A learning evaluation method and system based on artificial intelligence

Abstract

The present invention discloses a learning evaluation method and system based on artificial intelligence, which relates to the technical field of learning data processing. The method comprises: obtaining learning state features according to the learner's facial image; adopting a federated learning architecture to construct a local prediction model and a global prediction model according to the learning state features; inputting the learning state features obtained in real time into the pre-trained local prediction model to obtain a learning state prediction result. The present invention inputs the acquired learning state feature data set into the local prediction model obtained by federated learning, predicts the learner's learning state in real time from multiple dimensions, improves the accuracy of the learner's learning state prediction, and thus provides reliable data support for educational decision-making. Education managers can provide timely guidance and assistance based on accurate learning state prediction results, optimize educational policies and teaching models, and improve education quality.

Description

Technical Field

[0001] The present invention relates to the technical field of learning data processing, and in particular to a learning evaluation method and system based on artificial intelligence. Background Art

[0002] At present, education is a common concern of the public. In today's society where the learning environment and teaching quality are gradually improving, learners need to maintain their own learning state to achieve good learning results. The learning state includes the learner's emotions, motivation, and attention, which will affect the learner's ability to absorb knowledge. A good learning state will enable learners to better understand the content of learning. In recent years, artificial intelligence has attracted widespread attention from the public, and some people have gradually used artificial intelligence in the evaluation of learners' learning.

[0003] The existing artificial intelligence evaluation of learners is to analyze the learners' answers at a certain stage, such as analyzing the completion data of a certain homework or a certain exam, and predicting the learners' learning status at that stage. This evaluation method has the following problems: evaluating the learners' learning status based on a single answer has certain errors and cannot accurately predict the learners' learning status; secondly, the predicted learning status spans too long and cannot be adjusted when the learners' learning status is poor, and the real-time nature of the prediction is low. Summary of the invention

[0004] The purpose of the present invention is to provide a learning evaluation method and system based on artificial intelligence. The present invention inputs the acquired learning status feature data set into the local prediction model obtained through federated learning, predicts the learner's learning status in real time from multiple dimensions, and improves the accuracy of predicting the learner's learning status.

[0005] The purpose of the present invention is achieved by the following technical means:

[0006] In a first aspect, the present invention provides a learning evaluation method based on artificial intelligence, comprising:

[0007] According to the learner's facial image, learning state features are obtained;

[0008] Adopting the federated learning architecture, local prediction models and global prediction models are constructed based on the learning state characteristics;

[0009] Inputting the learning state features acquired in real time into the pre-trained local prediction model to obtain a learning state prediction result; the learning state features acquired in real time are acquired through the learner's facial image, answering process data, online learning behavior data, online discussion behavior data and / or interactive audio data;

[0010] Generate a personalized learning suggestion report based on the learning status prediction result; the personalized learning suggestion includes a real-time learning status suggestion report and a learning habit suggestion report.

[0011] Preferably, the acquiring of learning status features according to the learner's facial image includes:

[0012] capturing the facial image of the learner in real time;

[0013] Using a facial detection algorithm, locating a facial region in the facial image;

[0014] Performing cropping and standardization processing on the facial region to obtain a facial region image;

[0015] Using a facial feature point detection algorithm to locate feature points in the facial region image;

[0016] Obtaining facial posture data according to the feature points;

[0017] Performing time series analysis on the facial posture data to obtain the learning state characteristics.

[0018] Preferably, the performing time series analysis on the facial posture data to obtain the learning state features comprises:

[0019] Dividing the facial gesture data into windows of fixed length;

[0020] Obtaining statistical indicators of the facial posture data of each window;

[0021] According to the statistical indicators, obtaining the learner's concentration score, fatigue score and emotional state score;

[0022] The concentration score, the fatigue score and the emotional state score are calculated to obtain the learning state characteristics.

[0023] Preferably, the obtaining of statistical indicators of the facial posture data of each window includes:

[0024] The facial posture data includes: head state data, eye state data and mouth state data;

[0025] Calculating the standard deviation of the head angle data according to the head state data;

[0026] Calculating the number of blinks per unit time according to the eye state data;

[0027] According to the mouth state data, the average value of the mouth corner arc data is calculated.

[0028] Preferably, the method of adopting the federated learning architecture and constructing the local prediction model and the global prediction model according to the learning state features further includes: obtaining the time correlation features of the learning state features; specifically including:

[0029] Sampling the learning state feature at a first time interval to obtain a first state time series;

[0030] Acquire a first time correlation feature according to the first state time series;

[0031] Sampling the learning state feature at a second time interval to obtain a second state time series;

[0032] Acquire a second time correlation feature according to the second state time series;

[0033] fusing the first time correlation feature and the second time correlation feature to obtain the time correlation feature;

[0034] The time correlation feature is expressed as:

[0035]

[0036] in, is the time-dependent feature, is the first time correlation feature, is the second time correlation feature, is the weight coefficient of the first time correlation feature, is the weight coefficient of the second time correlation feature, is the Hadamard product.

