Personalized fitness data management method and system based on cloud computing

Abstract

The invention relates to the technical field of cloud computing, solves the technical problems of low accuracy and effectiveness of a fitness data analysis result due to difficulty in transverse comparison and analysis and lack of a unified processing mode in the prior art, and particularly relates to a personalized fitness data management method and system based on cloud computing. The method comprises the steps of collecting original data of personalized fitness, obtaining preprocessed data through preprocessing on the basis of the original data, guaranteeing safe uploading of the data through transmission encryption and user token authentication, analyzing a data mode according to a device identifier after cloud decryption, and obtaining the preprocessed data according to the analyzed data mode. Then, the multi-user original data is subjected to space-time standardization, and a three-dimensional data cube is constructed, so that the uniformity of the multi-user and multi-batch data is ensured, a unified analysis method flow is constructed, and the problem of difficulty in transverse comparison of the data due to equipment difference in the prior art is effectively solved; and a data unified processing mode is provided for group analysis and personalized modeling.

Description

Technical Field

[0001] This invention relates to the field of cloud computing technology, and in particular to a personalized fitness data management method and system based on cloud computing. Background Technology

[0002] Cloud computing is a type of distributed computing that refers to breaking down massive data processing programs into countless smaller programs through a network cloud. In fitness data management scenarios, users can collect diverse data using various smart fitness devices, such as exercise trajectories, heart rates, and movement accuracy, and upload them to the cloud in real time. By using algorithms such as machine learning and data mining, valuable information can be accurately extracted from massive amounts of data, such as analyzing users' exercise habits, physical fitness trends, and weak training areas, thereby creating highly personalized fitness training plans for users.

[0003] Existing technologies exhibit inconsistencies in sampling rates and data dimensions for fitness data, making it difficult to conduct cross-sectional comparative analysis. Furthermore, there is a lack of unified processing methods for group and individual analysis of fitness data, resulting in low accuracy and effectiveness of fitness data analysis results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a cloud-based personalized fitness data management method and system, which solves the technical problems of existing technologies, such as difficulty in conducting horizontal comparative analysis and lack of a unified processing method, resulting in low accuracy and effectiveness of fitness data analysis results.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a personalized fitness data management method based on cloud computing, the steps of which are: collecting raw data of personalized fitness, and obtaining preprocessed data based on the raw data through preprocessing; A three-dimensional data cube is obtained by cloud synchronization and data integration based on the preprocessed data; Based on the three-dimensional data cube, an optimized classification model and model parameters are obtained through personalized state assessment and dynamic modeling. Based on the optimized classification model and model parameters, personalized training schemes and scheme generation models are generated through adaptive processing. Implement personalized training programs and acquire real-time fitness data. Based on the real-time fitness data, dynamically optimize the personalized training programs to obtain new training programs.

[0006] Preferably, preprocessed data is obtained from the original data through preprocessing, including: obtaining normalized and filtered data from the original data through window normalization and filtering. Cleaned data is obtained by removing outliers from normalized filtered data; The normalized filtered data and the cleaned data are aligned and differentially compressed to obtain preprocessed data.

[0007] Preferably, based on the preprocessed data, cloud synchronization and data integration include: performing transport layer encryption on the preprocessed data to obtain an encrypted data stream; The encrypted data stream is decrypted and processed in the cloud to obtain the decrypted dataset; A three-dimensional data cube is obtained by constructing a three-dimensional data cube based on the decrypted dataset.

[0008] Preferably, personalized state evaluation and dynamic modeling based on the three-dimensional data cube include: constructing a multi-level LSTM coding layer based on the three-dimensional data cube to obtain temporal feature vectors; A Gaussian mixture clustering layer is constructed to perform probabilistic clustering on the temporal feature vectors, thereby obtaining the probability distribution of state categories. Soft classification optimization is performed based on the state category probability distribution and temporal feature vectors to obtain the optimized classification model and model parameters.

