Teaching abnormity identification prompting method and system of teaching robot

Abstract

The invention discloses a teaching abnormity identification prompting method and system of a teaching robot. The method comprises the following steps: acquiring a real-time learning behavior vector of a to-be-identified student in a set time window; inputting the real-time learning behavior vector into a pre-trained abnormal level identification model, and outputting an abnormal level label corresponding to the real-time learning behavior vector; wherein the training of the abnormal level identification model is based on the sorting deviation and the reconstruction deviation of a historical behavior sequence so as to fit the evolution of a learning behavior vector in historical data; according to the abnormity level label, triggering a teaching abnormity prompt of the teaching robot; according to the method, dependence on artificial experience and labels is effectively eliminated, and the description capability of the behavior evolution trajectory is remarkably enhanced; the overall reconstruction deviation of the behavior sequence in the feature space is measured, the sorting deviation of the behavior sequence in the time sequence structure is further evaluated, and the prompt pertinence of anomaly recognition is improved.

Description

Technical Field

[0001] The present invention relates to the field of teaching anomaly recognition and prompting, and specifically to a method and system for teaching anomaly recognition and prompting of a teaching robot. Background Art

[0002] With the in-depth application of artificial intelligence in the field of education, teaching robots have gradually taken on the task of dynamically recognizing students' learning states and providing personalized interventions. Especially in intelligent teaching systems based on big data, how to use behavioral data to real-time recognize students' abnormal learning behaviors has become a key issue in improving learning effects.

[0003] Currently, common anomaly recognition methods are mostly based on rule extraction or traditional classification models, which mainly rely on manually set threshold conditions, static behavioral features, or pre-labeled anomaly tags for training. Such methods have the following problems: Firstly, the construction of supervision labels depends on manual annotation or static rules, which has problems of strong subjectivity and poor scalability. Since learning anomalies in teaching scenarios are often hidden in temporal behavioral changes, it is difficult for humans to accurately identify different levels of anomaly degrees, resulting in strong model training dependence and high label distortion rate, and it is difficult to meet the classification requirements under large-scale data.

[0004] Secondly, existing models are difficult to comprehensively fit the temporal evolution law of behavioral sequences. In actual teaching, students' learning behaviors not only have "spatial deviations" from standard behaviors (i.e., inconsistent behavioral features), but also have "sorting structure differences" (i.e., different behavioral change orders). That is, existing technologies generally ignore the modeling of such sorting deviations, resulting in the recognition model being difficult to capture the implicit evolution trajectory and phased learning regression features in learning behaviors. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides a method and system for teaching anomaly recognition and prompting of a teaching robot, which solves the technical problems proposed in the background art by fitting the sorting deviation and reconstruction deviation in historical data.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A method and system for teaching anomaly recognition and prompting of a teaching robot, comprising the following steps: Obtain the real-time learning behavior vector of the student to be recognized within a set time window; Input the real-time learning behavior vector into a pre-trained anomaly level recognition model, and output the anomaly level label corresponding to the real-time learning behavior vector; Among them, the training of the anomaly level recognition model is based on the sorting deviation and reconstruction deviation of historical behavior sequences to fit the evolution of learning behavior vectors in historical data; Trigger the teaching exception prompt of the teaching robot according to the exception level label.

[0007] In some of these embodiments, the modeling steps of the exception level recognition model include: S1. Obtain the first-stage scores of N target students; S2. Construct a historical behavior sequence according to the first-stage scores of N target students; the historical behavior sequence includes: K high-performance behavior sequences and Q low-performance behavior sequences; S3. Define the K high-performance behavior sequences as an autoencoder training set; S4. Iteratively train the temporal autoencoder model using the autoencoder training set to obtain a temporal behavior reconstruction model that can fit the high-performance behavior sequence; S5. Input the Q low-performance behavior sequences into the temporal behavior reconstruction model and output Q reconstructed behavior sequences; S6. Screen non-extreme abnormal students from the target students according to the Q reconstructed behavior sequences and the corresponding low-performance behavior sequences; S7. Use the high-performance behavior sequences and the low-performance behavior sequences of non-extreme abnormal students to construct a self-supervised training set; S8. Perform supervised training on the temporal modeling model using the self-supervised training set to obtain the exception level recognition model.

