Intelligent training evaluation method based on multi-source data analysis

Abstract

The invention discloses an intelligent training evaluation method based on multi-source data analysis, and belongs to the technical field of intelligent training. According to the method, data of multi-source training and examination equipment is collected, then effective training and examination information analysis is conducted on multi-source heterogeneous type data, training and examination information such as the health condition, the ability level, the shortness and the strength of each person is analyzed and mined in real time, and the action level, the damage level and the like are displayed through three-dimensional reconstruction comparison. According to the invention, comprehensive, accurate and personalized evaluation of training and examination conditions is realized, intelligent grouping training can be carried out for different people, a scientific training scheme is given, real-time monitoring on the execution process of the training scheme is supported, the training scheme is optimized and adjusted in time, and the training effect is continuously improved. The method is high in reasoning analysis capability and has high market popularization value in the field of intelligent training.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent training technology, and specifically relates to an intelligent training and assessment method based on multi-source data analysis. Background Art

[0002] Due to the different domestic and international scenarios, the composition of training personnel, training equipment, and examination training subjects is different. There are obvious differences in the research points and technologies in physical training, and there are also differences in intelligent training and assessment methods.

[0003] At present, the technical means of intelligent training assessment in the field of intelligent training are mainly the assessment of single data type indicators, which requires the wearing of wearable devices, lacks data analysis, lacks comprehensiveness in the assessment angle, cannot be accurate, and lacks strong reasoning and analysis capabilities. After the assessment, it is impossible to give scientific and personalized training plans, nor is it possible to monitor the execution effect of the training plans in real time and adjust the training plans. Summary of the invention

[0004] The present invention aims to address the deficiencies of the prior art and proposes an intelligent training and testing assessment method based on multi-source data analysis. By collecting data from multi-source training and testing equipment, effective training and testing information analysis can be performed on multi-source heterogeneous data, and real-time analysis and mining of each person's health status, ability level, weaknesses, strengths and other training and testing information can be performed. By contrasting and displaying action levels, injury levels, etc. through three-dimensional reconstruction, a comprehensive, accurate and personalized assessment of the training and testing situation can be achieved. Intelligent group training can be performed for different groups of people, and scientific training plans can be given. This supports real-time monitoring of the execution process of the training plan, timely optimization and adjustment of the training plan, and continuous improvement of training results.

[0005] To achieve the above object, the present invention provides the following solution: an intelligent training and assessment method based on multi-source data analysis, comprising the following steps:

[0006] Step 1: Collect historical training and test information and real-time training and test information to obtain multi-source training and test data;

[0007] Step 2: Process the multi-source training and test data using a data analysis algorithm to obtain multi-source valid training and test data; the multi-source valid training and test data includes: training and test text content, training and test documents, and training and test videos;

[0008] Step 3: Based on the evaluation algorithm, the multi-source effective training and test data are integrated and analyzed to obtain an analysis report; and based on the analysis report, intelligent grouping is performed to provide personalized training program guidance and physical abnormality warnings for different groups;

[0009] Step 4: Optimize and adjust the training plan using reinforcement learning, and evaluate the effectiveness of the training plan based on real-time training and testing data;

[0010] Step 5: Conduct questions and answers on the training assessment content through voice and text, quickly obtain individual and group training reports, analysis charts and training videos, and recommended web pages to visit, and give assessment results. The assessment results are used to help students check the training status and answer questions in a timely manner.

[0011] Further preferably, in step 2, the steps of the data analysis algorithm include:

[0012] Step 1: parse the text data of the training and examination documents in the multi-source training and examination data, extract tables, seals, column paragraphs, pictures, and formula data, and obtain the training and examination text content;

[0013] Step 2: Use the YOLO algorithm to detect blurry frames in the video information of the multi-source training data, and introduce an attention mechanism to capture blurry frame information; use the RNN network to process the blurry frames to obtain clear training videos;

[0014] Step 3: Perform data governance on the parsed data to obtain the multi-source effective training and testing data.

