An artificial intelligence-based sports teaching data management system

Abstract

The application relates to the field of educational informatization and big data processing technology, and discloses a sports teaching data management system based on artificial intelligence, which comprises a processor and a memory and is used for executing the following steps: acquiring original logs of sports teaching process data, parsing the original logs into an event record table and an identity attribute table, and generating a multi-source session identifier; in a teaching statistical period, calculating an identity redundancy coefficient according to the total amount of the multi-source session identifier and a real user base, and calculating an access source distribution index and a cross-tenant switching frequency as noise parameters; constructing a multi-dimensional vector comprising an access source domain name, a tenant context and a time dimension and inputting the multi-dimensional vector into a neural network model, the model is trained by using a loss function with the identity redundancy coefficient as a dynamic weight, so as to output a normalized user primary key; generating sports process indicators by using the normalized user primary key for aggregation, and calculating a data calibration coefficient by combining the identity redundancy and the noise parameters, and outputting sports teaching management reports after the indicators are corrected.

Description

Technical Field

[0001] This invention relates to the field of educational informatization and big data processing technology, and more specifically, to an artificial intelligence-based sports teaching data management system. Background Technology

[0002] With the deepening development of educational informatization, digital teaching systems have been widely used in educational institutions at all levels. Especially in the field of physical education, the widespread adoption of smart wearable devices (such as wristbands), AI cameras, and mobile teaching terminals has generated massive amounts of process data during the teaching process. This data covers students' movement trajectories, physiological indicators (such as heart rate), classroom interaction records, and skill achievement status, containing extremely high analytical value and providing a basis for teaching quality assessment and personalized guidance. Despite the increasing hardware data collection capabilities, existing physical education teaching data management systems still face the following serious challenges in practical applications, making it difficult to meet the needs of refined management:

[0003] Data heterogeneity and the silo effect: Physical education teaching data often comes from different manufacturers and different types of terminal devices (such as third-party wristbands, campus cards, WeChat mini-programs, etc.), and the data formats are unstructured and inconsistent. Existing management systems mostly use static tables or simple databases for storage, lacking the ability to uniformly parse and deeply clean these multi-source heterogeneous logs, resulting in data being scattered in different silos, making it difficult to perform global correlation analysis.

[0004] In real-world educational management scenarios, there are complex situations involving multiple entry points and multiple tenants. The same student may access the system via WeChat, an app, or a web browser, and may be in different classes or tenant environments across different semesters. Current technology often tags these behavioral records from different sources with different identifiers (such as OpenID, LoginName, etc.), lacking an effective identity normalization mechanism. This makes it impossible for the system to recognize that these fragmented IDs actually belong to the same real user, resulting in problems such as inflated user base statistics and fragmented personal profile data.

[0005] Drawing inspiration from Customer Experience Management (CEM) in enterprise management, the value of data lies in closed-loop improvement. However, existing education management systems lack the ability to assess data noise. For example, when there are numerous cross-tenant handovers or frequent changes in access sources, the generated data often contains redundancy and uncertainty. Existing systems typically only perform simple summation and statistics on raw data, without considering the impact of identity redundancy on indicators, and lack data calibration mechanisms based on environmental parameters (such as access source distribution and handover frequency). This leads to biased final teaching reports that fail to accurately reflect the effectiveness of the teaching process, misleading management decisions. Summary of the Invention

[0006] This invention provides an artificial intelligence-based sports teaching data management system, which solves the technical problems mentioned in the background art.

[0007] This invention provides an artificial intelligence-based sports teaching data management system, including a processor and a memory, wherein the processor executes instructions stored in the memory to achieve the following steps:

[0008] Obtain the raw logs of physical education teaching process data, parse them into an event record table and an identity attribute table, and generate a multi-source session identifier corresponding to each record according to the field parsing rules;

[0009] Within the preset teaching statistics period, the identity redundancy coefficient is calculated based on the total number of multi-source session identifiers and the actual number of users in the system, and the access source distribution index and cross-tenant switching frequency are calculated simultaneously as noise characteristic parameters.

[0010] A multi-dimensional feature vector containing the access source domain name, tenant context, and time dimension is constructed and input into a neural network model. The neural network model is trained using a loss function with the identity redundancy coefficient as the dynamic weight parameter, thereby outputting a deterministic normalized user primary key for each event record.

[0011] The event records are aggregated using the normalized user primary key to generate sports process indicators. Data calibration coefficients are calculated based on the identity redundancy coefficient and the noise characteristic parameters to numerically correct the sports process indicators and output a sports teaching management report.

[0012] The beneficial effects of this invention include: by constructing a multi-dimensional vector containing access source, tenant context, and time dimension, and utilizing a neural network model with identity redundancy as dynamic weights, a normalized mapping of user identities across scenarios is achieved. Simultaneously, the system introduces an identity redundancy coefficient and noise parameters to construct a data calibration mechanism, which can automatically eliminate the interference of architectural noise on statistical indicators and numerically correct process indicators, thereby significantly improving the accuracy and decision-making effectiveness of sports teaching management data. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the architecture of a sports teaching data management system based on artificial intelligence according to the present invention. Detailed Implementation

[0014] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0015] like Figure 1 As shown, a sports teaching data management system based on artificial intelligence includes a processor and a memory. The processor executes instructions stored in the memory to achieve the following steps:

[0016] Obtain the raw logs of physical education teaching process data, parse them into an event record table and an identity attribute table, and generate a multi-source session identifier corresponding to each record according to the field parsing rules;

[0017] Within the preset teaching statistics period, the identity redundancy coefficient is calculated based on the total number of multi-source session identifiers and the actual number of users in the system, and the access source distribution index and cross-tenant switching frequency are calculated simultaneously as noise characteristic parameters.

[0018] A multi-dimensional feature vector containing the access source domain name, tenant context, and time dimension is constructed and input into a neural network model. The neural network model is trained using a loss function with the identity redundancy coefficient as the dynamic weight parameter, thereby outputting a deterministic normalized user primary key for each event record.

[0019] The event records are aggregated using the normalized user primary key to generate sports process indicators. Data calibration coefficients are calculated based on the identity redundancy coefficient and the noise characteristic parameters to numerically correct the sports process indicators and output a sports teaching management report.

[0020] In a preferred embodiment, the raw log of the physical education teaching process data is obtained and parsed into an event record table and an identity attribute table, including:

[0021] The original log is parsed into an event record table with the following structure. With identity attribute table :

[0022] ;

[0023] ;

[0024] in, Represents a timestamp. Indicates course identification. Indicates class identification. Indicates the event type, Represents numerical values. Indicates a multi-source session identifier; Indicates the tenant context. Indicates the source domain name being accessed. Indicates identity token, Indicates the login name. This represents the global user identifier; for fields missing in the data source, it is filled with NULL.