[0037] Preferably, before inputting the learning state features acquired in real time into the pre-trained local prediction model to obtain the learning state prediction result, the method further includes: updating the local prediction model and the global prediction model according to the learning state features acquired in real time; specifically including:

[0038] Inputting the learning state features acquired in real time into the local prediction model;

[0039] Updating the parameters of the local prediction model by gradient descent method;

[0040] Fusing all updated parameters of the local prediction models, and updating the global prediction model using a federated averaging algorithm;

[0041] Updating the local prediction model according to the updated model parameters of the global prediction model;

[0042] The federated averaging algorithm is expressed as:

[0043]

[0044] in, are the model parameters of the updated global prediction model, For the The model parameters of the local prediction model, is the number of local prediction models.

[0045] Preferably, updating the local prediction model according to the updated model parameters of the global prediction model includes:

[0046] Screening learners' personalized characteristics through the superadditivity criterion of cooperative games;

[0047] Constructing a local feature prediction matrix according to the personalized features;

[0048] Outputting and splicing the local feature prediction matrix with the updated global prediction model to obtain the local prediction model;

[0049] Training the local prediction model through the personalized features, and updating model parameters of the local prediction model;

[0050] The local feature prediction matrix is ​​expressed as:

[0051]

[0052] in, is the output of the local feature prediction matrix, is the weight matrix, is the personalized feature vector, is the bias vector, is the activation function.

[0053] Preferably, the training of the local prediction model by using the personalized features and updating the model parameters of the local prediction model include:

[0054] In the forward propagation, the local prediction model is used to generate a prediction result;

[0055] Calculate the error between the prediction result and the target feature through a loss function;

[0056] In back propagation, a weight matrix and a bias vector gradient are calculated, and model parameters of the local prediction model are updated by the weight matrix and the bias vector gradient;

[0057] The loss function is expressed as:

[0058]

[0059] in, is the number of samples, For the The prediction results of samples are For the The target value for each sample.

[0060] In a second aspect, the present invention provides a learning evaluation system based on artificial intelligence, comprising:

[0061] Learning state feature acquisition module: used to acquire learning state features based on the learner's facial image;

[0062] Prediction model building module: used to adopt the federated learning architecture to build local prediction models and global prediction models based on learning state characteristics;

[0063] Prediction model updating module: used to update the local prediction model and the global prediction model according to the learning state characteristics acquired in real time;

[0064] Learning state prediction module: used to input the learning state features acquired in real time into the pre-trained local prediction model to obtain the learning state prediction result.

[0065] In a third aspect, the present invention provides an electronic device comprising a processor and a memory, wherein the memory is used to store computer program code, and the computer program code comprises computer instructions. When the processor executes the computer instructions, the electronic device executes the above-mentioned artificial intelligence-based learning evaluation method.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] The present invention inputs the acquired learning status feature data set into the local prediction model obtained through federated learning, and predicts the learner's learning status in real time from multiple dimensions, thereby improving the accuracy of the learner's learning status prediction, thereby providing reliable data support for educational decision-making. Education managers can provide timely guidance and assistance based on accurate learning status prediction results, optimize educational policies and teaching models, and improve education quality;

[0068] The present invention obtains the first time correlation feature and the second time correlation feature of the learning state feature, and fuses them to obtain the time correlation feature of the learning state feature, thereby improving the coverage of the learning state feature on the time dimension and significantly improving the prediction accuracy and system responsiveness.

[0069] The present invention constructs and continuously iterates and optimizes the low prediction model and the global prediction model through a federated learning architecture, thereby improving the security of model use and the overall performance of prediction;

[0070] The present invention incorporates concentration, fatigue level and emotional state into the learning state characteristics as independent dimensions, which is conducive to concisely expressing the learning state characteristics and processing and analyzing the learning state characteristics;

[0071] The present invention constructs a local feature prediction matrix through the learner's personalized features, and obtains a local prediction model by splicing the output with the global prediction model, thereby improving the efficiency of Non-IID data training and the accuracy of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0073] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0074] Figure 1 A flowchart of a learning evaluation method based on artificial intelligence provided in this embodiment;

[0075] Figure 2 A schematic diagram of a process for obtaining learning status features based on a learner's facial image provided in this embodiment;