[0008] In some of these embodiments, the construction steps of the K high-performance behavior sequences and the Q low-performance behavior sequences include: S2-1. Compare the first-stage scores of the target students with the excellent score threshold and the passing score threshold respectively; If the first-stage score of a target student is higher than the excellent score threshold, mark this target student as a first-class student; If the first-stage score of a target student is lower than the passing score threshold, mark this target student as a second-class student; S2-2. Obtain the second-stage scores and the second-stage learning behavior vectors of the first-class students and the second-class students; S2-3. Compare the second-stage scores with the excellent score threshold and the passing score threshold respectively; If the second-stage score is higher than the excellent score threshold, mark the first-class student or the second-class student to which this second-stage score belongs as a third-class student; If the second-stage score is lower than the passing score threshold, mark the first-class student or the second-class student to which this second-stage score belongs as a fourth-class student; S2-4. Respectively screen out M learning behavior vectors of the first-class students and the second-class students that have been marked as third-class students and fourth-class students from the learning behavior vectors; S2-5. Sort the M learning behavior vectors of the third type of students and the fourth type of students respectively to generate corresponding high-performance behavior sequences and low-performance behavior sequences; S2-6. Repeat S2-1 to S2-5 until K high-performance behavior sequences and Q low-performance behavior sequences are obtained.

[0009] In some embodiments, obtaining the second-stage scores and the second-stage learning behavior vectors of the first type of students and the second type of students includes: S2-2-1. Define a time series window with a fixed length and slidable in the second stage; S2-2-2. Slide the time series window with a set sliding step size, and sequentially obtain the learning behavior vectors of the first type of students and the second type of students within M time series windows in the second stage; S2-2-3. When the time series window slides to the end of the second stage, collect the second-stage scores of the first type of students and the second type of students.

[0010] In some embodiments, generating the corresponding high-performance behavior sequences and low-performance behavior sequences includes: S2-5-1. Obtain the central timestamps of M time series windows; S2-5-2. Encode the central timestamps of M time series windows into time series numbers; S2-5-3. Sort the learning behavior vectors of the third type of students and the fourth type of students respectively according to the time series numbers of M time series windows to obtain the high-performance behavior sequences and the low-performance behavior sequences.

[0011] In some embodiments, screening non-extremely abnormal students from the target students includes: S6-1. Calculate the sorting deviation between the Q reconstructed behavior sequences and the corresponding low-performance behavior sequences to obtain the Q sorting deviations of the fourth type of students; The expression of the sorting deviation is: ; Wherein, is the sorting deviation, M is the number of time windows, is the sorting rank in the low-performance behavior sequence within the t-th time window, is the sorting rank in the reconstructed behavior sequence within the t-th time window; S6-2. Select J sorting deviations lower than the sorting deviation threshold from the Q sorting deviations; S6-3. Mark the fourth type of students to which the J sorting deviations belong as non-extremely abnormal students.

[0012] In some of these embodiments, constructing the self-supervised training set includes: S7-1. Screening out J reconstructed behavior sequences of students marked as non-extremely abnormal; S7-2. Calculating the reconstruction deviation between the J reconstructed behavior sequences and the corresponding low-performance behavior sequences to obtain J reconstruction deviations of non-extremely abnormal students; The expression of the reconstruction deviation is: ; where is the reconstruction deviation, is the learning behavior vector in the low-performance behavior sequence within the t-th time window, is the reconstructed behavior vector in the reconstructed behavior sequence within the t-th time window; is the square of the Euclidean distance between the learning behavior vector in the low-performance behavior sequence and the reconstructed behavior vector in the reconstructed behavior sequence in all feature dimensions; S7-3. Sorting the J reconstruction deviations by size and sequentially assigning J anomaly level labels to the low-performance behavior sequences of the corresponding non-extremely abnormal students; among them, the anomaly level label with the smallest reconstruction deviation is set to 1; S7-4. Uniformly setting the anomaly level label of the high-performance behavior sequences of the K third-category students to 0 to form the corresponding K anomaly level labels; S7-5. Pairing the J anomaly level labels with the corresponding low-performance behavior sequences, and at the same time pairing the K anomaly level labels with the corresponding high-performance behavior sequences to construct a self-supervised training set containing S anomaly supervision samples; where S = J + K.

