Screen-out teaching feedback optimization method and related equipment

Abstract

The invention discloses an out-of-screen teaching feedback optimization method and related equipment, and the method comprises the steps: synchronously collecting a multi-mode signal through a multi-source heterogeneous sensor, carrying out the space-time fusion feature extraction, pre-training a utility evaluation model, outputting a quantized value, and generating an optimization instruction through a DRL strategy network in combination with a content attribute parameter, and finally, dynamically adjusting the visual attribute of the out-of-screen content and performing closed-loop optimization. Firstly, a real-time closed loop is constructed, the feedback lag problem is solved, and it is ensured that an optimization instruction acts on current teaching instantly; secondly, features are covered and fused with time and space, the defects of single data acquisition and one-sided feature extraction are overcome, and a basis is provided for accurate evaluation; and thirdly, the pre-training model quantitatively outputs key indexes, and the DRL strategy network is dynamically optimized by taking a long-term teaching effect as a target, so that the evaluation accuracy and the optimization scientificity are improved. Finally, the content is enabled to adapt to the real-time state of the learner, the learning interest is improved, the cognitive burden is reduced, the interaction participation degree is improved, the screen teaching advantage is played, and long-term teaching effect optimization is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of smart teaching technology, more particularly, to a screen-out teaching feedback optimization method and related equipment. BACKGROUND

[0002] The current education digital transformation process is accelerating. Screen-out teaching (such as AR / VR stereoscopic screen-out presentation, 3D interactive teaching, etc.) has gradually been applied to various teaching scenarios because it can convert abstract knowledge into an intuitive and stereoscopic presentation form, which has shown significant potential in improving teaching immersion and enhancing knowledge transmission efficiency. However, in actual application, the effectiveness of screen-out teaching is greatly affected by the adaptability of content and the real-time state of learners. The degree of concentration of attention, the depth of knowledge understanding, and the enthusiasm of interactive participation of learners in the teaching process dynamically change. If these changes cannot be captured and used to adjust the screen-out content in a timely manner, problems such as declining interest in learning, exceeding the scope of cognitive load, and poor knowledge absorption may occur. Therefore, how to establish a feedback mechanism that can respond to the state of learners in real time and accurately optimize screen-out content has become a key requirement for the further development of screen-out teaching technology.

[0003] The existing feedback technology for screen-out teaching still has many shortcomings and cannot meet the above-mentioned requirements. Firstly, the timeliness of feedback is insufficient. Current mainstream methods rely on post-class questionnaires, teachers' subjective judgments based on observation and recording, or indirect evaluation of teaching effectiveness through periodic tests. These methods cannot obtain real-time state data of learners during screen-out teaching, resulting in a lag in adjusting screen-out content relative to the teaching process and failing to timely adapt to the dynamic needs of learners. Secondly, the data collection dimension is limited. Although some technical solutions introduce sensors to assist in data collection, they are mostly limited to a single type of signal, making it difficult to comprehensively obtain complete data reflecting the degree of attraction, understanding, and interactive participation of learners to screen-out content. Thirdly, there are shortcomings in evaluation and optimization capabilities. When extracting features from collected data, most solutions only analyze a single dimension, and the models used to evaluate teaching effectiveness are mostly simple models based on pre-set rules, which cannot output quantifiable evaluation results that accurately reflect key teaching indicators. Moreover, the optimization strategies for screen-out content are mostly based on fixed rules, such as adjusting content complexity at fixed time intervals, without considering dynamic optimization targeting long-term teaching effectiveness, resulting in a lack of pertinence and scientificity in optimization measures and difficulty in continuously improving the overall effectiveness of screen-out teaching.

[0004] Therefore, there is an urgent need for a new screen-out teaching feedback optimization method to address the shortcomings of existing technology, fully leverage the technical advantages of screen-out teaching, and ensure the continuous optimization of long-term teaching effectiveness. SUMMARY

[0005] The application provides a screen-out teaching feedback optimization method and related equipment. The method synchronously collects multi-modal signals through multi-source heterogeneous sensors, extracts spatio-temporal fusion features, and outputs quantitative values through a pre-trained utility evaluation model. Combined with content attribute parameters, an optimization instruction is generated by a DRL policy network to dynamically adjust the visual attributes of the screen-out content to complete closed-loop optimization. This method can solve the problems of feedback lag, single data collection, one-sided feature extraction, evaluation deviation, and fixed optimization in the prior art. Ultimately, it can adapt the content to the real-time state of the learner, improve learning interest, reduce cognitive burden, and improve interactive participation, fully exert the advantages of screen-out teaching, and ensure long-term teaching effect optimization.

[0006] A screen-out teaching feedback optimization method, comprising:

[0007] In the screen-out teaching process, a multi-source heterogeneous sensor group integrated in the client synchronously collects multi-modal teaching signals from the same time base point;

[0008] Spatio-temporal fusion features representing the attraction, understanding, and interaction of the screen-out content are extracted from the multi-modal teaching signals, and the spatio-temporal fusion features are input into a pre-trained utility evaluation model to output a quantitative evaluation value of the current screen-out effect. The quantitative evaluation value includes a screen-out attraction index, a cognitive load index, and an interaction conversion rate.

[0009] The quantitative evaluation value and the attribute parameters of the current screen-out content are jointly used as a state vector, which is input into a DRL policy network. The DRL policy network is configured with a reward function aiming to maximize the long-term teaching effect. The state vector is processed, and real-time optimization instructions for the screen-out content attributes are output.

[0010] Based on the optimization instructions, the visual presentation attributes of the screen-out content being demonstrated are dynamically adjusted, and the rendering engine is driven to update and output to the display device in real time, completing closed-loop feedback optimization.

[0011] Optionally, the multi-modal teaching signals include visual attention signals, interaction behavior signals, and context teaching signals, wherein:

[0012] The visual attention signals are obtained through a real-time eye tracking module based on a convolutional neural network (CNN) and include a distribution heat map of fixation points in the screen-out area and the non-screen-out area, and a sequence of pupil diameter changes.

[0013] The interaction behavior signals include the frequency of touch operations on the screen-out content, the operation force, and the mimic gesture operation trajectory in the air for the screen-out content.

[0014] The context teaching signals include an identifier of the current teaching knowledge point and the attribute parameters of the screen-out content.

[0015] Optionally, the calculation process of the quantitative evaluation value comprises:

[0016] The off-screen attraction index is calculated according to the ratio of the residence time of the fixation point in the off-screen area to the total viewing time in the visual attention signal, and the expansion rate of the pupil diameter at the off-screen moment.

[0017] The cognitive load index is calculated according to the saccade speed of the fixation point jumping between different sub-components of the off-screen content in the visual attention signal, and the hesitation time of the operation for the corresponding knowledge point of the off-screen content in the interaction behavior signal.

[0018] The interaction conversion rate is calculated according to the ratio of the number of effective interaction operations after triggering the off-screen content to the total viewing number in the interaction behavior signal.

[0019] Optionally, the construction framework of the DRL policy network comprises:

[0020] The state vector constitutes the state space of the DRL policy network;

[0021] The action space of the DRL policy network is defined as a set of operation instructions for adjusting the attributes of the off-screen content, including off-screen depth fine-tuning instructions, model visual saliency increase / decrease instructions, and additional explanatory label display / hide instructions.

[0022] The reward function is composed of a short-term reward and a long-term reward, wherein the calculation basis of the short-term reward is the immediate positive change of the quantitative evaluation value after the action is executed, and the calculation basis of the long-term reward is the knowledge point examination pass rate associated with the off-screen content in the subsequent teaching session.

[0023] By selecting the action in the action space and combining the reward value fed back by the reward function, the optimization learning of the DRL policy network is implemented until the goal of maximizing the long-term teaching effect is achieved.

[0024] Optionally, the DRL policy network is trained using a proximal policy optimization algorithm, and the network weight is initialized offline according to historical teaching session data, and is continuously fine-tuned and updated in the online teaching process through the multi-modal teaching signal of the student, to adapt to the teaching style of different teachers and the cognitive habits of different audiences.

[0025] Optionally, the adjustment of the visual presentation attribute comprises:

[0026] When it is detected that the off-screen attraction index is lower than a first preset threshold, the highlight reflection attribute of the off-screen content material surface is enhanced while the highlight displacement amount of the off-screen content is increased.

[0027] when the cognitive load index is detected to be higher than a second preset threshold, reducing the rendering complexity of the out-screen model or reducing the animation speed of the out-screen model, and triggering generation of a guiding annotation arrow pointing to a core component of the out-screen model;

[0028] when the interaction conversion rate is detected to be lower than a third preset threshold, dynamically generating a virtual interaction control on the side of the out-screen content, the virtual interaction control having a semi-transparent attribute and carrying a feedback metaphor guiding the user to perform an interactive operation on the out-screen content.

[0029] Optionally, further comprising:

[0030] identifying a target audience identifier and a field knowledge theme of the current teaching, obtaining an audience cognitive baseline level and an interaction preference parameter from a pre-built audience portrait library based on the target audience identifier, and loading an associated knowledge point network from a pre-built field knowledge graph based on the field knowledge theme;

[0031] The quantitative evaluation value, the current out-screen content attribute parameter, the audience cognitive baseline level and interaction preference parameter, and the associated point and confusion point feature vector of the current knowledge point extracted from the associated knowledge point network are collectively used as a state vector and input to the DRL policy network.

