Artificial intelligence-based preschool education evaluation target data processing method and system

Abstract

The invention discloses a preschool education evaluation target data processing method and system based on artificial intelligence, and belongs to the technical field of education informatization. The method comprises the following steps: collecting preschool education evaluation target data, preprocessing the data, associating the data with an infant identifier, and uploading the data to an edge computing node; generating an edge calculation result file at the edge calculation node and transmitting the edge calculation result file to the central server; generating a classification result of the infant on each preschool education evaluation target dimension on the central server; and storing in an evaluation target relation graph database. According to the method, multi-source heterogeneous data is collected in a preschool education scene, and preprocessing and association are carried out on child social behavior video streams, exercise physiological data, picture book creation image data and teacher observation keywords, so that a complete preschool education evaluation data chain is constructed; therefore, the development conditions of children in all dimensions are accurately described, and the problem that preschool education evaluation data acquisition is scattered and incomplete in the prior art is effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of educational informatization technology, and in particular to a method and system for processing target data of preschool education evaluation based on artificial intelligence. Background Art

[0002] As a critical stage in life development, the scientific and comprehensive nature of preschool education's evaluation system is directly related to the healthy growth of young children and the rational allocation of educational resources. In recent years, with the continuous advancement of science and technology, the education field has gradually introduced various technical means to improve the accuracy and objectivity of evaluation. Currently, preschool education evaluation mainly relies on methods such as daily teacher observation, simple tests, and parent feedback. While these methods have certain reference value, they suffer from strong subjectivity, incomplete data collection, and a lack of systematicity. In recent years, with the continuous development of information technology, some educational institutions have begun to explore the introduction of technical means such as video surveillance and wearable devices to assist in preschool education evaluation. However, these attempts are mostly limited to the collection and analysis of a single data source and lack the ability to comprehensively process multi-source heterogeneous data.

[0003] However, the existing preschool education evaluation system has significant limitations in data collection. On the one hand, data collection is fragmented, lacking a unified mechanism for collecting and integrating multi-source, heterogeneous data. Various collection devices and technologies are independent of each other, preventing the formation of a complete data chain. On the other hand, existing technologies focus solely on collecting data from a single data source, such as video surveillance or wearable devices, failing to fully cover the multi-dimensional information of young children, resulting in one-sided and inaccurate evaluation results. Furthermore, existing technologies lack the ability to systematically process and integrate collected multi-source data. The formats of various data sources are inconsistent, timestamps are misaligned, and feature vectors are difficult to integrate, making it difficult to effectively utilize the data. Summary of the Invention

[0004] The embodiments of the present application solve the problem of scattered and incomplete preschool education evaluation data collection in the prior art by providing a preschool education evaluation target data processing method and system based on artificial intelligence, and achieve accurate portrayal of the development status of young children in various dimensions.

[0005] The embodiment of the present application provides a method for processing preschool education evaluation target data based on artificial intelligence, comprising the following steps: S1. collecting preschool education evaluation target data, including video streams of children's social behavior, children's motor physiological data, picture book creation image data, and observation keywords entered by a teacher terminal, pre-processing the preschool education evaluation target data, associating it with a child identifier, and uploading it to an edge computing node;

[0006] S2. At the edge computing node, generate a behavioral feature vector based on the social behavior video, synchronously process the motion physiological data to generate a motion state vector, align and fuse the behavioral feature vector and the motion state vector by timestamp, generate an edge computing result file, and transmit it to the central server;

[0007] S3. On the central server, the behavior feature vectors and motion state vectors in the edge computing result file are parsed and extracted. An image feature vector is generated based on the image data from the picture book creation. A text keyword vector is generated based on the teacher's observation keywords. The behavior feature vector, motion state vector, image feature vector, and text keyword vector are concatenated and passed through a multimodal Transformer encoder to generate a fused feature vector. This fused feature vector is input into the target classifier to generate the classification results of the children on various preschool education evaluation target dimensions.

[0008] S4. Store the fused feature vectors, classification results and original preschool education evaluation target data in the evaluation target relationship graph database according to the three-dimensional index structure.

[0009] Furthermore, the video stream of the children's social behavior is captured by a multi-camera array to capture a panoramic view of the children's activity area;

[0010] Compress the captured data and record the camera position and orientation information to generate panoramic video data;

[0011] The panoramic video data and the locally cached original video stream are used as input data for behavioral feature vector recognition of the children's social behavior video stream.