[0015] Further preferably, step 1 comprises the following steps:

[0016] A1. Using a neural network algorithm to detect the content of the training text, and identifying complex training text and simple training text;

[0017] A2. Constructing a generative network and a discriminative network to process the complex training text, wherein the discriminative network is used to distinguish the training text and the training format; and the generative network is used to generate false data similar to the training text and the training format;

[0018] A3. Train and optimize the constructed generative network and discriminative network to obtain the optimal u-type text content generated by the generative network;

[0019] A4. Use the CRNN algorithm to perform text recognition on the text content and the simple training text, draw lines according to the coordinate sequence of the characters, and use a neural network to classify the drawn lines into common reading sequences to obtain text content with a correct reading sequence.

[0020] Further preferably, the objective function in A2 is:

[0021]

[0022] Where u represents the number of training and test texts and training and test formats; i represents the weight; D is the mapping function for discriminating the training text and the training format; G is the mapping function for the generator training text and the training format; xi The data consists of real training and test text and training and test format; data is the distribution of data composed of real training and test texts and training and test formats; i represents the noise of random training text and training format; p z Represents the noise distribution of random training text and training format.

[0023] Further preferably, the evaluation algorithm in step 3 comprises the following steps:

[0024] Step 1: Obtain the trainee's training and examination video in real time, and use computer vision technology to analyze the training and examination video to obtain video analysis results; the video analysis results include: the trainee's emotional state, physical injury status, and athletic ability;

[0025] Step 2: Based on the text content, construct a portrait of the trainees and the group from the aspects of training indicators and five-dimensional radar charts to obtain health status values; and analyze the strength and change trend of the training indicators of the trainees and the group over time based on the correlation between physical fitness, skills and qualities; and predict the risk threshold based on the training and test video data, and warn of physical abnormalities based on the risk threshold;

[0026] Step 3: Use big model technology to analyze the students' historical training and examination information, and then optimize the risk threshold;

[0027] Step 4: Intelligently group the students, provide personalized training plans for different groups, monitor their physical condition in real time, and issue warnings for physical abnormalities.

[0028] Further preferably, step 1 comprises the following steps:

[0029] B1. Use neural network algorithms to classify history training videos into three categories, including face, lower limb gait, and whole body movement videos; use Transformer neural network to extract 3D features of head facial expressions, lower limb gait movements, and whole body movement movements;

[0030] B2. Use 3D human posture estimation algorithm to identify movement changes, and identify the standard state, emotional state and injury state of movement based on graph convolutional neural network. Use 3D reconstruction technology to display the position of real movement and the amplitude of standard movement, optimize the details of movement generation, and then compare the gap between real movement and standard movement.

[0031] B3. Use the standard state of action, emotional state, injury state, and the difference between the real action and the standard action as feature fields, mark the performance scores, and use the lightGBM algorithm for training and optimization;

[0032] B4. Obtain real-time training and test videos, repeat B1 and B2 to obtain the difference between the real action and the standard action, the emotional state, and the field information of the injury state, and input them into the trained lightGBM algorithm in B3 for regression to give a real-time score.

[0033] Further preferably, the loss function of the three-dimensional reconstruction technology includes:

[0034] L(o,q)=L o +λ re L q ,

[0035] in,

[0036]

[0037]

[0038] In the formula, o represents the pixel set of each batch; q represents the pixel set of each batch; L o Indicates the pixel loss between blurred pixels and real pixels; L q Represents the distance loss between the 3D reconstructed model and the real model; re represents the distance loss weight; K represents the pixel set; H o Indicates blurred pixels; Represents real pixels; I q Represents the reconstruction model; Represents the true model.