[0009] Preferably, soft classification optimization is performed based on the probability distribution of state categories and time-series feature vectors, including: obtaining physiological index sequences from the original data, and obtaining load index values ​​by trend fitting based on the physiological index sequences; A personalized state vector is obtained by weighted fusion based on the probability distribution of state categories and load index values. Based on temporal feature vectors and personalized state vectors, the parameters of the multi-level LSTM encoding layer and Gaussian mixture clustering layer are optimized to obtain an optimized classification model and model parameters.

[0010] Preferably, the user classification identifier is obtained by using an adaptive scheme based on the optimized classification model and model parameters. Candidate solutions are collected based on user classification identifiers to obtain a candidate solution set. The candidate solution set and personalized state vector are processed to generate a solution, resulting in a personalized training solution and a solution generation model.

[0011] Preferably, the candidate solution set and personalized state vector are processed to generate a solution, including: processing the candidate solution set and personalized state vector through a policy network to obtain the original decision action; Perform security verification on the original decision-making actions to obtain the secure actions; Based on safety actions, a structured transformation is used to obtain personalized training schemes and scheme generation models.

[0012] Preferably, dynamic optimization of personalized training programs based on real-time fitness data includes: obtaining real-time feature data by performing lightweight LSTM feature processing on real-time fitness data; The feature data from the previous time step is obtained, and Kalman filtering is performed on the real-time feature data and the feature data from the previous time step to obtain optimized real-time features. Real-time guidance instructions are generated based on optimized real-time features and error mapping. Based on all optimized real-time features and corresponding real-time guidance instructions, a quantitative performance score is obtained through quantitative evaluation. The program generation model is adjusted based on quantitative performance scores and real-time fitness data to obtain an optimized new training program.

[0013] This technical solution also provides a system for applying the aforementioned cloud-based personalized fitness data management method, the system comprising: The preprocessing module is used to collect raw data for personalized fitness, and obtain preprocessed data based on the raw data through preprocessing. The cloud synchronization module is used to obtain a three-dimensional data cube based on preprocessed data through cloud synchronization and data integration; The dynamic building module is used to obtain an optimized classification model and model parameters based on a 3D data cube through personalized state evaluation and dynamic modeling. The scheme generation module is used to generate personalized training schemes and scheme generation models based on the optimized classification model and model parameters through adaptive scheme processing. The program fine-tuning module is used to implement personalized training programs and acquire real-time fitness data. Based on the real-time fitness data, the personalized training programs are dynamically optimized to obtain new training programs.

[0014] By employing the above technical solution, the present invention provides a personalized fitness data management method and system based on cloud computing, which has at least the following beneficial effects: 1. This invention ensures secure data upload through transmission encryption and user token authentication. After decryption in the cloud, the data pattern is parsed based on the device identifier. Subsequently, the original data of multiple users is spatiotemporally standardized and a three-dimensional data cube is constructed, ensuring the uniformity of data from multiple users and batches. A unified analysis methodology is constructed, effectively solving the problem of difficulty in horizontal data comparison due to device differences in existing technologies. It provides a unified data processing method for group analysis and personalized modeling.

[0015] 2. This invention uses LSTM to accurately capture the dynamic changes in motion state, effectively handles long-term dependencies in long-sequence data, and combines Gaussian mixture clustering layers to quantify uncertainty, significantly improving the accuracy and interpretability of action completion quality and real-time fatigue assessment. It can also dynamically adapt to the long-term evolution of user capabilities through incremental online learning, providing users with a more scientific and personalized motion state assessment.

[0016] 3. This invention utilizes clustering to quickly locate similar user groups, combines the sampling of successful historical solutions for the group to provide an empirical basis, avoids the problem of solution mismatch, dynamically adjusts training parameters based on the user's real-time state vector through a policy network, and introduces fitness safety restrictions to ensure the compliance of the solution with the user. This overcomes the limitations of existing technologies that rely on fixed presets and cannot be optimized in the long term, and greatly improves the effectiveness and sustainability of the solution.