[0012] Furthermore, the specific process of generating the behavior feature vector based on the social behavior video is as follows:

[0013] Collect a dataset of children's social behavior pictures in preschool education scenarios and annotate various behavior types. Then divide the dataset into training set, validation set, and test set according to the preset ratio.

[0014] Use the Adam optimizer, set the initial learning rate and adopt a learning rate decay strategy, train for a preset number of training cycles, and complete model training to generate a behavior recognition model when the accuracy on the validation set and test set reaches the preset standards.

[0015] The pre-processed video stream of children's social behavior is input into the trained behavior recognition model to identify the behavioral features in the video, mark the category and confidence on the video frame, and generate a behavior feature vector.

[0016] Furthermore, the processing of the motion physiological data to generate the motion state vector comprises the following steps:

[0017] The acceleration measurement value at the moment the bracelet is turned on is taken as the initial state estimate;

[0018] The initial estimated error covariance matrix is ​​obtained by multiplying the unit matrix by the preset error coefficient;

[0019] Generate process noise covariance matrix and measurement noise covariance matrix through experimental calibration;

[0020] Initialize the Kalman filter based on the initial state estimate and the initial estimation error covariance matrix;

[0021] The process noise covariance matrix and the measurement noise covariance matrix are used to set the parameters of the initialized Kalman filter;

[0022] Input the exercise physiological data into the set Kalman filter for filtering to obtain filtered exercise physiological data;

[0023] Extract features from the filtered motion physiological data;

[0024] The extracted features are input into the pre-trained decision tree classifier for classification to generate the motion state vector.

[0025] Furthermore, generating a fused feature vector by a multimodal Transformer encoder includes the following steps:

[0026] Normalize the image data of picture book creation, extract the RGB value, position and area ratio of the color block, and generate the image feature vector;

[0027] Encode the teacher observation keywords into word vectors to generate text keyword vectors;

[0028] The behavior feature vector, the motion state vector, the image feature vector, and the text keyword vector are spliced ​​according to a preset dimension to form a spliced ​​vector;

[0029] The concatenated vector is input into the multimodal Transformer encoder, which fuses multimodal information through the self-attention mechanism to generate a fused feature vector.

[0030] Furthermore, the method for obtaining the target classifier is:

[0031] Initialize the global model parameters on the central server and broadcast them to the edge computing nodes of each kindergarten participating in the training;

[0032] Each kindergarten's edge computing node calculates the gradient of the model parameters based on local data and sends the calculated local gradient to the central server through a secure aggregation protocol;

[0033] The central server aggregates local gradients to obtain global gradients and updates global model parameters based on the global gradients.

[0034] The updated global model parameters are broadcasted by the central server to the edge computing nodes of each kindergarten, replacing the local model parameters of each kindergarten.

[0035] Determine whether the iteration termination condition is met. If not, return to the gradient calculation step to continue iterative training. If it is met, end the training to generate the final target classifier.

[0036] Furthermore, the specific steps of storing the fused feature vector, classification results and original preschool education evaluation target data in the evaluation target relationship map database according to the three-dimensional index structure are as follows:

[0037] Construct a target relationship graph database with evaluation targets as nodes, including several first-level nodes and second-level targets;

[0038] Construct a three-dimensional index structure based on development areas, age segments, and time windows;

[0039] The fused feature vectors, classification results and original data files are stored in the database according to the three-dimensional index structure, and each data entry is associated with the corresponding index information.

[0040] Furthermore, the three-dimensional index structure includes:

[0041] The developmental domain, as the first dimension, includes multiple first-level nodes, including language, cognitive, social, emotional, and physical development;

[0042] Age segmentation is the second dimension, which is divided into several intervals according to the physical and mental development of children;

[0043] As the third dimension, the time window is set according to preset time units to record the development and changes of young children in different time periods.

[0044] Furthermore, the evaluation target relationship map database is obtained in the following manner:

[0045] Divide the development field into a number of first-level nodes, each of which includes a number of second-level goals, each of which corresponds to specific behavioral performance characteristics and evaluation criteria;

[0046] Age groups are divided into several intervals based on the physical and mental development of children;

[0047] The time window is set according to the preset time unit;

[0048] The fused feature vectors, classification results and original data files are stored in the database according to the three-dimensional index structure to generate data entries with associated index information.