[0076] Figure 3 A schematic diagram of a process for performing time series analysis on facial posture data to obtain learning state features provided in this embodiment;

[0077] Figure 4 A schematic diagram of a process for obtaining statistical indicators of facial posture data of each window provided in this embodiment;

[0078] Figure 5 A schematic diagram of a process for obtaining time correlation features of learning state features provided in this embodiment;

[0079] Figure 6 A schematic diagram of a process for updating a local prediction model and a global prediction model according to learning state features acquired in real time provided in this embodiment;

[0080] Figure 7 A schematic diagram of a process for updating a local prediction model according to model parameters of an updated global prediction model provided in this embodiment;

[0081] Figure 8A schematic diagram of a process for training the prediction model of this example through personalized features and updating the model parameters of the local prediction model provided in this embodiment;

[0082] Fig. 9 A schematic diagram of the structure of an artificial intelligence-based learning evaluation system provided in this embodiment;

[0083] Fig.10 A schematic diagram of the structure of an electronic device provided in this embodiment. DETAILED DESCRIPTION

[0084] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0085] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0086] In addition, the descriptions of "first", "second", etc. in the present invention are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0087] At present, education is a common concern of the public. In today's society where the learning environment and teaching quality are gradually improving, learners need to maintain their own learning state to achieve good learning results. The learning state includes the learner's emotions, motivation, and attention, which will affect the learner's ability to absorb knowledge. A good learning state will enable learners to better understand the content of learning. In recent years, artificial intelligence has attracted widespread attention from the public, and some people have gradually used artificial intelligence in the evaluation of learners' learning.

[0088] The existing artificial intelligence evaluation of learners is to analyze the learners' answers at a certain stage, such as analyzing the completion data of a certain homework or a certain exam, and predicting the learners' learning status at that stage. This evaluation method has the following problems: evaluating the learners' learning status based on a single answer has certain errors and cannot accurately predict the learners' learning status; secondly, the predicted learning status spans too long and cannot be adjusted when the learners' learning status is poor, and the real-time nature of the prediction is low.

[0089] The present invention inputs the acquired learning status feature data set into the local prediction model obtained through federated learning, and predicts the learner's learning status in real time from multiple dimensions, thereby improving the accuracy of predicting the learner's learning status; thereby providing reliable data support for educational decision-making, and educational managers can provide timely guidance and assistance based on accurate learning status prediction results, optimize educational policies and teaching models, and improve education quality;

[0090] The present invention obtains the first time correlation feature and the second time correlation feature of the learning state feature, and fuses them to obtain the time correlation feature of the learning state feature, thereby improving the coverage of the learning state feature on the time dimension and significantly improving the prediction accuracy and system responsiveness.

[0091] The present invention constructs and continuously iterates and optimizes the low prediction model and the global prediction model through a federated learning architecture, thereby improving the security of model use and the overall performance of prediction;

[0092] The present invention incorporates concentration, fatigue level and emotional state into the learning state characteristics as independent dimensions, which is conducive to concisely expressing the learning state characteristics and processing and analyzing the learning state characteristics;

[0093] The present invention constructs a local feature prediction matrix through the learner's personalized features, and obtains a local prediction model by splicing the output with the global prediction model, thereby improving the efficiency of Non-IID data training and the accuracy of prediction.

[0094] Example 1

[0095] A learning evaluation method based on artificial intelligence, such as Figure 1 As shown, including:

[0096] S1, obtain learning state features based on the learner’s facial image;

[0097] The facial image is a picture containing the learner's face during the learning process. The format of the facial image can be, but is not limited to, JPEG, BMP, PNG, and ICO. The learning state feature is a data indicator that reflects the learner's attention state during the learning process, such as attention concentration and attention stability. By capturing and analyzing the learner's facial image, the learning state feature is generated to provide a data basis for the subsequent analysis of the learner's learning state.

[0098] S3, adopts a federated learning architecture to build local prediction models and global prediction models based on learning state characteristics;

[0099] The federated learning architecture is a distributed machine learning framework that builds a neural network model by allowing multiple local devices to work together to train a global model. By building local and global prediction models through the federated learning architecture, the overall performance of local and global prediction models is improved while protecting the privacy of student data, promoting the intelligent development of the entire education platform.

[0100] S5, inputting the learning state features acquired in real time into a pre-trained local prediction model to obtain a learning state prediction result. The learning state features acquired in real time are acquired through the learner's facial image, answering process data, online learning behavior data, online discussion behavior data and / or interactive audio data.