[0013] Compared with the prior art, the teaching anomaly recognition and prompting method of a teaching robot according to the present invention is based on an autoencoder model trained with historical high-performance behaviors, utilizes its reconstruction ability for low-performance behavior sequences, automatically quantifies the deviation degree between the behavior sequence and the ideal behavior, and generates multi-level anomaly labels accordingly. It effectively gets rid of the dependence on manual experience and labels and has good self-supervised training ability.

[0014] Furthermore, the present invention significantly enhances the ability to depict the behavior evolution trajectory by introducing the dual fitting indexes of sorting deviation and reconstruction deviation. It not only measures the overall reconstruction deviation of the behavior sequence in the feature space, but also further evaluates its sorting deviation in the time sequence structure, and can accurately identify non-extremely abnormal students with stable behaviors but strategy deviations, thereby improving the pertinence of anomaly recognition and prompting.

[0015] In a second aspect, a teaching anomaly recognition and prompting system of a teaching robot according to the present invention includes: A vector acquisition module, configured to acquire the real-time learning behavior vector of the trainee to be recognized within a set time window An abnormal level output module, configured to input the real-time learning behavior vector into a pre-trained abnormal level recognition model, and output an abnormal level label corresponding to the real-time learning behavior vector; Wherein, the training of the abnormal level recognition model is based on the sorting deviation and reconstruction deviation of the historical behavior sequence to fit the evolution of the learning behavior vector in the historical data A prompt module, configured to trigger a teaching anomaly prompt of the teaching robot according to the abnormal level label.

[0016] Compared with the prior art, the beneficial effects of the teaching anomaly recognition and prompt system of a teaching robot according to the present invention are the same as those of the teaching anomaly recognition and prompt method of a teaching robot described above, so details are not described herein again. Description of the Drawings

[0017] Figure 1 is a schematic flowchart of a teaching anomaly recognition and prompt method of a teaching robot according to the present invention; Figure 2 is a schematic flowchart of the modeling process of the abnormal level recognition model described in the present invention; Figure 3 is a schematic flowchart of the marking process of non-extremely abnormal trainees described in the present invention; Figure 4 is a structural block diagram of a teaching anomaly recognition and prompt system of a teaching robot according to the present invention. Detailed Embodiment

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] Please refer to Figure 1 , the present invention provides a teaching anomaly recognition and prompt method for a teaching robot, including the following steps: A1. Acquire the real-time learning behavior vector of the trainee to be recognized within a set time window; The learning behavior vector is a multi-dimensional learning behavior feature combination collected within the current time window, and the feature dimensions include but are not limited to answer response time, page stay duration, click frequency, error rate distribution, and behavior pause duration, etc.

[0020] A2. Input the real-time learning behavior vector into a pre-trained anomaly level recognition model to output the anomaly level label corresponding to the real-time learning behavior vector; Among them, the training of the anomaly level recognition model is based on the sorting deviation and reconstruction deviation of historical behavior sequences to fit the evolution of learning behavior vectors in historical data; The anomaly level recognition model is a temporal classification model trained based on historical high-performance and low-performance student behavior samples, and can output the level labels of behavior sequences under different anomaly levels. The anomaly level label is a discrete value, indicating the degree of deviation between the student's behavior and high-performance behavior. The larger the label value, the higher the degree of deviation. Exemplarily, it includes: level 0 (normal), level 1 (mild anomaly), level 2 (moderate anomaly), level 3 (severe anomaly).

[0021] A3. Trigger the teaching anomaly prompt of the teaching robot according to the anomaly level label.

[0022] For example, when the recognized anomaly level label is 1-3, the teaching robot generates personalized teaching intervention prompt content corresponding to the recognized level according to different level strategies, including learning strategy suggestions, behavior reminders, or teacher assistance requests, etc., for dynamically guiding students to adjust their learning paths and improving learning effects.

