independence tests or related algorithms, the node structures of this information are structured into "behavior nodes", "physiological nodes", "test result nodes", etc., respectively representing the different performances of students in the actual teaching process. When constructing a causal graph, directed edges are configured between the nodes, and the causal strength (such as the confidence based on conditional independence tests) is calculated to form a traceable data structure to judge the influence relationship of a certain behavior or physiological state on the teaching result.

[0101] Previously, several evaluation dimensions (such as knowledge understanding, experimental practice, interaction and collaboration) have been preset in the Rubric and preliminary weights have been assigned. Here, the quantum multi-population search method is used to iterate repeatedly in the weight search space, and the strength of the key causal edges in the causal graph and the performance gap of the students as a whole are used as the fitness measurement criteria to find the weight combination that makes the Rubric more accurately reflect the true performance of the students. In this process, the particle swarm is represented by a quantum state and performs a jumping search to avoid local extreme stagnation. Each particle encodes the Rubric weight vector w, and the fitness F(w) is calculated jointly with the causal graph structure to screen the optimal one. Finally, the adaptive Rubric configuration data is output, as well as the stage student scoring data calculated for each student's comprehensive performance according to this configuration.

[0102] After obtaining the new Rubric weights and stage student scoring data, readable text and warning reports are generated through neuro-fuzzy summarization; the neuro-fuzzy summarization is based on the text generation model of the neural network and supplemented by fuzzy logic rules to verbalize the important causal chains or abnormal patterns in the causal graph, generating a comprehensive description of the teaching quality of students or classes; when outputting, the high-strength causal relationships and the corresponding Rubric weight tendencies will be marked out so that teachers or managers can quickly understand which behaviors lead to low-scoring groups or performance differentiation and provide guidance for subsequent interventions.

[0103] Preferably, when constructing the causal graph, the causal strength between the behavior nodes, emotional nodes, and teaching result nodes is extracted from the labeled multimodal data based on statistical independence tests, and a confidence threshold is assigned to each causal edge.

[0104] In the dynamic evaluation of the online and offline hybrid teaching performance based on the PDCA cycle, the construction of the causal graph is an important part of the "Check" stage. The present invention regards the previously obtained labeled multimodal data (such as behavior data, emotional data, teaching result data) as the basic source of the causal graph nodes, uses the statistical independence test method to identify the causal associations between the nodes, and assigns a confidence threshold to each causal edge as an important reference for subsequent weight tuning and teaching intervention.

[0105] Node classification includes:

[0106] Behavior nodes: mainly include students' learning behaviors (such as the number of times of watching videos, homework submission habits, etc.) or practical operation characteristics recorded by motion capture;

[0107] Emotion nodes: mostly come from the emotional states reflected by heart rate sensors or facial expression recognition, and also include the recognition results of emotional tendencies in online text discussions;

[0108] Teaching result nodes: indicators that can measure students' learning achievements, such as test scores, stage ratings, or final grades.

[0109] Regard each feature in the labeled multi-modal data as a candidate variable X i , for any two variables X i , X j Conduct a conditional independence test. If it is determined that there is a statistical dependence relationship between X i and X j under certain conditions, a directed edge may be established in the causal graph; use a certain causal algorithm (such as the PC algorithm based on CITest) to hierarchically exclude irrelevant edges, and only retain the edges with significant statistical dependence and closely related to the teaching outcome. Eventually, in the causal graph, some connections from behavior nodes and emotion nodes to teaching result nodes will be marked as possible "behavior-result" and "emotion-result" causal edges.

[0110] For each causal edge, the system will set a causal strength score α according to the significance level of the statistical test, and on this basis, assign a confidence threshold θ to each edge to distinguish between "high-confidence causal chains" and "low-confidence causal chains"; if α exceeds the threshold θ, then regard this causal edge as a key association, which will be preferentially cited during subsequent quantum multi-group fuzzy weight tuning to help optimize the Rubric fitness.

[0111] In practical applications, after collecting labeled multi-modal data (including the behavior emotions, test scores, etc. of each student at multiple time periods), create corresponding nodes for each type of information, such as "experimental operation error rate", "peak heart rate", "knowledge test score"; use statistical independence tests to measure the relationships between different nodes. If certain dependence conditions are met, then construct directed edges. The direction determination can be combined with the chronological order or use a mature causal discovery algorithm in more complex scenarios. The generated causal graph usually contains several key paths, such as "motion capture error → heart rate fluctuation → final test score loss". The system will assign a causal strength α and set a confidence level θ ≤ 1 for each edge.