[0025] In detail, the physical education teaching data acquisition system has completed communication integration with campus physical education teaching terminals (including teacher tablets, student smart bracelets, and venue AI cameras). Each terminal uploads raw log data streams of the physical education teaching process at a preset frequency of 1 second per upload. This data stream is in unstructured key-value pair format and needs to be converted into a structured event log table through a preset parsing program. With identity attribute table In designing the structure of these two tables, the behavioral events that occur during physical education teaching are distinguished from the user identity information that triggers those events. This decouples the storage of behavioral data from identity data, facilitating subsequent multi-dimensional correlation analysis. Specifically, It carries the timing and attribute information of behavioral events. This information serves as a linking identifier for the user's identity across different scenarios, linked by a timestamp. With multi-source session identifiers Establish the association. During the parsing process, the parser first extracts the time field from the original log and converts it into an integer timestamp in seconds. This parameter is used to chronologically sort all events. Its preferred value is a second-level timestamp based on Coordinated Universal Time (UTC), obtained by synchronizing the system clocks of each terminal device via Network Time Protocol (NTP). Subsequently, the encoded field identifying the current teaching course is extracted from the log and mapped to... This parameter is used to associate the corresponding teaching syllabus content, and its value is a unique character code pre-assigned in the campus course management system; then, the associated class code is extracted from the log and mapped to... This parameter is used to define the teaching organization unit to which the event belongs, and its value is a unique character code pre-assigned in the campus student registration management system. For The parsing program maps the enumerated values ​​of the behavior description fields in the logs (such as attendance, training, and achievement test) to preset integer enumerated values ​​as... Its preferred enumeration range is This corresponds to 10 preset types of physical education teaching events; then, the quantitative data corresponding to the behavioral events in the logs are extracted and directly used as... The value type of this parameter depends on Dynamic adaptation, for example, when When counting jump ropes, For integer values, when When scoring according to the correct action specifications, The value is a floating-point number; at this time Fields are reserved initially and will be populated after the identity attribute information is parsed. (For...) The parsing program extracts the field identifying the current teaching application deployment tenant from the logs and maps it to... This parameter is used to distinguish independent deployment environments in different schools or teaching areas. Its preferred value is a character-based tenant code, obtained based on the unique identifier assigned to each teaching unit during system deployment. Then, the domain name field identifying the terminal device's network access is extracted from the logs as... This parameter is used to distinguish the network domain to which the triggering terminal of the event belongs. Its value is a preset domain name enumeration value (e.g., teacher.school.edu.cn, student.school.edu.cn), obtained based on the system's pre-configured list of terminal access domain names. Then, the identity token obtained by the terminal device through third-party authentication is extracted from the logs and used as... This parameter is used to associate the user's identity with the third-party system. If this field is not present in the log, it is filled with NULL. Then, the user's login account field within the current system is extracted from the log and used as... This parameter is used to associate local user accounts within the system. If this field is not present in the log, it is filled with NULL. Finally, the system-pre-assigned unique user identifier field is extracted from the log and used as... This parameter is used to associate global user profiles within the system. If this field does not exist in the log, it is filled with NULL. After extracting and mapping the above fields, the parsing program writes the values ​​of each field to... and In the corresponding columns, for any missing field, a NULL value is directly written. This ensures the integrity of the table structure and prevents the parsing program from terminating abnormally due to missing fields. Simultaneously, the parsing program will... and In Perform consistency checks to ensure that the two tables corresponding to the same event are consistent. If the values ​​are exactly the same, and the verification passes, the parsing result will be stored in the corresponding table in the distributed database.

[0026] In a preferred embodiment, a multi-source session identifier corresponding to each record is generated according to field parsing rules, including:

[0027] Define the multi-source session identifier The generation rules are as follows:

[0028] ;

[0029] in, Indicates identity token, Indicates the tenant context. Indicates the source domain name being accessed. This represents the login name; this rule is used to structure identity fragments generated by multiple entry points, ensuring that each record obtains a unique multi-source session identifier.

[0030] In detail, in the completed event log sheet With identity attribute table After parsing the basic fields, the parsing program will generate a corresponding multi-source session identifier for each record. Specifically, considering the fragmented characteristics of user identity identifiers in multi-entry, multi-tenant scenarios, a unique session association identifier is constructed through hierarchical combinations of identity parameters. First, the parsing program will... In The field is checked for nullability. The preferred logic for this check is to detect that the field value is not NULL and the character length is greater than 0. This logic is based on the database's standard for storing null values. When not empty, The generation method is ,in Provides a unique identifier for users within third-party systems. Limit the tenant environment in which the identity is located to avoid the same identity being used by different tenants. confusion, Distinguish the domain name to which the access terminal belongs to the identity to avoid identity confusion between different access sources under the same tenant; when When empty, the parser will call In Field, at this time The generation method is ,in Provides the user's local login identifier within the current system. Also used to limit the tenant environment and prevent the same [conditions / conditions] from being used by different tenants. The confusion arises from the fact that, in a multi-entry scenario, the same user might use different entry points. In access systems and multi-tenant scenarios, the same identity may be used by different... The corresponding tenants are reused, therefore the user's identity identifier ( or ) and environmental parameters ( , Combine them, and prioritize those with greater global uniqueness. As a core identity parameter, only when The only one in the system is used when it is unavailable. This ensures that each record Maintain uniqueness across multiple scenarios while achieving structured association of identity fragments.

[0031] In a preferred embodiment, within a preset teaching statistical period, an identity redundancy coefficient is calculated based on the total number of multi-source session identifiers and the actual user base of the system, including...

[0032] The identity redundancy coefficient is calculated using the following formula. :

[0033] ;

[0034] in, This indicates the teaching statistics period; This indicates the multi-source session identifier; Represents the global user identifier; This indicates the deduplication count within the set;

[0035] The identity redundancy coefficient is used to quantify the degree of identity fragmentation within the teaching statistical period.

[0036] In detail, before calculating the identity redundancy coefficient, the preset teaching statistical period should be determined first. The preferred value is a continuous 604,800 seconds (corresponding to the length of a week). The logic behind setting this value is to match the weekly progress management rhythm of school physical education teaching. Subsequently, from the event log table... Extracting from All records within Fields, these are deduplicated through database lookup operations. Filter to obtain only those containing unique values. set , This refers to the number of elements in the set, which corresponds to the total number of all independent structured identity fragments within the teaching statistical period. Simultaneously, from the identity attribute table... Extract the same All records corresponding to The field, also through a deduplication query, yields only fields containing unique values. set , This refers to the number of elements in the set, which corresponds to the total number of real users actually participating in physical education teaching within the teaching statistics period. Specifically, considering multi-entry, multi-tenant scenarios, the same real user may generate multiple different... In this case, the cardinality ratio of the two sets is used to associate the correspondence between identity fragments and real users, thereby reflecting the distribution of identity fragmentation within a period. Specifically, the calculation involves the system's data analysis module calling the database aggregation interface to obtain the two aforementioned quantitative values, which are then substituted into the formula:

[0037] The division operation yields the result. The entire calculation process is completed in the module's memory operation unit, and the result will be stored in the corresponding field of the analysis result table.

[0038] In a preferred embodiment, calculating the access source distribution index and cross-tenant handover frequency as noise characteristic parameters includes:

[0039] The access source distribution index is calculated using the following formula. With the cross-tenant handover frequency :

[0040] ;

[0041] ;

[0042] in, This indicates the teaching statistics period; This represents the count of a specific access source domain name. Indicates the total count. This indicates the distribution ratio of the access source domain name; This represents a global user identifier. Indicates the tenant context; This represents the total number of deduplicated global user identifiers; the access source distribution index reflects the degree of dispersion of entry points, and the cross-tenant switching frequency reflects the frequency with which the same user crosses different tenant contexts. Within a preset teaching statistical period... (The preferred value is a continuous 604,800 seconds, which matches the weekly progress management rhythm of school physical education teaching.) Within this timeframe, the access source distribution index is first calculated. From the identity attribute table Extract all records within that period. Fields, for each specific Count the number of records corresponding to it within this period, and record this number as . At the same time, statistics are compiled for all data within this period. The corresponding total number of records, this total number is denoted as ,at this time For that specific The distribution ratio among all access sources is determined by and The ratio is obtained, that is The design concept of this index is to use information entropy to quantify the dispersion of access sources. The more dispersed the access sources are, the more dispersed they become. The smaller the numerical difference, the higher the corresponding entropy value will be, and then all of them will be... corresponding Substitute into the formula The calculation is completed by calling the natural logarithm operation function and the accumulation operation function through the data analysis module. Indicates all different corresponding Perform an accumulation operation. Then calculate the cross-tenant handover frequency. From the identity attribute table Extract All records within and The association pairs are for each specific global user identifier. Statistical analysis of the different corresponding periods The quantity, which is denoted as At the same time, statistics The total number of unique user identifiers deduplicated within the global scope, denoted as Specifically, through the different... The average number is used to reflect the switching status of the same user in a multi-tenant environment, and then all of them are... corresponding Add up, then combine Take the ratio, that is:

[0043] ,in Indicates all different corresponding The accumulation operation is performed, and this calculation is first performed by the data analysis module through grouped queries to obtain each... corresponding After removing duplicates, summation and division operations are performed to complete the process, resulting in the final result. and The calculation results will be stored in the corresponding fields of the analysis results table.