[0101] It should be noted that S1 obtains learning state features based on the learner's facial image, and its acquisition process is only through facial images, rather than integrating all possible other perceptual data sets. This is because facial expressions and reactions are direct indicators of emotions and cognitive states, and can provide immediate and accurate information about the individual's learning state. At the same time, by limiting the type of data input, the possibility of processing irrelevant and redundant information can be reduced, thereby ensuring the accuracy and efficiency of model training. In the early stage of model training, only obtaining learning state features through facial images can improve the focus and accuracy of the model when building and training the model. At the same time, in the subsequent application of the model, the learning state features are obtained through other perceptual pathways, achieving good flexibility and scalability of the local prediction model.

[0102] The learning state features acquired in real time can be acquired in the same way as S1, based on the learner's facial image, or the learner's learning state features can be acquired in real time through other perception pathways. Other perception pathways include answering process data, online learning behavior data, online discussion behavior data, and interactive audio data. After the collected learning state features are denoised and standardized, the feature vectors from different perception pathways are spliced ​​to form a comprehensive learning state feature, which is then converted into a format and data structure that the model can process and input into the local prediction model. By acquiring learning state features from multiple pathways, when a data source is unavailable or of poor quality, other data sources can provide supplementary information to increase the robustness of the evaluation system.

[0103] In this embodiment, by inputting the acquired learning status features into the local prediction model obtained through federated learning, the learner's learning status is predicted in real time from multiple dimensions, thereby improving the accuracy of the learner's learning status prediction, thereby providing reliable data support for educational decision-making. Education managers can provide timely guidance and assistance based on accurate learning status prediction results, optimize educational policies and teaching models, and improve education quality.

[0104] S6. Generate a personalized learning suggestion report based on the learning status prediction result; the personalized learning suggestion includes a real-time learning status suggestion report and a learning habit suggestion report.

[0105] After obtaining the learning status prediction results, this embodiment, on the one hand, classifies the real-time learning status of the students according to the learning status prediction results in order to generate a real-time learning status recommendation report; in a preferred embodiment, the learning status includes three evaluation dimensions: concentration, fatigue level and emotional state. The learning status is stored in a numerical manner. If the value of the learning status is lower than the preset state threshold, the real-time learning status of the student is input into the learning suggestion rule library, and the real-time recommendation data is obtained through the learning suggestion rule library. A real-time learning status recommendation report is generated based on the implementation recommendation data and pushed to the user. Among them, the mapping relationship between the learning status and the implementation recommendation data is stored in the learning suggestion rule library.

[0106] Secondly, in order to generate a study habit recommendation report, the collected learning status data is analyzed within the set time window, and the data below the preset status threshold is marked and classified to obtain the reasons for the decline in students' learning status, thereby providing optimization suggestions for mid- and long-term study habits, such as adjusting study time planning, optimizing learning methods, etc.

[0107] This embodiment uses a dual feedback mechanism to balance immediate adjustment with long-term optimization, which not only solves students' problems in real-time learning, but also provides mid- to long-term improvement suggestions, thereby comprehensively improving students' learning experience and results.

[0108] In some embodiments, S1, based on the learner's facial image, obtain learning state features, such as Figure 2 As shown, including:

[0109] S11, real-time capture of learner’s facial image;

[0110] Learners’ facial images can be captured by deploying high-definition cameras in the learning environment to ensure that the captured images have a sufficiently high resolution. The resolution of the images is usually above 1080p. High-resolution facial images can clearly capture learners’ facial features. High-definition cameras should be able to continuously record in real time to monitor facial changes throughout the learning process.

[0111] S12, using a face detection algorithm to locate a face region in the face image;

[0112] The face detection algorithm can use OpenCV's Haar cascade classifier or Dlib library. OpenCV's Haar cascade classifier is a target algorithm based on Haar features and cascade structure. It extracts Haar features and uses the algorithm for training to generate a cascade structure, thereby achieving efficient and accurate face detection. Dlib library is an open source machine learning library that contains a variety of computer vision and machine learning algorithms. The face detection function of Dlib library can easily complete the task of locating the position of the face from the picture.

[0113] S13, cropping and standardizing the facial region to obtain a facial region image;

[0114] After using the facial detection algorithm to obtain the facial region of the facial image, the facial region is cropped to remove the interference of background debris to ensure that the processed image pixels are only related to the learner's face. The cropped facial image is standardized so that the obtained facial region image has a consistent size and resolution, such as unifying the facial region image to 64×64 or 128×128 pixels, to facilitate subsequent processing and analysis.

[0115] S14, using a facial feature point detection algorithm to locate feature points in the facial region image;

[0116] The feature points of the facial region image are points with unique shapes or located in special positions on the face, including eyes, nose, mouth, eyebrows, etc. The facial feature point detection algorithm can use Dlib's 68-point feature detection method, which uses calibrated face images as training sets and obtains a model that can find the feature points on the face through training.