[0112] In subsequent Rubric weight optimization, edges with high causal strength and high confidence can indicate the factors that most affect teaching outcomes, thereby increasing or decreasing the weights of relevant dimensions in the Rubric accordingly; if the causal graph shows that certain features (such as reading duration) are not significantly associated with the final grade, the evaluation weight of this dimension can be correspondingly reduced in the next round of the loop. This structure based on the causal graph can enable teachers to accurately identify the root causes of problems in the "Act" phase. For example, if it is determined that there is a significant causal association between experimental operations and emotional fluctuations, the teacher can arrange differential reinforcement training to specifically relieve the psychological pressure of students during the experiment.

[0113] Through the causal graph, it is possible to effectively discover those "potential but crucial" influencing factor nodes, and better map the multi-modal data such as behavior and emotion in the teaching process to the final teaching effect; by assigning a confidence threshold to each causal edge, key association paths can be screened during subsequent algorithms or manual interpretation, reducing ineffective attempts on low-confidence relationships; the causal analysis with confidence support provides a quantitative and visual guiding basis for the next round of "Act" and "Recheck" in PDCA, improving the dynamic management and control of the performance of online and offline hybrid teaching.

[0114] In one embodiment, in the construction of the causal graph for a course, statistical independence tests are applied to find the correlation between "discussion text sentiment score → stage test score", with a causal strength α = 0.76 and a confidence threshold θ = 0.7. This means that if the discussion text of students tends to be positive, it has a relatively high positive impact on the test scores. The system will then give a greater evaluation weight to "online discussion activity" during subsequent Rubric optimization. At the same time, another node "abnormally elevated heart rate → experimental operation error rate" also exceeds the threshold θ = 0.8. Based on this, the teacher specifically designs relaxation training and hierarchical experimental strategies in the "Act" phase to reduce experimental mistakes caused by heart rate fluctuations. This embodiment demonstrates the key guiding role of the causal graph in improving teaching performance in the present invention.

[0115] Preferably, quantum multi-population fuzzy weight optimization screens out the optimized adaptive Rubric configuration data and phased student scoring data through multiple iterations of particles in the Rubric parameter search space and in combination with the causal strength evaluation index.

[0116] In the dynamic evaluation of the performance of online and offline hybrid teaching based on the PDCA cycle, quantum multi-population fuzzy weight optimization is a key step in searching for and optimizing Rubric parameters in the "Check" phase. Rubric parameters usually include the weights of each dimension and the fuzzy scoring threshold. Through the iterative method of quantum multi-population search, multiple evolutions are carried out in this parameter space, and in combination with evaluation indexes such as causal strength, the adaptive Rubric configuration data and phased student scoring data that can more accurately reflect the actual teaching effect are gradually screened out.

[0117] In the initial Rubric of the present invention, a number of evaluation dimensions are set (such as knowledge mastery, practical operation, emotional investment, etc.), and each dimension is assigned a weight w i and possible fuzzy thresholds, such as "qualified interval [a_i, b_i]". The search space is all possible w = (w1, w2,..., w n ) and threshold combinations; the goal is to enable the optimal parameter combination to maintain a high degree of matching with the labeled multi-modal data and causal strength information in subsequent evaluations, so as to improve the discrimination ability of the Rubric for students' true performance.

[0118] When the system is initialized, a number of populations are created, each population contains a number of particles, and the state vector of each particle corresponds to a set of Rubric parameters p j . If the number of a certain dimension is n, then p j ∈R n ; compared with the traditional particle swarm algorithm, the present invention introduces the idea of quantum state, enabling the particles to jump out of the local extreme value in an iterative manner in a "quantum jump" way; that is, in each round of search, the update of the probability distribution of the particle's position is no longer only linearly weighted, but will refer to the evolution of the quantum superposition state; through this mechanism, the risk of falling into the local optimum is reduced, ensuring to find a solution with more global optimality in the vast Rubric parameter space.