[0032] Optionally, further comprising:

[0033] When the utility evaluation model detects that there is a target knowledge point whose cognitive load index is continuously higher than a diagnostic trigger threshold, a root cause diagnosis is triggered, a root cause diagnosis result is formed and input to the DRL policy network, wherein the root cause diagnosis queries a set of pre-existing basic knowledge points and a set of easily confused knowledge points that cause the target knowledge point to be difficult to understand by accessing the field knowledge graph, and forms the root cause diagnosis result.

[0034] Optionally, further comprising:

[0035] Receiving identification information of a target learning content version of an audience of the current teaching, and retrieving a knowledge mastery graph formed for a previous basic version of the target learning content from a learning history record of the audience;

[0036] By comparing the difference between the target learning content and the knowledge mastery graph formed by the previous basic version, an incremental knowledge difference set is generated, which contains the knowledge points added, deleted and significantly changed in the target learning content relative to the previous basic version;

[0037] Based on the incremental knowledge difference set, the reward function of the DRL policy network is adjusted, wherein:

[0038] For the first knowledge point included in the incremental knowledge difference set, a higher reward coefficient is assigned to the first knowledge point than other knowledge points to guide the optimization strategy to prioritize the off-screen teaching effect of the first knowledge point;

[0039] For the knowledge points that have not changed in the target learning content version and the previous base version and have been mastered by the user, a simplified off-screen mode is adopted for demonstration or the detailed off-screen demonstration process of the knowledge points is skipped.

[0040] Optionally, it further comprises:

[0041] Through a preset cross-system interface, a diagnostic event generated by at least one external teaching auxiliary system is listened to, the diagnostic event including an event type identifier and an associated knowledge point identifier, for indicating that the learning or operation effect on a specific knowledge point is not as expected;

[0042] In response to the listened diagnostic event, a target knowledge point identifier to be reinforced is parsed out;

[0043] According to the target knowledge point identifier, a reinforcement teaching instruction is generated and input to the DRL policy network, so that the visual presentation attributes of the off-screen content associated with the target knowledge point are optimized through the real-time optimization instruction output by the DRL policy network.

[0044] An off-screen teaching feedback optimization device, comprising:

[0045] A signal acquisition module is configured to synchronously acquire multi-modal teaching signals originating from the same time base through a multi-source heterogeneous sensor group integrated in a client during off-screen teaching;

[0046] A quantitative evaluation module is configured to extract spatio-temporal fusion features representing off-screen content attraction, understanding and interaction from the multi-modal teaching signals, and input the spatio-temporal fusion features into a pre-trained utility evaluation model to output a quantitative evaluation value of the current off-screen effect, the quantitative evaluation value including an off-screen attraction index, a cognitive load index and an interaction conversion rate;

[0047] A strategy analysis module is configured to input the quantitative evaluation value and the attribute parameters of the current off-screen content as a state vector into a DRL policy network, the DRL policy network being configured with a reward function aiming to maximize long-term teaching effect, processing the input state vector and outputting real-time optimization instructions for off-screen content attributes;

[0048] An optimization execution module is configured to dynamically adjust the visual presentation attributes of the off-screen content being demonstrated based on the optimization instructions, and drive a rendering engine to update and output to a display device in real time, completing a closed-loop feedback optimization.

[0049] An out-screen teaching feedback optimization device comprises a memory and a processor;

[0050] The memory is configured to store a program.

[0051] The processor is configured to execute the program to implement the steps of the out-screen teaching feedback optimization method according to any one of the preceding embodiments.

[0052] A readable storage medium has a computer program stored thereon, and the computer program, when executed by a processor, implements the steps of the out-screen teaching feedback optimization method according to any one of the preceding embodiments.

[0053] A computer program product comprises a computer program, and the computer program, when executed by a processor, implements the steps of the out-screen teaching feedback optimization method according to any one of the preceding embodiments.

[0054] As can be seen from the technical solutions described above, the out-screen teaching feedback optimization method and related device provided by the embodiments of the present application have the core of synchronously collecting multi-modal signals through multi-source heterogeneous sensors, extracting spatio-temporal fusion features and outputting quantitative evaluation values through a pre-trained utility evaluation model, combining content attribute parameters to generate real-time optimization instructions through a DRL strategy network, and finally dynamically adjusting the visual attributes of out-screen content and completing closed-loop optimization.

[0055] The present application can solve the defects of the prior art and achieve significant beneficial effects. First, for the feedback lag problem, the multi-source heterogeneous sensor group can synchronously collect multi-modal signals at the same time base point in the off-screen teaching process in real time, combine subsequent real-time feature extraction, model calculation and optimization adjustment, and construct an end-to-end real-time closed loop from data acquisition to content optimization to ensure that the optimization instructions can act on the current teaching process in real time and avoid lag effects. Second, for the deficiencies of data acquisition and feature extraction, the multi-source heterogeneous sensor covers multiple dimensions such as physiology, action and vision, and can obtain complete data reflecting attraction, understanding and interaction. The spatio-temporal fusion feature integrates the time dynamic change and spatial distribution information of the signal to provide data and feature basis for accurate evaluation. Third, the present application improves the evaluation accuracy and optimization scientificity. The pre-trained utility evaluation model can stably and quantitatively output key indicators such as off-screen attraction index, cognitive load index and interaction conversion rate, avoiding the evaluation deviation of traditional rule models. The reward function configured by the DRL strategy network aims to maximize the long-term teaching effect, which can break through the limitations of preset fixed rules and dynamically learn and optimize strategies according to real-time quantitative evaluation values and content attributes. For example, when the cognitive load index is too high, the content visual complexity is automatically adjusted to reduce the load, while ensuring the long-term knowledge transmission efficiency, rather than short-term local optimization, to ensure that the optimization measures adapt to the current teaching state and continuously improve the overall teaching effect. Finally, the visual presentation attributes are dynamically adjusted and the rendering engine is updated in real time to make the off-screen content always highly adaptive to the learner's state, significantly improving the learner's learning interest, reducing the cognitive burden and improving the interaction participation, fully utilizing the technical advantages of off-screen teaching and ensuring the continuous optimization of long-term teaching effect. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0057] Figure 1 A flowchart of an off-screen teaching feedback optimization method disclosed by an embodiment of the present application;

[0058] Figure 2 A schematic diagram of an off-screen teaching feedback optimization device disclosed by an embodiment of the present application;

[0059] Figure 3 A hardware structure block diagram of an off-screen teaching feedback optimization device disclosed by an embodiment of the present application. DETAILED DESCRIPTION

[0060] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0061] The present application can be used in many general or special-purpose computing device environments or configurations. For example: personal computers, server computers, handheld or laptop devices, tablet devices, multiprocessor systems, distributed computing environments that include any of the above systems or devices, and the like.

[0062] Next, the technical solutions of the present application will be described. The present application proposes the following technical solutions, which are described below.

[0063] Figure 1 A flowchart of a screen-out teaching feedback optimization method disclosed in an embodiment of the present application.

[0064] As shown in Figure 1 , the method can include:

[0065] Step S1, in the screen-out teaching process, a multi-source heterogeneous sensor group integrated in a client is used to synchronously collect multi-modal teaching signals originating from the same time base point.

[0066] Specifically, the multi-source heterogeneous sensor group is integrated in a client (such as AR glasses, VR headsets, smart teaching terminals, etc.) used for screen-out teaching, and each sensor establishes a real-time data transmission link with the master module of the client to ensure that the collected multi-modal teaching signals have a unified time base point. Screen-out teaching includes AR / VR stereoscopic screen-out presentation, 3D interactive teaching (such as 3D presentation documents), and other forms. The master module sends a synchronization trigger instruction to each sensor to start data collection at the same timestamp, avoiding signal correlation deviation caused by collection time difference. The multi-modal teaching signals specifically include three types of visual attention signals, interactive behavior signals, and context teaching signals:

[0067] Visual attention signal: acquired by a convolutional neural network (CNN) based real-time eye tracking module integrated in the client, which includes an infrared camera, a light source assembly, and a CNN processing unit. The infrared camera captures eye images of the learner with the assistance of the light source assembly, and the CNN processing unit performs real-time feature extraction (such as pupil center positioning and eyelid opening recognition) on the eye images, outputting two types of core signals: one is a distribution heat map of the fixation point in the on-screen area and the non-on-screen area, which is based on the pixel coordinates of the on-screen display interface and uses different color gradients to represent the distribution density of the fixation point per unit time, directly reflecting the learner's attention area on the on-screen content; the other is the pupil diameter change sequence, which records the dynamic change data of the pupil diameter with time at a millisecond level sampling frequency, which is used for subsequent correlation analysis of the learner's cognitive load state (such as pupil diameter increase usually corresponding to cognitive load increase);

[0068] Interaction behavior signal: acquired by the touch sensor, pressure sensor, and three-dimensional motion capture sensor integrated in the client: the touch sensor records the frequency of the learner's touch operation on the on-screen content (such as the number of times the learner clicks or slides the virtual controls on the screen); the pressure sensor synchronously collects the operation force during the touch operation, reflecting the initiative of the interaction operation; the three-dimensional motion capture sensor captures the learner's gestural operation trajectory in the air for the on-screen content, recording the starting point, motion path, and endpoint of the gesture in the form of three-dimensional coordinate sequence, such as the gesture trajectory data of "grabbing" and "zooming" the virtual model on the screen;

[0069] Context teaching signal: acquired through real-time communication between the client and the teaching management system, including two types of core information: one is the identifier of the current teaching knowledge point, which is a unique code (such as "junior high school physics - mechanics - principle of buoyancy - 001"), used to associate the on-screen content with the teaching goal; the other is the attribute parameter of the on-screen content, including the visual complexity (such as the number of model polygons), the dynamic presentation frequency (such as the animation frame rate), and the information density (such as the number of words and graphics per unit interface) of the on-screen content, which are used as basic parameters for subsequent optimization and adjustment.