[0039] Further preferably, the gap includes: an angle gap and an arc gap;

[0040] The calculation method of the angle difference includes:

[0041]

[0042] In the formula, (x aa ,y aa 、z aa ) is the coordinate of the real action joint in three-dimensional space; (x bb ,y bb 、z bb ) is the coordinate of the joint point where the real action and the standard action intersect in three-dimensional space; (x cc ,y cc 、z cc ) are the coordinates of the standard action joints in three-dimensional space;

[0043] The calculation method of the arc difference includes:

[0044]

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The present invention can realize data collection of multi-source training and testing equipment, can effectively analyze training and testing information of multi-source heterogeneous data, analyze and mine each person's health status, ability level, shortcomings and strengths in real time, and compare and display action level, injury level, etc. through three-dimensional reconstruction to achieve comprehensive, accurate and personalized assessment of training and testing conditions. It can conduct intelligent group training for different groups of people, and provide scientific training plans, support real-time monitoring of the execution process of training plans, timely optimize and adjust training plans, and continuously improve training results. It provides students with efficient and intelligent training and testing methods, and has high market promotion value in the field of intelligent training. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0048] Figure 1 The figure is a flow chart of an intelligent training and assessment method based on multi-source data analysis according to an embodiment of the present invention;

[0049] Figure 2 Schematic diagram of the technical implementation of the intelligent training and assessment method based on multi-source data analysis in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] 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.

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] Embodiment 1:

[0053] This embodiment provides an intelligent training and assessment method based on multi-source data analysis. Figure 1 , Figure 2 As shown, the following steps are included:

[0054] Step 1: Collect historical training and test information and real-time training and test information to obtain multi-source training and test data.

[0055] Specifically, according to different protocols, multi-source hardware devices such as training equipment, assessment equipment, cameras, physical measurement instruments, watches, scanners, etc. are connected to collect data, and training and examination information from other training and examination systems is connected to obtain historical training and examination data and real-time training and examination data.

[0056] Step 2: Use a data analysis algorithm to process the multi-source training and testing data to obtain multi-source effective training and testing data, thereby realizing the fusion of multi-source training and testing data.

[0057] The specific steps of the data parsing algorithm include:

[0058] Step 1: Analyze the text data of the training and examination documents, and support data extraction of tables, stamps, column paragraphs, pictures, formulas, etc.

[0059] The specific steps include:

[0060] A1. Use neural network algorithms to detect text content and identify complex training texts (such as training tables, seals, column paragraphs, pictures, etc.) and simple training texts (such as common training text content, formulas, etc.).

[0061] A2. For complex training texts, they are processed by constructing a generative network and a discriminative network. The discriminative network is used to distinguish the content and format of the training texts; the generative network is used to generate false data similar to the content of the training texts.

[0062] The objective function is:

[0063]

[0064] In the formula, u represents the number of types of complex training and test texts; i represents the weight; D is the mapping function for distinguishing the content of the training test text and the format of the training test; G is the mapping function for generating the content of the training test text; x i The data consists of the real training and test text content and training and test format; data The distribution of data consisting of real training and test text content and training and test format; i represents the noise of random training text content; p z Represents the noise distribution of random training text content.

[0065] A3. Continuously train and optimize the constructed generative network and discriminative network, and finally obtain the optimal u-type text content generated by the generative network, from which the training and test format content has been removed.

[0066] A4. Use the CRNN algorithm to perform text recognition on the text content obtained in A3 and the simple training text, draw lines according to the coordinate sequence of the characters, and use the neural network to classify the drawn lines into common reading sequences, and map the recognized text content to obtain text content that matches the reading sequence.

[0067] Step 2: Use the YOLO algorithm as the detection network to detect blurry frames in the video information of multi-source training data, and introduce the attention mechanism to capture important blurry frame information; use the RNN network as the deblurring network to process blurry frames. Among them, the detection network and the deblurring network are integrated into a pipeline to process a blurry video into a clear video, reducing the impact of the environment such as lighting and darkness. Its loss function is expressed as:

[0068]

[0069] In the formula, Z represents the total number of training samples; Q and Represent the real image and the estimated image respectively; L b represents the loss of detecting blurry frames; b Represents a hyperparameter.