[0044] In a preferred embodiment, a multi-dimensional feature vector is constructed, comprising the access source domain name, tenant context, and time dimension, including:

[0045] The multidimensional feature vector is constructed using the following formula. :

[0046] ;

[0047] in, This represents the embedding vector generated by embedding mapping of discrete fields; This represents the hash bucket count vector generated by hash mapping a string; Represents a numeric field; Indicates the time dimension The periodic encoding of is calculated using the following formula:

[0048] ;

[0049] in, It represents the number of seconds in a day. The total number of seconds in a day;

[0050] The multidimensional feature vectors are used to explicitly encode architectural information.

[0051] Detailed, within the pre-set teaching statistics period Internally, construct multi-dimensional feature vectors. At that time, the discrete fields are processed first. , , , , Call the embedding mapping function for each field , The preferred output is a 32-dimensional dense vector, generated by a shallow neural network trained on historical sports teaching data. During training, the category value of the field is used as input, and the corresponding teaching event correlation degree is used as the supervision signal. Because the discrete fields are categorical data, directly inputting them into the model in numerical form would introduce irrelevant order relations. Embedding mapping transforms them into a low-dimensional unbiased vector representation, facilitating subsequent feature learning by the model. This process yields... , , , , Next, process the string field. and Call the hash mapping function , The preferred output is a 2048-dimensional hash bucket count vector. This is generated by mapping the string to an integer index using the SHA-256 hash function, then setting the bucket position corresponding to that index to 1 and the rest to 0. Because string fields have variable lengths and no unified encoding rules, hash mapping can transform them into a fixed-dimensional vector while preserving the uniqueness of the fields. , Then process the time dimension. Calculate its periodic code ,in The preferred value is 86400 seconds (corresponding to the total number of seconds in a day). for right The number of seconds in a day obtained after taking the modulus will be... and Substitute into the formula:

[0052] The system obtains a two-dimensional vector through its trigonometric function operation unit. Since the time dimension has periodic characteristics, it can be directly used... Numerical values ​​can mask temporal relationships within the same time period; trigonometric function encoding can preserve the periodicity of time. Finally, numerical fields... These vectors are directly used as one-dimensional vectors in the concatenation process. The vectors obtained above are then concatenated sequentially using a vector concatenation operation (denoted as...). The combination of features yields a multidimensional feature vector. The generation process of this vector is automatically executed by the feature engineering module of the system, and the generated result is stored in the corresponding record of the feature vector table.

[0053] In a preferred embodiment, the pre-training of the neural network model includes:

[0054] Weighted cross-entropy is used as the loss function for the neural network model. :

[0055] ;

[0056] in, Indicates training samples, To monitor the target, Output probabilities for the model; For teaching statistics period The dynamic weights are calculated using the following formula:

[0057] ;

[0058] in, Represents the identity redundancy coefficient. Indicates the access source distribution index. Indicates the frequency of cross-tenant handover; It is a preset fixed constant;

[0059] The dynamic weights are used to reduce the impact of high-noise cycles on model training.

[0060] In detail, before pre-training the neural network model, pre-defined fixed constants are first determined. , , The preferred values ​​are, where Take 0.7, Take 0.2, The value is set to 0.5. This value is chosen to match the impact weights of identity redundancy, access source distribution, and cross-tenant switching on training noise, which meets the noise suppression requirements in multi-entry, multi-tenant scenarios. Subsequently, the teaching statistical period is obtained. Training samples within ,in For the constructed multidimensional feature vector, The supervision target corresponding to this sample is the global user identifier. The corresponding category label, which is generated by the system from the identity attribute table. Extract and map these values ​​to integer category values. Next, calculate the teaching statistical period. Dynamic weights First, call the data already stored in the analysis results table. , , Substitute it into the formula:

[0061] Specifically, when Increase (the degree of identity fragmentation) Increase (the degree of dispersion of access sources) When the frequency of cross-tenant handover increases, The value of will be reduced accordingly to suppress the weights of high-noise periodic samples. Then, the training samples are input into the neural network model, and the model outputs the probability corresponding to the sample through forward propagation. This probability is the result of the model's output layer being processed by the softmax activation function, representing the model's prediction of the supervised target corresponding to this sample. The confidence level was determined. Finally, weighted cross-entropy was used as the loss function. Through the All training samples Performing the accumulation operation, we get:

[0062] Specifically, through dynamic weights Differential weights are assigned to samples from different statistical periods to prevent samples from high-noise periods from dominating model parameter updates. Throughout the pre-training process, the model minimizes the model using the backpropagation algorithm. The system performs iterative optimization of the parameters. The optimization process is executed by the system's model training module by calling GPU computing resources. After training, the model parameters are stored in the model parameter library.

[0063] In a preferred embodiment, a multidimensional feature vector is input into a neural network model to output a normalized user primary key for each event record, including:

[0064] The normalized user primary key is calculated using the following formula. :

[0065] ;

[0066] ;

[0067] in, Represents a multidimensional feature vector; Represents a multilayer perceptron network; and Indicates network parameters; This represents the normalized exponential function; Indicates belonging to a certain global user identifier The probability of; This indicates the operation of selecting the independent variable that maximizes the function value.

[0068] In detail, after the neural network model completes pre-training, the first step is to determine... The preferred structure is a three-layer fully connected structure, with 1024, 512, and 256 neurons in each layer, respectively. The logic behind this structure is to match the dimensions of the multi-dimensional feature vector, compressing the feature dimensions layer by layer to retain key identity-related information, and employing the ReLU activation function for non-linear transformation in each layer. Subsequently, the constructed multi-dimensional feature vector... Enter to middle, Call the parameters obtained from pre-training Perform forward computation to output a fixed-dimensional abstract feature vector, allowing the model to learn the relationships between the dimensions of the feature vector and extract features that can distinguish different user identities. Then, for... The output of performs a linear transformation operation, where The output layer weight matrix obtained through pre-training has dimensions corresponding to... Output dimensions and global user identifier Total quantity The output layer bias vector obtained from pre-training has the same dimension as the global user identifier. The total number is consistent, and the function of this linear transformation is to map the abstract features to each The corresponding score space. Then, the softmax function is called to normalize the result of the above linear transformation, resulting in... This function can convert each The corresponding scores are transformed into a probability distribution, such that all corresponding The sum of these values ​​equals 1, thus transforming the score into a quantifiable confidence level, which facilitates the identification of the user identity with the highest match to the current feature vector. Finally, for all global user identifiers... corresponding implement Operation, selection The largest value ,get:

[0069] Specifically, the probability distribution is used to determine the current multidimensional feature vector. The most relevant global user identifier enables the normalization mapping of multi-source identity fragments, with each event record corresponding to... It will be associated and stored in the event log table. In the corresponding fields, the entire process is automatically executed by the system's identity normalization module calling the pre-trained model.

