An intelligent exercise system and method based on AI intelligent recommendation

Abstract

The application discloses an intelligent exercise system and method based on AI intelligent recommendation, relates to the technical field of health management and sports science, and comprises a data acquisition module, which collects the heart rate variation of joint flexibility, muscle strength reading and metabolic rate index of a user in an exercise process through a sensor device, obtains original exercise data and physiological response information from multiple source channels, and obtains an initial data set containing a heterogeneous format; a normalization processing module adopts a normalization conversion method to process the dimensional difference and uneven acquisition frequency in the initial data set, unifies all indexes to a standard scale, and obtains a standardized data set; the intelligent exercise system and method based on AI intelligent recommendation significantly improve the pertinence and effectiveness of an exercise scheme for an old population through data fusion and accurate modeling, and provide intelligent support for health management.

Description

Technical Field

[0001] This invention relates to the fields of health management and exercise science technology, specifically to an intelligent exercise system and method based on AI intelligent recommendation. Background Technology

[0002] Improving the physical function and quality of life of the elderly has become an important research topic in the fields of health management and exercise science. Due to significant individual differences in physiological function, physical responsiveness, and chronic disease characteristics among the elderly, traditional, standardized, experience-based exercise programs are no longer sufficient to meet their diverse health needs.

[0003] Current technologies for developing exercise plans for older adults primarily rely on manual assessment and static analysis of basic physical data (such as age, weight, and heart rate), lacking in-depth modeling of dynamic movement responses. Some studies attempt to collect physiological data such as heart rate and cadence using wearable devices; however, processing data from different types of sensors generally suffers from inconsistencies in units of measurement, varying collection frequencies, and inconsistent data accuracy, making effective data integration difficult and impacting the accuracy of exercise effect analysis. Furthermore, when faced with high-dimensional, multi-type physiological response data from older adults, existing methods generally lack a unified normalization mechanism, making it difficult to construct high-quality, standardized datasets for modeling and analysis. In addition, personalized exercise recommendations generally fail to consider the interactions between key indicators such as core muscle strength and cardiopulmonary function among different older adults, lacking in-depth modeling of the comprehensive impact of these indicators. These deficiencies prevent exercise recommendations from accurately reflecting the specific degree of improvement in physical function (such as joint flexibility, balance, or metabolic capacity) brought about by various training activities. In terms of physiological response modeling, current methods mostly rely on static labels or coarse-grained clustering, failing to effectively group and label individuals based on their dynamic response characteristics, thus hindering quantitative prediction of training effects. Especially when dealing with the coordinated training of core muscle groups and cardiopulmonary function, existing technologies cannot construct fused feature vectors to support subsequent regression modeling, nor can they generate detailed training impact matrices, thereby limiting the ability to provide precise, data-driven training recommendations. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent exercise system and method based on AI intelligent recommendation, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-based intelligent exercise system, comprising a data acquisition module that collects heart rate changes, muscle strength readings, and metabolic rate indicators of the user's joint flexibility during exercise using sensor devices, obtaining raw exercise data and physiological response information from multiple sources to obtain an initial dataset containing heterogeneous formats; a normalization processing module that uses a normalization transformation method to process the dimensional differences and uneven collection frequencies in the initial dataset, unifying all indicators to a standard scale to obtain a normalized dataset; and a clustering analysis module that applies a clustering algorithm to the normalized dataset to group similar physiological response patterns, and determines the appropriate exercise system based on the grouping results. The system categorizes users' physical states to obtain a set of category labels. A feature fusion module, if the heterogeneity of the category label set exceeds a preset threshold, extracts and fuses indicators from different body parts to obtain a fused feature vector. A regression modeling module, based on the fused feature vector, constructs a regression model to predict the quantitative impact of exercise on physiological function, obtaining an impact score matrix. A contribution evaluation module, using the impact score matrix, calculates the contribution of each exercise method to different body parts; if the contribution exceeds a preset threshold, it is marked as a valid mode, obtaining a list of valid modes. An intelligent recommendation module matches the list of valid modes with the user's current physical state, generating an exercise sequence optimized for joint flexibility, resulting in a personalized fitness plan.

[0006] Preferably, the data acquisition module collects heart rate changes, muscle strength readings, and metabolic rate indicators related to joint flexibility during exercise using sensor devices. It acquires raw exercise data and physiological response information from multiple sources to obtain an initial dataset containing heterogeneous formats. This includes collecting heart rate changes, muscle strength readings, and metabolic rate indicators from sensor devices during exercise and acquiring raw data from multiple sources to generate the initial dataset. If the collected data is missing or abnormal, linear interpolation is used to complete the heart rate changes, muscle strength readings, and metabolic rate indicators to obtain a complete dataset. Standardization processing is used to unify the heterogeneous formats in the complete dataset, and a data field mapping table is used to map heart rate changes, muscle strength readings, and metabolic rate indicators. Heart rate and metabolic rate indicators are converted to a unified unit to generate a standardized dataset. Principal component analysis (PCA) is used to reduce the dimensionality of heart rate changes, muscle strength readings, and metabolic rate indicators in the standardized dataset, extracting key features related to joint flexibility to obtain a feature dataset. If the correlation coefficient between heart rate changes and muscle strength readings in the feature dataset exceeds a preset threshold, K-means clustering is used to group the feature dataset to determine the user's exercise mode classification. Based on the exercise mode classification, time series analysis is used to model the heart rate change trend of each group, generating a dynamic change model for each group. Using the dynamic change model, the correlation strength between metabolic rate indicators and joint flexibility for each group of users is calculated to obtain the exercise effect evaluation results.

[0007] Preferably, the normalization processing module uses a normalization transformation method to handle the dimensional differences and uneven collection frequencies in the initial dataset, unifying all indicators to a standard scale to obtain a normalized dataset. This includes processing the dimensional differences and uneven collection frequencies in the initial dataset through min-max normalization, mapping heart rate changes, muscle strength readings, and metabolic rate indicators to a unified scale interval to obtain a normalized dataset; using data cleaning techniques to process noise and outliers in the normalized dataset, replacing outliers if data points deviate from a preset statistical threshold using a median imputation method to obtain a cleaned dataset; and using feature selection methods to filter indicators related to joint flexibility in the cleaned dataset, based on an information gain algorithm. The weights of heart rate variability, muscle strength readings, and metabolic rate indicators are calculated to determine the key indicator set. If the correlation between heart rate variability and metabolic rate indicators in the key indicator set exceeds a preset threshold, a linear regression algorithm is used to model the key indicator set to obtain an indicator association model. Based on the indicator association model, a time series decomposition method is used to analyze the periodic patterns of heart rate variability, extract periodic trends, and obtain a trend dataset. Using the trend dataset, a sliding window technique is used to dynamically analyze the muscle strength readings of each group of users, calculate the strength change characteristics within each time window, and obtain a dynamic feature set. Based on the dynamic feature set, a decision tree algorithm is used to classify the users' exercise patterns, determine the joint flexibility level of each group, and obtain the classification results.

[0008] Preferably, the clustering analysis module applies a clustering algorithm to group similar physiological response patterns in the normalized dataset, determines the user's physical state classification based on the grouping results, and obtains a classification label set. This includes using the K-means clustering algorithm to group heart rate changes, metabolic rate, and blood oxygen saturation in the normalized dataset, calculating the distance between each data point and the cluster center based on Euclidean distance, determining the physiological response category, and obtaining the classification label set. Based on the classification label set, principal component analysis is used to reduce the dimensionality of the feature vectors of each group of physiological response categories, extracting feature components, and obtaining a dimensionality-reduced feature set. If the principal component variance contribution rate in the dimensionality-reduced feature set exceeds a preset threshold, linear discriminant analysis is used to divide the classification boundary of the dimensionality-reduced feature set and determine the classification of each group. Physiological state labels are generated to obtain a state label set. Based on the state label set, the time series data of heart rate changes are segmented using the sliding window technique, and the statistical characteristics within each window are calculated to obtain a time series feature set. If the mean heart rate change in the time series feature set deviates from a preset threshold, the time series feature set is smoothed using a moving average filtering method to eliminate fluctuations, resulting in a smoothed feature set. Based on the smoothed feature set, the physiological state of each group is classified using a decision tree algorithm, and the feature weights are calculated based on information gain to determine the health state category, resulting in a health classification set. Through the health classification set, the metabolic rate indicators in the smoothed feature set are periodically analyzed using a time series decomposition method to extract long-term trends, resulting in a trend feature set.

[0009] Preferably, in the feature fusion module, if the heterogeneity of the classification label set is higher than a preset value, the module extracts and fuses indicators from different parts to obtain a fused feature vector. This includes: if the heterogeneity of the classification label set is higher than a preset threshold, principal component analysis is used to reduce the dimensionality of cardiopulmonary function data and core muscle group data to obtain a dimensionality-reduced feature set; based on the dimensionality-reduced feature set, kernel density estimation is used to analyze the probability density of physiological indicators to obtain density distribution characteristics; if the peak value of the density distribution characteristics deviates from a preset range, wavelet transform is used to decompose the time-frequency characteristics of cardiopulmonary function data to obtain a time-frequency feature set; based on the time-frequency feature set, a support vector machine algorithm is used to classify the dynamic patterns of core muscle group data to determine dynamic pattern labels; if the dynamic pattern label indicates an abnormal state, a sliding window technique is used to segment the time series of physiological indicators to obtain a segmented feature set; based on the segmented feature set, statistical characteristics within each window are calculated to obtain a statistical feature set; based on the statistical feature set, time series analysis is used to decompose the periodic changes of physiological indicators to obtain a periodic feature set.

[0010] Preferably, the regression modeling module constructs a regression model based on the fused feature vectors to predict the quantitative impact of exercise on physiological function and obtains the impact score matrix. This includes: standardizing the feature vector set using a data preprocessing method to obtain a standardized feature set; if the variance of the standardized feature set exceeds a preset threshold, reducing the dimensionality of the standardized feature set using principal component analysis to obtain a dimensionality-reduced feature set; training the regression model using a random forest algorithm based on the dimensionality-reduced feature set, calculating feature weights to obtain training model parameters; evaluating the performance of the regression model using cross-validation based on the training model parameters to obtain cross-validation results; if the mean squared error of the cross-validation results is lower than a preset threshold, predicting the dimensionality-reduced feature set using the regression model to obtain the impact score matrix; grouping the physiological function data using cluster analysis based on the impact score matrix to obtain a grouped result set; and extracting the trend features of each group using time series analysis based on the grouped result set to obtain a trend feature set.