[0049] The embodiment of the present application provides a preschool education evaluation target data processing system based on artificial intelligence, including: a data acquisition module, an edge data transmission module, a feature fusion module, and a data classification and storage module;

[0050] The data acquisition module is used to collect preschool education evaluation target data, including video streams of children's social behavior, children's motor and physiological data, picture book creation image data, and observation keywords entered by teachers' terminals. The preschool education evaluation target data is pre-processed and associated with the children's identifiers before being uploaded to the edge computing node.

[0051] The edge data transmission module is used to generate a behavior feature vector based on the social behavior video at the edge computing node, synchronously process the motion physiological data to generate a motion state vector, align and fuse the behavior feature vector and the motion state vector according to the timestamp, generate an edge computing result file, and transmit it to the central server;

[0052] The feature fusion module is used to parse and extract the behavior feature vector and motion state vector from the edge computing result file on the central server, generate an image feature vector based on the picture book creation image data, generate a text keyword vector based on the teacher's observation keywords, and concatenate the behavior feature vector, motion state vector, image feature vector, and text keyword vector to generate a fused feature vector through a multimodal Transformer encoder; input the fused feature vector into the target classifier to generate the classification results of the children in various preschool education evaluation target dimensions;

[0053] The data classification storage module is used to store the fused feature vector, classification results and original preschool education evaluation target data in the evaluation target relationship map database according to the three-dimensional index structure.

[0054] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0055] 1. By deploying cameras, three-axis acceleration bracelets, high-precision scanners and other equipment in preschool education scenarios to collect multi-source heterogeneous data, and pre-processing and associating children's social behavior video streams, sports physiological data, picture book creation image data and teacher observation keywords, a complete preschool education evaluation data chain is constructed, thereby achieving an accurate portrayal of the development status of children in various dimensions, effectively solving the problem of scattered and incomplete preschool education evaluation data collection in existing technologies.

[0056] 2. By adopting the federated learning framework to train the target classifier on the central server, using the local data of each kindergarten to calculate the gradient and aggregating and updating the global model on the central server, we can achieve continuous optimization of the model while protecting the privacy of kindergarten data, and then realize data collaboration and resource sharing across kindergartens, effectively solving the problem of cross-kindergarten model optimization in existing technologies due to data privacy protection restrictions.

[0057] 3. By using the multimodal Transformer encoder to process the fused feature vector, the behavioral feature vector, motion state vector, image feature vector and text keyword vector are fused, thereby achieving a comprehensive analysis and classification of the multi-dimensional behavioral characteristics of young children, and further achieving an accurate evaluation of young children in various target dimensions of preschool education. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A flowchart of a method for processing target data for preschool education evaluation based on artificial intelligence provided in an embodiment of the present application;

[0059] Figure 2 This is a structural diagram of the artificial intelligence-based preschool education evaluation target data processing system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] The embodiments of the present application solve the problem of scattered and incomplete preschool education evaluation data collection in the prior art by providing a preschool education evaluation target data processing method and system based on artificial intelligence. By deploying cameras, three-axis acceleration bracelets, high-precision scanners and other equipment in preschool education scenarios to collect multi-source heterogeneous data, and pre-processing and associating children's social behavior video streams, sports physiological data, picture book creation image data and teacher observation keywords, a complete preschool education evaluation data chain is constructed, thereby achieving an accurate portrayal of the development status of children in various dimensions, and effectively solving the problem of scattered and incomplete preschool education evaluation data collection in the prior art.

[0061] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0062] like Figure 1 As shown, it is a flow chart of a method for processing target data of preschool education evaluation based on artificial intelligence provided by an embodiment of the present application. The method is applied to a preschool education evaluation target data processing system based on artificial intelligence. The method comprises the following steps: S1. collecting preschool education evaluation target data, including video streams of children's social behavior, children's movement and physiological data, picture book creation image data, and observation keywords input by the teacher terminal, pre-processing the preschool education evaluation target data, associating it with the child identifier, and uploading it to the edge computing node;

[0063] In this embodiment, video streams of children's social behaviors are collected and stored in a local cache; a three-axis acceleration bracelet is used to collect children's movement and physiological data and store them as time series files; a high-precision scanner is used to obtain picture book creation images and analyze the color block distribution to generate image feature data files; the teacher's terminal enters observation keywords and stores them in text form, with each keyword accompanied by a timestamp.