[0070] In a preferred embodiment, the event records are aggregated using the normalized user primary key to generate sports process indicators, and a data calibration coefficient is calculated based on the identity redundancy coefficient and the noise characteristic parameters to numerically correct the sports process indicators, outputting a sports teaching management report, including:

[0071] The following formula is used to aggregate and generate sports process indicators:

[0072] ;

[0073] ;

[0074] ;

[0075] in, Representing users respectively In the cycle The number of times students attended classes, the number of training records, and the number of achievement reports within the specified timeframe; For indicator functions; For the event Normalized user primary key;

[0076] Calculate the data calibration coefficient using the following formula. And revise the indicators:

[0077] ;

[0078] ;

[0079] in, It is a fixed constant; This is the class average indicator. The revised indicator;

[0080] Generate computational signature And output a report:

[0081] ;

[0082] in, Represents a hash function; This indicates string concatenation.

[0083] Detailed, within the pre-set teaching statistics period First, aggregate and generate process-related indicators for physical education. For example, the number of class attendance... traversal Every event within Call the event log table China Event corresponding Fields and (event (normalized user primary key), through indicator function determination Is it and Equal to global user identifier This function outputs 1 if the condition is met, and 0 otherwise, for all events. The judgment result is accumulated to obtain:

[0084] Specifically, this is achieved by normalizing user primary keys. The same user The system aggregates and associates class attendance events for different user identities to avoid duplicate or missed statistics for the same user's class attendance events in multi-entry scenarios. Next, it calculates the number of training records. Similarly, traversal Every event within ,determination Is it and equal The output of the indicator function is accumulated to obtain:

[0085] This process can aggregate the same user All training events recorded within the period. The number of qualifying texts is then calculated. traversal Every event within ,determination Is it ,event corresponding Is it 1 and equal The output of the indicator function is accumulated to obtain:

[0086] To count users The number of events meeting the criteria within the period is used to calculate the data calibration coefficient. First, determine the fixed constant. , The preferred values ​​are, where Take 0.1, The value is set to 0.3. The logic behind this setting is to match the weighting of the impact of access source distribution and cross-tenant handover on the indicator deviation, and then call the data stored in the analysis results table. , , Substitute into the formula:

[0087] Specifically, through The reciprocal of the formula offsets the metric inflation caused by identity redundancy. Furthermore, by using the access source distribution index and the index term of cross-tenant switching frequency, the metric bias in multi-entry, multi-tenant scenarios is further corrected. Then, the class average metric is calculated. This metric represents the global user identifiers within the class. corresponding (or , The arithmetic mean of ) and then and Perform the multiplication operation to obtain the corrected class average. This step allows for the correction of deviations in class-level physical education process indicators. Finally, a calculation signature is generated. Call the hash function (preferably SHA-256) to... , , The string form, and All users corresponding The sum of , The sum of , The sum of The strings are concatenated sequentially, and a hash operation is performed on the concatenated string to obtain:

[0088] Specifically, the core parameters involved in the indicator calculation are solidified through hash signatures, ensuring the traceability of the output sports teaching management reports. All indicators, correction results, and calculation signatures are integrated into the report template, and automatically exported as structured documents by the system's report generation module, stored in the corresponding period directory of the report database.

[0089] It is a structured event log table specifically adapted for physical education teaching scenarios. By stripping and standardizing behavioral events from unstructured logs, the structure is defined as follows:

[0090] This table is used to implement a three-dimensional binding of behavior time sequence, teaching attributes, and identity association. For example, a record could represent 2024-09-01 14:30. Basketball class =PE-B2024), Grade 8 (1) ( =G801), training events ( =2), jump rope 120 times ( =120), associated identity identifier ( =(o123,fid001,student.school.edu.cn)). Is with The accompanying identity attribute table is structured as follows: This is used to decouple behavioral events from identity information, avoiding redundancy in single-table fields. This table is accessed through... and of Alignment ensures that events occurring at the same time can be accurately associated with the corresponding user identity and context, such as a certain... The record can be represented as 2024-09-01 14:30 ( ), XX Middle School tenant ( =fid001), Student App Access ( =student.school.edu.cn), third-party authentication token ( =o123), Student ID ( =20240001), Global User Identifier ( =u8765), which fully records the user's access environment and identity dimensions. It involves enumerating and encoding the core behaviors in physical education teaching, and optimizing the range of enumerated values. Each integer value corresponds to a specific type of teaching event. For example, the enumeration value 1 corresponds to ATTEND, 2 corresponds to TRAIN_TEXT, 3 corresponds to PASS_TEXT, and 4 corresponds to SCORE. Fixed codes enable the system to quickly identify event types. It is an adaptation The dynamic quantification field is designed with the numerical type determined by the event type, ensuring the rationality and consistency of the quantified data. For example, when... When =2 (training events) and the specific behavior is skipping rope, For integers (e.g., 120, representing the number of jump ropes); when =3 (standard test) and the specific behavior is standing long jump, It is a floating-point number (e.g., 2.35, representing the long jump distance in meters); when When =4 (classroom integral), It is an integer (e.g., 5 represents the integral score). This is a third-party identity token extracted from the raw logs, used to adapt to the identity association requirements of third-party authentication in multi-entry access scenarios (such as WeChat Mini Programs, Apps, and Web pages). For example, when a student accesses the physical education teaching system through a WeChat Mini Program, the WeChat Open Platform will return a unique... (e.g., o123456), the system stores the token in This field serves as one of the core criteria for identifying the same user across platforms; if a student logs in using a local system account (without third-party authentication), then... The field is filled with NULL to ensure that identity information from different access methods can be stored in a structured manner. It is a globally unique user identifier pre-assigned by the system, used to provide a stable user baseline across scenarios and periods, unaffected by changes in access methods or tenant environments. For example, a unique identifier is generated when a student registers with the system. (e.g., u8765), regardless of whether it is subsequently generated through a WeChat mini-program ( =o123), local account ( =20240001) or different tenant environments ( (Change) access, This consistent approach serves as the core basis for statistically analyzing the actual user base, preventing duplicate or missed user counts due to fragmented user identification.

[0091] Parse the raw logs into structured format. and It involves behavior-identity decoupling and field standardization. First, regular expression matching is used to extract behavior-related fields such as time, course, class, behavior description, and quantitative data from unstructured key-value pair logs, and then mapping them to... The corresponding columns; then extract identity-related fields such as tenant, access domain name, third-party token, local login name, and global user identifier, and map them to... The corresponding column. Map the enumerated values ​​of the behavior description field to... The integer enumeration values ​​establish a fixed mapping relationship between natural language behavior descriptions and machine-recognizable integer codes. These mapping rules are pre-stored in the system configuration file and are automatically matched during parsing. For example, the configuration rules might be set to: 1. Class → 1. Training → 2. Achievement Test → 3. Class Points → 4. Leave of Absence → 5. etc. When the behavior description in the original log is "training," the system automatically queries the mapping rules and assigns the corresponding values. The value is assigned as 2; if the behavior description is a compliance test, the value is assigned as 3 to ensure coding consistency for the same behavior in logs from different sources. Dynamic adaptation The value type is determined by a preset event type-numeric type mapping rule during parsing. Automatically converts data types to avoid numerical storage errors. For example, the preset rule is: until class (1) → none. (or NULL), Training (2) → Integer, Standard Test (3) → Floating Point, Classroom Integral (4) → Integer, when When =2 (training), regardless of whether the quantized data in the original log is the string "120" or the number "120", it is converted to integer storage; when When the value is 3 (compliance test), the original data 2.35 is converted to floating-point storage to ensure... The type matches the event quantification requirements.