[0117] S15, obtaining facial posture data according to the feature points;

[0118] Facial posture data indirectly reflects the learning status of learners. Facial posture data can be obtained by analyzing feature points. Specifically, the head state data is obtained by analyzing the pitch angle, yaw angle and roll angle of the head according to the feature points; the eye position change is analyzed to identify the blinking frequency to obtain the eye state data; the mouth corner upward angle and opening and closing frequency are analyzed to obtain the mouth state data.

[0119] S16, performing time series analysis on the facial posture data to obtain learning state features.

[0120] During the entire learning process, facial posture data is continuously recorded, and by using data processing and analysis methods, such as sliding window method or statistical method, the facial posture data is quantified to obtain learning state characteristics.

[0121] In some embodiments, S16, time series analysis is performed on the facial posture data to obtain learning state features, such as Figure 3 As shown, including:

[0122] S161, dividing the facial posture data into windows of fixed length;

[0123] S162, obtaining statistical indicators of facial posture data of each window;

[0124] Specifically, the facial gesture data is divided into windows of fixed length according to time, the fixed length window is set to 30 seconds, and the facial gesture data of each time window is obtained after the division is completed. In some embodiments, S162, the statistical indicators of the facial gesture data of each window are obtained, such as Figure 4 As shown, including:

[0125] S1621, calculating the standard deviation of the head angle data according to the head state data;

[0126] S1622, calculating the number of blinks per unit time according to the eye state data;

[0127] S1623, calculating the average value of the mouth corner curvature data according to the mouth state data.

[0128] S163, obtaining the learner's concentration score, fatigue score, and emotional state score based on the statistical indicators;

[0129] By analyzing facial gesture data, the user's concentration, fatigue level and emotional state can be inferred. Specifically, smaller changes in head angles usually indicate that the user is in a state of high concentration. In addition, less frequent eye blinking and highly open eyes are also signs of concentration. On the contrary, a higher eye blinking frequency and prolonged eye opening and closing, as well as frequent up and down head tilts may indicate that the user is in a state of fatigue, especially an increased frequency of yawning is usually a sign of sleepiness. In terms of emotional state, a higher angle of the corners of the mouth often reflects that the user is in a positive emotion, while wide-open eyes may indicate surprise or other strong emotions. Therefore, through the statistical indicators of these facial gesture data, the user's psychological and physiological state can be comprehensively evaluated.

[0130] The random forest model can be used to obtain the user's concentration, fatigue and emotional state through the statistical indicators of facial posture data. Random forest reduces the overfitting and bias of a single model by introducing diversity. Each decision tree uses different samples and features during training. Combining the results of multiple decision trees can improve the accuracy and robustness of the overall prediction, which is particularly effective for dealing with noise and outliers in facial data. In addition, there is an obvious nonlinear relationship between facial posture data and concentration, fatigue and emotional state. Random forest combines multiple decision trees through ensemble learning methods to effectively capture these complex relationships, thereby improving the recognition ability of the model.

[0131] S164, calculating the concentration score, fatigue score and emotional state score to obtain learning state characteristics.

[0132] Specifically, the learning state feature is represented as a three-dimensional feature vector, where each dimension corresponds to attention (Attention, (A)), fatigue (Fatigue, (F)), and emotion (Emotion, (E)). The learning state feature is represented as:

[0133]

[0134] in, is the learning state feature, To rate the concentration, For the fatigue level rating, Rating of emotional state, concentration , fatigue rating and emotional state The scores are all values ​​between 0 and 1, indicating the intensity of the corresponding state.

[0135] In this embodiment, concentration, fatigue level and emotional state are incorporated into the learning status characteristics as independent dimensions. Through the feature vector representation method, the student's learning status can be clearly and comprehensively described, which not only concisely expresses the learning status characteristics, but also facilitates subsequent complex processing and analysis.

[0136] In some embodiments, S3, before constructing a local prediction model and a global prediction model based on learning state features using a federated learning architecture, it also includes: S2, obtaining a time correlation feature of the learning state features; Figure 5 As shown, specifically including:

[0137] S21, sampling the learning state feature at a first time interval to obtain a first state time series;

[0138] In a preferred embodiment, the value range of the first time interval is [5min, 60min]. Since the learning state feature data at adjacent moments are usually highly correlated, the learning state feature of the previous moment can often well predict the learning state feature of the next moment. Therefore, the first time interval is set to a smaller value. By modeling the data at adjacent moments, the changes in the learning state feature are quickly reflected to achieve short-term prediction. Within the first time interval, the learning state feature is extracted from the data source to form a first state time series. For example, the learning state data is recorded every 10 minutes and added to the first state time series.