[0064] S2. At the edge computing node, decode, grayscale, and histogram-equalize the social behavior video stream. Generate a behavior feature vector based on the social behavior video. Synchronously process the motion physiological data to generate a motion state vector. Align and fuse the behavior feature vector with the motion state vector based on timestamps. Generate an edge computing result file and transmit it to the central server.

[0065] S3. On the central server, the behavior feature vectors and motion state vectors in the edge computing result file are parsed and extracted. An image feature vector is generated based on the image data from the picture book creation. A text keyword vector is generated based on the teacher's observation keywords. The behavior feature vector, motion state vector, image feature vector, and text keyword vector are concatenated and passed through a multimodal Transformer encoder to generate a fused feature vector. This fused feature vector is input into the target classifier to generate the classification results of the children on various preschool education evaluation target dimensions.

[0066] S4. Store the fused feature vectors, classification results and original preschool education evaluation target data in the evaluation target relationship graph database according to the three-dimensional index structure.

[0067] Furthermore, the video stream of the children's social behavior is captured by a multi-camera array to capture a panoramic view of the children's activity area;

[0068] Compress the captured data and record the camera position and orientation information to generate panoramic video data;

[0069] The panoramic video data and the locally cached original video stream are used as input data for behavioral feature vector recognition of the children's social behavior video stream.

[0070] Furthermore, the specific process of generating the behavior feature vector based on the social behavior video is as follows:

[0071] Collect a dataset of children's social behavior pictures in preschool education scenarios and annotate various behavior types. Then divide the dataset into training set, validation set, and test set according to the preset ratio.

[0072] Use the Adam optimizer, set the initial learning rate and adopt a learning rate decay strategy, train for a preset number of training cycles, and complete model training to generate a behavior recognition model when the accuracy on the validation set and test set reaches the preset standards.

[0073] The pre-processed video stream of children's social behavior is input into the trained behavior recognition model to identify the behavioral features in the video, mark the category and confidence on the video frame, and generate a behavior feature vector.

[0074] In this example, the Adam optimizer is used to set the initial learning rate.

[0075] Train the model on the training set, calculate the loss function value and update the model weight parameters through backpropagation;

[0076] After each training cycle (epoch), the learning rate is adjusted according to the preset learning rate decay strategy;

[0077] Evaluate the model performance on the validation set and record the accuracy of the model on the validation set;

[0078] When the accuracy of the model on the validation set reaches the preset standard or the training reaches the preset number of training cycles, the final model performance evaluation is performed on the test set; the model training is completed to generate a behavior recognition model.

[0079] Furthermore, the processing of the motion physiological data to generate the motion state vector comprises the following steps:

[0080] The acceleration measurement value at the moment the bracelet is turned on is taken as the initial state estimate;

[0081] The initial estimated error covariance matrix is ​​obtained by multiplying the unit matrix by the preset error coefficient;

[0082] Generate process noise covariance matrix and measurement noise covariance matrix through experimental calibration;

[0083] Initialize the Kalman filter based on the initial state estimate and the initial estimation error covariance matrix;

[0084] The process noise covariance matrix and the measurement noise covariance matrix are used to set the parameters of the initialized Kalman filter;

[0085] Input the exercise physiological data into the set Kalman filter for filtering to obtain filtered exercise physiological data;

[0086] Perform feature extraction on the filtered motion physiological data to extract the magnitude and rate of change of acceleration;

[0087] The extracted features are input into the pre-trained decision tree classifier for classification to generate the motion state vector.

[0088] Furthermore, generating a fused feature vector by a multimodal Transformer encoder includes the following steps:

[0089] Normalize the image data of the picture book creation, extract the RGB value, position information and area ratio of the color block, and convert the position information of the color block into polar coordinates to generate the image feature vector;

[0090] Encode the teacher observation keywords into word vectors to generate text keyword vectors;

[0091] The behavior feature vector, the motion state vector, the image feature vector, and the text keyword vector are spliced ​​according to a preset dimension to form a spliced ​​vector;

[0092] The concatenated vector is input into the multimodal Transformer encoder, which fuses multimodal information through the self-attention mechanism to generate a fused feature vector.

[0093] In this embodiment, the training data of the multimodal Transformer encoder is prepared as follows:

[0094] Collect multimodal data samples including videos, sensors, images, and text from multiple kindergartens;

[0095] Convert all types of data into vector representations of the same dimension and align them by timestamp to form a fused feature vector;

[0096] Each of the fused feature vectors corresponds to a label of a preschool education evaluation target dimension.