[0092] right and The same event corresponding to Performing consistency checks ensures that behavioral events and identity information are synchronized in time, avoiding association errors caused by log transmission delays. The check logic is as follows: Search each record middle If identical records exist, the verification passes; if they do not exist, they are identified as records with inconsistent times and marked, requiring manual verification or automatic discarding.

[0093] This is a unique code for campus physical education courses. Its value follows the coding specifications of the campus course management system and is used to connect physical education teaching data with the school's overall curriculum system. The code usually includes dimensions such as course type, academic year, and semester. For example, PE-B2024 represents a basketball course in the first semester of the 2024 academic year (PE=physical education, B=basketball, 2024=academic year), and PE-R2024 represents a running course in the first semester of the 2024 academic year (R=running). Through standardized coding, the system can aggregate indicators by course dimension, such as statistics on student attendance rate and training frequency for the PE-B2024 course. It is a unique code for a class on campus, and its value follows the coding standard of the campus student registration management system. Its core function is to limit the scope of the organizational unit of teaching events. The code usually includes dimensions such as grade and class. For example, G801 represents the second year of junior high school (1) class (G8 = eighth grade, 01 = class 1), and G903 represents the third year of junior high school (3) class. Through this code, the system can accurately count the physical education teaching indicators of a specific class, such as the average number of training sessions and the pass rate of class G801. This is a unique identifier for the tenant deploying the teaching application. Its value is a character code assigned to each school or teaching area during system deployment, used to adapt to deployment scenarios where multiple schools share a single system. For example, assigning this value to XX Middle School... Assignment to YY Primary School ,pass It can isolate data from different schools, ensuring that the teaching data of XX Middle School is not mixed with that of YY Primary School, and also supports statistical data by tenant dimension, such as viewing data separately. (XX Middle School) and Overall physical education teaching indicators for (YY Primary School). This is the domain name for the terminal device to access the network. The value is taken from the system's pre-configured list of campus sports teaching terminal access domain names, used to distinguish the type of terminal that triggers the event. For example, student.school.edu.cn corresponds to the student-side app, teacher.school.edu.cn corresponds to the teacher-side tablet, and pc.school.edu.cn corresponds to the PC web interface. It can track the usage frequency of different access terminals, such as the percentage of training records accessed by students via an app and the percentage of guidance records accessed by teachers via tablets, providing data support for terminal optimization. This is the user's local login account within the system, typically a student ID or employee ID from a campus network, used to adapt to local login scenarios without third-party authentication. For example, a student's... =20240001 (student ID), teacher's =T202401 (employee number), when When empty (no third-party authentication), It serves as an identity identifier for local users in the associated system, ensuring that the behavioral events of locally logged-in users can be associated with the corresponding accounts.

[0094] Extract the course code field from the original log and map it to This involves converting the fields describing courses in the original logs (which may include course names, custom codes, etc.) into a unified format used by the campus course management system. The encoding and mapping rules are pre-configured by the school administrator. Extract the class code field from the original logs and map it to... By pre-setting the original class description - uniform The mapping rules convert class information in different formats in the logs into standardized encoding. For example, if the original logs contain the class field as Grade 8 Class 1 and Grade 8 (1), and the configuration rules are set to map both to G801, the system will automatically convert the two types of descriptions to a unified encoding. To avoid statistical bias caused by inconsistent class description formats, the tenant deployment field in the original logs is extracted and mapped to... It identifies the fields in the raw logs that identify the deployment tenant (which may be the school name, tenant code, etc.) and converts them into system-assigned fields. For example, if the tenant field in the original log is "XX Middle School", and the system's default mapping rule is "XX Middle School → fid001", then... Assign the value fid001; if the log already contains it. If so, the value is directly extracted and assigned to ensure that the data streams of different tenants are accurately isolated. The terminal access domain name field is extracted from the original logs and matched against a preset domain name list as... First, maintain a list of legitimate physical education teaching terminal access domains on campus (such as student.school.edu.cn, teacher.school.edu.cn, etc.). During DNS resolution, extract the access domains from the logs and match them with the list. If a match is found, assign it a value. If a match fails, it is marked as an illegal access and logged. For example, if the access domain name field in the original log is student.school.edu.cn, and it matches the preset list successfully, Assign the domain name as the value; if the field is unknown.school.edu.cn, it is considered unauthorized access, ensuring the compliance of the access terminal. Extract the system local login account field from the original log as... The process involves extracting the user's local login account (usually a student ID or employee ID) from the original logs. If the log does not contain this field, it is filled with NULL. For example, if the local login account field in the original log is 20240001 (student ID), then... The value is assigned to 20240001; if the user logs in through third-party authentication and there is no local login account field in the log, then... Fill with NULL to ensure that identity information in both local and third-party login scenarios is fully recorded.

[0095] It is a structured identity fragment identifier adapted for multi-entry, multi-tenant scenarios, ensuring the uniqueness of the identity association for each record. Generation rules are divided into two categories: when... When not empty, It combines three dimensions: third-party identity token, tenant environment, and access domain name; when When empty, This combines two dimensions: local login name and tenant environment. (Detection) In Whether a field is not null (field value is not NULL and character length > 0) is... The pre-judgment logic for generating rules prioritizes rules with stronger global uniqueness. As a core identity parameter, it is checked first during the determination. Check if the field is NULL. If it is not NULL, then check if its character length is greater than 0 (to avoid empty strings). If both conditions are met, it is determined to be non-empty; otherwise, it is determined to be empty. When not empty, combination , , generate ,pass Ensure cross-platform identity uniqueness. Isolate the tenant environment, By differentiating access terminals, the combination of these three elements can cover all identity fragmentation risks in third-party login scenarios. For example... =o123 (Unique identifier on WeChat). =fid001 (Tenant of XX Middle School) =student.school.edu.cn (student-side APP), generate Even if another user is in YY Elementary School ( =fid002) also has =o123 (Theoretically, a third-party platform) Unique, but tenant isolation must still be maintained), or the same user via teacher's tablet ( Log in at teacher.school.edu.cn and generate the file. They will also differ, ensuring the unique distinction of identity fragments. When empty, combination , generate ,pass Associate with local system accounts. Isolate tenant environments to avoid identity confusion caused by duplicate local login names under different tenants. For example =20240001 (Student ID of XX Middle School). =fid001 (XX Middle School tenant), generated If YY Primary School ( =fid002) also has =20240001, generated by the student Unlike the former, it ensures that the identity fragments of local login users under different tenants are not confused.

[0096] pass The generation rules for realizing the structured association of identity fragments are to integrate the scattered identity identifiers under multiple entry points and multiple tenants. , ) and environmental parameters ( , This structured combination transforms fragmented identity information into statistically measurable and associative identifiers. For example, the same student can generate [identification information] through a WeChat mini-program within a week. ), PC web version (generated) ), local account (generated) Three ways to access the system, three lines Although different, they are all related to the same thing. =u8765. In In the generation rules, Its purpose is to isolate tenant environments and solve the problem of identity identification when different schools or teaching areas share a system. , (This may be a recurring problem. For example, XX Middle School) =fid001) and YY Primary School ( =fid002) all have =20240001 students, if only using generate This could lead to duplicate student identities, and adding... After that, the generated They are respectively and This achieves the isolation of fragmented identities among tenants. In non-empty generation rules, Its function is to differentiate access terminals, resolving the issue of unique identification of fragmented identities when the same user under the same tenant accesses the network through different terminals. For example, the same student at XX Middle School ( =fid001) via the student app ( =student.school.edu.cn) and teacher-side tablets ( Access via teacher.school.edu.cn Both are o123, if only o123 is used and generate This will result in duplicate identity identifiers on the two terminals. After that, the generated They are respectively and This enables the differentiation of identity fragments among different access terminals under the same tenant. In the empty generation rule, Its function is local identity association. As a core identity parameter in scenarios without third-party authentication, it ensures that the behavioral events of locally logged-in users can be structurally correlated. For example, if a student has not bound a third-party account, they can be identified through their student ID 20240001 (…). =20240001) at XX Middle School ( =fid001) Local login, generate This identifier can be associated with all the student's behavioral events in the local login scenario (such as class attendance and training records), ensuring that the identity of users without third-party authentication is not overlooked.