[0011] Preferably, the contribution evaluation module calculates the contribution of each exercise method to different body parts using an influence score matrix. If the contribution exceeds a preset threshold, it is marked as an effective mode. Obtaining the effective mode list includes: quantifying the contribution of each exercise method to each physiological part using a weighted calculation method based on the influence score matrix to obtain a contribution vector; using a threshold judgment method based on the contribution vector, marking an exercise method as an effective mode if its contribution exceeds a preset threshold, thus generating a preliminary mode set; grouping the effective modes from the preliminary mode set using cluster analysis based on the correlation of physiological parts to obtain mode group sets; calculating the average contribution of each group of effective modes using a mean aggregation method based on the mode group sets to obtain a grouped contribution set; extracting the effective mode with the highest contribution from each group from the grouped contribution set to generate a preferred mode set; sorting the effective modes from the preferred mode set according to their contribution from high to low using a sorting algorithm to obtain a final effective mode list; and extracting multiple modes from the final effective mode list to generate a mode list.

[0012] Preferably, the intelligent recommendation module matches the user's current physical state with an effective pattern list to generate an exercise sequence optimized for joint flexibility, thereby obtaining a personalized fitness plan. This includes using physical state data and a standardized processing method to convert joint flexibility-related indicators into normalized feature vectors to obtain a user feature dataset. If the joint flexibility indicators in the user feature dataset are lower than a preset threshold, a classification algorithm is used to select exercise patterns related to micro-joints from a preset effective pattern list to obtain a candidate pattern set. Based on the candidate pattern set, a weighted scoring method is used, combined with the normalized feature vectors in the user feature dataset, to calculate the matching degree of each exercise pattern, resulting in a matching degree ranking list.

[0013] Preferably, the intelligent recommendation module uses a list of effective patterns to match the user's current physical state to generate an exercise sequence optimized for joint flexibility. Obtaining a personalized fitness plan also includes extracting the highest-matching exercise patterns from a matching ranking list to generate a preliminary personalized sequence. If the number of exercise patterns in the preliminary personalized sequence is insufficient, cluster analysis is used to supplement patterns related to joint flexibility from the candidate pattern set to obtain an optimized sequence. Based on the optimized sequence, time series analysis is used to adjust the execution order of the exercise patterns to generate a final personalized fitness plan. Finally, using an iterative update method, the exercise patterns in the final personalized fitness plan are adjusted according to changes in the user's subsequent physical state data to obtain a dynamic fitness plan.

[0014] A smart exercise method based on AI intelligent recommendation, employing the aforementioned AI-based smart exercise system, includes the following steps: Collecting heart rate changes, muscle strength readings, and metabolic rate indicators related to joint flexibility during exercise using sensor devices; acquiring raw exercise data and physiological response information from multiple sources to obtain an initial dataset containing heterogeneous formats; using a normalization transformation method to handle dimensional differences and uneven collection frequencies in the initial dataset, unifying all indicators to a standard scale to obtain a normalized dataset; applying a clustering algorithm to the normalized dataset to group similar physiological response patterns, determining the user's body state classification based on the grouping results, and obtaining a classification label set; if the heterogeneity of the classification label set is higher than a preset threshold, fusing indicators from different body parts through feature extraction to obtain a fused feature vector; constructing a regression model based on the fused feature vector to predict the quantitative impact of exercise on physiological function, obtaining an influence score matrix; calculating the contribution of each exercise method to different body parts using the influence score matrix, marking those with a contribution higher than a preset threshold as valid modes, and obtaining a list of valid modes; matching the list of valid modes with the user's current body state to generate an exercise sequence optimized for joint flexibility, thus obtaining a personalized fitness plan.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects: This AI-based intelligent exercise system and method addresses the business problem of accurately analyzing the physiological responses of the elderly from multi-source heterogeneous data and developing targeted exercise plans. The invention collects multi-dimensional data such as heart rate, muscle strength, and metabolic rate using sensor devices, and uses normalization to unify the dimensions and frequencies, generating a standardized dataset. Based on this, a clustering algorithm is used to group similar physiological response patterns and determine body state classification labels. If the labels show high heterogeneity, feature extraction is performed to fuse core muscle group and cardiopulmonary function data to construct a fused feature vector. A regression model is then used to predict the quantitative impact of exercise on physiological function, generating an impact score matrix. Based on the matrix, the invention calculates the contribution of each exercise method to joint flexibility, selects effective patterns, matches them to the user's status, and ultimately generates a personalized fitness plan. Through data fusion and precise modeling, this invention significantly improves the targeting and effectiveness of exercise plans for the elderly, providing intelligent support for health management. Attached Figure Description

[0016] Figure 1 This is a connection diagram of the AI-based intelligent exercise system of the present invention; Figure 2 This is a flowchart of the intelligent exercise method based on AI intelligent recommendation of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1As shown, this invention provides a technical solution: an AI-based intelligent exercise system, comprising a data acquisition module that collects heart rate changes, muscle strength readings, and metabolic rate indicators related to joint flexibility during exercise using sensor devices, obtaining raw exercise data and physiological response information from multiple sources to obtain an initial dataset containing heterogeneous formats; a normalization module that uses a normalization transformation method to process the dimensional differences and uneven collection frequencies in the initial dataset, unifying all indicators to a standard scale to obtain a normalized dataset; and a clustering analysis module that applies a clustering algorithm to the normalized dataset to group similar physiological response patterns, determining the user's physical condition based on the grouping results. The system is divided into several modules: a body state classification module to obtain a classification label set; a feature fusion module to extract and fuse indicators from different body parts if the heterogeneity of the classification label set is higher than a preset threshold, resulting in a fused feature vector; a regression modeling module to construct a regression model based on the fused feature vector to predict the quantitative impact of exercise on physiological function, resulting in an impact score matrix; a contribution evaluation module to calculate the contribution of each exercise method to different body parts using the impact score matrix, marking it as an effective mode if the contribution is higher than a preset threshold, resulting in a list of effective modes; and an intelligent recommendation module to match the user's current body state with the list of effective modes, generating an exercise sequence optimized for joint flexibility, resulting in a personalized fitness plan.

[0019] This system collects physiological data in real time during user exercise using multi-dimensional sensors (including but not limited to heart rate monitors, electromyography sensors, and metabolic rate sensors), forming an initial dataset. The initial data suffers from heterogeneous data formats, inconsistent sampling frequencies, and different units of measurement. The normalization module converts this data to a unified standard using methods such as Z-score, Min-Max, or Log transformation, making different physiological parameters usable for subsequent unified analysis. The clustering analysis module employs unsupervised learning methods such as K-means, DBSCAN, or hierarchical clustering to cluster the standardized data, identifying the physiological response patterns exhibited by users under different exercise states, and labeling the user's current physical state accordingly. To address the issue of high heterogeneity between labels, the system introduces a feature fusion mechanism to fuse information from physiological indicators of different body parts corresponding to the classification label set, such as using principal component analysis (PCA) and autoencoders to extract fused feature vectors. Using these fused feature vectors, the system builds regression models (such as linear regression, XGBoost, or deep neural networks) to quantify and predict the specific impact of various exercises on physiological functions, generating a multi-dimensional impact score matrix. Next, the contribution evaluation module assesses the local effectiveness of the exercise methods, retaining only those with high contribution. Finally, the intelligent recommendation module matches appropriate exercise combinations based on the list of effective patterns and the user's current state, outputting customized, phased goal-oriented exercise plans, which are particularly helpful in improving joint flexibility.

[0020] This invention enables real-time multi-source data collection, standardized processing, cluster modeling, and personalized exercise recommendations for users' physiological states, significantly improving the scientific rigor and relevance of exercise programs. Through quantitative regression model construction, it can clearly predict the specific impact of various exercises on users' physical functions (such as joint flexibility and muscle strength), helping to select efficient exercise methods and thus improve fitness efficiency. Simultaneously, the feature fusion and contribution evaluation mechanism effectively avoids biases caused by single indicators, improving the accuracy and reliability of the recommended programs. Compared with traditional experience-based recommendation methods, this system is more systematic, intelligent, and adaptable.

[0021] The data acquisition module collects heart rate changes, muscle strength readings, and metabolic rate indicators related to joint flexibility during exercise using sensor devices. It acquires raw exercise data and physiological response information from multiple sources to obtain an initial dataset containing heterogeneous formats. If the collected data is missing or abnormal, linear interpolation is used to complete the heart rate changes, muscle strength readings, and metabolic rate indicators, resulting in a complete dataset. Standardization is then applied to unify the heterogeneous formats in the complete dataset, using a data field mapping table to map heart rate changes, muscle strength readings, and metabolic rate indicators. Metabolic rate indicators are converted to uniform units to generate a standardized dataset. Principal component analysis (PCA) is used to perform dimensionality reduction on heart rate changes, muscle strength readings, and metabolic rate indicators in the standardized dataset, extracting key features related to joint flexibility to obtain a feature dataset. If the correlation coefficient between heart rate changes and muscle strength readings in the feature dataset exceeds a preset threshold, K-means clustering is used to group the feature dataset to determine the user's exercise mode classification. Based on the exercise mode classification, time series analysis is used to model the heart rate change trend of each group, generating a dynamic change model for each group. Using the dynamic change model, the correlation strength between metabolic rate indicators and joint flexibility for each group of users is calculated to obtain the exercise effect evaluation results.

[0022] This implementation first uses a data acquisition module to control various sensor devices to synchronously start data acquisition tasks according to a preset sampling frequency. The heart rate sensor records heart rate data once per second, measured in times per minute; the muscle strength sensor records the force value generated by muscle contraction or stretching once per second, measured in Newtons; and the metabolic rate sensor records energy consumption values ​​every 5 seconds based on user respiratory gas analysis, measured in kilocalories per minute. All sensor data includes a timestamp, sensor number, and measurement value. The system aligns all raw data by timestamp, forming an initial dataset with intervals of seconds. If a time point is missing, for example, if heart rate data is not recorded at the 20th second, the system calls a linear interpolation module to add the heart rate values ​​recorded at the 19th and 21st seconds and divide by 2 to obtain the complete heart rate value for the 20th second. If a certain indicator has no data for 3 consecutive seconds or more, the system marks that segment as unusable data and removes it from the analysis, ensuring the stability and reliability of subsequent model training.

[0023] Even after completion, the dataset still suffers from inconsistent units, prompting the system to invoke the standardization conversion module. First, it unifies the units of various data types based on the field mapping table, specifically converting all heart rate values ​​to beats per minute, muscle strength to Newtons, and metabolic rate to kilocalories per minute. Next, it normalizes the data values ​​using a max-min normalization method for each indicator. For example, if the heart rate range is 60 to 180, it is normalized by subtracting 60 from each value and then dividing by 120, thus mapping all heart rate values ​​to the range of 0 to 1. Muscle strength and metabolic rate are processed in the same way according to their maximum and minimum ranges. After normalization, the standardized dataset is complete, containing three indicators at the same scale: heart rate variability, muscle strength readings, and metabolic rate.