[0119] The previously constructed causal graph gives the causal strength and confidence threshold between nodes. For example, when the influence strength of a certain behavior node on the teaching result node is α, it may be more practical to increase the weight of the relevant dimension in the Rubric; in each search iteration, for the given Rubric parameter p j calculate a comprehensive fitness F(p j ), which may include the following factors: the difference degree from the labeled multi-modal data, if the Rubric score has a large deviation from the true observation, the fitness decreases; the causal strength matching degree, if the weight distribution maintains a good correspondence with the high-strength causal chain, the fitness increases accordingly; the quantum multi-population search adjusts the particle distribution according to F(p j ) in each iteration, gradually approaching the optimal solution.

[0120] After multiple iterations are completed, a set of optimal Rubric parameters w * is obtained at the global or better solution, and the system records it as the adaptive Rubric configuration data; and based on this, a comprehensive score is given to each student in each dimension, and the phased student score data is output, which is convenient for targeted intervention or difference strengthening in the next round of "Act" stage.

[0121] By virtue of the globality of quantum multi-population search, a better solution is found in the complex Rubric parameter space, improving the adaptability of Rubric to multi-modal teaching data. Compared with the approach of optimizing only based on single-dimensional indicators such as test scores, the causal strength is incorporated into the evaluation index, enabling the Rubric weights to closely fit the real teaching causal chain. The adaptive Rubric configuration data and the phased student scoring data can provide quantitative basis for subsequent teaching interventions (such as holographic augmented reality feedback and differential reinforcement training), and continuously enter the next "Act" and the next "Check" in the PDCA cycle for improvement iteration.

[0122] In one embodiment, in an online and offline hybrid course, the initial dimensions of Rubric are knowledge understanding w K , practical operation w P , and emotional interaction w C , with the constraint w K +w P +w C =1. When the quantum multi-population algorithm is initialized, multiple groups of candidate vectors p j =(w K , w P , w C ) are generated, and the fitness F(p j ) is calculated according to the causal graph strength information and the matching degree of labeled multi-modal data after each iteration.

[0123] For example, if the causal graph shows that the causal strength between experimental operation mistakes and test score losses is as high as α = 0.8, then when w P in the particle tends to increase and the scoring accuracy is improved simultaneously, the particle is regarded as the evolutionary direction. After dozens of generations of iteration and quantum leap, it finally converges to a group of w * approximately (0.3, 0.5, 0.2). This solution has the highest coupling degree with the causal graph nodes and the smallest mean square error, and outputs a new adaptive Rubric configuration and phased student scoring. Thus, teachers can pay more attention to practical operation problems in subsequent links, assisting students to improve their practical abilities and reduce the error rate.

[0124] Preferably, the neuro-fuzzy summary uses the phased student scoring data in the adaptive Rubric configuration data and the causal relationship constructed by the causal graph to generate readable summary information and warning report data and mark the key behavior defects or skill deficiency links.

[0125] In the dynamic evaluation of the performance of online-offline hybrid teaching based on the PDCA cycle, the "neural fuzzy summary" process aims to integrate the phased student scoring data in the adaptive Rubric configuration data with the causal relationship information represented by the causal diagram, and generate summary and warning report data in a readable language form. Briefly, the neural network part is used for natural language generation, and the fuzzy logic part can convert information such as uncertain or continuous scores and causal strengths into linguistic descriptions such as "possibility" or "higher / lower", so as to provide intuitive information for teachers and teaching decision-makers at the end of the "Check" stage or the beginning of the "Act" stage.

[0126] After the quantum multi-population fuzzy weight tuning is completed, the present invention obtains the adaptive Rubric configuration data and the phased scoring results of each student; these scores reflect the comprehensive performance of students in multiple aspects such as the knowledge dimension, practice dimension, and emotion dimension; the neural fuzzy summary needs to extract key indicators (such as score segments below the threshold, excellent performances above the average, etc.) from these scores, and conduct a comparative analysis with the high-confidence associations in the causal diagram.

[0127] There are several connections between the behavior nodes, emotion nodes and teaching result nodes in the causal diagram, and each connection has a confidence threshold and a causal strength; when a certain connection has a high verification degree or a score anomaly under the adaptive Rubric evaluation, the neural fuzzy summary will call the corresponding causal information, form a brief inference statement and mark the "key behavior defect" or "skill deficiency" link.