[0097] It is a quantitative teaching statistics cycle The parameter of the degree of fragmentation of internal identity is used to intuitively reflect the redundancy of identity identification in multi-entry and multi-tenant scenarios by the ratio of the total number of identity fragments to the actual user base. It is the teaching statistics period The deduplicated set of multi-source session identifiers is used to filter out all independent identity fragments within the period, avoiding duplicate counting. It is a set The number of elements, i.e., the period. The total number of independent identity fragments is The calculated molecular parameters. Used to accurately count the total number of identity fragments. It is the teaching statistics period The deduplicated global user identifier set is used to filter out the unique users who actually participated in physical education teaching within the period, avoiding duplicate user statistics.

[0098] This is a statistical time window set to adapt to the pace of physical education teaching management, with a preferred value of 604,800 consecutive seconds (corresponding to one week). It balances statistical granularity and data representativeness. A one-week cycle covers a complete teaching unit (such as a full course week) while avoiding data randomness caused by a cycle that is too short (such as one day) or noise accumulation caused by a cycle that is too long (such as one month). Determining the teaching statistical cycle. The preferred value of 604,800 seconds (one week) is a statistical period set according to the regular rhythm of school physical education teaching. The basis for this value is that primary and secondary school physical education teaching is mostly based on weekly teaching plans and course content arrangements. A one-week cycle can accurately match the time dimension requirements of teaching management. It uses information entropy to quantify the period. The parameter for the degree of dispersion of internal access sources is used to reflect the uniformity of the distribution of access sources through the magnitude of the entropy value. The higher the entropy value, the more dispersed the access sources are, and the higher the potential risk of identity fragmentation. It is a quantitative cycle The parameter of the frequency of cross-tenant context switching for the same user is used to reflect the fluidity of user identity in a multi-tenant scenario through the average number of switching. The higher the value, the more frequently the user switches between tenants, and the higher the risk of identity fragmentation.

[0099] This represents the proportion of records for a specific access source domain to the total number of records. It is used to convert the count of access sources into a proportion, providing standardized input for information entropy calculation. For example, Records of internal student app Total number of records ,but PC web page record count ,but By using a proportional conversion, the counts of access sources of different sizes can be made comparable. Substituting into the formula:

[0100] Calculating the access source distribution index utilizes the properties of information entropy to transform the distribution ratio of access sources into a scalar reflecting the degree of dispersion. During execution, the system calls the natural logarithm operation function to calculate each... The logarithm of the first, then the corresponding Multiply the results, sum them, and then take the negative of the result to obtain the entropy value. Substitute this into the formula:

[0101] The cross-tenant handover frequency is calculated by averaging the total number of handovers divided by the actual number of users, quantifying the overall frequency of cross-tenant handovers for all users. During execution, the data analysis module first... Grouping, for each user Deduplication count The total number of switches is obtained by summing the counts of all users, and then divided by the actual number of users. .

[0102] It is a unified input vector that integrates all related features. It is used to explicitly encode architectural information and business features, transforming multi-dimensional information such as discrete categories, string identities, time periods, and numerical quantization into a unified vector that the model can process, thus avoiding the limitations of a single feature dimension. This is a function that performs low-dimensional dense mapping on discrete fields, preferably outputting a 32-dimensional vector, generated by a shallow neural network trained on historical sports teaching data. It addresses the irrelevant ordering problem introduced when discrete category fields are directly input into the model, transforming category information into an unbiased vector representation. During training, the category values ​​of the discrete fields (such as...) are used as the basis for the model's operation. , Using the correlation between the category and the user's identity as input, and the degree of correlation between the category and the user's identity as the supervision signal, the vector distance between the same category is made closer, and the vector distance between different categories is made farther. This function performs a fixed-length vector mapping on string fields, ideally outputting a 2048-dimensional hash bucket count vector, implemented using the SHA-256 hash function. It is used to solve string field ( , This addresses the issues of variable length and lack of standardized encoding rules, while preserving the uniqueness of the string. During execution, the string is first converted into a 64-bit integer index using SHA-256 hashing. Then, the hash bucket position corresponding to that index is set to 1, and the remaining positions are set to 0. timestamp A two-dimensional vector that undergoes periodic encoding is used to preserve the periodicity of time, avoiding direct use. The numerical values ​​mask the correlation characteristics within the same time period. For example, 14:30 every day ( The encoding vector corresponding to (seconds) is:

[0103] ;

[0104] Regardless of whether the date is Monday or Friday, the encoding vector for that time period is exactly the same, enabling the model to learn the pattern of similar identity associations among physical education teaching events within the same time period. (Call) Generating a 32-dimensional embedding vector from a discrete field is to , , , , Discrete category fields are transformed into low-dimensional dense vectors using pre-trained embedding functions. During execution, the system retrieves pre-trained vectors for the corresponding field category from the embedding vector library, eliminating the need for real-time training and ensuring efficiency. For example, The corresponding pre-trained embedding vector is (32-dimensional) The corresponding vector is This operation transforms discrete categories, whose similarity could not be directly calculated, into quantifiable vectors. (Call...) Generating a 2048-dimensional hash bucket count vector for a string field involves converting a variable-length string into a fixed-length vector using hash operations, ensuring the uniqueness of the string and the consistency of the model's input dimensions. During execution, the string is first hashed using SHA-256 to obtain a 64-bit integer, then modulo 2048 to obtain the hash bucket index (range 0-2047), and finally, a sparse vector is generated corresponding only to index positions where 1 is present. (Calculation...) Implementing time-periodic encoding is stripping The date information retains the periodic characteristics of intraday time periods. During execution, it first... Calculate the number of seconds in a day ( Then substitute the sine and cosine functions to calculate the two-dimensional vector. Concatenate the vectors in order. generate This involves integrating discrete embedding vectors, string hash vectors, temporal encoding vectors, and numerical features into a unified multidimensional feature vector in a fixed order, ensuring the structure and consistency of the model input. The concatenation order is fixed as follows:

[0105] .