[0097] Furthermore, the method for obtaining the target classifier is:

[0098] Initialize the global model parameters on the central server and broadcast them to the edge computing nodes of each kindergarten participating in the training;

[0099] Each kindergarten's edge computing node calculates the gradient of the model parameters based on local data and sends the calculated local gradient to the central server through a secure aggregation protocol;

[0100] The central server aggregates local gradients to obtain global gradients and updates global model parameters based on the global gradients.

[0101] The updated global model parameters are broadcasted by the central server to the edge computing nodes of each kindergarten, replacing the local model parameters of each kindergarten.

[0102] Determine whether the iteration termination condition is met. If not, return to the gradient calculation step to continue iterative training. If it is met, end the training to generate the final target classifier.

[0103] Furthermore, the specific steps of storing the fused feature vector, classification results and original preschool education evaluation target data in the evaluation target relationship map database according to the three-dimensional index structure are as follows:

[0104] Construct a target relationship graph database with evaluation targets as nodes, including several first-level nodes and second-level targets;

[0105] Construct a three-dimensional index structure based on development areas, age segments, and time windows;

[0106] The fused feature vectors, classification results and original data files are stored in the database according to the three-dimensional index structure, and each data entry is associated with the corresponding index information.

[0107] In this embodiment, data is stored in the target relational graph database using a distributed storage architecture:

[0108] Distribute and store data from different kindergartens, different development areas, different age groups, and different time windows on multiple server nodes;

[0109] The index position of the data block is calculated by the hash algorithm Index = Hash (Data) mod N to achieve fast data positioning;

[0110] The database management system regularly backs up and archives data, combines multiple storage media and establishes a data redundancy mechanism. When a node fails, data can be quickly restored from other nodes to generate a complete database backup for data retrieval.

[0111] The data retrieval process is:

[0112] Receive user search requests and analyze the development areas, age groups, time windows, and specific evaluation target requirements;

[0113] Quickly locate matching data blocks in the database based on the three-dimensional index structure;

[0114] An approximate nearest neighbor search algorithm based on locality sensitive hashing is used to generate retrieval results and sort them according to their similarity to the query target;

[0115] The sorted data items, including fused feature vectors, classification results, and related raw data segments are returned to the user.

[0116] In this embodiment, by adopting a distributed storage architecture, data from different kindergartens, different development fields, different age segments, and different time windows are distributed and stored in multiple server nodes, and a hash algorithm is used to calculate the index position of the data block to achieve fast data positioning, thereby constructing an efficient and reliable preschool education evaluation data storage system. The database management system regularly backs up and archives the data, and adopts a combination of multiple storage media and establishes a data redundancy mechanism. When a node fails, data can be quickly restored from other nodes and a complete database backup can be generated, thereby achieving high data availability and fault tolerance, and effectively solving the single point failure problem caused by centralized data storage and the problem of low data retrieval efficiency in the existing technology.

[0117] Furthermore, the three-dimensional index structure includes:

[0118] The developmental domain, as the first dimension, includes multiple first-level nodes, including language, cognitive, social, emotional, and physical development;

[0119] Age segmentation is the second dimension, which is divided into several intervals according to the physical and mental development of children;

[0120] As the third dimension, the time window is set according to preset time units to record the development and changes of young children in different time periods.

[0121] Furthermore, the evaluation target relationship map database is obtained in the following manner:

[0122] Divide the development field into a number of first-level nodes, each of which includes a number of second-level goals, each of which corresponds to specific behavioral performance characteristics and evaluation criteria;

[0123] Age groups are divided into several intervals based on the physical and mental development of children;

[0124] The time window is set according to the preset time unit;

[0125] The fused feature vectors, classification results and original data files are stored in the database according to the three-dimensional index structure to generate data entries with associated index information.

[0126] like Figure 2 As shown, it is a structural diagram of the preschool education evaluation target data processing system based on artificial intelligence provided by an embodiment of the present application. The preschool education evaluation target data processing system based on artificial intelligence provided by an embodiment of the present application includes: a data acquisition module, an edge data transmission module, a feature fusion module, and a data classification and storage module;

[0127] The data acquisition module is used to collect preschool education evaluation target data, including video streams of children's social behavior, children's motor and physiological data, picture book creation image data, and observation keywords entered by teachers' terminals. The preschool education evaluation target data is pre-processed and associated with the children's identifiers before being uploaded to the edge computing node.