[0024] Subsequently, the system applies principal component analysis (PCA) to reduce the dimensionality of the standardized dataset. The first step calculates the average value of each indicator and the centered covariance matrix. The second step extracts the most significant eigenvalues ​​and corresponding eigenvectors from the covariance matrix. The third step uses the eigenvectors to perform matrix multiplication on each original sample data to obtain the principal component scores. The system extracts the first two principal components as the core features after dimensionality reduction. These two principal components are the two most strongly correlated combination variables of joint flexibility, forming a new feature dataset for subsequent clustering modeling.

[0025] Within the feature dataset, the system calculates the correlation coefficient between heart rate variability and muscle strength readings. The method involves calculating the mean of each indicator, then taking the difference between each data point and the mean, summing the products of these differences, and dividing by the product of the standard deviations of the two sets of data. The resulting correlation coefficient is then compared to a set threshold of 0.7. This threshold, determined by professional exercise physiology research, serves as a standard reference for measuring strong correlations between physiological indicators. If the correlation coefficient is greater than 0.7, the system considers heart rate and muscle strength to be highly correlated and activates the K-means clustering module to perform cluster analysis.

[0026] The K-means clustering process is as follows: First, all samples in the feature dataset are initialized into 3 classes, with an initial centroid for each class. Second, the Euclidean distance between each sample and the 3 centroids is calculated, and the sample is assigned to the class with the shortest distance. Third, the centroid of each class is updated to the average value of all samples within that class. Fourth, the assignment and update operations are repeated until the classification of all samples no longer changes or 30 iterations have been completed, with the earlier result being the stopping criterion. After clustering is complete, the system generates a unique training pattern classification label for each group of samples.

[0027] Based on the time series data of each sample type, the system extracts heart rate change data and sequentially builds a dynamic change model. The dynamic change modeling process is as follows: heart rate data is processed by moving average in 5-second time windows, and the average value of all heart rate data within each window is calculated. The system continuously moves the time windows to generate an average value sequence, which is used to construct a heart rate change trend model for each type of exercise. This model reflects the dynamic evolution of the user's physiological response under each exercise mode.

[0028] Finally, the system invokes the evaluation module to perform correlation analysis using the metabolic rate index and flexibility score within the corresponding time period of the dynamic model described above. The flexibility score is derived from a quantitative score of the user's flexibility performance before and after exercise, ranging from 0 to 100. The system calculates the mean and standard deviation of the metabolic rate and flexibility score for each category, and then evaluates the strength of the correlation between the two using the Pearson correlation coefficient. The correlation coefficient obtained for each category is the evaluation result of the exercise effect. This value is used to determine the strength of different exercise modes in improving joint flexibility, providing data support for subsequent recommendations and feedback.

[0029] The normalization module employs a normalization transformation method to address dimensional differences and uneven sampling frequencies in the initial dataset, unifying all indicators to a standard scale to obtain a normalized dataset. This includes processing the dimensional differences and sampling frequency unevenness in the initial dataset through min-max normalization, mapping heart rate changes, muscle strength readings, and metabolic rate indicators to a unified scale interval. Data cleaning techniques are used to remove noise and outliers from the normalized dataset. If data points deviate from a preset statistical threshold, outliers are replaced using median imputation, resulting in a cleaned dataset. Finally, feature selection methods are used to filter indicators related to joint flexibility in the cleaned dataset, and heart rate is calculated based on the information gain algorithm. The weights of heart rate changes, muscle strength readings, and metabolic rate indicators are assigned to determine the key indicator set. If the correlation between heart rate changes and metabolic rate indicators in the key indicator set exceeds a preset threshold, a linear regression algorithm is used to model the key indicator set, resulting in an indicator association model. Based on the indicator association model, a time series decomposition method is used to analyze the periodic patterns of heart rate changes, extracting periodic trends to obtain a trend dataset. Using the trend dataset, a sliding window technique is used to dynamically analyze the muscle strength readings of each group of users, calculating the strength change characteristics within each time window to obtain a dynamic feature set. Based on the dynamic feature set, a decision tree algorithm is used to classify the users' exercise patterns, determine the joint flexibility level of each group, and obtain the classification results.

[0030] In this implementation, the system first normalizes the heart rate variability, muscle strength readings, and metabolic rate indicators in the initial dataset to unify the dimensions and sampling frequencies of different indicators. To do this, the system extracts the maximum and minimum values ​​of each indicator in the current dataset. For example, the minimum value for heart rate variability is 60, and the maximum value is 180; the minimum value for muscle strength is 0, and the maximum value is 800; and the minimum value for metabolic rate is 0, and the maximum value is 12. The system subtracts the minimum value of each indicator from each original data value and then divides it by the difference between the maximum and minimum values, mapping the value to a closed interval between 0 and 1, thus achieving consistency in the dimensions of the indicators. When dealing with the problem of uneven sampling frequency, the system establishes a time index table with 1 second as the unit. If there are multiple records of a certain indicator in a certain second, the arithmetic mean of these records is calculated as the representative value of that second. If there is no data in a certain second, the interpolation function is called to perform linear interpolation based on the effective values ​​of the previous second and the next second. That is, the value of the previous second is added to the value of the next second and then divided by 2, which is used as the filling value for the missing second, thereby achieving data alignment on the time axis.

[0031] After normalization, the system performs data cleaning on the resulting normalized dataset to remove or correct outliers caused by sensor fluctuations, abnormal user actions, etc. The system calculates the overall mean and standard deviation for each indicator, and then performs absolute deviation detection on each data point. If a point's value deviates from the mean of its indicator by more than three times the standard deviation, it is considered an outlier. For example, if the mean heart rate is 90 and the standard deviation is 10, then data points below 60 or above 120 will be considered outliers. The system does not delete these points; instead, it replaces them with median-filled values. The median is determined as follows: all non-outlier values ​​in the current dataset for that indicator are extracted, sorted in ascending order, and if the number of data points is odd, the median is taken; if the number is even, the average of the two middle values ​​is taken as the median. This ensures the continuity and statistical representativeness of the cleaned data structure.

[0032] The cleaned dataset then enters the feature selection process. The system uses heart rate variability, muscle strength readings, and metabolic rate as candidate features and calculates their information gain. The information gain calculation process includes three steps: First, the system uses the user's flexibility rating level (e.g., low, medium, high) as the target variable and calculates its entropy value. The entropy value reflects the degree of confusion in flexibility classification within the current sample set. Second, the system divides the candidate indicators into multiple intervals, calculates the proportion of flexibility categories within each interval, and calculates the sum of the weighted entropy values ​​for each interval, i.e., the conditional entropy. Third, the information gain is the entropy value minus the conditional entropy. The system calculates the information gain values ​​for each of the three indicators and compares them with the overall average information gain value, selecting indicators greater than the average to form the key indicator set. In most cases, heart rate variability and metabolic rate are typically high information gain indicators.

[0033] If the correlation between the key indicators "heart rate variability" and "metabolic rate" exceeds a threshold of 0.6, the system performs a linear regression model for them. The correlation is assessed using the Pearson correlation coefficient, calculated as follows: first, the mean and standard deviation of each indicator are calculated; then, the difference between each pair of data and the mean is calculated, and the products are summed to obtain the covariance; finally, the covariance is divided by the product of the two standard deviations to obtain the correlation coefficient. The threshold of 0.6 is determined by the system after comparing the degree of synergy among multiple groups of indicators in experimental data, effectively identifying statistically significant linear relationships between indicator pairs. When the correlation exceeds this value, the system uses heart rate variability as the independent variable and metabolic rate as the dependent variable, employing the least squares method to solve for the slope and intercept of the optimal straight line, and outputs the regression equation. This regression model is called the indicator association model, used to capture the linear synergistic trend of two key physiological variables.

[0034] Based on this, the system applies a time series decomposition method to the heart rate change index. After generating a complete sequence of the index by time, the system uses a 7-day analysis period. First, it calculates the moving average within this period as the long-term trend term. After removing the trend term, it extracts the periodic component from the remaining part, analyzing whether there are regular fluctuations in the remaining sequence within the 7-day period. If so, these are extracted as periodic components. Finally, the system calculates the residual term to evaluate the model error. The trend dataset generated by the above processing contains trend data of periodic heart rate changes, which is used for subsequent dynamic analysis.

[0035] Next, the system uses a sliding window technique to dynamically extract features from the muscle strength readings of each user group. The sliding window parameters are set to a window length of 5 seconds and a sliding step size of 1 second. Within each window, the system calculates feature indicators such as maximum muscle strength, minimum muscle strength, average strength value, and maximum rate of increase (the difference between the current window's maximum value and the previous window's average value divided by time). The system slides the window second by second, repeating the extraction process to ultimately construct a dynamic feature set. This feature set reflects the rate of change, stability, and fluctuation trend of the user's strength during exercise, providing support for subsequent classification modeling.

[0036] Finally, the system invokes a decision tree algorithm to classify and analyze the dynamic feature set, determining the joint flexibility level of each user group. The decision tree generation process includes: first, the system calculates the information gain value of each feature when classifying flexibility levels; selects the feature with the largest information gain as the root node; then, based on this feature, the sample set is divided into several subsets; the above process is repeated for each subset until each leaf node corresponds to a specific flexibility level. Flexibility levels are divided into three categories: low, medium, and high, assigned based on the amplitude and accuracy of the user's flexible movements. Finally, after each group of user data undergoes decision tree path matching, the system outputs its flexibility classification result, completing the entire modeling and recognition process of this implementation method.

[0037] The clustering analysis module applies clustering algorithms to group similar physiological response patterns in a normalized dataset. Based on the grouping results, it determines the user's physical state classification and obtains a classification label set. This includes using the K-means clustering algorithm to group heart rate changes, metabolic rate, and blood oxygen saturation in the normalized dataset, calculating the distance between each data point and the cluster center based on Euclidean distance to determine the physiological response category, and obtaining the classification label set. Based on the classification label set, principal component analysis is used to reduce the dimensionality of the feature vectors of each group of physiological response categories, extracting feature components to obtain a dimensionality-reduced feature set. If the principal component variance contribution rate in the dimensionality-reduced feature set exceeds a preset threshold, linear discriminant analysis is used to divide the classification boundary of the dimensionality-reduced feature set to determine the physiological state of each group. The system first identifies state labels to obtain a state label set. Based on this set, a sliding window technique is used to segment the time-series data of heart rate changes, calculating the statistical characteristics within each window to obtain a time-series feature set. If the mean heart rate change in the time-series feature set deviates from a preset threshold, a moving average filtering method is used to smooth the feature set, eliminating fluctuations and obtaining a smoothed feature set. Based on the smoothed feature set, a decision tree algorithm is used to classify the physiological states of each group, calculating feature weights based on information gain to determine the health status category, resulting in a health classification set. Finally, using the health classification set, a time-series decomposition method is employed to perform periodic analysis on the metabolic rate indicators in the smoothed feature set, extracting long-term trends to obtain a trend feature set.