[0070] Step S2, extract the spatio-temporal fusion features representing the on-screen content attraction, understanding, and interaction from the multi-modal teaching signals, and input the spatio-temporal fusion features into the pre-trained utility evaluation model to output a quantitative evaluation value of the current on-screen effect, the quantitative evaluation value including the on-screen attraction index, the cognitive load index, and the interaction conversion rate.

[0071] Specifically, first, the multimodal teaching signal is preprocessed: the pixel-level normalization is performed on the distribution heat map of the visual attention signal to remove noise points caused by environmental light interference; the pupil diameter change sequence is smoothed to eliminate the interference of accidental actions such as blinking; the coordinate calibration is performed on the operation trajectory of the interactive behavior signal to ensure consistency with the virtual coordinate system of the off-screen content; and the attribute parameters of the context teaching signal are numerically encoded, and the non-numeric parameters are converted into One-Hot vectors.

[0072] Subsequently, the spatio-temporal fusion feature extraction is performed: from the time dimension, the dynamic features of each signal within a preset time window are extracted, such as the proportion of the duration of the fixation point in the off-screen area in the visual attention signal, the standard deviation of the pupil diameter change, the time interval of the touch operation in the interactive behavior signal, and the motion speed change rate of the gesture trajectory; from the spatial dimension, the spatial distribution features of the signal are extracted, such as the spatial overlap rate of the high attention area in the visual attention heat map and the core teaching content of the off-screen, and the spatial matching degree of the gesture trajectory and the interactive control of the off-screen; and then the spatio-temporal dimension features are fused into a unified dimension spatio-temporal fusion feature vector through a feature splicing module, wherein: the feature sub-vector representing the attraction degree is mainly based on the spatio-temporal features of the visual attention signal (such as the proportion of the fixation time in the off-screen area, and the overlap rate of the high attention area); the feature sub-vector representing the understanding degree is mainly based on the spatio-temporal features of the pupil diameter change (such as the standard deviation of the pupil diameter, and the associated change rate of the pupil diameter and the difficulty of the knowledge point); and the feature sub-vector representing the interaction degree is mainly based on the spatio-temporal features of the interactive behavior signal (such as the touch operation frequency, the gesture trajectory matching degree, and the operation force average).

[0073] The pre-trained utility evaluation model is a deep learning-based regression model, and the training process takes the spatio-temporal fusion features of the historical multimodal teaching signal as input, and takes the manually labeled off-screen effect evaluation value (such as attraction score, cognitive load score, and interactive conversion rate) as label, and optimizes the model parameters to convergence through gradient descent algorithm. After inputting the current spatio-temporal fusion features into the model, the model outputs three continuous numerical quantitative evaluation values, and the specific calculation process of each evaluation value is as follows:

[0074] The off-screen attraction index is calculated according to the ratio of the residence time of the fixation point in the off-screen area to the total viewing time in the visual attention signal, and the dilation change rate of the pupil diameter at the off-screen moment;

[0075] The cognitive load index is calculated according to the saccade speed of the fixation point jumping between different sub-components of the off-screen content in the visual attention signal, and the operation hesitation time for the corresponding knowledge point of the off-screen content in the interactive behavior signal;

[0076] The interaction conversion rate is calculated according to the ratio of the number of valid interaction operations to the total number of views in the interaction behavior signal after triggering the off-screen content.

[0077] Specifically, the off-screen attraction index is calculated according to two types of parameters in the visual attention signal. One is the fixation point residence time ratio, that is, the ratio of the cumulative residence time of the fixation point falling in the off-screen area (the cumulative time from the fixation point entering the off-screen area coordinate range to leaving) to the total viewing time of the learner in the teaching period (the total time from the beginning of the presentation of the off-screen content to the current time). The other is the pupil diameter expansion rate, that is, the ratio of the pupil diameter increment at the first presentation moment of the off-screen content (based on the off-screen rendering completion timestamp) to the average pupil diameter of the off-screen content. The model weights and sums the two types of parameters by a preset weight, and then maps them to the corresponding interval after normalization processing. The higher the value, the higher the attention of the learner to the off-screen content.

[0078] The cognitive load index is calculated according to two types of parameters in the visual attention signal and the interaction behavior signal. One is the fixation point saccade speed, that is, the average speed (ratio of jump distance to jump time) of the fixation point jumping between different sub-components (such as knowledge point icons, explanation text boxes, and other preset sub-regions) of the off-screen content in the visual attention signal. The faster the saccade speed, the faster the cognitive processing rhythm and the higher the load. The other is the operation hesitation time, that is, the interval time from the presentation of the corresponding knowledge point of the off-screen content (based on the knowledge point identifier loading completion timestamp) to the learner's first triggering of the interaction operation for the knowledge point in the interaction behavior signal. The longer the hesitation time, the more difficult the understanding of the knowledge point and the higher the cognitive load. After standardization processing of the two types of parameters, the model obtains the index in the corresponding interval by weighted summation and value mapping. The higher the value, the heavier the current cognitive burden of the learner.

[0079] The interaction conversion rate is calculated according to the ratio of the number of valid interaction operations to the total number of views in the interaction behavior signal after triggering the off-screen content, wherein the number of valid interaction operations refers to the number of interaction operations triggered by the learner and matched with the teaching goal of the off-screen content (such as clicking the knowledge point detail control, correctly completing the mimic gesture operation, and determining the validity by comparing the operation target and the knowledge point identifier); the total number of views refers to the total number of views of the current off-screen content by the learner in the teaching period (counted by the number of complete rendering presentations of the off-screen content, excluding incomplete presentations due to stuttering). The model directly calculates the ratio of the two, and the higher the ratio reflects the higher the interaction participation degree of the learner to the off-screen content.

[0080] Step S3, the quantitative evaluation value and the attribute parameter of the current on-screen content are jointly input as a state vector into a DRL policy network configured with a reward function aiming to maximize long-term teaching effect, the state vector is processed, and a real-time optimization instruction for the on-screen content attribute is output.

[0081] Specifically, the attribute parameter of the current on-screen content is a key characteristic parameter of the current teaching content in the presentation form. These parameters are quantitative descriptions of the presentation attributes of the on-screen content itself, including visual complexity (such as the number of picture elements, color levels, etc.), dynamic presentation frequency (such as animation switching speed, content update frequency, etc.), information density (such as the number of knowledge points displayed per unit time, the density of text and charts, etc.), as well as font size, contrast, interactive element type, and other content presentation-related attribute indicators. They collectively constitute a parameter set describing the attribute of the current on-screen content, providing basic information about the characteristics of the content itself for the construction of the state vector.

[0082] In step S3, the state vector is first constructed. The on-screen attraction index, cognitive load index, and interactive conversion rate quantitative evaluation values output in step S2 are vector-spliced with the obtained current on-screen content attribute parameters (including visual complexity, dynamic presentation frequency, information density, etc.) to form a state vector with fixed dimensions, which fully represents the correspondence between the on-screen effect and the content attribute in the current teaching scenario.

[0083] The DRL policy network is a policy generation network under the deep reinforcement learning framework, which has the ability to dynamically generate optimization strategies based on input state vectors. The training process of this network aims to maximize the long-term teaching effect: first, multiple rounds of on-screen teaching scenarios are constructed in the training environment, each round corresponding to on-screen content of different knowledge points; second, a reward function is configured, which is calculated based on the weighted sum of short-term reward and long-term reward. The short-term reward is calculated based on the current quantitative evaluation value (for example, a positive short-term reward is given when the on-screen attraction index is greater than 8 and the cognitive load index is less than 5; a negative short-term reward is given when the interactive conversion rate is less than 0.3); the long-term reward is calculated based on the average of the quantitative evaluation values in the preset teaching period and the knowledge point mastery test score, ensuring that the reward function is not limited to the optimization effect of the current step, but is directed towards the realization of the long-term teaching goal.

[0084] In the real-time processing stage, the DRL policy network performs feature mapping on the input state vector and outputs discrete or continuous real-time optimization instructions. The instructions specifically correspond to the adjustment direction and adjustment amplitude of the on-screen content attribute, and the instruction format is compatible with the content adjustment module of the client, ensuring that it can be directly parsed and executed.

[0085] In addition, during the analysis process of the DRL policy network, the external teaching auxiliary system diagnostic events can also be fused for processing, and the technical implementation details are as follows:

[0086] ①Through the preset cross-system interface, at least one diagnostic event generated by the external teaching auxiliary system is monitored, and the diagnostic event contains an event type identifier and an associated knowledge point identifier, which is used to indicate that the learning or operation effect of a specific knowledge point does not meet the expectation;

[0087] ②In response to the monitored diagnostic event, the target knowledge point identifier to be strengthened is parsed;

[0088] ③According to the target knowledge point identifier, a reinforcement teaching instruction is generated and input to the DRL policy network, so as to optimize the visual presentation properties of the off-screen content associated with the target knowledge point through the real-time optimization instruction output by the DRL policy network.