[0017] 4. This invention extracts real-time data streams using lightweight LSTM features, which can efficiently extract key features and reduce computational resource consumption. It uses Kalman filtering for state tracking, which can accurately optimize real-time state estimation, effectively deal with data noise interference, and quickly generate targeted guidance instructions to correct user actions in a timely manner. At the same time, the evaluation considers multiple indicators, and the quantitative evaluation is comprehensive and accurate. Finally, the model and scheme are iteratively updated based on the evaluation results to achieve dynamic optimization of the training scheme. This provides users with highly personalized, scientific and effective training guidance, which greatly improves training effect and user experience. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the cloud computing-based personalized fitness data management method of the present invention; Figure 2 This is a structural block diagram of the cloud-based personalized fitness data management system of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0020] Example 1: Due to the limitations of existing technologies in conducting comparative analysis and the lack of standardized processing methods, the accuracy and effectiveness of fitness data analysis results are relatively low. Please refer to [the relevant documentation / reference]. Figure 1This embodiment provides a cloud-based personalized fitness data management method that ensures the consistency of data from multiple batches, constructs a unified analysis method flow, effectively solves the problem of difficulty in horizontal data comparison due to equipment differences, and provides a unified data processing method. The steps of this method are as follows: S1. Collect raw data for personalized fitness, and obtain preprocessed data based on the raw data through preprocessing. Existing technologies for processing personalized fitness data on mobile devices face challenges due to the diverse data sources, making unified processing difficult. Furthermore, data collection is prone to noise and inconsistent formats due to sensor inaccuracies and environmental factors. To address these issues, the specific implementation steps are as follows: S11. Based on the original data, normalized and filtered data is obtained through window normalization and filtering. In this step, the mean and standard deviation of all data vectors within the current time window are calculated based on the original data. The original data includes physiological parameters, exercise data, environmental data, and user behavior data, such as heart rate, blood pressure, sleep, body fat percentage, daily steps, exercise distance, exercise type, temperature, humidity, light intensity, and browsing data. These can be integrated into the original data vector, which can be calculated using common formulas for the mean and standard deviation. Then, a sliding window with a sliding step size of A and a size of B is constructed. The mean of the data within the current time window is subtracted from the original data vector at this moment, and the difference is divided by the standard deviation of the data within the current time window. The final result is the normalized data at this moment. Then, with the current time point t as the center, N consecutive normalized data points are selected before and after it. If the boundary is insufficient, the window is dynamically adjusted, and the arithmetic mean is calculated as the filtered output value, thus obtaining the normalized filtered data.

[0021] S12. Based on the normalized filtered data, clean data is obtained by outlier removal. First, the mean and standard deviation of all data points in the current data window are calculated. Data points that exceed the range of mean plus or minus three times the standard deviation are identified as outliers and removed. The remaining data points that meet the range are retained as valid data after cleaning, which is the clean data.

[0022] S13. Alignment and differential compression are performed on the normalized filtered data and cleaned data to obtain preprocessed data. In this step, timestamp alignment is first used to ensure that the two types of data are strictly synchronized on the time axis. For example, normalized filtered data such as heart rate and step frequency are matched with cleaned environmental parameters such as temperature and humidity at the same time point. Then, differential encoding technology in lossy compression encoding is used to calculate the adjacent difference between data values ​​at consecutive time points. For example, the heart rate value at the current time is subtracted from the heart rate value at the previous time. Only the difference is stored instead of the original data to reduce storage space. At the same time, a difference threshold is set according to the data characteristics. Small fluctuations less than the threshold are rounded to further compress the data volume. For example, the heart rate difference threshold is 10. If the absolute value of the difference between the heart rate value at the current time and the heart rate value at the previous time is greater than 10, it can be retained. Otherwise, it indicates that the fluctuation is too small and is discarded. Finally, all data are integrated to obtain preprocessed data. This invention performs standardized preprocessing through normalization, outlier handling, and alignment compression, which significantly reduces transmission bandwidth consumption and provides a high-quality input database for subsequent analysis.