[0128] In principle, it combines the neural network text generation framework (such as the encoder-decoder model based on the attention mechanism) and fuzzy logic reasoning:

[0129] Analysis stage: Screen out outliers or nodes with particularly obvious causal strength relationships from the phased student scoring data;

[0130] Fuzzy reasoning: If the score is lower than a certain fuzzy threshold γ (γ is given by the adaptive Rubric or statistical analysis), then this dimension is marked as "possibly defective"; combining with the causal diagram, if the causal strength of "heart rate peak → experimental operation error rate" is high, then the association between heart rate and operation mistakes is further pointed out;

[0131] Language generation: Map the fuzzy logic judgment result into a natural language description, such as "Student A has a low score in the practice dimension, and there is a strong causal relationship with high heart rate fluctuations. The operation environment and stress reduction measures need to be focused on."

[0132] For the key behavior defect or skill deficiency links, they are highlighted in the output report text through the annotation module, and brief suggestions are attached, such as "experimental operation guidance" or "emotional communication counseling".

[0133] After using the neuro-fuzzy summary, teachers do not need to directly read large quantities of score data or complex causal diagrams. The system will automatically synthesize them into highly readable text reports and provide intuitive explanations for the most influential causal chains. For students with persistently low scores or multiple negative emotional nodes, the early warning report data can quickly point out the problem links (such as insufficient experimental operation skills, blind spots in knowledge understanding, etc.), helping teachers to conduct differentiated intensive training or flipped classrooms in the "Act" stage. Students can also understand the reasons for their own learning weaknesses through brief summaries. If they see "heart rate peak is highly correlated with experimental failure", they will realize the importance of emotional regulation and improve themselves in subsequent learning.

[0134] In one embodiment, suppose the system performs rubric evaluation on 30 students in this cycle, and the average "practical ability" score is 65 points, of which 8 students have a practical score <50, and have a high confidence θ=0.75 with the edge of the causal graph "laboratory operation error → final exam score loss". The neuro-fuzzy summary model generates a report through the following steps:

[0135] Selection of abnormalities: 8 students were detected with practical scores <50, and 5 of them had a correlation between emotional nodes (such as heart rate data) and practical operation errors >0.7;

[0136] Fuzzy reasoning: It is believed that the causal relationship between "tension" and "experimental failure rate" of these five students is strong, and they need to be marked as "objects that need special attention";

[0137] Text output: Output a brief description, such as "The system detects that some students have low performance in the practical dimension, which is highly correlated with the high heart rate state. It is recommended to introduce adjustment techniques or improve the guidance method in the experimental class." And list the five students in the key behavioral defect link table of the "early warning report data".

[0138] Therefore, teachers can provide personalized interventions (such as differentiated reinforcement training, psychological counseling, etc.) to these five students in the "measures" stage of the next PDCA cycle without having to personally review all the original test scores or heart rate records, thereby achieving accurate and efficient teaching interventions.

[0139] like Figure 3 As shown in the figure, based on the periodic student scoring data and early warning report data, differential reinforcement training and holographic augmented reality feedback are implemented, new teaching intervention process data are collected, and the adaptive rubric configuration data is adjusted to obtain iterative rubric configuration data, so as to realize dynamic improvement of online and offline hybrid teaching performance.

[0140] In the dynamic evaluation of the performance of online-offline hybrid teaching based on the PDCA cycle, this process belongs to the "Act" stage. Based on the phased student scoring data and early warning report data generated in the previous stage ("Check"), differential reinforcement training and holographic augmented reality feedback are implemented for students identified as key or having obvious problems. By collecting new teaching intervention process data, the system can further correct the adaptive Rubric configuration data, form the iterated Rubric configuration data, and complete a closed-loop from "problem discovery" to "precision intervention" and then to "evaluation adjustment".

[0141] The core idea is to set personalized training difficulties or specialized practice modules for students with different performance types or problem points based on multi-modal data analysis. In practical courses, if a student has a low experimental operation accuracy, the difference Δd can be calculated by comparing their motion capture trajectory with the standard operation. If Δd exceeds a certain threshold γ, reinforcement training materials or additional practice sessions will be automatically assigned;

[0142] For example, in the scenario of a physics experiment class, the deviation Δd of a student's multiple measurements is averaged. Once the result is higher than the threshold γ = 5mm (example), the student will be prompted that they need specialized instrument operation guidance, and the progress of this intervention will be recorded in the new teaching intervention process data.