[0089] Specifically, the preset cross-system interface adopts a modular design and supports bidirectional communication with at least one external teaching auxiliary system, such as an intelligent homework analysis system, a knowledge point mastery evaluation platform, a virtual practical operation examination system, etc. The interface needs to complete protocol adaptation and permission authentication in advance to ensure the security and legality of data transmission. The triggering conditions of the diagnostic event specifically include but are not limited to: the homework correct rate of a specific knowledge point in the external teaching auxiliary system is lower than a preset threshold, the number of incorrect answers in the knowledge point evaluation exceeds the set upper limit, the key operation steps are wrong due to knowledge point understanding deviation in the practical operation simulation training and retry fails, etc.; the event type identifier can be specifically divided into "homework diagnostic event", "evaluation diagnostic event", "practical operation diagnostic event", etc., and each type of identifier corresponds to a unique event processing priority, so that the DRL policy network can subsequently respond to high-urgency teaching needs first; the associated knowledge point identifier adopts a structured coding rule to ensure accurate association to specific knowledge points in the off-screen teaching content.

[0090] In response to the diagnostic event, the response mechanism adopts a combination of real-time triggering and batch verification. Specifically, when a diagnostic event is detected, the system first starts a real-time response process, extracts the "knowledge point identifier" field from the event data packet through a pre-set event analysis engine, and performs format verification on the identifier based on the built-in knowledge point coding rules to eliminate invalid codes (such as missing fields or incorrect code length identifiers). If multiple diagnostic events are detected at a time and the same knowledge point identifier is associated multiple times (such as a knowledge point triggering both job diagnostic events and evaluation diagnostic events), a batch verification process is started, the repeated knowledge point identifiers are processed through event fusion algorithms, and the reinforcement demand weight of the knowledge point is calculated according to the event type priority. Finally, the de-duplicated knowledge point identifier that meets the pre-set reinforcement threshold is determined as the target knowledge point identifier for reinforcement. At the same time, the system performs consistency comparison between the target knowledge point identifier obtained by analysis and the built-in teaching content database to confirm that the identifier corresponds to a knowledge point that already exists in the current on-screen teaching curriculum system, avoiding the generation of invalid reinforcement teaching instructions in the future.

[0091] The generation of reinforcement teaching instructions needs to be based on a pre-set knowledge point and reinforcement strategy mapping library. The mapping library pre-stores the optimal visual presentation reinforcement strategies corresponding to different types of knowledge points. For example, for theoretical concept type target knowledge points, the mapping library configures the reinforcement strategy of "improving the color contrast of core concept text and background, and increasing the insertion frequency of dynamic illustrations for concept explanation"; for practical step type target knowledge points, the mapping library configures the reinforcement strategy of "reducing the dynamic presentation rate of key operation steps, and increasing the visual prominence of operation guide annotations". In the specific generation process, the system first queries the mapping library according to the target knowledge point identifier to obtain the corresponding reinforcement strategy parameters (including the attribute type of the on-screen content to be adjusted, the attribute adjustment direction, and the recommended adjustment amplitude), and then generates a structured reinforcement teaching instruction based on the attribute parameter baseline value of the current on-screen content. The instruction at least includes "target knowledge point identifier, attribute type to be optimized, attribute reinforcement direction, and reinforcement priority" fields. Subsequently, the reinforcement teaching instruction is transmitted through the dedicated input port of the DRL strategy network, fused with the state vector containing the quantitative evaluation value and the attribute parameter of the current on-screen content constructed in the previous step, and an extended state vector is formed. At this time, the reward function configured by the DRL strategy network automatically adjusts the weight and sets an additional reward coefficient for the optimization effect of the on-screen content attributes related to the target knowledge point, so as to guide the DRL strategy network to prioritize the visual presentation reinforcement needs of the target knowledge point when outputting real-time optimization instructions, ensure that the on-screen content associated with the target knowledge point can be optimized through attribute adjustment to improve the students' attention focus and understanding acceptance of the knowledge point, further improve the closed-loop feedback optimization mechanism of on-screen teaching, and ensure the realization of the long-term teaching effect maximization goal.

[0092] Step S4, based on the optimization instruction, dynamically adjusting the visual presentation properties of the on-screen content being demonstrated, and driving the rendering engine to update the output to the display device in real time, completing the closed-loop feedback optimization.

[0093] Specifically, first, the optimization instruction output by the DRL policy network is parsed, and the visual presentation property type to be adjusted and the corresponding adjustment parameter in the instruction are extracted to generate a standardized property adjustment instruction. Subsequently, according to the property adjustment instruction, the visual presentation properties of the on-screen content being demonstrated are dynamically adjusted, and the specific adjustment rules are as follows:

[0094] When it is detected that the on-screen attraction index is lower than a first preset threshold, the highlight displacement amount of the on-screen content is increased, and the high light reflection property of the material surface of the on-screen content is enhanced;

[0095] When it is detected that the cognitive load index is higher than a second preset threshold, the rendering complexity of the on-screen model is reduced or the animation speed of the on-screen model is reduced, and a guiding annotation arrow is triggered to be generated, the guiding annotation arrow pointing to the core component of the on-screen model;

[0096] When it is detected that the interaction conversion rate is lower than a third preset threshold, a virtual interaction control is dynamically generated beside the on-screen content, the virtual interaction control having a semi-transparent property and carrying a feedback metaphor guiding the user to perform an interactive operation on the on-screen content.

[0097] Specifically, for the adjustment of the on-screen attraction index: when it is detected that the output on-screen attraction index is lower than a first preset threshold (the first preset threshold can be configured according to the type of teaching content), the content adjustment module performs two visual enhancement operations: first, the highlight displacement amount of the on-screen content is increased, and the original center position of the on-screen content is taken as the reference, and the on-screen content is offset by a preset distance in the depth direction perpendicular to the display interface; second, the high light reflection property of the material surface of the on-screen content is enhanced, and by adjusting the high light reflection coefficient of the material and expanding the high light reflection range, the material surface presents a more obvious luster under the environment light rendering, and the visual attraction is improved;

[0098] Adjustment for cognitive load index: when the cognitive load index is detected to be higher than a second preset threshold (the second preset threshold is configured according to the age of the learner), the content adjustment module performs cognitive load reduction operation: first, reduce the rendering complexity or animation speed of the out-screen model. If the rendering complexity is selected to be reduced, the polygon number of the out-screen three-dimensional model is reduced by calling the edge folding algorithm, or the model surface texture is simplified. If the animation speed is selected to be reduced, the animation playing rate of the out-screen model is reduced from the original rate, and the display time length of the key action is prolonged. Second, a guiding annotation arrow is triggered to be generated. The annotation arrow adopts a high-contrast color. The starting point of the arrow is the edge area of the out-screen content, and the end point points to the core component of the out-screen model. The arrow is accompanied by a slow flashing effect, guiding the learner to focus on the core knowledge and reducing invalid cognitive consumption.

[0099] Adjustment for interaction conversion rate: when the interaction conversion rate is detected to be lower than a third preset threshold (the third preset threshold is configured according to the interaction scene), the content adjustment module dynamically generates a virtual interaction control on the side of the out-screen content. The control adopts a semi-transparent property. The control form matches the interaction operation type, and carries a guiding feedback metaphor. For example, the control will appear color gradient when the learner's line of sight is close, and a slight vibration is sent through the vibration module of the client, or a short prompt sound is played through the audio module, prompting the learner to perform the interaction operation and reducing the interaction cognitive threshold.

[0100] After the content adjustment is completed, the main control module of the client sends an update instruction to the integrated rendering engine. The rendering engine performs rendering processing such as light calculation, texture mapping, and stereoscopic projection transformation according to the adjusted out-screen content data, generates out-screen image data conforming to the display device parameters, and outputs to the display device (such as the optical display component of AR glasses or the display screen of VR head-mounted display) in real time through the data transmission interface, realizing the real-time update of the out-screen content.

[0101] At the same time, the timestamp of this optimization adjustment, the attribute parameters before and after the adjustment, and the quantitative evaluation value are recorded to form a closed-loop feedback log. On the one hand, it is used for incremental training of the subsequent DRL policy network, and on the other hand, it realizes the traceability of the optimization process, so as to complete the complete closed-loop feedback optimization and ensure that the out-screen teaching effect is continuously improved with the teaching process.

[0102] As can be seen from the above technical solutions, the out-screen teaching feedback optimization method and related equipment provided by the embodiments of the present application core is to synchronously collect multi-modal signals through multi-source heterogeneous sensors, extract spatio-temporal fusion features, and output quantitative evaluation values through a pre-trained utility evaluation model. Then, the content attribute parameters are combined to generate real-time optimization instructions through a DRL policy network, and finally the visual attributes of the out-screen content are dynamically adjusted and closed-loop optimization is completed.