[0106] , , It is to adjust the dynamic weights The fixed constants are preferably taken as follows: , , Weights are used to match the impact of identity redundancy, access source distribution, and cross-tenant handover on training noise. The highest weight (0.7) is because identity redundancy is the most critical source of noise; The weight is the lowest (0.2) because the impact of the distribution of access sources is relatively minor; The weight is set to a middle value (0.5) to balance the impact of cross-tenant switching. The value is based on noise suppression experience in multi-entry, multi-tenant scenarios. Through extensive experimental verification, this combination can effectively suppress the impact of samples with high noise periods while retaining the effective information of samples with low noise periods. These are the data units used for model training, where For multidimensional feature vectors, For monitoring targets (global user identifiers) (Corresponding integer category label). Used to establish the mapping relationship between feature vectors and user identities, enabling the model to learn rules for identifying user identities from multi-dimensional features. The generation method is from the identity attribute table Extract And map it to consecutive integer labels (such as...) Mapped to label 100), avoid This addresses the issue of low model training efficiency caused by strings or non-continuous numerical values. The model predicts the sample. Corresponding supervision objectives The confidence score is obtained after processing with the softmax activation function, and its value range is [value range missing]. ,all The corresponding probability sums to 1. This is used to transform the linear output of the model into a quantifiable probability distribution. It follows the teaching statistics cycle The varying sample weights are used to reduce the weight of high-noise periods and increase the weight of low-noise periods, suppressing the gradient contribution of high-noise samples through an exponential decay function. It uses a weighted cross-entropy model to optimize the objective function, which is used to dynamically adjust the weights. Differential weights are assigned to samples from different periods, allowing the model to focus more on samples from low-noise periods while reducing the impact of samples from high-noise periods. , , The optimal values ​​are determined based on the noise impact of multi-entry, multi-tenant scenarios, setting the adjustment weights for each noise parameter in the dynamic weighting. The value selection process was completed through extensive comparative experiments: the identity recognition accuracy of the model was tested under different parameter combinations, ultimately finding... (Focus on suppressing identity redundancy noise) (Mild suppression of diffuse noise from the access source) The combination of (moderate suppression of cross-tenant switching noise) achieves optimal identity normalization under various noise scenarios, avoiding both underfitting due to excessively high weights for certain noise types and insufficient noise suppression due to excessively low weights. (Constructing training samples) It is a multidimensional feature vector With corresponding user identity tags Correlation ensures the accuracy of supervision signals during model training. During execution, the system extracts features from the feature vector table. within From the identity attribute table Extract the corresponding and through the label mapping table Convert to integer tags ,For example Mapped to , Mapped to Ensure that each Each has a unique counterpart To avoid confusion in monitoring signals. If a certain... corresponding If missing, then Set a special label (e.g., -1) and process it separately as an unsupervised sample or discard it.

[0107] Calculate dynamic weights It is integration , , Three types of noise parameters are used to generate periodic sample weights. During execution, the system retrieves the values ​​of the three types of parameters from the analysis results table, substitutes them into the formula to calculate the three exponentially decaying terms in turn, and then multiplies the results to obtain the final value. Construct a weighted cross-entropy loss function. It is to use dynamic weights By incorporating cross-entropy loss, model optimization is achieved to differentiate sample weights. During execution, the system iterates through... All training samples Calculate the value of each sample multiplied by Then take the opposite number and sum the results of all samples to get the total loss. . It is the computational unit of the neural network, with a preferred structure of 3 fully connected layers. The number of neurons in each layer is 1024, 512, and 256 respectively, and each layer uses the ReLU activation function. This hierarchical structure is used to adapt multidimensional feature vectors through high-dimensional feature compression → semantic extraction → identity mapping. The system progressively strips away redundant information, retaining only the core semantic features that distinguish user identities. For example, the first layer of 1024 neurons receives the input vector and performs a nonlinear transformation, compressing the 5205-dimensional features to 1024 dimensions; the second layer of 512 neurons further extracts feature associations; and the third layer of 256 neurons outputs abstract identity features, providing a foundation for subsequent linear transformations to the user identity space. The ReLU activation function, by introducing nonlinearity, prevents the model from falling into the problem of linear inseparability. yes The weight matrices and bias vector sets for each layer, obtained through optimization using the weighted cross-entropy loss function, are stored in the model parameter library. The mapping rules between feature vectors and user identities are used to solidify these rules; the parameter values ​​are determined by gradient descent optimization during model backpropagation to ensure... It can reliably extract identity-related features. It is a connection Output the weight matrix of the user identity space, obtained through pre-training, with corresponding dimensions. Output dimensions (256 dimensions) and global user identifier Total quantity (e.g.) (i.e., dimensional), that is The dimension is Used to The output 256-dimensional abstract features are linearly mapped to Each user identity corresponds to a score space, and each user identity corresponds to a column of weights, which measures the strength of the association between the abstract feature and the user. It is the bias vector of the output layer, obtained through pre-training, with dimensions equal to the global user identifier. Total number Consistency Used to compensate for systematic biases in linear mappings, avoiding [the effects of] [various factors]. The mean shift of the output features leads to bias in the calculation of identity scores. It is a multidimensional feature vector Belongs to a certain global user identifier The confidence level is determined by The calculated range is: The sum of the probabilities corresponding to all user identities is 1. This is used to transform the identity scores of a linear mapping into a quantifiable confidence distribution, making it easier to filter the most likely user identities. yes Global user identifier corresponding to the highest probability in the middle This is used to normalize fragmented identities generated by users through different access methods. Corresponding to different Vectors, but after model calculations, these... The vector corresponding All in the same The maximum value is obtained at this point. That is This allows for the association of fragmented identities with the same real user, thus resolving the issue of statistical bias caused by identity fragmentation in multi-entry scenarios.

[0108] Sure The optimized structure and adoption of the ReLU activation function are based on the input feature dimension and the complexity of the identity recognition task. This involves designing suitable network layers and the number of neurons, and introducing a non-linear activation function. The selection of layers and the number of neurons is based on a balance between efficiency and accuracy in input dimension, feature compression, and identity mapping: the input vector is 5205 dimensions, and the first layer has 1024 neurons to ensure no feature loss. Subsequent layers halve the dimensions to 256, ensuring both computational efficiency and preservation of core features. The ReLU activation function is chosen because it is less prone to gradient vanishing during gradient descent and is computationally simple, effectively improving model training speed and generalization ability. Specifically, the output of each neuron is... That is, when the input value is less than 0, the output is 0; when the input value is greater than 0, the output is the original input value, thus achieving a non-linear transformation. The operation is then determined. From probability distribution The user identity with the highest confidence level is selected from the options and used as the current feature vector. The normalized user primary key. During execution, the system iterates through the probability values ​​corresponding to all user identities and records the user identifier corresponding to the maximum value. That is This ensures the uniqueness and certainty of identity verification. Associated storage to This involves establishing a link between event records and normalized user primary keys, ensuring that each physical education teaching event is accurately linked to a real user. During execution, the system retrieves the corresponding data for each event record from the model output. And write it to the event log table. The new addition Fields, compared to the original , Fields are associated and stored together.

[0109] User In the teaching statistics period The total number of class attendance events within the specified timeframe, used to calculate the number of events based on the normalized user primary key. Aggregate class attendance events to ensure that class attendance records for the same user are not duplicated or omitted. For indicator functions, when an event occurs The type is Attend and Output 1 if the condition is met, otherwise output 0. The accumulated result is the user's answer. The number of times you attend class. User In the cycle The total number of training record events within the specified range is used to focus on training record events, filtered by an indicator function. For training records (TRAIN_TEXT) and The events were summed up to obtain the statistical results. User In the cycle The total number of qualified text events within the scope is used to add to the event type filtering. The criteria for compliance are determined to ensure that only events that meet the compliance standards are counted. It is a core parameter for correcting biases in sports process indicators. It is used to integrate three types of noise parameters to offset indicator inflation or bias caused by identity redundancy, dispersed access sources, and cross-tenant switching. This directly offsets the counting bias caused by identity redundancy. and Exponential decay corrects for additional biases caused by dispersed access sources and cross-tenant switching.