[0070] When there is blur in a training video frame t, the RNN network is used to extract hierarchical features f for video frames t-1, t-2, t, t+1, and t+2. t-2 , f t-1 , f t , f t+1 , f t+2 , we can get the fusion feature: F t =Conv(CAT(f t-2 , f t-1 , f t+1 , f t+2 , f t )). CAT stands for cascade operation; Conv stands for convolution operation; F t After upsampling through the deconvolution layer, images of t frames of the training and testing video are generated to replace the original blurred images and obtain clear training and testing videos.

[0071] Step 3: Perform data governance on the parsed data, filter out irrelevant information, and achieve multi-source data aggregation and fusion.

[0072] Step 3: Use the assessment algorithm to integrate and analyze the training and test text content and clear training and test videos obtained in step 2, capture and evaluate each student's health status, ability level, shortcomings and strengths in an all-round and real-time manner, and generate group and individual analysis reports; then, intelligently group different groups of people and provide differentiated and personalized training plans. By constructing a causal relationship network between risk warning influencing factors, the ability level can be assessed according to the training and test indicators, and alarms can be issued for physical abnormalities.

[0073] The evaluation algorithm includes the following steps:

[0074] Step 1: Obtain student training and test videos in real time, and use computer vision technology to analyze and evaluate the head and facial emotions, lower limb gait, and whole-body movement in the training and test videos to obtain video analysis results; video analysis results include: student emotional state, physical injury status, and athletic ability.

[0075] Specifically, the steps include:

[0076] B1. Use neural network algorithm to divide history training videos into three categories, namely face, lower limb gait, and whole body movement videos. Use Transformer neural network to extract 3D features of head facial expressions, lower limb gait movements, and whole body movement movements.

[0077] B2. Use 3D human posture estimation algorithm to identify movement changes, and capture the dependency between features based on graph convolutional neural network, identify the standard state of movement (low, medium, high), emotional state (normal and abnormal), and injury state (mild, medium, high) during exercise, use 3D reconstruction technology to show the position of real movement and the amplitude that standard movement should reach, use Gan to optimize the details of movement position generation, and then compare the gap between the trainees' real movement and standard movement.

[0078] Among them, the loss functions of 3D reconstruction technology include:

[0079] L(o,q)=L o +λ re L q ,

[0080] in,

[0081]

[0082]

[0083] In the formula, o represents the pixel set of each batch; K represents the pixel set; q represents the model set of each batch of 3D reconstruction; m represents the size of the 3D reconstruction model set; L oIndicates the pixel loss between blurred pixels and real pixels; L q Represents the distance loss between the 3D reconstructed model and the real model; re represents the distance loss weight; H o Indicates blurred pixels; Represents real pixels; I q Represents the reconstruction model; Represents the true model.

[0084] The above-mentioned gap includes: angle gap and arc gap. The calculation method of the angle gap LL between the real action angle and the standard action angle includes:

[0085]

[0086] In the formula, (x aa ,y aa 、z aa ) is the coordinate of the real action joint in three-dimensional space; (x bb ,y bb 、z bb ) is the coordinate of the joint point where the real action and the standard action intersect in three-dimensional space; (x cc ,y cc 、z cc ) are the coordinates of the standard action joints in three-dimensional space.

[0087] The calculation method of the arc difference LV between the real action angle and the standard action angle includes:

[0088]

[0089] B3. Use data such as standard action status, emotional state, injury status, and the difference between real action and standard action as feature fields, mark the performance scores, and use the lightGBM algorithm to continuously perform training and optimization.

[0090] B4. Obtain real-time training and test videos, repeat steps B1 and B2 to obtain data field information such as the difference between the real action and the standard action, emotional state (normal and abnormal), injury status (mild, moderate, and severe), etc., and input them into the trained lightGBM algorithm regression in B3 to give a real-time score and assess the ability level.

[0091] Step 2: Construct portraits of the training and examination texts obtained above (including: training and examination results, training and examination plans, training and examination personnel information, training and examination physical fitness information and skill subject information, etc.) according to individual and group training indicators (including: psychological indicators, morphological indicators, physiological indicators, biochemical indicators, mechanical indicators, sports injury type indicators) and five-dimensional radar charts (including: speed, endurance, strength, flexibility, and agility) to obtain health status values; analyze the relationship between physical fitness and skills, the relationship between quality and skills, and the relationship between quality and physical fitness, analyze weights, analyze the strengths and weaknesses of individuals and groups in various training indicators and their changing trends over time, and combine the analysis results of the training and examination videos in step 1 to predict risks and give physical abnormality warnings.