[0023] Refer to Figures 2 to 3 , this embodiment also discloses the modeling steps of the anomaly level recognition model, and the modeling steps include: S1. Obtain the first-stage scores of N target students; Exemplarily, the first-stage score can be the final exam score of the target student in the previous semester.

[0024] S2. Construct historical behavior sequences according to the first-stage scores of N target students; the historical behavior sequences include: K high-performance behavior sequences and Q low-performance behavior sequences; S3. Define the K high-performance behavior sequences as the autoencoder training set; S4. Iteratively train the temporal autoencoder model using the autoencoder training set to obtain a temporal behavior reconstruction model that can fit the high-performance behavior sequences; In this embodiment, the temporal autoencoder model can adopt a sequence modeling structure composed of an encoder and a decoder. Specifically: The encoder is used to receive and compress the time series features in the high-performance behavior sequence, and perform a non-linear mapping on the learning behavior vectors at different times to extract their potential temporal feature representations.

[0025] The decoder is used to sequentially reconstruct a sequence of behavior vectors with the same length as the input high-performance behavior sequence based on the temporal feature representations extracted by the encoder.

[0026] Both the encoder and decoder modules can adopt a stacked long short-term memory network (LSTM) or gated recurrent unit (GRU) structure to enhance the ability to model long-term dependencies in the behavior sequence. During the model training process, the high-performance behavior sequence itself is used as the supervision signal, and the model parameters are iteratively optimized by minimizing the reconstruction deviation between the input high-performance behavior sequence and the reconstructed sequence. Finally, the trained temporal behavior reconstruction model can accurately reconstruct the high-performance behavior sequence.

[0027] S5. Input Q low-performance behavior sequences into the temporal behavior reconstruction model, and output Q reconstructed behavior sequences; S6. Screen non-extremely abnormal students from the target students according to the Q reconstructed behavior sequences and the corresponding low-performance behavior sequences; S7. Use the high-performance behavior sequences and the low-performance behavior sequences of non-extremely abnormal students to construct a self-supervised training set; S8. Use the self-supervised training set to perform supervised training on the temporal modeling model to obtain the abnormal level recognition model.

[0028] Specifically, the temporal modeling model can adopt a bidirectional long short-term memory network (BiLSTM) structure to extract the time dependencies in the student behavior sequence; a fully connected classification layer and a Softmax activation function are set at the output end to realize the prediction of the abnormal level label for each behavior sequence. The modeling steps of supervised training can adopt general supervised training steps.

[0029] Exemplarily, the steps for obtaining the high-performance sequence and the low-performance sequence are as follows: S2-1. Compare the first-stage scores of the target students with the excellent score threshold and the passing score threshold respectively; If the first-stage score of the target student is higher than the excellent score threshold, mark the target student as the first type of student, which is used to represent the high-performance student group; If the first-stage score of the target student is lower than the passing score threshold, mark the target student as the second type of student; which is used to represent the low-performance student group.

[0030] S2-2. Obtain the second-stage scores and the second-stage learning behavior vectors of the first type of students and the second type of students; Specifically, the learning behavior vector essentially refers to the combination of a learner's learning behavior characteristics in multiple dimensions within a time window. That is, within each time window, the original feature data of multiple behavior dimensions are extracted from the learner's continuous learning behavior, and corresponding feature standardization or normalization strategies are adopted according to the nature of each type of feature (such as numerical type, proportional type, or categorical type) to be converted into feature vectors under a unified dimension. Finally, the processed multi-dimensional learning behavior characteristics within this time window constitute a learning behavior vector.

[0031] S2-3. Compare the scores of the second stage with the excellent score threshold and the passing score threshold respectively; If the score of the second stage is higher than the excellent score threshold, mark the first type of learners or the second type of learners to which this second-stage score belongs as the third type of learners; The third type of learners includes the following two categories: One category is: samples with excellent historical scores (i.e., belonging to the first type of learners) and whose scores in the second stage are still higher than the excellent score threshold, indicating the continuous consistency of their learning behavior and performance; The other category is: samples with poor historical scores (i.e., belonging to the second type of learners), but whose scores in the second stage have significantly improved to the excellent level, indicating significant progress in their learning in the second stage.