[0128] The edge data transmission module is used to generate a behavior feature vector based on the social behavior video at the edge computing node, synchronously process the motion physiological data to generate a motion state vector, align and fuse the behavior feature vector and the motion state vector according to the timestamp, generate an edge computing result file, and transmit it to the central server;

[0129] The feature fusion module is used to parse and extract the behavior feature vector and motion state vector from the edge computing result file on the central server, generate an image feature vector based on the picture book creation image data, generate a text keyword vector based on the teacher's observation keywords, and concatenate the behavior feature vector, motion state vector, image feature vector, and text keyword vector to generate a fused feature vector through a multimodal Transformer encoder; input the fused feature vector into the target classifier to generate the classification results of the children in various preschool education evaluation target dimensions;

[0130] The data classification storage module is used to store the fused feature vector, classification results and original preschool education evaluation target data in the evaluation target relationship map database according to the three-dimensional index structure.

[0131] In summary, the embodiments of the present application collect multi-source heterogeneous data by deploying cameras, three-axis acceleration bracelets, high-precision scanners and other equipment in preschool education scenarios, and pre-process and associate children's social behavior video streams, sports physiological data, picture book creation image data and teacher observation keywords, thereby constructing a complete preschool education evaluation data chain, and thus achieving accurate portrayal of the development status of children in various dimensions.

[0132] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0133] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0134] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0135] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0136] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0137] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for processing target data of preschool education evaluation based on artificial intelligence, characterized in that: The following steps are involved: S1. Collect preschool education evaluation target data, including video streams of children's social behavior, children's motor and physiological data, image data of picture book creation, and observation keywords entered by teachers' terminals. Pre-process the preschool education evaluation target data, associate it with children's identifiers, and upload it to the edge computing node. S2. At the edge computing node, generate a behavioral feature vector based on the social behavior video, synchronously process the motion physiological data to generate a motion state vector, align and fuse the behavioral feature vector and the motion state vector by timestamp, generate an edge computing result file, and transmit it to the central server; S3. On the central server, the behavior feature vectors and motion state vectors in the edge computing result file are parsed and extracted. An image feature vector is generated based on the image data from the picture book creation. A text keyword vector is generated based on the teacher's observation keywords. The behavior feature vector, motion state vector, image feature vector, and text keyword vector are concatenated and passed through a multimodal Transformer encoder to generate a fused feature vector. This fused feature vector is input into the target classifier to generate the classification results of the children on various preschool education evaluation target dimensions. S4. Store the fused feature vectors, classification results and original preschool education evaluation target data into the evaluation target relationship graph database according to the three-dimensional index structure.

2. The method for processing target data of preschool education evaluation based on artificial intelligence according to claim 1, characterized in that: The video stream of the children's social behavior is captured by a multi-camera array to capture a panoramic view of the children's activity area; Compress the captured data and record the camera position and orientation information to generate panoramic video data; The panoramic video data and the locally cached original video stream are used as input data for behavioral feature vector recognition of the children's social behavior video stream.

3. The method for processing target data of preschool education evaluation based on artificial intelligence according to claim 1, characterized in that: The specific process of generating a behavior feature vector based on a social behavior video is as follows: Collect a dataset of children's social behavior pictures in preschool education scenarios and annotate various behavior types. Then divide the dataset into training set, validation set, and test set according to the preset ratio. Use the Adam optimizer, set the initial learning rate and adopt a learning rate decay strategy, train for a preset number of training cycles, and complete model training to generate a behavior recognition model when the accuracy on the validation set and test set reaches the preset standards. The pre-processed video stream of children's social behavior is input into the trained behavior recognition model to identify the behavioral features in the video, mark the category and confidence on the video frame, and generate a behavior feature vector.

4. The method for processing target data of preschool education evaluation based on artificial intelligence according to claim 1, characterized in that: The processing of the motion physiological data to generate the motion state vector comprises the following steps: The acceleration measurement value at the moment the bracelet is turned on is taken as the initial state estimate; The initial estimated error covariance matrix is ​​obtained by multiplying the unit matrix by the preset error coefficient; Generate process noise covariance matrix and measurement noise covariance matrix through experimental calibration; Initialize the Kalman filter based on the initial state estimate and the initial estimation error covariance matrix; The process noise covariance matrix and the measurement noise covariance matrix are used to set the parameters of the initialized Kalman filter; Input the exercise physiological data into the set Kalman filter for filtering to obtain filtered exercise physiological data; Extract features from the filtered motion physiological data; The extracted features are input into the pre-trained decision tree classifier for classification to generate the motion state vector.