[0038] In this implementation, the clustering analysis module first calls the K-means clustering algorithm to cluster the normalized dataset. This dataset contains three normalized physiological parameters: heart rate variability, metabolic rate, and blood oxygen saturation. The values ​​of each indicator have been normalized to between 0 and 1 in the preprocessing stage. The clustering algorithm initially sets the number of clusters to 3. The rationale is that by calculating the silhouette coefficient variation curves under different numbers of clusters in the training samples, it was found that the silhouette coefficient is the largest when the number of clusters is 3, indicating that the intra-cluster compactness and inter-cluster separation are optimal at this time. Therefore, the number of clusters is 3 as the default parameter. The specific steps of the K-means algorithm include the following parts: The system first randomly selects 3 sample points as initial cluster centers, and then calculates the Euclidean distance between each sample and all cluster centers. The Euclidean distance is calculated by subtracting the corresponding values ​​of heart rate variability, metabolic rate, and blood oxygen saturation from the corresponding values ​​of the center points, squaring them, adding them together, and finally taking the square root to obtain the distance value. The system compares the distances of the sample to the 3 centers and selects the center with the smallest distance as the category to which the sample belongs. After all samples have been assigned, the system averages the three metrics of all samples in each class to obtain new cluster centers, and then assigns them again. This process iterates until the cluster centers no longer change or the maximum number of iterations (30) is reached. Finally, the system outputs the category to which each sample belongs, forming a classification label set.

[0039] Subsequently, the system treats each category in the classification label set as a subset and applies principal component analysis (PCA) to each category of samples for dimensionality reduction. The dimensionality reduction process first performs mean-zeroing on the three indicators for each sample group, calculates the average value of each indicator in that category, and subtracts the mean from the original value to obtain centered data. Next, the system constructs a 3×3 covariance matrix and performs eigenvalue decomposition on this matrix, extracting the two eigenvectors corresponding to the largest eigenvalues ​​as principal component directions. The data for each sample is projected onto these two principal component directions to obtain the dimensionality-reduced principal component scores, forming the dimensionality-reduced feature set. The system then calculates the principal component variance contribution rate, which is the proportion of variance explained by the first two principal components relative to the total variance of the original data. If this cumulative variance contribution rate is greater than 0.85, it indicates that the dimensionality reduction effect is significant, and the two principal components have covered more than 85% of the data feature information, meeting the requirements for subsequent classification. The threshold of 0.85 is set based on the principle of information preservation. Referring to the theoretical standards of principal component analysis, a value between 0.8 and 0.9 is usually considered a high-quality dimensionality reduction. The choice of 0.85 is to achieve a balance between information preservation and model simplification.

[0040] If the above conditions are met, the system further applies linear discriminant analysis to the dimensionality-reduced feature set, aiming to delineate the boundaries between different classes and generate physiological state labels. The linear discriminant analysis steps include: calculating the class mean vector for the principal component scores of each class; then calculating the sum of squared differences between all samples and their corresponding class means to obtain the intra-class divergence; next, calculating the sum of squared differences between all class means to obtain the inter-class divergence; the system determines the optimal linear projection direction based on the principle of maximizing the ratio of inter-class divergence to intra-class divergence, and projects all samples along this direction, delineating boundaries based on the projection values, completing the assignment of state labels for each class, and obtaining the state label set.

[0041] Next, based on the status label set, the system segments each group of heart rate change data into sliding windows in chronological order. The window parameters are set to a length of 10 seconds and a sliding step size of 2 seconds. This setting is based on the fact that the heart rate data sampling frequency is once per second; 10 seconds can form a relatively stable short-term response range, while the 2-second step size balances coverage and feature resolution. Within each window, the system calculates the average, maximum, minimum, and standard deviation of heart rate changes as statistical features for that window, forming a time-series feature set. The system determines whether the mean heart rate of each window deviates from the resting heart rate threshold. The resting heart rate is set by the user during registration or by the average value during long-term monitoring, and the threshold deviation range is set to 20. That is, if the mean heart rate of a certain window is higher than 85 or lower than 45, it is considered an abnormal fluctuation. This threshold is set based on clinical physiological studies; 65 is the standard resting heart rate for common adults, and a range of ±20 covers most normal activity heart rate fluctuation ranges. If deviations exist, the system applies a moving average filtering method, sets the window width to 5, calculates the average value of 5 windows (2 windows before and after the current window), and uses this value to replace the average heart rate value of the current window to achieve data smoothing and output a smooth feature set.

[0042] Using a smoothed feature set as input, the system calls a decision tree algorithm to classify physiological states. During decision tree training, the system first calculates the information gain value of all statistical features for classifying the target variable "health status," selects the feature with the highest information gain as the root node, divides the data at the optimal split point, and recursively constructs branches until all leaf nodes correspond to only one health status. The health status categories are set to three: "healthy," "sub-healthy," and "high-risk," defined by medical experts based on physiological indicator ranges and risk assessment standards. Each sample in the training data is labeled by a doctor. The system outputs the classification results as a health category set.

[0043] Finally, the system performs periodic trend extraction on the metabolic rate time series data in the health classification set. The system constructs a complete sequence for each group of metabolic rate data in chronological order and performs three processes using a time series decomposition method: first, it performs a 5-point moving average to extract the long-term trend; second, it sets the period parameter to 7 days and identifies periodic components through daily difference analysis; finally, it calculates the residual components, i.e., the short-term disturbance term, and retains the trend term as a trend feature set, which is used to assist in the judgment and feedback of long-term health change trends.

[0044] The system uses a decision tree algorithm to classify these feature samples, aiming to assign a specific health status category label to each sample. The decision tree construction process is as follows: First, the system calculates the information gain of each feature with respect to the target variable "health status category". The information gain is calculated as follows: First, the entropy value of the "health status" variable in the training set (i.e., the information uncertainty of all categories) is calculated. Then, using each feature as a condition, the conditional entropy of the health status under different value intervals of that feature is calculated. The information gain is the original entropy value minus the conditional entropy. The higher the information gain, the more helpful the feature is in classification. Second, the system selects the feature with the largest information gain as the current decision node, selects an optimal split point from that feature, and divides the samples into two subsets. Third, the system repeats the above process in each subset, recursively building child nodes until each leaf node contains only one health status label or the number of samples in the node is less than a preset minimum sample threshold (e.g., 10 samples), at which point the splitting stops. After constructing the complete decision tree, the system judges new smooth feature samples. Starting from the root node, the system sequentially compares the feature values ​​of samples with the threshold values ​​set for each node, proceeding along the path into the left or right subtree until a leaf node is reached. The category corresponding to the leaf node is the health status label for that sample. Finally, the system assigns health status labels to all samples, including three categories: "healthy," "sub-healthy," or "high-risk," forming a "health category set." Then, using this health category set as an index, the system extracts time-series data of metabolic rate indicators from the smoothed feature set and performs time-series decomposition analysis on it. The goal of this process is to extract long-term trends from the metabolic rate data to assess whether the user's physiological state is showing continuous improvement or deterioration.

[0045] The specific steps of time series decomposition are as follows: First, the system arranges the metabolic rate sequence of each user in chronological order to construct a complete sequence. Second, the system performs a moving average process, using a time window of length 5, taking the average of 5 consecutive time points each time as the trend estimate of the center point, and sliding the entire sequence sequentially to obtain a curve representing the long-term trend change. Next, the system analyzes the direction and magnitude of the trend curve, such as continuously rising or falling trend segments, and the slowdown in the degree of fluctuation, to assess whether the user's metabolic level is improving in a healthy direction. If periodic analysis is enabled in the system configuration, the periodic difference method is further applied to detect whether there is a recurring fluctuation pattern with a 7-day cycle, used to identify rhythmic metabolic response characteristics. The above processing results constitute a "trend feature set," which represents the trend data of each user's metabolic rate evolution over a period of time, and can be used by the system to further generate intervention suggestions or train personalized recommendation models.

[0046] The feature fusion module, if the heterogeneity of the classification label set is higher than a preset value, then it fuses indicators from different parts through feature extraction to obtain a fused feature vector. If the heterogeneity of the classification label set is higher than a preset threshold, then it performs dimensionality reduction processing on cardiopulmonary function data and core muscle group data through principal component analysis to obtain a dimensionality-reduced feature set. Based on the dimensionality-reduced feature set, it uses kernel density estimation to analyze the probability density of physiological indicators to obtain density distribution characteristics. If the peak value of the density distribution characteristics deviates from a preset range, then it uses wavelet transform to decompose the time-frequency characteristics of cardiopulmonary function data to obtain a time-frequency feature set. Based on the time-frequency feature set, it uses a support vector machine algorithm to classify the dynamic patterns of core muscle group data to determine dynamic pattern labels. If the dynamic pattern label indicates an abnormal state, then it uses sliding window technology to segment the time series of physiological indicators to obtain a segmented feature set. Based on the segmented feature set, it calculates the statistical characteristics within each window to obtain a statistical feature set. Based on the statistical feature set, it uses time series analysis to decompose the periodic changes of physiological indicators to obtain a periodic feature set.

[0047] In this embodiment, the feature fusion module is activated when the heterogeneity of the classification label set exceeds a preset threshold. The system first calculates the heterogeneity of the classification label set. Specifically, this involves grouping all samples by label, calculating the intra-class covariance matrix for each group, calculating the inter-class covariance matrix for the class mean vectors of all groups, and then dividing the determinant of the inter-class covariance matrix by the determinant of the intra-class covariance matrix to obtain the heterogeneity value. If this heterogeneity value is below 0.3, the system considers the distinction between different categories insufficient, indicating cross-confusion, and requires feature fusion to improve the modeling foundation. The threshold of 0.3 is determined by testing the relationship between heterogeneity and classification accuracy in 50 sets of experimental data. It was found that when the heterogeneity is below 0.3, the classification accuracy drops by an average of more than 20 percentage points; therefore, this value is set as the lower limit for triggering the fusion operation.

[0048] After the fusion trigger condition is met, the system invokes principal component analysis (PCA) to perform dimensionality reduction on the two sets of indicators: "cardiopulmonary function data" and "core muscle group data." Cardiopulmonary function data includes normalized heart rate, respiratory rate, and blood oxygen saturation; core muscle group data includes electromyographic signal amplitude, signal frequency, and contraction duration. The PCA calculation process is as follows: the system centers each data set by subtracting the mean of each indicator from the mean of the sample set, constructs a covariance matrix, and performs eigenvalue decomposition on this matrix. The eigenvectors corresponding to the first two eigenvalues ​​are extracted, and the original multidimensional data are projected onto these two directions to form the dimensionality-reduced principal component data, called the dimensionality-reduced feature set. The criterion for selecting eigenvectors is a cumulative variance contribution rate exceeding 0.85, meaning the first two principal components cumulatively explain more than 85% of the original data variance, ensuring sufficient information retention. This threshold is set according to statistical standards, and experiments have shown that when it is above 0.85, the performance of the dimensionality-reduced model is close to that of the full-dimensional data.