[0092] Specifically, the steps include:

[0093] C1. Analyze historical and real-time physical fitness and skill training and test data and personal information data for individuals and groups, and analyze physical fitness test subjects T n Subjects for performance and skills assessment J m Performance and quality indicators x Use linear regression algorithm to calculate the weight W corresponding to the quality index in the physical fitness assessment subject score xn And the corresponding weight W of the quality index in the skill assessment subject score xm , thus obtaining the historical and real-time five-dimensional radar chart. The five-dimensional radar chart value can be calculated according to the following formula:

[0094]

[0095] In the formula, M1 and M2 represent the number of physical fitness assessment subjects and the number of skill assessment subjects respectively.

[0096] At the same time, according to historical and real-time psychological indicators, morphological indicators, physiological indicators, biochemical indicators, mechanical indicators, sports injury type indicator data S y , using the neural network algorithm to supervise the training weights W aa and W bb , get the health status value: S = W aa S x +W bb S y .

[0097] C2. Use the significance test to conduct correlation analysis on the quality indicators of physical fitness assessment subjects and skill assessment subjects respectively. According to the historical training and test data, use the linear regression algorithm to calculate the weight W corresponding to the quality indicators of physical fitness assessment subjects. nx And the weight W corresponding to the quality index of the skill assessment subject mx, and then obtain the relationship between the scores of physical fitness assessment subjects and skill assessment subjects.

[0098] in,

[0099] Both

[0100] Both

[0101] In the formula, M3 represents the number of quality indicators; S nx Indicates the quality value of the physical fitness assessment subject corresponding to the quality index; S mx Indicates the quality value of the skill assessment subject corresponding to the quality indicator.

[0102] C3. By monitoring the real-time training and test data of individual and group students, we analyze the strengths and weaknesses of individual and group indicators and their changing trends over time. Combined with the analysis results of the training and test videos in step 1, we use the support vector machine algorithm to predict the risk threshold. By comparing the relationship between the predicted value and the risk threshold, we give a physical risk warning.

[0103] Step 3: Use big model technology to intelligently generate relationships, attributes and entities for students' historical training and examination information, such as training and examination subjects, scores, videos, personnel, indicators, plans, etc., build knowledge graphs, establish causal networks, enhance reasoning capabilities, and analyze potential connections through correlation mining. Analyze each person's health status, ability level, shortcomings and strengths, and then optimize risk thresholds. The big model can give risk warnings and guidance suggestions, and continuously adjust the results based on the questions and answers between users and the big model.

[0104] Step 4: Intelligently group different groups of people and provide personalized training plans for different groups. The training plans can be divided into adaptation period, development period, sprint period and assessment period. Correct exercise posture, training plan and dietary suggestions can be given. Knowledge services such as text, video and pictures as well as targeted guidance and exercise rehabilitation prescriptions can be provided to injured persons. Real-time monitoring of physical condition can be carried out and warnings can be given.

[0105] Specifically, the steps include:

[0106] D1. Based on the five-dimensional radar charts, psychological indicators, morphological indicators, physiological indicators, biochemical indicators, mechanical indicators, and sports injury indicators of individuals and groups obtained above, the random forest algorithm is used to group the trainees according to individuals and groups for training. The grouping categories include: sports injury group SG, psychological training group XG, basic training group CG, skill training group JG, and comprehensive training group ZH. Each group is divided into primary pr, intermediate mi, and advanced hi. Each level can be divided into adaptation period sy, development period fz, sprint period cc, and assessment period kh. The Y level P period under the X group can be recorded as X YP ,X∈(SG,XG,CG,JG,ZH),Y∈(pr,mi,hi),P∈(sy,fz,cc,kh). The training program includes dietary advice recorded as u1, medicine recorded as u2, training plan recorded as u3, exercise courseware u4, and correct exercise posture video recorded as u5. u∈(u1,u2,u3,u4,u5). The training program u under X group, Y level and P period can be recorded as: X YPu .