[0103] This application addresses the shortcomings of existing technologies and achieves significant beneficial effects: First, regarding the feedback lag problem, the multi-source heterogeneous sensor array can synchronously collect multimodal signals at the same time point during the on-screen teaching process. Combined with subsequent real-time feature extraction, model calculation, and optimization adjustments, it constructs an end-to-end real-time closed loop from data acquisition to content optimization, ensuring that optimization instructions can be applied to the current teaching process immediately and avoiding the impact of lag. Second, addressing the deficiencies in data acquisition and feature extraction, the multi-source heterogeneous sensors cover multiple dimensions such as physiology, movement, and vision, and can acquire complete data reflecting attractiveness, comprehension, and interactivity. Spatiotemporal fusion features integrate the temporal dynamic changes and spatial distribution information of the signals, providing a data and feature foundation for accurate evaluation. Third, this application improves the accuracy of evaluation and the scientific nature of optimization. The pre-trained utility evaluation model can stably quantify the output of the on-screen results. By utilizing key indicators such as attraction index, cognitive load index, and interaction conversion rate, and avoiding the evaluation biases of traditional rule-based models, the DRL strategy network, configured with a reward function aimed at maximizing long-term teaching effectiveness, can break free from the limitations of preset fixed rules. It dynamically learns and optimizes strategies based on real-time quantitative evaluation values ​​and content attributes. For example, when the cognitive load index is too high, it automatically adjusts the visual complexity of the content to reduce the load, while ensuring long-term knowledge transfer efficiency, rather than short-term localized optimization. This ensures that optimization measures are both adapted to the current teaching state and continuously improve the overall teaching effectiveness. Finally, by dynamically adjusting visual presentation attributes and driving real-time updates to the rendering engine, the out-of-screen content is always highly adapted to the learner's state, significantly enhancing learners' learning interest, reducing cognitive burden, and increasing interactive participation. This fully leverages the technological advantages of out-of-screen teaching and ensures continuous optimization of long-term teaching effectiveness.

[0104] In some embodiments of this application, the DRL policy network described in step S3 is introduced.

[0105] The framework for constructing the DRL policy network includes:

[0106] The state vectors constitute the state space of the DRL policy network.

[0107] The action space of the DRL policy network is defined as a set of operation instructions for adjusting the attributes of the out-of-screen content, including instructions for fine-tuning the out-of-screen depth, instructions for increasing or decreasing the visual saliency of the model, and instructions for showing or hiding additional explanatory annotations.

[0108] The reward function consists of a short-term reward and a long-term reward. The short-term reward is calculated based on the immediate positive change of the quantitative evaluation value after the action is performed. The long-term reward is calculated based on the pass rate of the knowledge points related to the content displayed on the screen in the subsequent teaching process.

[0109] By selecting the action in the action space, the optimization learning of the DRL policy network is implemented in combination with the reward value fed back by the reward function until the goal of maximizing the long-term teaching effect is achieved.

[0110] Specifically, the construction framework of the DRL policy network includes state space definition, action space definition, reward function design and optimization learning mechanism, and each part cooperates to realize the strategy generation with the goal of long-term teaching effect, as follows:

[0111] 1. State space definition

[0112] The state vector constitutes the state space of the DRL policy network, and the dimension of the state vector is determined by the number of quantitative evaluation values and out-screen content attribute parameters. For example, if the quantitative evaluation values include out-screen attraction index, cognitive load index and interaction conversion rate, and the out-screen content attribute parameters include out-screen depth, model visual saliency and the number of additional explanatory annotations, the dimension of the state vector is 6, and the state space is a 6-dimensional real vector space.

[0113] At the same time, in order to avoid the dimensional difference of different dimension parameters affecting network training, normalization processing is required for each parameter in the state vector; and the boundary of the state space is limited by the measured data of the teaching scene, such as the out-screen depth not exceeding the maximum depth of 10 centimeters supported by the display device, and the model visual saliency not being less than 20 to ensure basic visibility, so as to ensure the rationality and practicality of the state space.

[0114] 2. Action space definition

[0115] The action space of the DRL policy network is defined as a set of operation instructions for adjusting the out-screen content attributes, which includes three types of core operation instructions, each of which clearly adjusts the direction, amplitude and operation boundary, as follows:

[0116] Out-screen depth fine-tuning instruction: used to adjust the position of the out-screen content in the vertical direction of the display interface depth, and the adjustment range is consistent with the value range of the out-screen depth in the state space;

[0117] Model visual saliency increase / decrease instruction: used to adjust the visual prominence of the out-screen model, the adjustment objects include the high light reflection coefficient and contrast of the model, and the visual saliency after adjustment needs to be maintained within the interval of 20-100;

[0118] Additional explanatory annotation display / hide instruction: used to control the display or hiding of the explanatory annotations associated with the out-screen content, the annotation types include text notes and schematic icons, the instruction includes "display specified annotations" and "hide specified annotations", and the annotations need to maintain a preset spatial distance from the out-screen content when displayed.

[0119] All instructions in the action space are converted into a vector form recognizable by the network using One-Hot encoding, for example, three types of instructions include 7 specific operations (2 depth fine-tuning, 2 saliency increase and decrease, 3 label display and hide), the action space dimension is 7, and each action corresponds to a 7-dimensional One-Hot vector.

[0120] 3. Reward function design

[0121] The reward function is composed of the weighted sum of short-term reward and long-term reward, and the total reward value is equal to the short-term reward multiplied by the short-term reward weight, plus the long-term reward multiplied by the long-term reward weight. The calculation basis and method of the two types of rewards are as follows:

[0122] Short-term reward: the calculation basis is the immediate positive change of the quantitative evaluation value after the action is executed, which is specifically calculated by the difference between the quantitative evaluation values before and after the action is executed. Among them, the change value of the off-screen attraction index is the index after the action is executed minus the index before the action is executed, the positive change value of the cognitive load index is the index before the action is executed minus the index after the action is executed (the decrease of cognitive load is converted into a positive index), and the change value of the interaction conversion rate is the conversion rate after the action is executed minus the conversion rate before the action is executed; if the contribution weights of the three types of indicators are set to 0.4, 0.3, and 0.3 respectively, the short-term reward is equal to the change value of the off-screen attraction index multiplied by 0.4, plus the positive change value of the cognitive load index multiplied by 0.3, and plus the change value of the interaction conversion rate multiplied by 0.3. When the calculation result is greater than or equal to 0, it is determined that the immediate positive change is retained, and the calculation value is retained; when the calculation result is less than 0, the short-term reward is taken as 0 to avoid excessive influence of negative feedback on network exploration.

[0123] Long-term reward: the calculation basis is the passing rate of the knowledge point examination associated with the off-screen content in the subsequent teaching process, which is obtained from the teaching management system. The long-term reward value is the passing rate of the knowledge point examination multiplied by 10, which maps the passing rate from the range of 0-1 to the interval of 0-10, keeping the same order of magnitude as the short-term reward, for example, when the passing rate of the knowledge point examination is 80%, the long-term reward value is 8; if there is no direct examination data in the subsequent process, the average interaction conversion rate of the knowledge point within 1 hour after the action is executed is used instead of the passing rate to ensure that the long-term reward can be normally calculated.

[0124] 4. Optimization learning mechanism

[0125] Through the iterative process, the optimization learning of the DRL policy network is implemented until the goal of maximizing the long-term teaching effect is achieved, and the specific process is as follows:

[0126] Load the pre-trained network parameters trained based on historical teaching data, set the iteration rounds (each round corresponds to a complete knowledge point teaching process), and set the convergence threshold. At the current time of each iteration, collect the real-time state vector from the state space and perform normalization processing. The action is selected using the ε-greedy strategy. When the randomly generated value is greater than the ε value, the action with the maximum network output is selected. When the randomly generated value is less than or equal to the ε value, an action is randomly selected from the action space. After executing the selected action, the short-term reward value is obtained according to the calculation method of the short-term reward. After the subsequent teaching session is completed, the long-term reward value is obtained according to the calculation method of the long-term reward, and then the total reward value is calculated. Based on the total reward value and the time difference error, the network weight is updated using the gradient descent algorithm to minimize the strategy loss function. If the average long-term reward of consecutive multiple iterations fluctuates by no more than the convergence threshold, it is determined that the network training converges, and the iteration is stopped. Otherwise, return to continue iteration until the target is reached.

[0127] Further, the DRL strategy network is trained using a proximal policy optimization algorithm, and the network weight is initialized offline based on historical teaching session data and continuously updated online during the teaching process based on the multi-modal teaching signals of the student to adapt to different teaching styles of teachers and cognitive habits of different audiences.

[0128] Specifically, the DRL policy network adopts a proximal policy optimization algorithm for training. The algorithm realizes stable updating of the policy by constructing a clipped surrogate objective function. The core lies in that, in each round of training, the current policy is first allowed to interact with the teaching environment to generate complete trajectory data containing states, actions, and rewards. Then, the probability ratio of the new and old policies selecting the same action under the same state is calculated, and the ratio is constrained within a reasonable range to avoid policy divergence due to excessive parameter update amplitude in a single round, thereby ensuring the stability and convergence of the training process. For network weights, offline initialization is first completed based on historical teaching session data. Specifically, teaching session data covering different teaching scenarios, different teacher styles, and different target groups are collected. After cleaning and removing samples with sensor abnormalities and labeling errors, the data are divided into a training set and a validation set in proportion. Then, appropriate hyperparameters are set according to the data size and network complexity. The state vector is taken as the input, the action instruction is taken as the output, and the reward function result is taken as the optimization target for multiple rounds of iterative training until the long-term reward average of the validation set tends to be stable. Then, the trained weights are stored in the parameter library according to the scene, style, and group labels for calling and matching the initial weights when a new teaching scenario is started. In the online teaching process, the network starts weight fine-tuning in two modes, i.e., a mode triggered at a fixed time according to the teaching rhythm and a mode triggered by an abnormal quantitative evaluation value. During each fine-tuning, recent real-time multi-modal teaching signals are collected, and the features are aligned with the offline historical data to form a mixed data set containing historical and real-time features. Meanwhile, a decaying learning rate mechanism is adopted, and the parameter update proportion is controlled to avoid excessive deviation from the basic policy. Through this mechanism, the network can automatically capture the teaching rhythm of the teacher, the key labeling habits in the context teaching signals, and the cognitive characteristics of different target groups in the multi-modal signals, continuously adjust the policy output to adapt to individual needs, and ensure stable teaching effect optimization capability under different teaching styles and cognitive habits.