[0049] Subsequently, the system applies kernel density estimation to model the probability distribution of the dimensionality-reduced feature set. The Gaussian kernel function is chosen as the basis for kernel density estimation due to its continuity and natural decay characteristics for outliers. The bandwidth parameter is set to 0.2, which is automatically optimized using the Silverman rule combined with the current sample standard deviation to ensure that the estimated curve is neither overly smooth nor generates multi-peak interference. The Silverman rule is used to automatically optimize the bandwidth parameter of the kernel function in conjunction with the current sample standard deviation. Specifically, the system first calculates the standard deviation for each principal component data dimension, reflecting the degree of fluctuation of the current sample in that dimension. Then, the total number of samples in that dimension is obtained, and these two parameters are substituted into the Silverman rule. The bandwidth value is determined by the standard deviation and the number of samples, where the bandwidth is directly proportional to the standard deviation and inversely proportional to the first power of the number of samples. This ensures that a narrower bandwidth is used when the number of samples is large or the data fluctuation is small, and a wider bandwidth is used when the samples are sparse or the fluctuation is severe, balancing the risks of overfitting and oversmoothing. The system then generates a Gaussian kernel function centered on each sample point using the aforementioned bandwidth, and superimposes the kernel functions corresponding to all samples to construct the overall probability density distribution curve. This method requires no manual parameter setting and relies entirely on the statistical characteristics of the samples themselves for dynamic adjustment. It adapts to the differences in amplitude and distribution between cardiopulmonary function and muscle group data, while also meeting the stability and repeatability requirements of automated processing. This ensures that the final density estimation result maintains the continuity of distribution while possessing the ability to sensitively identify the main peak, concentrated regions, and abnormal regions. The system superimposes a Gaussian curve centered on each sample point, sums all sample curves, and normalizes them to form the overall probability density curve. The system extracts the main peak position and maximum value from the density curve. If the main peak position is not within the interval of 0.4 to 0.6, it is considered a peak shift. This range is set through statistical analysis of the central tendency of normally distributed samples, indicating that the distribution center of most samples is within this range. If it deviates, it indicates that the sample distribution is abnormal and further time-frequency analysis is required.

[0050] The system then performs wavelet transform on the cardiopulmonary function data, selecting the Daubechies 4 type wavelet with a three-layer transform. The process for each layer is as follows: the system performs discrete wavelet decomposition on the original signal, obtaining low-frequency approximation coefficients and high-frequency detail coefficients. Signal components with frequencies ranging from high to low are extracted from layers 1 to 3. After reconstruction, time signals in different frequency bands are obtained, and information such as energy distribution, peak frequency, and fluctuation period are extracted to form a time-frequency feature set. The parameter setting of three layers is based on empirical verification, ensuring both frequency resolution and controlling time accuracy loss, making it suitable for analyzing non-stationary signals such as heart rate.

[0051] After obtaining the time-frequency feature set, the system feeds the core muscle group data into a support vector machine (SVM) classification model for dynamic pattern recognition. The SVM uses radial basis functions as the kernel function. The sample input consists of multidimensional features of the electromyographic signal over a short period, including average amplitude, maximum frequency, and duration. The output is either "normal" or "abnormal." The training process uses 10-fold cross-validation for model evaluation, ultimately determining the optimal penalty parameter to be 1 and the kernel width to be 0.5. These two parameters control the tolerance of the classification boundary and the response range of the kernel function, respectively. After model training, each new data input is classified, and the corresponding dynamic pattern label is output.

[0052] If the dynamic pattern label of a sample is "abnormal," the system further performs sliding window processing. The window length is set to 10 seconds, and the step size is set to 2 seconds. This is because the effective period of most electromyographic signals is between 8 and 12 seconds; a 10-second setting can cover a complete physiological event, and the 2-second step size is used to improve temporal accuracy and feature density. Within each window, the system extracts six statistical features, including maximum value, minimum value, mean, standard deviation, rate of change (the difference between the mean of the current window and the mean of the previous window), and coefficient of variation (the ratio of standard deviation to mean), forming a segmented feature set.

[0053] After completing the segmented feature set, the system calls the time series analysis module to perform periodic analysis. The periodic analysis process consists of three steps: First, the system uses a 5-point moving average to extract the long-term trend component; second, it performs seasonal differencing on each indicator sequence, with the period set to 7, to identify whether there is a recurring fluctuation pattern every 7 time points; third, the residual part is extracted as short-term perturbation features. The period is set to 7 based on experimental data showing that metabolic changes in exercise behavior exhibit a high-frequency recurrence trend within 7 periods, thus facilitating the identification of rhythmic behavior. Finally, the trend and periodic components are retained, and the periodic feature set is output as the final analysis result after fusion, supporting subsequent regression modeling and intelligent recommendation module calls.

[0054] The regression modeling module constructs a regression model based on fused feature vectors to predict the quantitative impact of exercise on physiological function, obtaining an impact score matrix. This includes standardizing the feature vector set using data preprocessing methods to obtain a standardized feature set; if the variance of the standardized feature set exceeds a preset threshold, dimensionality reduction is performed using principal component analysis to obtain a dimensionality-reduced feature set; based on the dimensionality-reduced feature set, a random forest algorithm is used to train the regression model, calculating feature weights to obtain training model parameters; based on the training model parameters, cross-validation is used to evaluate the performance of the regression model, obtaining cross-validation results; if the mean squared error of the cross-validation results is lower than a preset threshold, the regression model is used to predict the dimensionality-reduced feature set, obtaining an impact score matrix; based on the impact score matrix, cluster analysis is used to group the physiological function data, obtaining a grouped result set; based on the grouped result set, time series analysis is used to extract trend features for each group, obtaining a trend feature set.

[0055] In this implementation, the main function of the regression modeling module is to construct a regression model based on fused feature vectors to quantify and predict the specific impact of exercise on users' physiological functions, thereby forming an impact score matrix. First, the system performs standardization on the input fused feature vector set. Specifically, it uses zero-mean unit variance standardization, subtracting the mean of each feature data from the entire sample set and dividing by the standard deviation to ensure that different features have a uniform numerical scale, thus eliminating interference from dimensional differences in model training. The standardized feature set is then called the standardized feature set. The system calculates the overall variance of this set. If the variance is greater than 0.25, the system determines that there is a strong correlation or redundant information among the features, and therefore performs principal component analysis for dimensionality reduction. This threshold is determined based on experimental data analysis. When the standardized feature variance exceeds 0.25, the sample distribution in the multidimensional space exhibits a significant linear aggregation trend, leading to a decrease in the model's predictive ability; therefore, it is set as the boundary for redundancy judgment.

[0056] The dimensionality reduction process is as follows: The system first constructs the covariance matrix of the standardized feature set, then performs eigenvalue decomposition, extracting the top principal components with a cumulative variance contribution rate greater than 90% to form the dimensionality-reduced feature set. This effectively compresses data dimensionality while retaining key information, improving modeling efficiency. The contribution rate threshold is set to 0.90 to balance information retention and model complexity. Tests on most sample sets have shown that this value maintains prediction accuracy and training stability.

[0057] The dimensionality-reduced feature set is then fed into the regression modeling process. The regression model constructs a mathematical function by analyzing the mapping relationship between input variables (independent variables) and output variables (dependent variables), ensuring that the input feature values, when substituted, can output a predicted value as accurately as possible. The essence of the model is to minimize the error between the predicted value and the true value. In this system, the input is the dimensionality-reduced fused feature vector, where each dimension represents the principal component of different physiological dimensions. The output is the predicted value of the exercise effect index, such as the percentage increase in metabolic rate. The influence score matrix output by the regression model represents the predicted effect of each exercise program on multiple physiological dimensions. For example, the value in the 3rd row and 2nd column represents the expected improvement for the 3rd user in the 2nd physiological function dimension. This matrix can serve as the basis for evaluating subsequent exercise recommendations. A regression model is constructed based on the obtained dimensionality-reduced feature set, trained using the random forest algorithm, and the weights of each feature are calculated to obtain the final training model parameters. The entire implementation process is as follows: First, the system uses the dimensionality-reduced feature set as the input dataset, where each sample consists of several principal components of the reduced-dimensional features. The target variable is the quantitative result of one or more physiological indicators of the user after exercise, such as heart rate recovery time, muscle strength increase, or percentage change in metabolic rate. The system initializes a random forest model, setting the number of trees to 100, the maximum tree depth to 10 layers, and the minimum number of leaf node samples per tree to 5. These parameters are determined through grid search combined with training and validation errors. The 100 trees cover most feature combinations, the 10-layer depth is sufficient to express non-linear structures, and the 5 leaf node samples avoid overfitting. During the model training phase, the system performs bootstrap sampling with replacement on the training data, that is, randomly sampling from the original training samples to construct a training subset for each tree, and simultaneously randomly extracting several dimensions from the dimensionality-reduced features to form a feature subset for that tree, thus ensuring the diversity of each tree. Each tree is constructed recursively. At each split node, the system selects the feature with the highest information gain within the current feature subset and determines the optimal split point to divide the data into two subsets. This process continues until the maximum depth or insufficient sample conditions are met. After the entire forest is constructed, the system statistically analyzes the split node information of each tree, calculates the frequency with which each feature is used for splitting in the forest, and accumulates the total reduction in prediction error for the target variable, which is used as the importance score for that feature. The system normalizes the importance scores of all features to obtain a complete feature weight vector. This vector reflects the contribution of each dimensionality-reduced feature to predicting changes in physiological function. Finally, the system saves the structural parameters, splitting conditions, feature weight distribution, and residual error information recorded during training for all decision trees as training model parameters for subsequent validation, prediction, and model updates.Through the above steps, the system achieves random forest regression modeling based on dimensionality-reduced feature sets, effectively obtaining the feature influence strength under multidimensional nonlinear relationships, and providing a high-precision modeling foundation for personalized exercise program recommendations. The system selects the random forest algorithm to construct the regression model. This algorithm consists of multiple regression decision trees, each using different feature subsets and sample subsets during training to avoid overfitting. During training, the system uses the fused features in the training samples to establish a mapping relationship with the corresponding target variables (such as changes in heart rate before and after exercise, muscle strength improvement ratio, etc.), outputs the prediction results of each tree, and forms the final prediction value by averaging the outputs of multiple trees. The system statistically analyzes the frequency with which each feature is used for splitting nodes across all trees and the degree of error reduction after splitting, thereby calculating the importance weight of each feature to form the training model parameters.