[0139] S22, acquiring a first time correlation feature according to the first state time series;

[0140] S23, sampling the learning state feature at a second time interval to obtain a second state time series;

[0141] In a preferred embodiment, the second time interval is 24 hours. By utilizing the data in the longer second time interval, the periodic change pattern of the learning state characteristics is captured. The periodic pattern of the learning state characteristics is very obvious, such as the peak and trough periods of the learning state characteristics every day, or the weekday and weekend patterns of each week. In the second time interval, the learning state characteristics are extracted from the data source to form a second state sequence. For example, the learning state data is recorded every 24 hours and added to the second state time series.

[0142] S24, acquiring a second time correlation feature according to the second state time series;

[0143] In some embodiments, the first time correlation is obtained according to the first state time series, and the second time correlation feature is obtained according to the second time interval. Both feature extraction is achieved through autocorrelation coefficient, partial autocorrelation coefficient, Fourier transform and time series decomposition. Specifically, the autocorrelation coefficient is expressed as:

[0144]

[0145] in, is the length of the time series, is the tth value in the first state time series, is the mean of the first state time series. The autocorrelation coefficient (ACF) measures the correlation between different lag steps of the time series, indicating the similarity between the series value and its own different time lag values; the autocorrelation coefficient can reveal the periodicity and trend of the time series, and provide information to determine whether the series has significant time dependence.

[0146] The partial autocorrelation coefficient (PACF) quantifies the direct correlation between different lag steps of the time series without the influence of the intermediate lag steps. The partial autocorrelation coefficient can help identify the direct causal relationship in the time series and avoid being confused by the indirect influence of the intermediate lag steps.

[0147] The Fourier transform converts the time series from the time domain to the frequency domain, representing it as a superposition of a series of sine and cosine waves of different frequencies. The frequency domain data processed by the Fourier transform can reveal the frequency components in the time series and help identify periodic and oscillatory behavior.

[0148] Time series decomposition is to decompose the time series into trends, seasonality, periodicity and random noise, thereby providing a structured understanding of complex time series and helping to identify and separate various components.

[0149] S25, fusing the first time correlation feature and the second time correlation feature to obtain a time correlation feature. In the process of model training and application, the time correlation feature will be input into the prediction model as an additional dimension of the learning state feature.

[0150] The time correlation feature is expressed as:

[0151]

[0152] in, is the time-dependent feature, is the first time correlation feature, is the second time correlation feature, is the weight coefficient of the first time correlation feature, is the weight coefficient of the second time correlation feature, is the Hadamard product.

[0153] In this embodiment, sampling at different time intervals and feature fusion are helpful to comprehensively and accurately capture the changes in the learner's learning status, improve the coverage and understanding of the learning status features on the time dimension, and also improve the prediction accuracy and system responsiveness of the model, providing solid support for personalized and dynamic learning evaluation.

[0154] In some embodiments, S5, inputting the learning state features acquired in real time into a pre-trained local prediction model to obtain the learning state prediction result, further comprising: S4, updating the local prediction model and the global prediction model according to the learning state features acquired in real time; Figure 6 As shown, specifically including:

[0155] S41, inputting the learning state features acquired in real time into the local prediction model; it should be noted that, similar to step S1, the learning state features acquired in real time in step S41 are acquired through facial images.

[0156] S42, updating the parameters of the local prediction model by gradient descent method;

[0157] Gradient descent is an optimization algorithm used to find the optimal values ​​of model parameters when training models in machine learning and deep learning. It finds the optimal values ​​of the model by continuously iteratively updating the parameters, helping the model to better fit the data and achieve efficient and accurate predictions.

[0158] S43, fusing all updated parameters of local prediction models and updating the global prediction model using a federated average algorithm;

[0159] The federated averaging algorithm is expressed as:

[0160]

[0161] in, are the model parameters of the updated global prediction model, For the The model parameters of the local prediction model, is the number of local prediction models.

[0162] S44, updating the local prediction model according to the model parameters of the updated global prediction model;

[0163] In this embodiment, the above method is continuously executed to update the model parameters. With the addition of new learning state features in each round, the model after each round of training is continuously optimized, thereby improving the adaptability of the model and the accuracy of prediction.

[0164] In real-world applications, in the application scenarios of personalized education platforms, the learning data of different students show obvious non-independent and identically distributed (Non-IID) characteristics, including differences in data quality, which are manifested in significant differences in learning behaviors, learning habits, knowledge mastery, etc. among different students; and uneven data volume, which is manifested in the different amounts of data that each student may generate. Some students are highly active and generate more data, while other students have less data. This Non-IID characteristic makes the model face challenges when training, because the model needs to maintain efficient training and accurate prediction on this unbalanced and heterogeneous data. However, traditional federated learning models often perform poorly when processing Non-IID data and are difficult to adapt to the personalized needs of each student.