[0110] , It involves adjusting the access source distribution index and the cross-tenant handover frequency to adjust the calibration coefficient. The fixed constants affecting the weights are preferably taken as follows: , Used to match the degree of influence of the two types of noise on the index deviation. The value is relatively small (0.1) because the impact of dispersed access sources on the indicator is weaker than that of identity redundancy; The value is moderate (0.3), which balances the impact of cross-tenant switching. The value is based on a large number of experimental verifications. This combination can make the calibrated index closest to the real situation and avoid the excessive influence of a single noise factor on the calibration result. It refers to all users within the class. corresponding , or The arithmetic mean is used to reflect the overall teaching indicator level at the class level, and is calculated based on the actual user base. The denominator is used to ensure that the average value is not affected by identity fragments. It is a calibration coefficient The revised class average index is used to eliminate index bias caused by noise, resulting in a class-level index that is closer to the actual situation. It is a hash signature that ensures the traceability of reports, preferably implemented using the SHA-256 hash function. It is used to solidify the core input parameters for metric calculations, generating a unique signature through hash operations to verify the consistency and integrity of metric calculations and prevent data tampering. , , They are cycles All users corresponding , , The cumulative sum is used to provide the raw total data for indicator calculation, and is also used to generate the calculation signature. This can also serve as a reference for total indicators at the class level. (Aggregation generation) , , Events are filtered and accumulated based on event type and normalized user primary key to ensure the accuracy of single-user metric statistics. During execution, the system iterates through the cycle. All events , call In , , Fields, via indicator functions Determining whether an event meets statistical criteria: Statistics Time determination and ;statistics Time determination and ;statistics Additional The criteria for achieving the target are determined by summing the output values ​​of all indicator functions that meet the criteria, resulting in a single-user metric. Calculation And correct get It integrates noise parameters to generate calibration coefficients and corrects the class average index through multiplication operations. During execution, the system first retrieves data from the analysis results table. , , Substitute into the formula to calculate Then calculate the class average index. Finally, multiply the two results to obtain the corrected index. (Generation) Integrating report data solidifies the indicator calculation chain, ensuring report traceability and verifiability. During execution, the system first assembles the data in a fixed order. , , , , , The string format (e.g., 1.20.61.29012060) is then used to perform a SHA-256 hash function on the concatenated string to generate the hash value. Finally, Noise characteristic parameters , The data is integrated into a preset report template, which organizes the data according to the structure of class-indicator type-original value-calibration value-signature, ensuring that managers can clearly see the indicator correction process and the basis for verification.

[0111] It should be noted that, Figure 1 The Chaoxing Teaching System (Superior CEM) is a sports teaching data management system.

[0112] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A sports teaching data management system based on artificial intelligence, characterized in that, Includes a processor and memory, the processor executing instructions stored in the memory to perform the following steps: Obtain the raw logs of physical education teaching process data, parse them into an event record table and an identity attribute table, and generate a multi-source session identifier corresponding to each record according to the field parsing rules; Within the preset teaching statistics period, the identity redundancy coefficient is calculated based on the total number of multi-source session identifiers and the actual number of users in the system, and the access source distribution index and cross-tenant switching frequency are calculated simultaneously as noise characteristic parameters. A multi-dimensional feature vector containing the access source domain name, tenant context, and time dimension is constructed and input into a neural network model. The neural network model is trained using a loss function with the identity redundancy coefficient as the dynamic weight parameter, thereby outputting a deterministic normalized user primary key for each event record. The event records are aggregated using the normalized user primary key to generate sports process indicators, and the data calibration coefficient is calculated based on the identity redundancy coefficient and the noise characteristic parameters to numerically correct the sports process indicators and output a sports teaching management report. Pre-training of neural network models includes: The loss function of the neural network model is configured to include a dynamic weight that changes with the teaching statistical period. The dynamic weight is calculated by an exponential decay function based on the identity redundancy coefficient, access source distribution index, and cross-tenant handover frequency. During training, the dynamic weight is used to suppress the gradient contribution of samples with high identity redundancy coefficient, high access source distribution index, or high cross-tenant handover frequency, so that the neural network model learns a normalized user primary key generation rule for architectural noise. The event records are aggregated using the normalized user primary key to generate sports process indicators. Data calibration coefficients are calculated based on the identity redundancy coefficient and the noise characteristic parameters to numerically correct the sports process indicators. A sports teaching management report is then output, including: Within the normalized user primary key space, the occurrence counts of class attendance events, training record events, and achievement text events are counted to generate sports process indicators. The reciprocal of the identity redundancy coefficient is calculated and multiplied by an exponential decay term based on the access source distribution index and cross-tenant switching frequency to obtain a data calibration coefficient. The sports process indicators are multiplied by the data calibration coefficient to obtain the corrected indicators. The identity redundancy coefficient, noise feature parameters, and corrected indicators are concatenated and hashed to generate a computational signature, and a sports teaching management report containing the computational signature is output.

2. The sports teaching data management system based on artificial intelligence according to claim 1, characterized in that, Obtain the raw logs of the physical education teaching process data and parse them into an event record table and an identity attribute table, including: The original logs are parsed into an event record table containing timestamps, course identifiers, class identifiers, event types, numerical values, and multi-source session identifiers; and into an identity attribute table containing timestamps, tenant contexts, access source domains, identity tokens, login names, and global user identifiers; for fields missing in the data source, null values ​​are filled in according to preset rules.

3. The sports teaching data management system based on artificial intelligence according to claim 1, characterized in that, Generate a multi-source session identifier for each record based on the field parsing rules, including: Check if a non-empty identity token exists in the identity attribute table; if the identity token is not empty, combine the identity token, tenant context, and access source domain name to generate the multi-source session identifier; if the identity token is empty, combine the login name and tenant context to generate the multi-source session identifier.

4. The sports teaching data management system based on artificial intelligence according to claim 1, characterized in that, Within a preset teaching statistical period, the identity redundancy coefficient is calculated based on the total number of multi-source session identifiers and the actual user base of the system, including: The number of deduplicated multi-source session identifiers and the number of deduplicated global user identifiers within the teaching statistical period are counted as the real user base of the system; the ratio of the number of multi-source session identifiers to the real user base of the system is calculated to obtain the identity redundancy coefficient.

5. The sports teaching data management system based on artificial intelligence according to claim 1, characterized in that, The access source distribution index and cross-tenant handover frequency are calculated as noise characteristic parameters, including: For the access source distribution index, the proportion of the occurrences of each access source domain name to the total occurrences within the teaching statistical period is calculated, and the negative of the sum of the products of the proportion and the natural logarithm of the proportion is calculated. For the cross-tenant switching frequency, for each global user identifier, the number of different tenant contexts associated with it within the teaching statistical period is calculated, and the arithmetic mean of the number corresponding to all global user identifiers is calculated.

6. The sports teaching data management system based on artificial intelligence according to claim 1, characterized in that, Construct a multi-dimensional feature vector that includes the access source domain name, tenant context, and time dimension, including: The event type, course identifier, class identifier, tenant context, and access source domain name in the event log table are mapped to fixed-dimensional embedding vectors; the identity token and login name in the identity attribute table are hashed to generate a fixed-length hash bucket count vector; the time dimension is converted into a periodic encoded vector based on sine and cosine functions; and the numerical fields are directly used as numerical features; finally, all the above vectors and numerical features are concatenated to form the multidimensional feature vector.

7. The sports teaching data management system based on artificial intelligence according to claim 1, characterized in that, The multidimensional feature vector is input into the neural network model to output a normalized user primary key for each event record, including: The multidimensional feature vector is nonlinearly transformed using a neural network model, and the probability distribution of the vector belonging to each global user identifier is calculated using a normalized exponential function; the global user identifier corresponding to the maximum probability in the probability distribution is selected and determined as the normalized user primary key of the event record.

Images