[0058] After model training is complete, the system uses five-fold cross-validation to evaluate the regression model's performance. This involves dividing all samples into five subsets, selecting one subset as the validation set in each round, and using the remaining four subsets as the training set. This process is repeated five times, and the mean squared error (MSE) between the predicted results and the true values ​​in each round is calculated. The average of these MSE values ​​is then used as the cross-validation result. If this MSE is less than 0.15, the model demonstrates high fitting and generalization ability and is suitable for practical prediction. This threshold is set based on the fact that most users' physiological indicators fall between 0 and 1. An error controlled within 0.15 corresponds to a prediction error of no more than 15%, which is acceptable for clinical applications and is therefore set as the evaluation criterion.

[0059] After successful validation, the system uses the trained model to predict the entire dimensionality-reduced feature set, obtaining an influence score matrix representing the strength of each feature's impact across different physiological function dimensions. Each row in the influence score matrix corresponds to a user sample, and each column corresponds to a specific physiological function indicator, such as cardiorespiratory endurance, muscle power, and metabolic efficiency. The numerical value represents the predicted quantitative score of the training effect.

[0060] Next, the system applies cluster analysis to group the influence score matrix. The algorithm used is K-means clustering, with the number of clusters K set to 5. The K value was determined by evaluating the grouping effect of the model within a range of 2 to 10 based on the silhouette coefficient. It was found that the silhouette coefficient was largest when K was 5, indicating optimal intra-group compactness and inter-group separation; therefore, this value was adopted. The system calculates the Euclidean distance between each sample and each cluster center, assigning the samples to the nearest cluster, and finally generating the grouping result set.

[0061] Finally, the system performs time series analysis on each group to extract the physiological indicator trend characteristics of each group of users at different time periods. The analysis process includes: first, smoothing the physiological data of each user using a 5-point moving average method to remove the influence of short-term fluctuations; then, performing periodic decomposition using a seasonal trend decomposition method, with a period length set to 7 to represent a one-week exercise cycle; the system extracts trend items and analyzes their slope changes and directions to form a trend feature set. This trend feature set is used by the subsequent intelligent recommendation module to identify the user's exercise response trends at different times, achieving more accurate optimization of exercise plans.

[0062] The contribution evaluation module calculates the contribution of each exercise method to different body parts using an influence score matrix. Modes with contribution values ​​exceeding a preset threshold are marked as effective. The effective mode list is generated by: quantifying the contribution of each exercise method to each physiological part using a weighted calculation method based on the influence score matrix, resulting in a contribution vector; using a threshold judgment method based on the contribution vector, marking exercise methods as effective modes if their contribution exceeds a preset threshold, generating a preliminary mode set; grouping the effective modes from the preliminary mode set using cluster analysis based on the correlation of physiological parts, resulting in mode group sets; calculating the average contribution of each effective mode group using a mean aggregation method, resulting in a grouped contribution set; extracting the effective mode with the highest contribution from each group from the grouped contribution set, generating a preferred mode set; sorting the effective modes from the preferred mode set using a ranking algorithm, from highest to lowest contribution, resulting in a final effective mode list; and extracting multiple modes from the final effective mode list to generate a mode list.

[0063] The system first obtains the influence score matrix output from the regression modeling module. This matrix is ​​a two-dimensional array, with rows representing different exercise methods and columns representing different physiological indicators, such as muscle strength, joint flexibility, metabolic rate, and cardiorespiratory endurance. The value in each cell is the quantitative influence score of that exercise method on a specific physiological area. To more accurately reflect the comprehensive effectiveness of exercise methods in multidimensional physiological performance, the system introduces a weighting mechanism. Each column of physiological indicators is assigned a weight value, determined through extensive experimental data and expert evaluation; for example, muscle strength has a weight of 0.4, joint flexibility 0.3, metabolic rate 0.2, and cardiorespiratory endurance 0.1. These weights reflect the relative contribution of different indicators to overall health improvement. For each exercise method, its contribution vector is generated by multiplying each indicator value in the corresponding row by its corresponding weight and summing the results, forming a vector containing the weighted score for each area. This vector is used for subsequent threshold determination and cluster analysis.

[0064] After calculating the contribution vector, the system compares and filters the weighted total contribution of each exercise method. The system sets a fixed threshold to determine the effectiveness of each exercise method. This preset threshold is typically set to 0.5, derived from the median contribution of all exercise methods in a large-sample experiment, while also considering that a contribution exceeding this value significantly increases the likelihood of positive improvements in at least three indicators. If the contribution value of each exercise method is equal to or greater than this threshold, it is considered an "effective mode" with substantial improvement effects on the body, and its contribution vector and its number are added to the initial mode set. This step eliminates inefficient exercise methods, ensuring that subsequent analysis only processes programs with significant effects.

[0065] After obtaining the initial pattern set, the system uses the contribution vectors of all exercise methods within it as input samples to perform cluster analysis to identify the internal structure. The K-means algorithm is chosen as the clustering method due to its good stability and computational efficiency. The system first sets the number of clusters K; in this embodiment, K is set to 3, a value determined based on the number of clusters with the largest silhouette coefficients in the training data. The system initializes K centroids and calculates the Euclidean distance between the contribution vector of each exercise method and each centroid. Samples are then classified into the cluster of the centroid with the smallest distance. The centroid positions are then iteratively updated until all samples are classified stably and the error is minimized. After clustering, each type of exercise method exhibits similar contribution structures in certain physiological parts, such as one group primarily effective for the muscular system, and another group primarily improving cardiopulmonary function. This forms a clearly structured and targeted pattern grouping set for subsequent refined recommendations.

[0066] After completing clustering and generating pattern group sets, the system uses each cluster as a basic unit for statistical analysis. Specifically, the system iterates through each cluster, extracts the contribution vector of all exercise methods within that group, and calculates its weighted total contribution, which is the single value synthesized by weighting the indicators as mentioned above. For each group, the system sums all contribution values ​​and divides them by the number of samples in that group to obtain the average contribution value of that group, which is recorded as the representative value of that group. This processing is called the mean aggregation method, which is simple to operate and computationally efficient. Taking K equals 3 as an example, the system will eventually obtain a dataset containing three average values, namely the group contribution set. This set is used to measure the overall effectiveness of exercise methods under different physiological response modes and provides a quantitative basis for the next step of selecting the optimal pattern.

[0067] The system backtracks to find all valid patterns within each cluster based on the index of each cluster in the group contribution set. For each cluster, the system selects the exercise method with the highest contribution value, i.e., the one with the largest total contribution value in the group. This process is implemented through linear scanning and maximum value comparison to ensure that the selected pattern has the strongest representativeness and physiological improvement ability. Only one optimal sample is extracted from each group, and the final set of multiple optimal samples constitutes the preferred pattern set. This set represents the exercise methods that perform best in different physiological response types, balancing diversity with ensuring outstanding exercise effects.

[0068] Each exercise method in the preferred pattern set has a clearly defined total contribution value. The system uses these values ​​as the sorting criteria and arranges them in descending order using a sorting algorithm. In this embodiment, the system selects the merge sort algorithm because it has the characteristics of good stability and low time complexity in small to medium-sized sample data. During the sorting process, the system arranges the patterns from high to low contribution to ensure that the sorting results reflect the actual effect level of each pattern. After sorting, a final list of effective patterns is generated, where each element is an exercise method, and they are arranged from front to back according to the strength of their comprehensive contribution in different physiological dimensions, providing a weighting basis for subsequent calls or recommendations.

[0069] After constructing the final list of effective training patterns, the system selects multiple exercise methods as recommended outputs according to preset rules, forming the final pattern list. There are two extraction methods: one is the fixed-quantity method, directly extracting the first 3 or 5 patterns; the other is based on a contribution threshold method, for example, setting the extraction condition to a contribution greater than 0.7. The current implementation adopts the second method. The 0.7 threshold is derived from the system's statistical analysis of the contribution distribution of exercise methods in a large number of user samples, using the upper quartile as the screening criterion to ensure that the extracted exercise patterns are all high-contribution schemes, representing the most effective training strategies for improving users' health in practice. After extraction, the system provides this pattern list as output for downstream modules to use, such as for generating personalized exercise sequences or for recommendation display in the user interface.

[0070] The intelligent recommendation module matches the user's current physical condition with a list of effective patterns to generate an exercise sequence optimized for joint flexibility, thus obtaining a personalized fitness plan. This involves: using physical condition data, standardizing joint flexibility-related indicators into normalized feature vectors to obtain a user feature dataset; if the joint flexibility indicators in the user feature dataset are below a preset threshold, a classification algorithm is used to filter exercise patterns related to micro-joints from a preset list of effective patterns, resulting in a candidate pattern set; based on the candidate pattern set, a weighted scoring method is used, combined with the normalized feature vectors from the user feature dataset, to calculate the matching degree of each exercise pattern, resulting in a matching degree ranking list; from the matching degree ranking list, the top few exercise patterns with the highest matching degrees are extracted to generate a preliminary personalized sequence; if the number of exercise patterns in the preliminary personalized sequence is insufficient, cluster analysis is used to supplement the candidate pattern set with patterns related to joint flexibility, resulting in an optimized sequence; based on the optimized sequence, time series analysis is used to adjust the execution order of the exercise patterns, generating a final personalized fitness plan; and based on the final personalized fitness plan, an iterative update method is used to adjust the exercise patterns in the plan according to changes in the user's subsequent physical condition data, resulting in a dynamic fitness plan.

[0071] In this implementation, the intelligent recommendation module first receives the user's current physical state data, which includes multiple indicators related to joint flexibility, such as joint range of motion (in degrees), flexibility test scores (out of 100 or a maximum of 100), and muscle length measurements (in centimeters). The system extracts the mean and standard deviation for each indicator from the training sample set. The mean is the arithmetic mean of the measurements for that indicator from all training users, and the standard deviation is the square root of the average of the squared deviations of these values ​​from the mean. For each user, the standardized score is obtained by subtracting the mean of that indicator from their current indicator value and then dividing by the standard deviation. All standardized scores constitute the user's normalized feature vector. Multiple such vectors form the user feature dataset, representing the user's current state in the flexibility dimension.

[0072] The system then checks whether the "joint flexibility index" in the user's feature dataset is below a preset threshold, which is set to -1, meaning the standardized score is below negative one standard deviation. This threshold is based on the standardized score distribution of people with normal flexibility in the training samples, where scores below negative one standard deviation indicate an abnormally low level. This standard has been consistently adopted from medical exercise physiology literature and user health feedback. When the system finds that one or more of a user's joint flexibility indices are below -1, it determines that the user has insufficient flexibility.