[0032] If the score of the second stage is lower than the passing score threshold, mark the first type of learners or the second type of learners to which this second-stage score belongs as the fourth type of learners; The fourth type of learners also includes the following two situations: One category is: samples with poor historical scores (i.e., belonging to the second type of learners) and whose scores in the second stage are still lower than the passing score threshold, indicating the continuous poor learning effect; The other category is: samples with excellent historical scores (i.e., belonging to the first type of learners), but whose scores in the second stage have significantly dropped to fail, indicating possible abnormal fluctuations or regression behaviors in the current learning stage.

[0033] S2-4. Respectively screen out M learning behavior vectors marked as the third type of learners and the fourth type of learners from the learning behavior vectors of the first type of learners and the second type of learners; S2-5. Sort the M learning behavior vectors of the third type of learners and the fourth type of learners respectively to generate corresponding high-performance behavior sequences and low-performance behavior sequences; S2-6. Repeat the execution of S2-1 to S1-5 until K high-performance behavior sequences and Q low-performance behavior sequences are obtained.

[0034] Furthermore, step S2-2 specifically includes: S2-2-1. Define a time window with a fixed length and that can slide in the second stage; the time window is used to segment and extract the learning behavior data of the trainees.

[0035] S2-2-2. Slide the time window with a set sliding step size, and sequentially obtain the learning behavior vectors of the first type of trainees and the second type of trainees in the second stage within M time windows; The multi-dimensional behavior feature vectors composed of the answer response time, page stay duration, click frequency, error rate distribution, behavior pause duration, etc. within each time window.

[0036] S2-2-3. When the time window slides to the end of the second stage, collect the second-stage scores of the first type of trainees and the second type of trainees; the second-stage scores can be used for the correlation between learning behavior and learning performance. Exemplarily, the second-stage score can be the final semester score of the target trainee.

[0037] Furthermore, step S2-5 specifically includes: S2-5-1. Obtain the central timestamps of the M time windows; used to represent the position of each time window within the second stage.

[0038] S2-5-2. Encode the central timestamps of the M time windows into time sequence numbers, used to clarify the chronological order of each learning behavior vector.

[0039] S2-5-3. According to the time sequence numbers of the M time windows, sort the learning behavior vectors of the third type of trainees and the fourth type of trainees respectively, to obtain the high-performance behavior sequence and the low-performance behavior sequence.

[0040] In this embodiment, the set formed by arranging multiple time windows in sequence constitutes a complete behavior sequence. The high-performance behavior sequence and the low-performance behavior sequence respectively refer to the ordered vector sequences formed by splicing the learning behavior vectors within multiple consecutive time windows according to the time sequence numbers, used to describe the dynamic characteristic trajectory of the learning behavior of trainees evolving over time within the second stage. This behavior sequence not only retains the multi-dimensional attributes of the behavior characteristics but also reflects the time-dependent relationship.

[0041] Exemplarily, the screening steps for non-extremely abnormal trainees include: S6-1. Calculate the sorting deviation between the Q reconstructed behavior sequences and the corresponding low-performance behavior sequences to obtain the Q sorting deviations of the fourth type of trainees; The expression for the sorting deviation is: ; Where, It is the sorting deviation, which is used to measure the difference degree of the sorting structure of the reconstructed behavior sequence and the low-performance behavior sequence in the learning behavior vector within the time window; M is the number of time windows. is the sorting rank in the low-performance behavior sequence within the t-th time window. is the sorting rank in the reconstructed behavior sequence within the t-th time window. Specifically, the way to obtain the sorting rank is as follows: First, for the learning behavior vectors in each time window calculate their Euclidean norms respectively: ; Among them, is the learning behavior vector in the low-performance behavior sequence within the t-th time window, D is the number of feature dimensions of the learning behavior vector, is the eigenvalue of the d-th dimension in the t-th time window; Then sort all the Euclidean norms in the sequence to obtain their sorting numbers (1 - M) in the sequence, and this number is the sorting rank ; The same method is used to calculate for the reconstructed behavior sequence.

[0042] Therefore, although the whole sequence is a whole, the sorting operation is carried out on the basis of the vector norm length in each time window. Therefore, the sorting rank reflects the intensity or relative position of the behavior characteristics at each time step.