[0165] In some embodiments, S44, the local prediction model is updated according to the model parameters of the updated global prediction model, such as Figure 7 As shown, including:

[0166] S441, Screening learners’ personalized characteristics through the superadditivity criterion of cooperative games;

[0167] In cooperative game theory, superadditivity means that multiple individuals can obtain greater benefits through cooperation than acting alone. Specifically, when screening personalized features, we can evaluate their importance by calculating the marginal contribution of each feature to the target, such as calculating the marginal contribution of each feature to the predicted learning effect. Through the superadditivity criterion of cooperative games, combined with the simultaneous collection and association of various types of data, we can efficiently screen out important personalized features and provide decision-making basis for personalized local prediction models. In some embodiments, personalized features include grade characteristics, subject characteristics, and learning mode characteristics, and the learning mode includes four behavioral modes: answering questions, online classes, communication, and reading textbooks. Among them, grade data, subject data, and learning mode data related to grade characteristics, subject characteristics, and learning mode characteristics are synchronously collected and associated with learning status characteristics.

[0168] S442, constructing a local feature prediction matrix based on the personalized features;

[0169] The local feature prediction matrix is ​​expressed as:

[0170]

[0171] in, is the output of the local feature prediction matrix, is the weight matrix, is the personalized feature vector, is the bias vector, is the activation function.

[0172] S443, outputting and splicing the local feature prediction matrix with the updated global prediction model to obtain a local prediction model;

[0173] After the local feature prediction matrix is ​​normalized by filling missing values ​​and normalizing numerical values, the global prediction result can be concatenated with the output of the local feature prediction matrix to form a complete input vector for subsequent personalized training.

[0174] In this embodiment, a local feature prediction matrix is ​​constructed through the learner's personalized features, and the local prediction model is obtained by splicing the output with the global prediction model, thereby improving the efficiency of the student's Non-IID data training and the accuracy of prediction.

[0175] S444, training the local prediction model through the personalized features, and updating the model parameters of the local prediction model.

[0176] In some embodiments, S444, the local prediction model is trained by the personalized features, and the model parameters of the local prediction model are updated, such as Figure 8 As shown, including:

[0177] S4441, in the forward propagation, using the local prediction model to generate prediction results;

[0178] S4442, calculating the error between the prediction result and the target feature through a loss function;

[0179] S4443, in the back propagation, calculating the weight matrix and the bias vector gradient, and updating the model parameters of the local prediction model through the weight matrix and the bias vector gradient;

[0180] The loss function is expressed as:

[0181]

[0182] in, is the number of samples, For the The prediction results of samples are For the The target value for each sample.

[0183] Example 2

[0184] A learning evaluation system based on artificial intelligence, such as Fig. 9 As shown, including:

[0185] Learning state feature acquisition module 1: used to acquire learning state features based on the learner's facial image;

[0186] Prediction model building module 3: used to adopt the federated learning architecture to build local prediction models and global prediction models according to learning state characteristics;

[0187] Learning state prediction module 5: used to input the learning state features acquired in real time into the pre-trained local prediction model to obtain the learning state prediction results.

[0188] In this embodiment, the acquired learning status features are input into a local prediction model obtained through federated learning, so that the learning status of the learner is predicted in real time from multiple dimensions, thereby improving the accuracy of the prediction of the learner's learning status.

[0189] It should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, each functional module can be integrated into a processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software functional modules.

[0190] Example 3

[0191] An electronic device 2, such as Fig.10 As shown, a processor 21 and a memory 24, the memory 24 is used to store computer program codes, the computer program codes include computer instructions, when the processor 21 executes the computer instructions, the electronic device executes the above-mentioned learning evaluation method based on artificial intelligence.

[0192] The electronic device 2 includes a processor 21, a memory 22, an input device 23, and an output device 24. The processor 21, the memory 22, the input device 23, and the output device 24 are coupled via a connector, and the connector includes various interfaces, transmission lines, or buses, etc., which are not limited in the embodiments of the present invention. It should be understood that in various embodiments of the present invention, coupling refers to mutual connection in a specific manner, including direct connection or indirect connection through other devices, for example, through various interfaces, transmission lines, buses, etc.

[0193] The processor 21 may be one or more graphics processing units (GPUs). When the processor 21 is a GPU, the GPU may be a single-core GPU or a multi-core GPU. Optionally, the processor 21 may be a processor group consisting of multiple GPUs, and the multiple processors are coupled to each other via one or more buses. Optionally, the processor may also be other types of processors, etc., which are not limited in the embodiments of the present invention.