[0073] If the above judgment is true, the system filters exercise patterns related to "microjoints" from the list of valid patterns. To this end, each valid pattern has a pre-stored joint influence vector in the system, which indicates the contribution value of the pattern to parts such as fingers, toes, and wrist joints under the "microjoints" category. The system maps the corresponding microjoint indices in the user's standardized feature vector to the corresponding items in the joint influence vectors of these patterns, and calculates the Pearson correlation coefficient between these corresponding items (by multiplying the product of the differences between the user's standardized value and the pattern's influence value, calculating the covariance, and then dividing by the product of the standard deviations of both) to assess the matching between the pattern and the user's microjoint flexibility. The system selects the top 3 patterns with the highest correlation coefficients, or all patterns with correlation coefficients greater than or equal to 0.6, as the candidate pattern set. The threshold of 0.6 is derived from the training observation that when the correlation coefficient is ≥0.6, the flexibility improvement effect of the corresponding pattern on the user is significant in actual testing.

[0074] Next, the system calculates a matching score for each candidate pattern. The matching score method is as follows: the influence value of each micro-joint component in the joint influence vector of the pattern is multiplied by the standardized score of the corresponding component in the user's standardized feature vector, and then multiplied by the weight of that component (these weights, such as 0.3 for finger joints, 0.3 for toe joints, and 0.4 for wrist joints, are determined by statistically analyzing the improvement magnitude of each micro-joint and the user's subjective satisfaction in the training samples). Finally, the sum is obtained to obtain the total matching score of the pattern. After obtaining matching scores for all candidate patterns, the system sorts these scores from high to low to form a matching score ranking list. The system sets the rule for the number of patterns to be extracted as follows: extract the 3 patterns with the highest matching scores or all patterns with matching scores greater than or equal to 0.7 as the initial personalized training sequence. The threshold of 0.7 comes from the upper quartile of the matching score distribution in the training samples, representing a high matching score threshold.

[0075] If the number of patterns extracted from the initial personalized sequence is less than the set minimum (3 patterns), the system will perform a supplementary step: The system uses cluster analysis to group the remaining unselected patterns in the candidate pattern set, with the number of clusters set to 3. This is because the silhouette coefficient is the largest and the clustering structure is optimal when the number of clusters is 3 in the candidate pattern training samples. The clustering method is K-means clustering. Using the matching degree vector of these unselected patterns (i.e., the matching degree term between the pattern and each of the user's micro-joint indicators) as input samples, the system initializes 3 centroids and repeats the assignment and centroid update until the cluster centroids no longer change or the iteration reaches the upper limit (e.g., 30 times). Then, the pattern with the highest correlation to the joint flexibility index or the highest contribution of the matching degree to the user's micro-joint indicators is selected from each cluster group and added to the initial sequence to form the optimized sequence.

[0076] The system establishes time-series characteristic data for each exercise mode within the optimized sequence, specifically including changes in key physiological indicators before and after each execution of the mode, such as improvements in joint flexibility, changes in fatigue index, and muscle recovery time. This data is recorded chronologically according to the user's exercise dates, forming a time-series dataset. Next, the system employs time-series analysis to model and analyze the changing trends of the effects of each exercise mode. The analysis method uses a sliding window approach, with a window length of 3. This means that data points from three consecutive exercises are used for local trend assessment, and the window is slid forward one step at a time, covering the entire sequence to ensure clear identification of short-term trends. By analyzing the mean and rate of change of physiological indicators within each window, the system identifies the optimal exercise frequency and timing for that exercise mode.

[0077] After identifying the trend characteristics of each exercise mode, the system further compares the interdependencies between the modes. If the recovery process of mode A within 24 hours after execution has a positive or negative impact on the effect of mode B (e.g., continued flexibility improvement or failure to recover from fatigue), the system optimizes the order of the two modes accordingly. For example, if mode A requires significant joint recovery after execution, while mode B places high demands on the joints, the system avoids executing B immediately after A, and schedules B for the next day or later. By constructing an inter-mode influence matrix, and combining mode improvement efficiency and recovery cycle, the system reorders the exercise modes in the optimized sequence according to the principle of maximizing flexibility gain and minimizing fatigue accumulation, ultimately generating a logically sound and physiologically appropriate personalized fitness plan.

[0078] After the final personalized fitness plan is implemented by the user, the system continues to monitor changes in the user's physical condition data in real time. Specifically, after each complete training cycle (such as every 7 days or after completing a cycle of all training modes), the system collects and standardizes the user's joint flexibility, muscle strength, and related indicators to obtain a new user feature vector. The system compares the new data with the previous pre-training status data, and if some indicators do not improve significantly or decline, an iterative update mechanism is triggered.

[0079] During the iterative update process, the system recalculates the contribution and matching degree of each exercise mode in the current personalized fitness plan, and determines whether a new mode has a higher matching degree in the current state, or whether the performance of the existing mode has declined. If the above conditions are met, the system supplements, replaces, or adjusts the existing modes from the original candidate mode set, and re-sorts the time series to form an updated dynamic fitness plan. This process continuously iterates and updates to ensure that the plan is always highly matched with the user's current state, truly realizing a personalized, dynamically adaptable fitness guidance system with flexibility optimization as its core goal.

[0080] like Figure 2 As shown, an AI-based intelligent exercise method is also provided. Employing the AI-based intelligent exercise system, the method includes: collecting heart rate changes, muscle strength readings, and metabolic rate indicators related to joint flexibility during exercise using sensor devices; acquiring raw exercise data and physiological response information from multiple sources to obtain an initial dataset containing heterogeneous formats; using a normalization transformation method to handle dimensional differences and uneven collection frequencies in the initial dataset, unifying all indicators to a standard scale to obtain a normalized dataset; applying a clustering algorithm to the normalized dataset to group similar physiological response patterns, determining the user's body state classification based on the grouping results, and obtaining a classification label set; if the heterogeneity of the classification label set is higher than a preset threshold, then feature extraction is used to fuse indicators from different body parts to obtain a fused feature vector; a regression model is constructed based on the fused feature vector to predict the quantitative impact of exercise on physiological function, obtaining an influence score matrix; the contribution of each exercise method to different body parts is calculated using the influence score matrix, and if the contribution is higher than a preset threshold, it is marked as an effective mode, obtaining a list of effective modes; the list of effective modes is used to match the user's current body state to generate an exercise sequence optimized for joint flexibility, obtaining a personalized fitness plan.

[0081] This method first uses devices such as heart rate sensors, electromyography (EMG) sensors, and metabolic analysis devices to continuously collect three types of physiological indicators during the user's actual exercise: heart rate variability, muscle strength readings, and metabolic rate data. Data sources may include real-time devices, historical records, and synchronization information from third-party platforms, resulting in significant differences in data format, units, and sampling frequency, forming a heterogeneous initial dataset. Subsequently, the system preprocesses this dataset using a min-max normalization method, mapping each indicator to a standard range of 0 to 1. To address the differences in sampling frequency across different devices, the system performs time alignment using linear interpolation and generates a normalized dataset after standardizing the unit format.

[0082] On the normalized dataset, the K-means clustering algorithm is used to group the data samples. The K value is dynamically determined based on the silhouette coefficient evaluation results, and is usually set between 3 and 7. Euclidean distance is used to calculate the differences between samples during clustering, ultimately resulting in multiple sets of classification labels representing different physiological response patterns. If the mean difference between samples within each category in the clustering results exceeds the heterogeneity threshold (set to 0.15, determined by the statistical distribution of clustering in the training set), it indicates significant heterogeneity between different physiological responses. In this case, the system uses principal component analysis to reduce the dimensionality of multiple indicators in the classified samples, extracts principal component vectors, and then fuses them into a unified fusion feature vector.

[0083] The fused feature vectors are used as input, standardized, and then fed into a random forest regression model for modeling. The model is trained using five-fold cross-validation, with a mean squared error below 0.05 as the convergence criterion. It outputs predicted values ​​for the impact of each exercise mode on various physiological functions, summarizing these values ​​to form an impact score matrix. On this matrix, the system calculates the weighted contribution of each mode to flexibility, muscle strength, and metabolic level by body part, and filters out modes with significant physiological benefits as effective modes by setting a threshold (e.g., 0.6). All effective modes are compiled into an effective mode list.

[0084] During the recommendation process, the system first collects the user's current state data, which is then standardized to form a user feature vector. By calculating the matching degree between this vector and the physiological indicators of each effective mode, a matching degree list is generated based on a weighted scoring mechanism. Multiple modes with a matching degree higher than 0.7 are selected to form a personalized initial exercise sequence. If the number of modes is less than three, other efficient modes similar to the user's state are supplemented through clustering. Finally, this initial sequence is optimized and rearranged through time series analysis, and combined with a fatigue recovery model and a target flexibility improvement curve, a final fitness plan with the optimal execution order is generated.

[0085] This method integrates multi-source physiological data and employs AI-driven algorithms such as clustering, regression, classification, and time series optimization to achieve a fully intelligent training process from data collection to personalized recommendations. Compared to traditional training recommendation methods, its advantages are: first, it achieves standardized data fusion, adapting to various sensors and data platforms; second, it introduces dynamic clustering and feature fusion mechanisms to adapt to individual differences in physiological responses; third, it uses a regression prediction mechanism to accurately quantify training effects and assist in scientific decision-making; and fourth, it provides iteratively optimized dynamic solutions that continuously adapt to changes in user status, thereby improving training effectiveness, reducing the risk of sports injuries, and significantly enhancing the user experience.

[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent exercise system based on AI intelligent recommendation, characterized in that, include: The data acquisition module collects heart rate changes, muscle strength readings, and metabolic rate indicators of the user's joint flexibility during exercise through sensor devices, and obtains raw exercise data and physiological response information from multiple sources to obtain an initial dataset containing heterogeneous formats. The normalization module uses a normalization transformation method to handle the differences in units and the unevenness of collection frequency in the initial dataset, unifying all indicators to a standard scale to obtain a normalized dataset. The clustering analysis module applies a clustering algorithm to the normalized dataset to group similar physiological response patterns, determines the user's physical state classification based on the grouping results, and obtains a set of classification labels. The feature fusion module, if the heterogeneity of the classification label set is higher than a preset threshold, will extract and fuse indicators from different parts to obtain a fused feature vector. The regression modeling module constructs a regression model based on the fused feature vectors to predict the quantitative impact of exercise on physiological function and obtains the impact score matrix. The contribution assessment module calculates the contribution of each exercise method to different body parts through the influence score matrix. If the contribution is higher than a preset threshold, it is marked as an effective mode, and a list of effective modes is obtained. The intelligent recommendation module uses a list of effective patterns to match the user's current physical condition, generates an exercise sequence optimized for joint flexibility, and provides a personalized fitness plan.