[0107] D2: According to step D1, the training video and text are supervised by the neural network algorithm, and irrelevant training videos and texts are filtered out to obtain (u1, u2, u3, u4, u5). According to the given design template data, the training plan X is formed. YPu .

[0108] D3. Further expand the knowledge graph of the generated training plan, deeply explore the correlation between students' training and testing data, generate individual and group analysis reports, and provide scientific and customized training plans based on the situations of different personnel.

[0109] D4. Monitor the execution process of the personalized training plan in real time, update the five-dimensional radar chart value and health value in step 3 step 2, and give a real-time alarm when data abnormalities occur.

[0110] Step 4: Use reinforcement learning to find the optimal training plan (u1, u2, u3, u4, u5), optimize and adjust the training plan, and continuously improve the training results. Each person's training and testing process can be defined as an intelligent agent Actor-Critic architecture. The Actor network gives a corresponding training plan based on the status of real-time training and testing analysis data, and the Critic network gives an effect evaluation based on the status of real-time training and testing analysis data.

[0111] The training and testing status is a collection of five-dimensional radar chart values ​​and health values, and the action is that the intelligent agent gives a corresponding training plan based on the status of real-time training and testing analysis data.

[0112] Total reward: r = w1r1 + w2r2 + w3r3. The closer to the training target value, the larger the value.

[0113] Where r1=-(V A -V s ) 2 ,

[0114]

[0115] r3=-γ t ,

[0116] Where V A represents the actual training evaluation value, V s represents the training target value, γ t is the optimal time coefficient. If the time exceeds a certain range, a negative reward value is given. The farther away from the ideal time range, the smaller the value.

[0117] Step 5: You can use voice and text to conduct questions and answers on the training assessment content, quickly obtain individual and group training reports, analysis charts and training videos, and recommended web pages to visit, give assessment results, and help students check the training status and answer questions in a timely manner.

[0118] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. An intelligent training and assessment method based on multi-source data analysis, characterized in that: The following steps are involved: Step 1: Collect historical training and test information and real-time training and test information to obtain multi-source training and test data; Step 2: Process the multi-source training and testing data using a data analysis algorithm to obtain multi-source effective training and testing data; The multi-source valid training and examination data includes: training and examination text content, training and examination documents, and training and examination videos; Step 3: Based on the evaluation algorithm, the multi-source effective training and test data are integrated and analyzed to obtain an analysis report; and based on the analysis report, intelligent grouping is performed to provide personalized training program guidance and physical abnormality warnings for different groups; Step 4: Optimize and adjust the training plan using reinforcement learning, and evaluate the effectiveness of the training plan based on real-time training and testing data; Step 5: Conduct questions and answers on the training assessment content through voice and text, quickly obtain individual and group training reports, analysis charts and training videos, and recommended web pages to visit, and give assessment results. The assessment results are used to help students check the training status and answer questions in a timely manner.

2. According to claim 1, the intelligent training and assessment method based on multi-source data analysis is characterized in that: In step 2, the steps of the data analysis algorithm include: Step 1: parse the text data of the training and examination documents in the multi-source training and examination data, extract tables, seals, column paragraphs, pictures, and formula data, and obtain the training and examination text content; Step 2: Use the YOLO algorithm to detect blurry frames in the video information of the multi-source training data, and introduce an attention mechanism to capture blurry frame information; use the RNN network to process the blurry frames to obtain clear training videos; Step 3: Perform data governance on the parsed data to obtain the multi-source effective training and testing data.