[0023] S2. Based on the preprocessed data, a three-dimensional data cube is obtained through cloud synchronization and data integration. Existing technologies exhibit inconsistencies in data sampling rates and data dimensions, making it difficult to conduct cross-sectional comparative analysis. Furthermore, there is a lack of unified processing methods in group and individual analysis of fitness data, resulting in low accuracy and effectiveness of fitness data analysis results. To address these issues, the specific implementation steps are as follows: S21. Preprocessed data is encrypted using transport layer encryption to obtain an encrypted data stream. In this step, a secure communication tunnel is first established between the user terminal and the cloud server via the TLS protocol. High-strength encryption suites, such as AES 256, are used to encrypt the transmitted data, ensuring that the data cannot be eavesdropped on or tampered with during transmission. Simultaneously, user token authentication is introduced. When a user logs in, the server generates a unique identity token. Each subsequent data upload requires this token for authentication. The server verifies the validity of the token to confirm the user's legitimacy, and data upload is only allowed when both verifications pass. For example, when a user uploads preprocessed data such as heart rate and steps collected by a smart bracelet to the cloud fitness management platform, the system automatically adds a TLS encryption layer to the data packet and verifies the dynamic token stored in the user's mobile app. If the token is valid and the data integrity verification passes, the data is securely transmitted to the cloud in encrypted stream form, thus creating an encrypted data stream. This ensures the confidentiality of data transmission and prevents unauthorized access. AES 256 is used in this process. 256 is a commonly used method for encrypting data transmission. User token authentication is a token-based identity verification and authorization mechanism used to verify the legitimacy of a user's identity and control their access permissions when a user interacts with a system, such as an application, API interface, or cloud service. It will not be elaborated on here.

[0024] S22. The encrypted data stream is decrypted via the cloud to obtain the decrypted dataset. In this step, a pre-configured TLS decryption tool, such as the algorithm and key used when encrypting with the client, is used to decrypt the data stream layer by layer to recover the original binary data packets. Then, based on the device identifier embedded in the packet header, such as the unique ID of the smart bracelet, the corresponding data parsing method is matched from a pre-defined device mode library. The method includes a clearly defined data sampling rate, such as collecting heart rate data once per second, and data dimensions, such as including three indicators: heart rate, steps, and blood oxygen. The binary data is parsed in byte order and then converted into a structured set of values ​​to generate a decrypted dataset. For example, after receiving an encrypted data stream from user A's smart bracelet with ID HW001, the cloud first decrypts the binary data using the AES256 algorithm. Then, based on the device ID HW001, it matches the data pattern as 1 time per second, 3 dimensions, including 2 bytes of heart rate, 2 bytes of steps, and 1 byte of blood oxygen. The binary string 0055001E01 is then parsed into decrypted data from a decrypted dataset with a heart rate of 85 beats per minute, 30 steps, and 97% blood oxygen.

[0025] S23. Based on the decrypted dataset, a three-dimensional data cube is constructed using three-dimensional data. In this step, all data is first grouped by user ID through time alignment processing and then uniformly mapped to a preset base time axis, such as one time point every 0.1 seconds, to ensure that the data collected by different devices are strictly synchronized in the time dimension. Then, the feature data of each time point is resampled. If the original data sampling rate is higher than the base frequency, such as 100Hz, key features are retained by downsampling. If it is lower than the base frequency, such as 5Hz, the data is completed by interpolation. Finally, a standardized data cube is generated with user ID as the first dimension, unified time series as the second dimension, and multi-dimensional features, such as motion acceleration, heart rate, and ambient temperature, as the third dimension. For example, user A's smartwatch collects heart rate at 5Hz, collecting one data value every 0.2 seconds, and smart running shoes collect step frequency at 20Hz, collecting one value every 0.05 seconds. After aligning the data of the two to a 10Hz time axis, the heart rate data is completed by linear interpolation to the intermediate time points, such as 0. The values ​​at 0.5 seconds and 0.1 seconds were used to retain key values ​​in the cadence data through mean downsampling. This resulted in the construction of a three-dimensional data cube containing 10 feature dimensions for user A, including heart rate, cadence, and ambient temperature every 0.1 seconds within a specific time period. Linear interpolation is a common method that uses the average of the two adjacent values ​​of a missing value as filler. Mean downsampling is a common data compression and smoothing technique that reduces data volume by calculating the average within a local data window while preserving key feature trends; details are omitted here. This invention ensures secure data upload through transmission encryption and user token authentication. After decryption in the cloud, the data pattern is parsed based on the device identifier. Subsequently, the original data from multiple users is spatiotemporally standardized, and a three-dimensional data cube is constructed. This ensures the uniformity of data from multiple users and batches, establishing a unified analysis methodology. This effectively solves the problem of difficulty in horizontal data comparison due to device differences in existing technologies, providing a unified data processing method for group analysis and personalized modeling.