5. The method for processing target data of preschool education evaluation based on artificial intelligence according to claim 1, characterized in that: Generating a fused feature vector by a multimodal Transformer encoder comprises the following steps: Normalize the image data of picture book creation, extract the RGB value, position and area ratio of the color block, and generate the image feature vector; Encode the teacher observation keywords into word vectors to generate text keyword vectors; The behavior feature vector, the motion state vector, the image feature vector, and the text keyword vector are spliced ​​according to a preset dimension to form a spliced ​​vector; The concatenated vector is input into the multimodal Transformer encoder, which fuses multimodal information through the self-attention mechanism to generate a fused feature vector.

6. The method for processing target data of preschool education evaluation based on artificial intelligence according to claim 1, characterized in that: The method for obtaining the target classifier is: Initialize the global model parameters on the central server and broadcast them to the edge computing nodes of each kindergarten participating in the training; Each kindergarten's edge computing node calculates the gradient of the model parameters based on local data and sends the calculated local gradient to the central server through a secure aggregation protocol; The central server aggregates local gradients to obtain global gradients and updates global model parameters based on the global gradients. The updated global model parameters are broadcasted by the central server to the edge computing nodes of each kindergarten, replacing the local model parameters of each kindergarten. Determine whether the iteration termination condition is met. If not, return to the gradient calculation step to continue iterative training. If it is met, end the training to generate the final target classifier.

7. The method for processing target data of preschool education evaluation based on artificial intelligence according to claim 1, characterized in that: The specific steps of storing the fused feature vector, classification results and original preschool education evaluation target data in the evaluation target relationship map database according to the three-dimensional index structure are as follows: Construct a target relationship graph database with evaluation targets as nodes, including several first-level nodes and second-level targets; Construct a three-dimensional index structure based on development areas, age segments, and time windows; The fused feature vectors, classification results and original data files are stored in the database according to the three-dimensional index structure, and each data entry is associated with the corresponding index information.

8. The method for processing target data of preschool education evaluation based on artificial intelligence according to claim 7, characterized in that: The three-dimensional index structure includes: The developmental domain, as the first dimension, includes multiple first-level nodes, including language, cognitive, social, emotional, and physical development; Age segmentation is the second dimension, which is divided into several intervals according to the physical and mental development of children; As the third dimension, the time window is set according to preset time units to record the development and changes of young children in different time periods.

9. The method for processing target data of preschool education evaluation based on artificial intelligence according to claim 7, characterized in that: The evaluation target relationship map database is obtained in the following manner: Divide the development field into a number of first-level nodes, each of which includes a number of second-level goals, each of which corresponds to specific behavioral performance characteristics and evaluation criteria; Age groups are divided into several intervals based on the physical and mental development of children; The time window is set according to the preset time unit; The fused feature vectors, classification results and original data files are stored in the database according to the three-dimensional index structure to generate data entries with associated index information.

10. An artificial intelligence-based preschool education evaluation target data processing system, used to implement the artificial intelligence-based preschool education evaluation target data processing method according to any one of claims 1 to 9, characterized in that: It includes data acquisition module, edge data transmission module, feature fusion module, and data classification and storage module; The data acquisition module is used to collect preschool education evaluation target data, including video streams of children's social behavior, children's motor and physiological data, picture book creation image data, and observation keywords entered by teachers' terminals. The preschool education evaluation target data is pre-processed and associated with the children's identifiers before being uploaded to the edge computing node. The edge data transmission module is used to generate a behavior feature vector based on the social behavior video at the edge computing node, synchronously process the motion physiological data to generate a motion state vector, align and fuse the behavior feature vector and the motion state vector according to the timestamp, generate an edge computing result file, and transmit it to the central server; The feature fusion module is used to parse and extract the behavior feature vector and motion state vector from the edge computing result file on the central server, generate an image feature vector based on the picture book creation image data, generate a text keyword vector based on the teacher's observation keywords, and concatenate the behavior feature vector, motion state vector, image feature vector, and text keyword vector to generate a fused feature vector through a multimodal Transformer encoder; input the fused feature vector into the target classifier to generate the classification results of the children in various preschool education evaluation target dimensions; The data classification storage module is used to store the fused feature vector, classification results and original preschool education evaluation target data in the evaluation target relationship map database according to the three-dimensional index structure.

Images