[0043] S6-2. Select J sorting deviations that are lower than the sorting deviation threshold from the Q sorting deviations; S6-3. Mark the J fourth-category students to whom the sorting deviations belong as non-extremely abnormal students; Specifically, the non-extremely abnormal students refer to those whose second-stage scores have not reached the passing grade, but their learning behaviors show relatively stable and orderly evolutionary characteristics in the time dimension. Although there are certain differences in the learning behavior patterns of this type of students from the high-performance behaviors, their behavior sequences have relatively high temporal consistency and continuity in structure, indicating that they may have certain learning motivations or goal orientations in the learning process, and only due to learning path selection, strategy deviations or phased adaptation problems, their current performance fails to meet expectations.

[0044] Exemplarily, the construction steps of the self-supervised training set include: S7-1. Select J reconstructed behavior sequences of students who have been marked as non-extremely abnormal students; S7-2. Calculate the reconstruction deviations between the J reconstructed behavior sequences and the corresponding low-performance behavior sequences to obtain J reconstruction deviations of non-extremely abnormal students; the reconstruction deviations are used to measure the degree of deviation between the behavior trajectory of non-extremely abnormal students and the high-performance behavior model.

[0045] The expression of the reconstruction deviation is: ; in, Reconstruction bias is used to measure the difference in behavior patterns between low-performance behavior sequences and reconstructed behavior sequences. is the learning behavior vector in the low-performance behavior sequence in the t-th time window, is the reconstructed behavior vector in the reconstructed behavior sequence within the t-th time window; is the square of the Euclidean distance between the learned behavior vector in the low-performance behavior sequence and the reconstructed behavior vector in the reconstructed behavior sequence in all feature dimensions, indicating the behavior deviation in the corresponding time window; S7-3, sorting the J reconstruction deviations by size, and sequentially assigning J abnormality level labels to the low-performance behavior sequences of the non-extreme abnormal students; wherein the abnormality level label of the minimum reconstruction deviation is set to 1; That is to say, the abnormal level label is positively correlated with the reconstruction deviation in value, and is used to reflect the degree of deviation between the behavior sequence and the high-performance behavior model. Specifically, multiple abnormal level division intervals are preset, and labels of corresponding levels are assigned according to the interval position of the reconstruction deviation corresponding to each low-performance behavior sequence; illustratively, it can be divided into multiple discrete level labels, such as mild abnormality (1), moderate abnormality (2), and severe abnormality (3).

[0046] S7-4, uniformly setting the abnormal level labels of the K high-performance behavior sequences of the third category students to 0, to form corresponding K abnormal level labels; Specifically, the high-performance behavior sequences of the K third-category students have been used to train the temporal behavior reconstruction model in the previous steps, and their high-performance behavior sequences can be regarded as ideal behaviors. Therefore, in the abnormal level recognition task, their corresponding labels are directly set to 0, indicating that there is no deviation in this type of behavior.

[0047] S7-5, pair J abnormal level labels with corresponding low-performance behavior sequences, and pair K abnormal level labels with corresponding high-performance behavior sequences, to construct a self-supervised training set containing S abnormal supervision samples; where S=J+K; The constructed abnormal level label system is composed of normal samples (label level is 0) and abnormal samples with different deviation degrees (label levels are 1, 2, and 3), forming a supervised classification label structure with hierarchical discrimination ability.

[0048] Specifically, each sample in the self-supervised training set consists of a "behavior sequence" and a "corresponding abnormal level label". The behavior sequence is a standardized vector sequence after splicing time windows. The value range of the abnormal level label is {0, 1, 2, 3}, representing normal, mild, moderate, and severe abnormal levels respectively.

[0049] A teaching anomaly recognition and prompting method for a teaching robot in this embodiment breaks through two major technical bottlenecks existing in existing teaching behavior recognition methods: First, traditional methods highly rely on manual annotation of abnormal behavior labels and are difficult to adapt to rapid deployment and dynamic update in large-scale teaching scenarios; Second, existing recognition models mostly perform behavior fitting from a single dimension and are difficult to capture the deviation trends of learning behaviors in both the time structure and the feature space simultaneously. By introducing self-supervised label generation and a dual-index fitting of ranking deviation - reconstruction deviation, the recognition accuracy of anomaly recognition is significantly improved.