[0143] In a teaching environment with conditions, teachers or the system can use holographic projection / AR glasses to show students their deficiencies or success points in a certain dimension in real time. When the system determines that there is a significant relationship between a student's emotional fluctuations and the failure rate in the hands-on practice session, the heart rate curve and the moment of operation errors will be prominently displayed on the AR interface, allowing the student to intuitively feel the connection between their peak tension and operation difficulties; at the same time, instant instructions or tips can be provided, such as "relax the grip strength" and "check the operation instructions", so as to help the student correct in time. The generated AR operation records will also be recorded by the system as intervention process data, including the operation improvement rate of the student in the AR environment, the number of error tolerances at each time period, etc.

[0144] In the above-mentioned difference-enhanced training or AR feedback, the system will continue to perform multimodal acquisition to generate new teaching intervention process data. For example, it will measure the student operation success rate, heart rate stability, feedback frequency in the discussion area again, etc.; incorporate these intervention process data into the original multimodal dataset, and make a re-comparison in combination with the phased student scoring data. If the student's score in a specific dimension significantly improves or remains low after the intervention, the system will correspondingly adjust that dimension in the adaptive Rubric: increase or decrease the weight; correct the passing or excellent threshold; add more accurate scoring factors to specific behavior or emotion nodes; through such fine-tuning, finally form the iterated Rubric configuration data and record it in the solution database for continued invocation when entering the "Check" link of the next PDCA cycle.

[0145] Preferably, during the implementation of difference-enhanced training and holographic augmented reality feedback, time alignment is performed on the new teaching intervention process data and compared with the adaptive Rubric configuration data, and the parameters of each dimension are adjusted according to the comparison results to generate the iterated Rubric configuration data.

[0146] In the online and offline hybrid teaching performance dynamic evaluation method based on the PDCA cycle, the new teaching intervention process data collected after the implementation of difference-enhanced training and holographic augmented reality feedback needs to be time-aligned again and then compared with the previously obtained adaptive Rubric configuration data. Through this comparison process, the system can determine whether the intervention is effective, which dimensions have significant improvements or still have deviations, so as to dynamically adjust the parameters of each dimension of the Rubric and form the iterated Rubric configuration data. This not only ensures the accurate evaluation of the student's current learning status, but also enables the Rubric itself to be continuously optimized with the teaching intervention results, realizing the continuous improvement of the online and offline hybrid teaching performance.

[0147] In the difference-enhanced training session, if the system sets up a personalized training module for the student's experimental operation errors or emotional fluctuations, data such as operation accuracy, heart rate changes, and AR interaction times will be collected during this process; the holographic augmented reality feedback records various action responses of the student in the AR scenario (such as the number of error corrections, the improvement range of operations), emotional indicators, etc.; these raw data all have timestamps, and after being aligned through the time benchmark established in the previous step, available multimodal intervention information is generated. For example, if a student participates in AR error correction at t = 200s and their heart rate drops to the normal level at t = 202s, the system will summarize these two records into the same intervention event window for the next comparative analysis.

[0148] The adaptive Rubric configuration data includes the weight and threshold settings for several dimensions (knowledge, practice, emotion, collaboration, etc.); if the new teaching intervention process data shows that the error rate of students in the "practical operation" dimension has decreased significantly, it can be determined that the intervention corresponding to this dimension is relatively successful, and it is also possible to evaluate whether the weight of this dimension is too low or too high; the comparison method often uses the error metric between the "pre-intervention scoring result" and the "post-intervention observation": Rubric prediction before intervention: the score S of the student in the practical operation dimension P ; Post-intervention observation: the actual improvement amplitude G. If G is greater than the Rubric expected value, it means that Rubric has an underestimation tendency in this dimension, and its weight can be slightly increased; otherwise, the threshold of this dimension can be decreased or tightened moderately; the system will accumulate such comparison results for each dimension and adjust the corresponding parameters in Rubric according to the overall performance distribution.

[0149] When the comparison conclusion indicates that the weights of some dimensions need to be fine-tuned, or a certain threshold needs to be increased or decreased, the system performs a local search in the Rubric parameter search space (the quantum multi-population search can be called again or a linear update can be directly performed) until the fitness of each dimension tends to be stable for the intervention process data; the updated Rubric is recorded as the Rubric configuration data after iteration and publicized to teachers and students, and enters a new PDCA cycle or is applied in subsequent teaching.