3. According to claim 2, the intelligent training and assessment method based on multi-source data analysis is characterized in that: Step 1 includes the following steps: A1. Using a neural network algorithm to detect the content of the training text, and identifying complex training text and simple training text; A2. Constructing a generative network and a discriminative network to process the complex training text, wherein the discriminative network is used to distinguish the training text and the training format; and the generative network is used to generate false data similar to the training text and the training format; A3. Train and optimize the constructed generative network and discriminative network to obtain the optimal u-type text content generated by the generative network; A4. Use the CRNN algorithm to perform text recognition on the text content and the simple training text, draw lines according to the coordinate sequence of the characters, and use a neural network to classify the drawn lines into common reading sequences to obtain text content with a correct reading sequence.

4. According to claim 3, the intelligent training and assessment method based on multi-source data analysis is characterized in that: The objective function in A2 is: Where u represents the number of training and test texts and training and test formats; i represents the weight; D is the mapping function for distinguishing the training text and the training format; G is the mapping function for the generator training text and the training format; x i The data consists of real training and test text and training and test format; data is the distribution of data composed of real training and test texts and training and test formats; i represents the noise of random training text and training format; p z Represents the noise distribution of random training text and training format.

5. According to claim 1, the intelligent training and assessment method based on multi-source data analysis is characterized in that: The assessment algorithm in step 3 includes the following steps: Step 1: Obtain the trainee's training and examination video in real time, and use computer vision technology to analyze the training and examination video to obtain video analysis results; the video analysis results include: the trainee's emotional state, physical injury status, and athletic ability; Step 2: Based on the text content, construct a portrait of the trainees and the group from the aspects of training indicators and five-dimensional radar charts to obtain health status values; and analyze the strength and change trend of the training indicators of the trainees and the group over time based on the correlation between physical fitness, skills and qualities; and predict the risk threshold based on the training and test video data, and warn of physical abnormalities based on the risk threshold; Step 3: Use big model technology to analyze the students' historical training and examination information, and then optimize the risk threshold; Step 4: Intelligently group the students, provide personalized training plans for different groups, monitor their physical condition in real time, and issue warnings for physical abnormalities.

6. According to claim 5, the intelligent training and assessment method based on multi-source data analysis is characterized in that: Step 1 includes the following steps: B1. Use neural network algorithms to classify history training videos into three categories, including face, lower limb gait, and whole body movement videos; use Transformer neural network to extract 3D features of head facial expressions, lower limb gait movements, and whole body movement movements; B2. Use 3D human posture estimation algorithm to identify movement changes, and identify the standard state, emotional state and injury state of movement based on graph convolutional neural network. Use 3D reconstruction technology to display the position of real movement and the amplitude of standard movement, optimize the details of movement generation, and then compare the gap between real movement and standard movement. B3. Use the standard state of action, emotional state, injury state, and the difference between the real action and the standard action as feature fields, mark the performance scores, and use the lightGBM algorithm for training and optimization; B4. Obtain real-time training and test videos, repeat B1 and B2 to obtain the difference between the real action and the standard action, the emotional state, and the field information of the injury state, and input them into the trained lightGBM algorithm in B3 for regression to give a real-time score.

7. According to claim 6, the intelligent training and assessment method based on multi-source data analysis is characterized in that: The loss function of the 3D reconstruction technology includes: L(o,q)=L o +λ re 50 q , in, In the formula, o represents the pixel set of each batch; q represents the pixel set of each batch; L o Indicates the pixel loss between blurred pixels and real pixels; L q Represents the distance loss between the 3D reconstructed model and the real model; re represents the distance loss weight; K represents the pixel set; H o Indicates blurred pixels; Represents real pixels; I q Represents the reconstruction model; Represents the true model.

8. According to claim 6, the intelligent training and assessment method based on multi-source data analysis is characterized in that: The gap includes: angle gap and arc gap; The calculation method of the angle difference includes: In the formula, (x aa ,y aa 、z aa ) is the coordinate of the real action joint in three-dimensional space; (x bb ,y bb 、z bb ) is the coordinate of the joint point where the real action and the standard action intersect in three-dimensional space; (x cc ,y cc 、z cc ) are the coordinates of the standard action joints in three-dimensional space; The calculation method of the arc difference includes:

Images