[0129] In some embodiments of the present application, considering that a single feedback optimization mechanism is difficult to adapt to the cognitive differences of different target groups, the characteristics of the knowledge field, and the dynamic changes of knowledge in the learning process, it may lead to insufficient targetedness of the optimization policy and limitations of the teaching effect improvement. Therefore, the present application can further include feedback optimization based on target group portraits and learning knowledge fields, feedback optimization based on root cause diagnosis, and feedback optimization based on incremental knowledge differences to improve the accuracy, adaptability, and dynamic adaptability of off-screen teaching feedback. The following will explain this in detail.

[0130] I. Feedback optimization based on target group portraits and learning knowledge fields

[0131] The core of this optimization mechanism is to fuse the individualized features of the target group and the associated information of the knowledge field to make the state vector of the DRL policy network more suitable for the exclusive needs of the current teaching scenario, thereby avoiding the adaptation deviation caused by the generalization of the optimization policy. The specific implementation process is as follows:

[0132] ①Identify the audience identifier and the field knowledge topic of the current teaching, obtain the audience cognitive baseline level and the interaction preference parameter from the pre-built audience portrait library based on the audience identifier, and load the associated knowledge point network from the pre-built field knowledge graph based on the field knowledge topic;

[0133] ②The quantitative evaluation value, the current out-screen content attribute parameter, the audience cognitive baseline level and the interaction preference parameter, and the associated point and confusion point feature vector of the current knowledge point extracted from the associated knowledge point network are collectively used as a state vector and input to the DRL policy network.

[0134] Specifically, first, the audience identifier and the field knowledge topic of the current teaching are identified, wherein the audience identifier is information for uniquely identifying a teaching object group (such as a class number, a student group label “graduate student”, “adult continuing education student”, “engineer”), which can be pre-entered or automatically read through a client teaching management module; the field knowledge topic is the knowledge field and core topic corresponding to the current out-screen teaching (such as “mechanical engineering-drawing-three-dimensional part assembly”), which is determined by analyzing the knowledge point identifier of the out-screen content. Based on the above-mentioned audience identifier, the corresponding audience cognitive baseline level and interaction preference parameter are retrieved from the pre-built audience portrait library, which is a structured database pre-built through historical teaching data, preliminary test results and interaction records, wherein the “cognitive baseline level” includes the basic mastery degree of the audience to the field knowledge (such as “weak foundation”, “moderate mastery”, “skilled application”), the average cognitive load tolerance threshold, etc.; the interaction preference parameter includes the preferred interaction mode of the audience (such as “touch operation priority”, “air gesture priority”), the preference for the visual style of the out-screen content (such as “simple annotation preference”, “dynamic animation preference”), etc. At the same time, based on the field knowledge topic, the associated knowledge point network is loaded from the pre-built field knowledge graph. The field knowledge graph is a graph model constructed according to the knowledge logic relationship, which contains the parent-child relationship, the dependency relationship, the easy-to-confuse relationship, etc. between knowledge points, and the associated knowledge point network is a network structure composed of the previous basic knowledge points, the subsequent extended knowledge points and the same level easy-to-confuse knowledge points of the current knowledge point, from which the associated features (such as “dependency knowledge point ‘triangle three-side relationship’”) and confusion features (such as “easy to be confused with ‘triangle similarity judgment’”) of the current knowledge point are extracted and converted into feature vector form.

[0135] Subsequently, the state vector dimension of the DRL policy network is expanded, and the quantified evaluation value, the current out-screen content attribute parameters, the audience cognitive baseline level obtained above, the interaction preference parameters, and the current knowledge point association point and confusion point feature vector extracted from the association knowledge point network are jointly spliced to form a new state vector. The state vector not only contains the original "effect-content" dimension information, but also adds "audience-knowledge" dimension information, so that the DRL policy network can fully consider the cognitive ability boundary of the audience (such as avoiding generating high complexity out-screen content adjustment instructions for audiences with weak cognitive baseline) and interaction habits (such as prioritizing optimizing the gesture interaction response attributes of out-screen content for audiences who prefer gesture interaction), while adapting to the association characteristics of the knowledge field (such as prioritizing strengthening the discriminative visual features of out-screen content for easily confused knowledge points), thereby improving the scene adaptability of the optimization strategy.

[0136] II. Feedback optimization based on root cause diagnosis

[0137] For the case where the cognitive load index is abnormally high, the optimization mechanism locates the root cause of the understanding difficulty, so that the optimization instruction of the DRL policy network upgrades from simply reducing complexity to addressing the root cause, avoiding incomplete knowledge transmission caused by blind load reduction. The specific implementation process is as follows:

[0138] When the utility evaluation model detects that there is a target knowledge point with a cognitive load index continuously higher than the diagnostic trigger threshold, root cause diagnosis is triggered, and the root cause diagnosis result is formed and input to the DRL policy network. The root cause diagnosis queries the set of prerequisite basic knowledge points and the set of easily confused knowledge points that cause the target knowledge point to be difficult to understand by accessing the domain knowledge graph, and forms the root cause diagnosis result.

[0139] Specifically, first, the trigger condition for root cause diagnosis is set. When the cognitive load index output by the utility evaluation model continuously exceeds the preset diagnostic trigger threshold (which is dynamically adjusted based on the audience cognitive baseline level) and the duration reaches the preset period (such as 3 consecutive data collection periods), it is determined that there is a target knowledge point with excessively high cognitive load (i.e., the core knowledge point corresponding to the current out-screen teaching), and the root cause diagnosis process is triggered.

[0140] Subsequently, root cause diagnosis is performed. Through the real-time data interface between the client and the domain knowledge graph, the domain knowledge graph is accessed to query two types of key information with the target knowledge point as the retrieval core: one is the set of prerequisite basic knowledge points, which must be mastered in the learning process of the target knowledge point; the other is the set of easily confused knowledge points, which are easily confused with the target knowledge point in content, form, or application scenario. The set of prerequisite basic knowledge points and the set of easily confused knowledge points are sorted by influence degree to form a structured root cause diagnosis result.

[0141] Finally, the root cause diagnosis result is input into the DRL policy network, and the root cause diagnosis result is converted into a numerical feature vector as a supplementary dimension into the state vector of the DRL policy network. At this time, the optimization instruction generated by the DRL policy network is no longer a single reduction of the rendering complexity of the out-screen model, but a targeted instruction that fundamentally solves the problem of excessive cognitive load, taking into account both load reduction and knowledge delivery integrity.

[0142] III. Feedback optimization based on incremental knowledge difference

[0143] This optimization mechanism focuses on the dynamic update of knowledge content in the learning process. By identifying the differences between the target learning content and the historical base version, the feedback optimization focuses on incremental knowledge and simplifies known knowledge reasonably, avoiding redundant teaching and insufficient optimization of key knowledge. The specific implementation process is as follows:

[0144] ① Receive the identification information of the target learning content version from the audience of this teaching, and retrieve the knowledge mastery map formed for the previous base version of the target learning content from the audience's learning history records;

[0145] ② By comparing the differences between the target learning content and the knowledge mastery map formed by the previous base version, generate an incremental knowledge difference set, which contains the newly added knowledge points, deleted knowledge points, and knowledge points with major changes of the target learning content relative to the previous base version;

[0146] ③ Adjust the reward function of the DRL policy network based on the incremental knowledge difference set, wherein:

[0147] For the first knowledge point included in the incremental knowledge difference set, give the first knowledge point a higher reward coefficient than other knowledge points to guide the optimization strategy to prioritize the out-screen teaching effect of the first knowledge point;

[0148] For the knowledge points that do not change between the target learning content version and the previous base version and that the user has mastered, use a simplified out-screen mode to demonstrate or skip the detailed out-screen demonstration process of the knowledge points.

[0149] Specifically, first, the target learning content version identifier and historical knowledge mastery data are acquired, a target learning content version identifier corresponding to the off-screen content used in this teaching is received, the identifier includes a content version number and a core knowledge point range, which can be extracted from the metadata of the off-screen content. At the same time, the learning history database of the audience is accessed through the client (the database stores the knowledge mastery of the audience after learning each version of the learning content), and the knowledge mastery graph formed for the previous basic version (such as V1.0 for the target version V2.0) of the target learning content is searched. The graph takes knowledge points as nodes and mastery degrees as node attributes, and directly reflects the audience's mastery state of the knowledge of the previous basic version.