2. The intelligent exercise system based on AI intelligent recommendation according to claim 1, characterized in that: The data acquisition module collects heart rate changes, muscle strength readings, and metabolic rate indicators related to joint flexibility during exercise using sensor devices. It acquires raw exercise data and physiological response information from multiple sources, resulting in an initial dataset containing heterogeneous formats, including: By collecting heart rate changes, muscle strength readings, and metabolic rate indicators of users during exercise using sensor devices, raw data is obtained from multiple sources to generate an initial dataset. If the collected data is missing or abnormal, linear interpolation is used to complete the heart rate changes, muscle strength readings and metabolic rate indicators to obtain a complete dataset. Standardization processing was used to unify the heterogeneous formats in the complete dataset. Based on the data field mapping table, heart rate variability, muscle strength readings and metabolic rate indicators were converted into uniform units to generate a standardized dataset. Principal component analysis was used to reduce the dimensionality of heart rate changes, muscle strength readings and metabolic rate indices in the standardized dataset, and key features related to joint flexibility were extracted to obtain a feature dataset. If the correlation coefficient between heart rate changes and muscle strength readings in the feature dataset is greater than a preset threshold, the feature dataset is grouped using the K-means clustering algorithm to determine the user's exercise mode classification. Based on the exercise mode classification, time series analysis was used to model the heart rate change trend of each group, generating dynamic change models for each group. By using a dynamic change model, the correlation strength between metabolic rate and joint flexibility for each group of users is calculated to obtain the evaluation results of the exercise effect.

3. The intelligent exercise system based on AI intelligent recommendation according to claim 1, characterized in that: The normalization module uses a normalization transformation method to handle the differences in units and uneven collection frequencies in the initial dataset, unifying all indicators to a standard scale to obtain a normalized dataset, including: By processing the dimensional differences and uneven collection frequencies in the initial dataset through min-max normalization, heart rate changes, muscle strength readings and metabolic rate indicators are mapped to a uniform scale interval to obtain a normalized dataset. Data cleaning techniques are used to process noise and outliers in the normalized dataset. If a data point deviates from a preset statistical threshold, the outlier is replaced by the median padding method to obtain a cleaned dataset. The key indicator set was determined by filtering indicators related to joint flexibility in the clean dataset using feature selection methods and calculating the weights of heart rate change, muscle strength readings and metabolic rate indicators based on the information gain algorithm. If the correlation between heart rate changes and metabolic rate indicators in the key indicator set exceeds a preset threshold, the key indicator set is modeled using a linear regression algorithm to obtain an indicator correlation model. Based on the index association model, the time series decomposition method is used to analyze the periodic patterns of heart rate changes, extract the periodic trends, and obtain a trend dataset. Using trend datasets, we employ sliding window technology to dynamically analyze muscle strength readings for each group of users, calculate the strength change characteristics within each time window, and obtain a dynamic feature set. Based on the dynamic feature set, the decision tree algorithm is used to classify the user's exercise patterns, determine the joint flexibility level of each group, and obtain the classification results.

4. The intelligent exercise system based on AI intelligent recommendation according to claim 1, characterized in that: The clustering analysis module applies a clustering algorithm to the normalized dataset to group similar physiological response patterns, determines the user's physical state classification based on the grouping results, and obtains a set of classification labels including: K-means clustering algorithm was used to group heart rate changes, metabolic rate and blood oxygen saturation in the normalized dataset. The distance between each data point and the cluster center was calculated based on Euclidean distance to determine the physiological response category and obtain the classification label set. Based on the classification label set, principal component analysis was used to reduce the dimensionality of the feature vectors of each group of physiological response categories, and feature components were extracted to obtain the dimensionality-reduced feature set. If the principal component variance contribution rate in the dimensionality reduction feature set exceeds a preset threshold, the dimensionality reduction feature set is divided into classification boundaries using linear discriminant analysis to determine the physiological state labels of each group, thus obtaining a state label set. Based on the status label set, the time series data of heart rate changes are segmented using the sliding window technique, and the statistical characteristics within each window are calculated to obtain the time series feature set. If the mean heart rate change in the time series feature set deviates from the preset threshold, the time series feature set is smoothed by the moving average filtering method to eliminate fluctuations and obtain a smoothed feature set. Based on the smooth feature set, the physiological state of each group is classified using the decision tree algorithm. The feature weights are calculated based on information gain to determine the health status category, thus obtaining the health classification set. Using a health classification set, a time series decomposition method is employed to perform periodic analysis on the metabolic rate indicators in the smooth feature set, extracting long-term trends to obtain a trend feature set.

5. The intelligent exercise system based on AI intelligent recommendation according to claim 1, characterized in that: If the heterogeneity of the classification label set is higher than a preset value, the feature fusion module extracts and fuses indicators from different parts to obtain a fused feature vector, including: If the heterogeneity of the classification label set is higher than the preset threshold, then principal component analysis is used to reduce the dimensionality of the cardiopulmonary function data and core muscle group data to obtain the dimensionality-reduced feature set. Based on the dimensionality reduction feature set, the probability density of physiological indicators is analyzed using the kernel density estimation method to obtain the density distribution characteristics; If the peak value of the density distribution characteristics deviates from the preset range, the time-frequency characteristics of the cardiopulmonary function data are decomposed by wavelet transform to obtain the time-frequency feature set; Based on the time-frequency feature set, the support vector machine algorithm is used to classify the dynamic patterns of the core muscle group data and determine the dynamic pattern labels. If the dynamic mode label indicates an abnormal state, the time series of physiological indicators is segmented using the sliding window technique to obtain segmented feature sets. Based on the segmented feature set, the statistical features within each window are calculated to obtain the statistical feature set; Based on the statistical feature set, time series analysis was used to decompose the periodic changes of physiological indicators to obtain the periodic feature set.

6. The intelligent exercise system based on AI intelligent recommendation according to claim 1, characterized in that: The regression modeling module constructs a regression model based on the fused feature vectors to predict the quantitative impact of exercise on physiological function, obtaining an impact score matrix including: Based on the fused feature vectors, the feature vector set is standardized using data preprocessing methods to obtain a standardized feature set. If the variance of the standardized feature set exceeds a preset threshold, the standardized feature set is dimensionality reduced by principal component analysis to obtain a dimensionality-reduced feature set. Based on the dimensionality-reduced feature set, a regression model is trained using the random forest algorithm, and feature weights are calculated to obtain the parameters of the trained model. Based on the parameters of the trained model, the performance of the regression model is evaluated using the cross-validation method, and the cross-validation results are obtained. If the mean squared error of the cross-validation results is lower than the preset threshold, the dimensionality-reduced feature set is predicted by the regression model to obtain the influence score matrix. Based on the influence score matrix, cluster analysis was used to group the physiological function data to obtain a grouping result set; Based on the grouping result set, time series analysis is used to extract the trend characteristics of each group, resulting in a trend characteristic set.

7. The intelligent exercise system based on AI intelligent recommendation according to claim 1, characterized in that: The contribution evaluation module calculates the contribution of each exercise method to different body parts using an influence score matrix. If the contribution exceeds a preset threshold, it is marked as a valid mode, and a list of valid modes is obtained, including: By using the influence score matrix and employing a weighted calculation method, the contribution of each exercise method to each physiological part is quantified to obtain a contribution vector. Based on the contribution vector, a threshold judgment method is used. If the contribution of the exercise method is higher than the preset threshold, it is marked as an effective mode, and a preliminary mode set is generated. From the initial pattern set, cluster analysis is used to group the effective patterns, and the pattern group set is obtained based on the correlation of physiological parts; Based on the pattern grouping set, the mean aggregation method is used to calculate the average contribution of each effective pattern group, thus obtaining the group contribution set. From the set of group contributions, extract the effective patterns with the highest contribution in each group to generate the preferred pattern set; Based on the preferred pattern set, a sorting algorithm is used to sort the effective patterns from high to low according to their contribution, and the final list of effective patterns is obtained. Extract multiple patterns from the final list of valid patterns to generate a pattern list.

8. The intelligent exercise system based on AI intelligent recommendation according to claim 1, characterized in that: The intelligent recommendation module uses a list of effective patterns to match the user's current physical condition, generates an exercise sequence optimized for joint flexibility, and obtains a personalized fitness plan, including: By using body condition data and employing standardized processing methods, joint flexibility-related indicators are converted into normalized feature vectors to obtain a user feature dataset. If the joint flexibility index in the user feature dataset is lower than the preset threshold, a classification algorithm is used to select exercise patterns related to micro-joints from the preset list of effective patterns to obtain a candidate pattern set. Based on the candidate pattern set, a weighted scoring method is used, combined with the normalized feature vectors in the user feature dataset, to calculate the matching degree of each training pattern, resulting in a matching degree ranking list.

9. The intelligent exercise system based on AI intelligent recommendation according to claim 8, characterized in that: The intelligent recommendation module uses a list of effective patterns to match the user's current physical condition, generates an exercise sequence optimized for joint flexibility, and obtains a personalized fitness plan, which also includes: From the matching score ranking list, extract the multiple training modes with the highest matching scores to generate preliminary personalized sequences; If the number of exercise patterns in the initial personalized sequence is insufficient, cluster analysis is used to supplement the candidate pattern set with patterns related to joint flexibility to obtain an optimized sequence. Based on the optimized sequence, time series analysis is used to adjust the execution order of exercise modes and generate a final personalized fitness plan. The final personalized fitness plan is then updated iteratively, adjusting the exercise mode based on changes in the user's subsequent physical condition data to obtain a dynamic fitness plan.

10. An AI-based intelligent exercise method, employing the AI-based intelligent exercise system according to any one of claims 1-9, characterized in that, The method includes: By collecting heart rate changes, muscle strength readings, and metabolic rate indicators related to joint flexibility during exercise using sensor devices, raw exercise data and physiological response information are obtained from multiple sources to obtain an initial dataset containing heterogeneous formats. The normalization transformation method is used to handle the differences in units and the unevenness of collection frequency in the initial dataset, and all indicators are unified to the standard scale to obtain a normalized dataset. Clustering algorithms are applied to normalized datasets to group similar physiological response patterns. Based on the grouping results, user physical status classification is determined, and a classification label set is obtained. If the heterogeneity of the classification label set is higher than a preset threshold, then the indicators of different parts are fused through feature extraction to obtain a fused feature vector; A regression model is constructed based on the fused feature vectors to predict the quantitative impact of exercise on physiological function, and an impact score matrix is ​​obtained. The contribution of each exercise method to different body parts is calculated by using the influence score matrix. If the contribution is higher than a preset threshold, it is marked as an effective mode, and a list of effective modes is obtained. By matching the user's current physical condition with a list of effective patterns, a training sequence optimized for joint flexibility is generated to obtain a personalized fitness plan.

Images