[0050] The embodiment of the present invention also provides a teaching anomaly recognition and prompting system for a teaching robot. This system is used to implement the above method embodiment, and those that have been described will not be repeated here. The following terms such as "module", "unit", "sub-unit", etc. can be a combination of software and / or hardware that can achieve a predetermined function. Although the system described in the following embodiments is preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0051] As Figure 4 shown, Figure 4 is a structural block diagram of a teaching anomaly recognition and prompting system for a teaching robot of the present invention. This system includes: A vector acquisition module, which is used to acquire the real-time learning behavior vector of the trainee to be recognized within a set time window An abnormal level output module, which is used to input the real-time learning behavior vector into a pre-trained abnormal level recognition model and output the abnormal level label corresponding to the real-time learning behavior vector; Among them, the training of the abnormal level recognition model is based on the ranking deviation and reconstruction deviation of historical behavior sequences to fit the evolution of learning behavior vectors in historical data A prompting module, which is used to trigger the teaching anomaly prompt of the teaching robot according to the abnormal level label.

[0052] In the above system, a real-time learning behavior vector is obtained through the vector acquisition module; an anomaly level label is obtained through the anomaly level output module; and a teaching anomaly prompt of the teaching robot is triggered through the prompt module, thereby solving the problem of ignoring sorting bias.

[0053] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as infrared, wireless, microwave, etc.).

[0054] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present application, and all should be covered by the protection scope of the present application.

Claims

1. A teaching abnormality recognition and prompting method for a teaching robot, characterized in that: include: Obtain the real-time learning behavior vector of the student to be identified within the set time window; Inputting the real-time learning behavior vector into a pre-trained abnormality level recognition model, and outputting an abnormality level label corresponding to the real-time learning behavior vector; The training of the abnormal level recognition model is based on the sorting deviation and reconstruction deviation of the historical behavior sequence to fit the evolution of the learning behavior vector in the historical data; According to the abnormal level label, the teaching robot's teaching abnormality prompt is triggered.

2. The teaching abnormality recognition and prompting method of a teaching robot according to claim 1, characterized in that: The modeling steps of the abnormal level recognition model include: S1. Obtain the first-stage scores of N target students; S2. Constructing a historical behavior sequence based on the first-stage scores of the N target students; the historical behavior sequence includes: K high-performance behavior sequences and Q low-performance behavior sequences; S3, defining the K high-performance behavior sequences as autoencoding training sets; S4, iteratively train the temporal autoencoder model using the autoencoder training set to obtain a temporal behavior reconstruction model that can fit high-performance behavior sequences; S5, inputting Q low-performance behavior sequences into the temporal behavior reconstruction model, and outputting Q reconstructed behavior sequences; S6. Screen non-extremely abnormal students from the target students based on the Q reconstructed behavior sequences and the corresponding low-performance behavior sequences; S7. Use high-performance behavior sequences and low-performance behavior sequences of non-extremely abnormal students to construct a self-supervised training set; S8. Use the self-supervised training set to perform supervised training on the time series modeling model to obtain the abnormality level recognition model.

3. The teaching abnormality identification and prompting method of a teaching robot according to claim 2, characterized in that: The steps of constructing K high-performance behavior sequences and Q low-performance behavior sequences include: S2-1. Compare the first-stage scores of the target students with the excellent score threshold and the passing score threshold respectively; If the target student's first-stage score is higher than the excellent score threshold, the target student is marked as a first-category student; If the target student's first-stage score is below the passing score threshold, the target student is marked as a second-category student; S2-2, obtaining the second-stage scores and the second-stage learning behavior vectors of the first and second-stage students; S2-3, comparing the second stage scores with the excellent score threshold and the passing score threshold respectively; If the second-stage score is higher than the excellent score threshold, the first-category student or the second-category student to whom the second-stage score belongs is marked as a third-category student; If the second stage score is lower than the passing score threshold, the first category student or the second category student to which the second stage score belongs is marked as a fourth category student; S2-4, selecting M learning behavior vectors marked as third-category students and fourth-category students from the learning behavior vectors of the first-category students and the second-category students respectively; S2-5, sorting the M learning behavior vectors of the third type of students and the fourth type of students respectively, generating corresponding high performance behavior sequences and low performance behavior sequences; S2-6. Repeat S2-1 to S2-5 until K high-performance behavior sequences and Q low-performance behavior sequences are obtained.