[0194] The memory 22 can be used to store computer program instructions and various computer program codes including program codes for executing the scheme of the present invention. Optionally, the memory includes but is not limited to random access memory (RAM), read-only memory (ROM), erasable programmable read only memory (EPROM), or portable read only memory (CD-ROM), which is used for related instructions and data.

[0195] The input device 23 is used to input data and / or signals, and the output device 24 is used to output data and / or signals. The output device 23 and the input device 24 may be independent devices or an integrated device.

[0196] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A learning evaluation method based on artificial intelligence, characterized in that: include: According to the learner's facial image, learning state features are obtained; Adopting the federated learning architecture, local prediction models and global prediction models are constructed based on the learning state characteristics; Inputting the learning state features acquired in real time into a pre-trained local prediction model to obtain a learning state prediction result; the learning state features acquired in real time include learners' facial images, answering process data, online learning behavior data, online discussion behavior data and / or interactive audio data; Generate personalized learning recommendation reports based on learning status prediction results; Before inputting the learning state features acquired in real time into the pre-trained local prediction model, the method further includes: inputting the learning state features acquired in real time into the local prediction model; updating the parameters of the local prediction model by a gradient descent method; fusing all updated parameters of the local prediction models and updating the global prediction model by a federated average algorithm; and updating the local prediction model according to the model parameters of the updated global prediction model; The updating of the local prediction model according to the model parameters of the updated global prediction model includes: screening the personalized features of the learner through the superadditivity criterion of the cooperative game; constructing a local feature prediction matrix according to the personalized features; outputting and splicing the local feature prediction matrix with the updated global prediction model to obtain a local prediction model; training the local prediction model through the personalized features and updating the model parameters of the local prediction model; The method of training the local prediction model through personalized features and updating the model parameters of the local prediction model includes: in forward propagation, using the local prediction model to generate a prediction result; calculating the error between the prediction result and the target feature through a loss function; in backward propagation, calculating the weight matrix and the bias vector gradient, and updating the model parameters of the local prediction model through the weight matrix and the bias vector gradient; Before constructing the local prediction model and the global prediction model according to the learning state characteristics, the method also includes: sampling the learning state characteristics at a first time interval, and obtaining a first time correlation characteristic according to the obtained first state time series; sampling the learning state characteristics at a second time interval, and obtaining a second time correlation characteristic according to the obtained second state time series; and fusing the first time correlation characteristic and the second time correlation characteristic to obtain a time correlation characteristic.

2. The artificial intelligence-based learning evaluation method according to claim 1, characterized in that: The step of obtaining learning status features according to the learner's facial image includes: capturing the facial image of the learner in real time; Using a facial detection algorithm, locating a facial region in the facial image; Performing cropping and standardization processing on the facial region to obtain a facial region image; Using a facial feature point detection algorithm to locate feature points in the facial region image; Obtaining facial posture data according to the feature points; Performing time series analysis on the facial posture data to obtain the learning state characteristics.

3. The learning evaluation method based on artificial intelligence according to claim 2, characterized in that: The performing time series analysis on the facial posture data to obtain the learning state feature includes: Dividing the facial gesture data into windows of fixed length; Obtaining statistical indicators of the facial posture data of each window; According to the statistical indicators, obtaining the learner's concentration score, fatigue score and emotional state score; The concentration score, the fatigue score and the emotional state score are calculated to obtain the learning state characteristics.

4. The artificial intelligence-based learning evaluation method according to claim 3, characterized in that: The step of obtaining the statistical indicators of the facial posture data of each window includes: The facial posture data includes: head state data, eye state data and mouth state data; Calculating the standard deviation of the head angle data according to the head state data; Calculating the number of blinks per unit time according to the eye state data; According to the mouth state data, the average value of the mouth corner arc data is calculated.

5. A learning evaluation system based on artificial intelligence, characterized in that: A method for implementing an artificial intelligence-based learning evaluation method as claimed in any one of claims 1 to 4, comprising: Learning state feature acquisition module: used to acquire learning state features based on the learner's facial image; Prediction model building module: used to adopt the federated learning architecture to build local prediction models and global prediction models based on learning state characteristics; Learning state prediction module: used to input the learning state features acquired in real time into the pre-trained local prediction model to obtain the learning state prediction results.

6. An electronic device, characterized in that: It includes a processor and a memory, the memory is used to store computer program code, the computer program code includes computer instructions, and when the processor executes the computer instructions, the electronic device executes an artificial intelligence-based learning evaluation method as described in any one of claims 1 to 4.

Images