[0150] By aligning the new teaching intervention process data in time and comparing it with the adaptive Rubric configuration data, the system can understand the immediate effectiveness or deficiencies of the intervention; if the intervention effectiveness is better than the original Rubric expectation, Rubric will pay more attention to this successful path in the next stage; if the intervention effectiveness is worse than expected, Rubric will automatically relax or tighten the threshold of this dimension to avoid continuous errors; as students continuously participate in the differential reinforcement training and AR error correction, Rubric continuously undergoes the adaptive iteration of "comparison - adjustment - re-comparison", so that the online and offline blended teaching performance shows a spiral improvement in the entire PDCA cycle.

[0151] In one embodiment, after a "practical operation" reinforcement training, a certain student repeated the operation many times in the AR scenario, and the peak heart rate decreased from 140 bpm to 120 bpm and remained stable, and the experimental success rate increased from 50% to 70%. Map this process record to the time interval [t1, t2], and align the heart rate sensor and the AR operation log. Rubric originally predicted that the experimental success rate of this student would only increase to 60%, indicating that the "practical operation" dimension underestimated the intervention effect relatively. Therefore, the system decides to adjust the weight w P of the "practical operation" dimension from 0.35 to 0.40 and increase the threshold by Δ threshold= 5. Finally, new iterated Rubric configuration data is formed and fed back to the teaching management terminal, entering the "Execute (Do)" and "Check (Check)" phases of the next cycle to test the fitness of the new Rubric.

[0152] As Figure 4 shown, an online and offline hybrid teaching performance dynamic evaluation system based on the PDCA cycle is used to implement the online and offline hybrid teaching performance dynamic evaluation method based on the PDCA cycle. The system includes:

[0153] A course analysis and Rubric generation module, which is used to analyze the type, syllabus, and student information of the target course, clarify the evaluation dimensions and initial weight intervals and thresholds, generate initial Rubric configuration data, and formulate a multi-source data collection plan to obtain collection configuration data; mainly relying on the server or cloud environment to run algorithms and manage the front end, teachers can enter the course type, syllabus, and student information through the PC or web page to generate the initial Rubric and collection configuration plan.

[0154] A multi-modal collection and preprocessing module, which is used to obtain multi-modal online and offline information based on the data collection plan and perform time alignment, noise filtering, and label annotation on the information to generate labeled multi-modal data and basic statistical data; use hardware such as cameras, heart rate sensors, and motion capture devices in the classroom to collect offline data, and embed a log collection script on the online platform at the same time. Optionally configure an edge computing unit to perform time alignment, noise filtering, and label annotation on the original data, and finally transmit it to the server to generate labeled multi-modal data and statistical information.

[0155] A causal graph and quantum tuning module, which is used to comprehensively evaluate the labeled multi-modal data and basic statistical data by using causal graph construction and quantum multi-population fuzzy weight tuning to obtain adaptive Rubric configuration data and stage student scoring data, and generate readable summary information and early warning report data through neuro-fuzzy summarization; usually deployed on one or more high-performance servers, combined with a graph database and parallel computing to perform causal graph construction and quantum multi-population search, and store or read labeled data to output an adaptive Rubric and stage student scoring. If the data volume is large, distributed storage or cloud clusters can be used.

[0156] The differential reinforcement and feedback module is used to implement differential reinforcement training and holographic augmented reality feedback based on phased student scoring data and early warning report data, collect new teaching intervention process data, and then adjust the adaptive Rubric configuration data to obtain the iterated Rubric configuration data and achieve dynamic improvement of the online-offline hybrid teaching performance. Differential reinforcement requires corresponding training equipment or AR equipment (such as AR glasses, holographic projection, etc.) to be used in the classroom in cooperation, and the intervention process data is immediately transmitted back to the server. Communicate with the aforementioned optimization module through a network interface or dedicated line, adjust the Rubric parameters according to the newly collected information, and generate the iterated Rubric configuration data.

[0157] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.