[0150] Secondly, the incremental knowledge difference set is generated, the differences between the target learning content and the knowledge structure of the previous basic version are compared through a preset content comparison algorithm, and three types of knowledge point changes are identified: first, new knowledge points, that is, knowledge points that are added in the target version and not included in the previous basic version; second, deleted knowledge points, that is, knowledge points included in the previous basic version and deleted in the target version; third, major change knowledge points, that is, knowledge points in the target version that have major adjustments in content, derivation process or application scenario of the knowledge points in the previous basic version. The above three types of knowledge points are classified and arranged according to the change type to form an incremental knowledge difference set, and each knowledge point is labeled with a change description.

[0151] Finally, the feedback optimization strategy is adjusted based on the incremental knowledge difference set. On the one hand, the reward function of the DRL strategy network is adjusted, and for the first knowledge point (that is, the new knowledge point and the major change knowledge point) in the incremental knowledge difference set, a higher reward coefficient than other knowledge points is assigned, so that the DRL strategy network prioritizes the optimization of the off-screen attributes of the first knowledge point (such as prioritizing the increase of the attraction degree of the off-screen content of the new knowledge point and the reduction of its cognitive load) when generating optimization instructions, to ensure that the audience focuses on and masters the incremental knowledge; on the other hand, the off-screen demonstration mode is adjusted, for the knowledge points in the incremental knowledge difference set that do not change (that is, the content of the target version is consistent with that of the previous basic version) and are marked as mastered in the knowledge mastery graph, a simplified off-screen mode (such as reducing the number of annotations of the off-screen content of the knowledge point and reducing the frequency of animation dynamics) is triggered, or the detailed off-screen demonstration process of the knowledge point is directly skipped (only the brief off-screen prompt of the knowledge point title is retained), to avoid redundant teaching and occupy time, and to concentrate optimization resources on incremental knowledge and unmastered knowledge, to improve teaching efficiency and feedback optimization accuracy.

[0152] Next, a kind of off-screen teaching feedback optimization device provided by the embodiments of the present application is described, and the off-screen teaching feedback optimization device described below can be correspondingly referred to the off-screen teaching feedback optimization method described above.

[0153] Referring to Figure 2 ,Figure 2 A schematic diagram of a screen-out teaching feedback optimization device disclosed in an embodiment of the present application.

[0154] As shown in Figure 2 The screen-out teaching feedback optimization device can include:

[0155] The signal acquisition module 110 is configured to synchronously acquire multi-modal teaching signals originating from the same time base point through a multi-source heterogeneous sensor group integrated in the client during the screen-out teaching process.

[0156] The quantitative evaluation module 120 is configured to extract spatio-temporal fusion features representing the screen-out content attraction, understanding and interaction from the multi-modal teaching signals, and input the spatio-temporal fusion features into a pre-trained utility evaluation model to output a quantitative evaluation value of the current screen-out effect, wherein the quantitative evaluation value includes a screen-out attraction index, a cognitive load index and an interaction conversion rate.

[0157] The strategy analysis module 130 is configured to input the quantitative evaluation value and the attribute parameter of the current screen-out content as a state vector into a DRL strategy network, wherein the DRL strategy network is configured with a reward function aiming to maximize the long-term teaching effect, processes the input state vector, and outputs a real-time optimization instruction for the screen-out content attribute.

[0158] The optimization execution module 140 is configured to dynamically adjust the visual presentation attribute of the screen-out content being demonstrated based on the optimization instruction, and drive a rendering engine to update and output to a display device in real time to complete the closed-loop feedback optimization.

[0159] As can be seen from the above technical solutions, the screen-out teaching feedback optimization method and related device provided by the embodiment of the present application is to synchronously acquire multi-modal signals through a multi-source heterogeneous sensor, extract spatio-temporal fusion features and output quantitative evaluation values through a pre-trained utility evaluation model, then combine the content attribute parameters to generate real-time optimization instructions through a DRL strategy network, and finally dynamically adjust the visual attribute of the screen-out content and complete the closed-loop optimization.

[0160] The present application can solve the defects of the prior art and achieve significant beneficial effects. First, for the feedback lag problem, the multi-source heterogeneous sensor group can synchronously collect multi-modal signals of the same time base in the off-screen teaching process in real time, combine subsequent real-time feature extraction, model calculation and optimization adjustment, construct an end-to-end real-time closed loop from data acquisition to content optimization, ensure that the optimization instruction can act on the current teaching process in time, and avoid the lagging effect. Second, for the lack of data acquisition and feature extraction, the multi-source heterogeneous sensor covers multiple dimensions such as physiology, action and vision, and can obtain complete data reflecting attraction, understanding and interaction. The spatio-temporal fusion feature integrates the time dynamic change and spatial distribution information of the signal, providing data and feature basis for accurate evaluation. Third, the present application improves the evaluation accuracy and optimization scientificity. The pre-trained utility evaluation model can stably and quantitatively output key indicators such as off-screen attraction index, cognitive load index and interaction conversion rate, avoiding the evaluation deviation of traditional rule models. The reward function configured by the DRL strategy network aims to maximize the long-term teaching effect, which can break through the limitation of preset fixed rules, dynamically learn and optimize the strategy according to the real-time quantitative evaluation value and content attribute, such as automatically adjusting the content visual complexity to reduce the load when the cognitive load index is too high, while ensuring the long-term knowledge transmission efficiency, rather than short-term local optimization, ensuring that the optimization measures not only adapt to the current teaching state, but also continuously improve the overall teaching effect. Finally, the visual presentation attribute is dynamically adjusted, and the rendering engine is updated in real time, so that the off-screen content is always highly adapted to the learner's state, significantly improving the learner's learning interest, reducing the cognitive burden, and improving the interaction participation, fully playing the technical advantages of off-screen teaching, and ensuring the continuous optimization of long-term teaching effect.

[0161] The off-screen teaching feedback optimization device provided by the embodiments of the present application can be applied to an off-screen teaching feedback optimization equipment. Figure 3 The hardware structure block diagram of the off-screen teaching feedback optimization equipment is shown, referring to Figure 3 The hardware structure of the off-screen teaching feedback optimization equipment can include at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4;

[0162] In the embodiments of the present application, the number of processors 1, communication interfaces 2, memories 3 and communication buses 4 is at least one, and the processor 1, the communication interface 2 and the memory 3 complete communication with each other through the communication bus 4;

[0163] The processor 1 can be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application, etc.

[0164] The memory 3 can comprise a high-speed RAM memory and possibly also a non-volatile memory, such as at least one disk memory;

[0165] The memory stores a program, and the processor can invoke the program stored in the memory, and the program is used for:

[0166] In the out-screen teaching process, a multi-source heterogeneous sensor group integrated in the client is used to synchronously collect multi-modal teaching signals from the same time base point;

[0167] The spatio-temporal fusion features representing the out-screen content attraction degree, understanding degree and interaction degree are extracted from the multi-modal teaching signals, and the spatio-temporal fusion features are input into a pre-trained utility evaluation model to output a quantitative evaluation value of the current out-screen effect, and the quantitative evaluation value includes an out-screen attraction index, a cognitive load index and an interaction conversion rate;

[0168] The quantitative evaluation value and the attribute parameter of the current out-screen content are jointly used as a state vector, which is input into a DRL policy network configured with a reward function with the goal of maximizing the long-term teaching effect, the state vector is processed, and a real-time optimization instruction for the out-screen content attribute is output;

[0169] Based on the optimization instruction, the visual presentation attribute of the out-screen content being demonstrated is dynamically adjusted, and a rendering engine is driven to update and output to a display device in real time, thereby completing a closed-loop feedback optimization.

[0170] Optionally, the refinement function and the extension function of the program can refer to the description above.

[0171] The embodiments of the present application also provide a readable storage medium, which can store a program suitable for processor execution, and the program is used for:

[0172] In the out-screen teaching process, a multi-source heterogeneous sensor group integrated in the client is used to synchronously collect multi-modal teaching signals from the same time base point;

[0173] The spatio-temporal fusion features representing the out-screen content attraction degree, understanding degree and interaction degree are extracted from the multi-modal teaching signals, and the spatio-temporal fusion features are input into a pre-trained utility evaluation model to output a quantitative evaluation value of the current out-screen effect, and the quantitative evaluation value includes an out-screen attraction index, a cognitive load index and an interaction conversion rate;

[0174] The quantitative evaluation value and attribute parameters of the current out-screen content are jointly taken as a state vector, input to a DRL policy network configured with a reward function aiming to maximize long-term teaching effect, the state vector is processed, and a real-time optimization instruction for the attribute of the out-screen content is output;

[0175] Based on the optimization instruction, the visual presentation attribute of the out-screen content being demonstrated is dynamically adjusted, and a rendering engine is driven to update output to a display device in real time, completing closed-loop feedback optimization.

[0176] Optionally, the refinement function and the extension function of the program can refer to the description above.

[0177] The embodiments of the present application also provide a computer program product comprising a computer program, which, when executed by a processor, performs the method.

[0178] In the out-screen teaching process, a multi-source heterogeneous sensor group integrated in the client synchronously collects multi-modal teaching signals from the same time base point;

[0179] Temporal-spatial fusion features representing the attraction, understanding and interaction of the out-screen content are extracted from the multi-modal teaching signals, and the temporal-spatial fusion features are input into a pre-trained utility evaluation model to output a quantitative evaluation value of the current out-screen effect, the quantitative evaluation value including an out-screen attraction index, a cognitive load index and an interaction conversion rate;

[0180] The quantitative evaluation value and attribute parameters of the current out-screen content are jointly taken as a state vector, input to a DRL policy network configured with a reward function aiming to maximize long-term teaching effect, the state vector is processed, and a real-time optimization instruction for the attribute of the out-screen content is output;

[0181] Based on the optimization instruction, the visual presentation attribute of the out-screen content being demonstrated is dynamically adjusted, and a rendering engine is driven to update output to a display device in real time, completing closed-loop feedback optimization.