4. The teaching abnormality recognition and prompting method of a teaching robot according to claim 3 is characterized in that: Obtaining the second-stage scores and the second-stage learning behavior vectors of the first- and second-stage students includes: S2-2-1, in the second stage, define a fixed-length and slidable timing window; S2-2-2, sliding the time sequence window with a set sliding step length, and sequentially obtaining the learning behavior vectors of the first type of students and the second type of students in the second stage in M ​​time sequence windows; S2-2-3. When the timing window slides to the end of the second stage, the second stage scores of the first and second category students are collected.

5. The teaching abnormality recognition and prompting method of a teaching robot according to claim 3 is characterized in that: Generating corresponding high performance behavior sequences and low performance behavior sequences includes: S2-5-1. Obtain the central timestamp of M time series windows; S2-5-2, encoding the central timestamps of the M time series windows into time series numbers; S2-5-3. According to the timing numbers of the M timing windows, the learning behavior vectors of the third category of students and the fourth category of students are sorted respectively to obtain the high-performance behavior sequence and the low-performance behavior sequence.

6. The teaching abnormality recognition and prompting method of a teaching robot according to claim 5, characterized in that: Screening non-extreme abnormal students from the target students includes: S6-1, calculating the ranking deviations between the Q reconstructed behavior sequences and the corresponding low-performance behavior sequences, and obtaining Q ranking deviations of the fourth type of students; The expression of the sorting deviation is: ； in, is the sorting deviation, M is the number of time windows, is the ranking rank of the low-performance behavior sequence in the t-th time window, Reconstruct the sorting rank in the behavior sequence within the t-th time window; S6-2, selecting J sorting deviations below the sorting deviation threshold from the Q sorting deviations; S6-3. Mark the fourth category of students to which J ranking deviations belong as non-extreme abnormal students.

7. The teaching abnormality recognition and prompting method of a teaching robot according to claim 6, characterized in that: Constructing a self-supervised training set includes: S7-1, select J reconstructed behavior sequences of students marked as non-extreme abnormal students; S7-2, calculating the reconstruction deviations between the J reconstructed behavior sequences and the corresponding low-performance behavior sequences, and obtaining J reconstruction deviations of non-extreme abnormal students; The expression of the reconstruction deviation is: ； in, To reconstruct the deviation, is the learning behavior vector in the low-performance behavior sequence in the t-th time window, is the reconstructed behavior vector in the reconstructed behavior sequence within the t-th time window; is the square of the Euclidean distance between the learned behavior vector in the low-performance behavior sequence and the reconstructed behavior vector in the reconstructed behavior sequence in all feature dimensions; S7-3, sorting the J reconstruction deviations by size, and sequentially assigning J abnormality level labels to the low-performance behavior sequences of the non-extreme abnormal students; wherein the abnormality level label of the minimum reconstruction deviation is set to 1; S7-4, uniformly setting the abnormal level labels of the K high-performance behavior sequences of the third category students to 0, to form corresponding K abnormal level labels; S7-5. Pair J anomaly level labels with corresponding low-performance behavior sequences, and pair K anomaly level labels with corresponding high-performance behavior sequences to construct a self-supervised training set containing S anomaly supervision samples; where S=J+K.

8. A teaching abnormality recognition and prompting system for a teaching robot, comprising: Vector acquisition module, used to obtain the real-time learning behavior vector of the student to be identified within the set time window An abnormality level output module, used for inputting the real-time learning behavior vector into a pre-trained abnormality level recognition model, and outputting an abnormality level label corresponding to the real-time learning behavior vector; The training of the abnormal level recognition model is based on the sorting deviation and reconstruction deviation of the historical behavior sequence to fit the evolution of the learning behavior vector in the historical data. The prompt module is used to trigger the teaching robot's teaching exception prompt according to the exception level label.

Images