[0158] The above are only the embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A dynamic evaluation method for the performance of online and offline hybrid teaching based on the PDCA cycle, characterized in that, The following steps are involved: Analyze the type, syllabus and learning information of the target course, clarify the evaluation dimensions and preliminary weight ranges and thresholds, generate initial Rubric configuration data, and formulate a multi-source data collection plan to obtain collection configuration data; based on the data collection plan, obtain multimodal online and offline information and perform preprocessing to generate labeled multimodal data and basic statistical data; The causal graph construction and quantum multi-group fuzzy weight tuning are used to comprehensively evaluate the labeled multimodal data and basic statistical data, and adaptive rubric configuration data and periodic student scoring data are obtained. Readable summary information and early warning report data are generated through neuro-fuzzy summarization. Based on the periodic student scoring data and early warning report data, differential reinforcement training and holographic augmented reality feedback are implemented, new teaching intervention process data are collected, and the adaptive rubric configuration data is adjusted to obtain iterative rubric configuration data, so as to achieve dynamic improvement of online and offline hybrid teaching performance.

2. The method according to claim 1, characterized in that After analyzing the type of target courses, preliminary evaluation dimensions are set for the knowledge objectives and practical aspects in the syllabus, and the preliminary weight range and threshold of each dimension are determined after the preliminary evaluation dimensions are associated with the learning information.

3. The method according to claim 1, wherein When formulating a multi-source data collection plan, a feasibility assessment is conducted on the collection equipment and privacy policy, and the acquisition frequency and time alignment rules for camera images, heart rate sensors, and discussion text information are determined for online and offline teaching environments respectively.

4. The method according to claim 1, wherein When acquiring multimodal online and offline information based on the data acquisition scheme, coordinate correction is performed on the trajectory of key body points generated by the motion capture device, and the trajectory after coordinate correction is synchronized with the heart rate sensor information according to the timestamp.

5. The method according to claim 2, wherein When performing preprocessing, key indicators in the video viewing log and discussion text are mapped to the preliminary evaluation dimensions through noise filtering and label annotation to generate labeled multimodal data and basic statistical data.

6. The method according to claim 1, characterized in that, When constructing a causal graph, the causal strength between behavioral nodes, emotional nodes, and teaching outcome nodes is extracted from labeled multimodal data based on statistical independence tests, and a confidence threshold is assigned to each causal edge.

7. The method according to claim 1, wherein Quantum multi-group fuzzy weight tuning selects the optimized adaptive Rubric configuration data and periodic student scoring data through multiple iterations of particles in the Rubric parameter search space and combined with the causal strength evaluation index.

8. The method according to claim 1, wherein The neuro-fuzzy summary utilizes the periodic student scoring data in the adaptive rubric configuration data and the causal relationship obtained by constructing the causal graph to generate readable summary information and early warning report data and mark key behavioral defects or skill deficiencies.

9. The method according to claim 1, wherein In the process of implementing differential reinforcement training and holographic augmented reality feedback, the new teaching intervention process data is time aligned and compared with the adaptive rubric configuration data. According to the comparison results, the parameters of each dimension are adjusted to generate the iterative rubric configuration data.

10. A dynamic evaluation system for the performance of online-offline hybrid teaching based on the PDCA cycle, which is used to implement the dynamic evaluation method for the performance of online-offline hybrid teaching based on the PDCA cycle according to any one of claims 1 to 9, characterized in that, The system includes: The course analysis and rubric generation module is used to analyze the type, syllabus and learning information of the target course, clarify the evaluation dimensions and preliminary weight range and threshold, generate initial rubric configuration data, and formulate a multi-source data collection plan to obtain collection configuration data; A multimodal acquisition and preprocessing module, which is used to acquire multimodal online and offline information based on the data acquisition scheme and perform time alignment, noise filtering and labeling on the information to generate labeled multimodal data and basic statistical data; The causal graph and quantum tuning module is used to comprehensively evaluate the labeled multimodal data and basic statistical data using causal graph construction and quantum multi-group fuzzy weight tuning to obtain adaptive rubric configuration data and staged student scoring data, and generate readable summary information and early warning report data through neuro-fuzzy summarization; The differential reinforcement and feedback module is used to implement differential reinforcement training and holographic augmented reality feedback based on periodic student scoring data and early warning report data, and to adjust the adaptive rubric configuration data after collecting new teaching intervention process data to obtain iterative rubric configuration data and achieve dynamic improvement of online and offline hybrid teaching performance.

Images