[0182] Optionally, the refinement function and the extension function of the program can refer to the description above.

[0183] Finally, it should be noted that the terms "first", "second", and the like, herein do not denote any order, quantity, combination, or importance, but rather are used to distinguish one element from another, and are not necessarily intended to denote the temporal or chronological sequence of the elements. Also, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the presence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0184] The various embodiments in the specification are described with progression in this order of description. Embodiments having the same or similar descriptions are referenced by the same reference numerals.

[0185] The above description of disclosed embodiments provides enabling teaching for a person skilled in the art to realize or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for optimizing out-of-screen teaching feedback, characterized in that, The application relates to a method for optimizing the content of an out-screen teaching process. In the out-screen teaching process, a multi-source heterogeneous sensor group integrated in a client is used to synchronously collect multi-modal teaching signals from the same time base point; spatial and temporal fusion features representing the out-screen content attraction, understanding and interaction are extracted from the multi-modal teaching signals, and the spatial and temporal fusion features are input into a pre-trained utility evaluation model to output a quantitative evaluation value of the current out-screen effect, the quantitative evaluation value including an out-screen attraction index, a cognitive load index and an interaction conversion rate; The quantitative evaluation value and the attribute parameters of the current out-screen content are used as a state vector to input into a DRL policy network, the DRL policy network is configured with a reward function aiming at maximizing the long-term teaching effect, the state vector is processed, and real-time optimization instructions for the out-screen content attributes are output; Based on the optimization instructions, the visual presentation attributes of the out-screen content being demonstrated are dynamically adjusted, and a rendering engine is driven to update and output to a display device in real time, thereby completing the closed-loop feedback optimization.

2. The method of claim 1, wherein, The multi-modal teaching signals include visual attention signals, interaction behavior signals and context teaching signals, wherein: The visual attention signals are obtained through a real-time eye tracking module based on a convolutional neural network (CNN) and include a distribution heat map of fixation points in the out-screen area and the non-out-screen area and a pupil diameter change sequence; The interaction behavior signals include the frequency of touch operations on the out-screen content, the operation force and the mimic gesture operation trajectory in the air for the out-screen content; The context teaching signals include the identifier of the current teaching knowledge point and the attribute parameters of the out-screen content.

3. The method of claim 1, wherein, The calculation process of the quantitative evaluation value includes: The out-screen attraction index is comprehensively calculated according to the ratio of the residence time of the fixation points in the out-screen area to the total watching time and the expansion change rate of the pupil diameter at the out-screen moment in the visual attention signals; The cognitive load index is comprehensively calculated according to the saccade speed of the fixation points jumping between different sub-components of the out-screen content in the visual attention signals and the operation hesitation time of the corresponding knowledge points of the out-screen content in the interaction behavior signals; The interaction conversion rate is calculated according to the ratio of the number of effective interaction operations after triggering the out-screen content to the total watching number in the interaction behavior signals.

4. The method of claim 1, wherein, The construction framework of the DRL policy network includes: The state vector constitutes the state space of the DRL policy network; The action space of the DRL policy network is defined as a set of operation instructions for adjusting the out-screen content attributes, including out-screen depth fine-tuning instructions, model visual saliency increase / decrease instructions and additional explanatory annotation display / hide instructions; The reward function is composed of a short-term reward and a long-term reward, wherein the calculation basis of the short-term reward is the immediate positive change of the quantitative evaluation value after the action execution, and the calculation basis of the long-term reward is the knowledge point examination pass rate associated with the out-screen content in the subsequent teaching link; The optimization learning of the DRL policy network is implemented by selecting the actions in the action space and combining the reward values fed back by the reward function until the goal of maximizing the long-term teaching effect is achieved.

5. The method of claim 4, wherein, The DRL policy network is trained by using a proximal policy optimization algorithm, and network weights thereof are initialized offline according to historical teaching session data and are continuously fine-tuned and updated online during teaching by using the multi-modal teaching signals of students, so as to adapt to teaching styles of different teachers and cognitive habits of different audiences.

6. The method of claim 1, wherein, The adjustment of the visual presentation attribute comprises: When it is detected that the off-screen attraction index is lower than a first preset threshold, the highlight displacement amount of the off-screen content is increased, and the gloss reflection attribute of the material surface of the off-screen content is enhanced; When it is detected that the cognitive load index is higher than a second preset threshold, the rendering complexity of the off-screen model is reduced or the animation speed of the off-screen model is reduced, and a guiding annotation arrow is triggered to be generated, the guiding annotation arrow pointing to a core component of the off-screen model; When it is detected that the interaction conversion rate is lower than a third preset threshold, a virtual interaction control is dynamically generated beside the off-screen content, the virtual interaction control having a semi-transparent attribute and carrying a feedback metaphor for guiding a user to perform an interactive operation on the off-screen content.

7. The method of claim 1, wherein, Further comprising: identifying an audience identifier and a domain knowledge theme of the current teaching, obtaining an audience cognitive baseline level and an interaction preference parameter from a pre-constructed audience portrait library based on the audience identifier, and loading an associated knowledge point network from a pre-constructed domain knowledge graph based on the domain knowledge theme; the quantitative evaluation value, the current off-screen content attribute parameter, the audience cognitive baseline level and interaction preference parameter, and the associated point and confusion point feature vector of the current knowledge point extracted from the associated knowledge point network are collectively used as a state vector and input to the DRL policy network.

8. The method of claim 1, wherein, Further comprising: triggering root cause diagnosis when the utility evaluation model detects that there is a target knowledge point whose cognitive load index is continuously higher than a diagnostic trigger threshold, forming a root cause diagnosis result and inputting the root cause diagnosis result to the DRL policy network, wherein the root cause diagnosis queries a set of prerequisite basic knowledge points and a set of easily confused knowledge points that cause the target knowledge point to be difficult to understand by accessing the domain knowledge graph and forms the root cause diagnosis result.

9. The method of claim 1, wherein, Further comprising: receiving identification information of a target learning content version of an audience of the current teaching, and retrieving a knowledge mastery graph formed for a previous basic version of the target learning content from a learning history record of the audience; generating an incremental knowledge difference set by comparing differences between the target learning content and the knowledge mastery graph formed for the previous basic version, the incremental knowledge difference set including knowledge points that are newly added, deleted and significantly changed in the target learning content relative to the previous basic version; adjusting a reward function of the DRL policy network based on the incremental knowledge difference set, wherein: for a first knowledge point included in the incremental knowledge difference set, a reward coefficient higher than that of other knowledge points is given to the first knowledge point, so as to guide the optimization strategy to preferentially guarantee the off-screen teaching effect of the first knowledge point; For the knowledge points that have not changed in the target learning content version and the previous base version and have been mastered by the user, a simplified off-screen mode is adopted for demonstration or skipping the detailed off-screen demonstration process of the knowledge points.

10. The method of claim 1, wherein, Also includes: Through a preset cross-system interface, a diagnostic event generated by at least one external teaching auxiliary system is listened to, the diagnostic event contains an event type identifier and an associated knowledge point identifier, which is used to indicate that the learning or operation effect on a specific knowledge point is not expected; In response to the diagnostic event listened to, the target knowledge point identifier to be reinforced is parsed out; According to the target knowledge point identifier, a reinforcement teaching instruction is generated and input to the DRL policy network, so as to optimize the visual presentation attribute of the off-screen content associated with the target knowledge point through the real-time optimization instruction output by the DRL policy network.

11. An out-of-screen teaching feedback optimization apparatus, comprising: Including: The signal acquisition module is used for synchronously collecting multi-modal teaching signals originating from the same time base through the multi-source heterogeneous sensor group integrated in the client during the off-screen teaching process; The quantitative evaluation module is used for extracting spatio-temporal fusion features representing off-screen content attraction, understanding and interaction from the multi-modal teaching signals, and inputting the spatio-temporal fusion features into a pre-trained utility evaluation model to output a quantitative evaluation value of the current off-screen effect, the quantitative evaluation value includes off-screen attraction index, cognitive load index and interaction conversion rate; The strategy analysis module is used for inputting the quantitative evaluation value and the attribute parameter of the current off-screen content as a state vector into a DRL policy network, the DRL policy network is configured with a reward function aiming to maximize the long-term teaching effect, processes the input state vector, and outputs real-time optimization instructions for off-screen content attributes; The optimization execution module is used for dynamically adjusting the visual presentation attribute of the off-screen content being demonstrated based on the optimization instructions, and driving the rendering engine to update the output to the display device in real time to complete the closed-loop feedback optimization.

12. An out-of-screen teaching feedback optimization device, comprising: Including a memory and a processor; The memory is used to store programs; The processor is used to execute the programs to realize the steps of the off-screen teaching feedback optimization method in any one of claims 1-10.

13. A readable storage medium, having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the off-screen teaching feedback optimization method in any one of claims 1-10.

14. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the off-screen teaching feedback optimization method in any one of claims 1-10.

Images