[0026] S3. Based on the 3D data cube, an optimized classification model and model parameters are obtained through personalized state assessment and dynamic modeling. Existing technologies typically use static classifiers for state assessment and modeling, which easily leads to low accuracy in dynamic assessment and classification. They also suffer from significant errors in judging situations with ambiguous state boundaries, cannot adapt to changes in motion states over time, and struggle to quantify uncertain indicators such as fatigue probability. To address these issues, the specific steps are as follows: S31. Construct a multi-level LSTM encoding layer based on the 3D data cube to obtain temporal feature vectors. In this step, the forget gate calculates the forgetting factor based on the current input features and the hidden state of the previous time step through an activation function to determine the proportion of information discarded from the previous cell state. The input gate generates the input factor in a similar manner and combines it with the candidate cell state, which is generated by the current input and the previous hidden state through a nonlinear transformation, to evaluate the value of new information. Then, when updating the cell state, the product of the forgetting factor and the previous cell state is added to the product of the input factor and the candidate cell state to form the updated cell state. Then, the output gate regulates the information flow from the cell state to the hidden state. Finally, the hidden state is obtained by multiplying the output factor and the nonlinear transformation of the cell state, and is passed to the next layer as the output of the current step. Through multi-layer stacking, the network gradually abstracts high-level temporal feature vectors, effectively capturing long-term temporal patterns. In complex dynamic relationships, the activation function is generally the Sigmoid function. The output vector after processing by the Sigmoid function is the forgetting factor. For example, in this invention, the user's daily heart rate, exercise duration, calorie consumption and other multi-dimensional fitness data constitute a standardized three-dimensional data cube. Its three-dimensional structure corresponds to the user dimension, time step dimension and feature dimension, respectively. The temporal feature slice for a certain user is a two-dimensional matrix extracted from it along the user dimension. The matrix includes the time step multiplied by the feature dimension. However, in order to adapt to the network input, it is often reconstructed into a three-dimensional data cube with a batch size of 1, that is, 1 multiplied by the time step multiplied by the feature dimension. Through multi-level long short-term memory network encoding, the temporal feature slice is transformed into a high-level temporal feature vector that reflects the user's fitness habits and physical fitness change trend. The multi-level LSTM encoding can be set to 3-layer LSTM encoding, and finally the temporal feature vector is obtained.

[0027] S32. Construct a Gaussian mixture clustering layer to perform probabilistic clustering on the temporal feature vectors, obtaining the state category probability distribution. In this step, based on Gaussian mixture clustering, the vector distribution in the feature space is modeled as a weighted combination of multiple Gaussian components, each component corresponding to a potential typical motion state. The model iteratively optimizes parameters such as mixture weights, mean vector, and covariance matrix through the expectation-maximization algorithm to maximize the likelihood probability of the observed data under this mixture distribution. For any input feature vector, the model calculates its posterior probability belonging to each Gaussian component. This probability is obtained by normalizing the product of the weight of the corresponding component and its probability density under the Gaussian distribution of that component. The output is a discrete probability distribution vector, where each element represents the probability that the input vector belongs to a certain preset state category. For example, after multiple LST layers... The monthly fitness feature vector extracted by M for a user may contain a comprehensive pattern of their training intensity, recovery efficiency, and adaptive changes. This vector is input into a trained Gaussian mixture model, which has previously summarized three typical states through a large amount of user data: standard training state, mild fatigue state, and overtraining risk state. The model will calculate the probability that the vector belongs to each state. Assuming the output probability distribution is 0.75, 0.20, and 0.05, it indicates that the user is currently in a 75% probability of being in an efficient standard training state, a 20% probability of exhibiting mild fatigue characteristics, and only a 5% risk of showing an overtraining tendency. Among them, Gaussian mixture clustering is a commonly used soft clustering method, which allows data points to belong to multiple clusters and gives the probability of each data point belonging to each cluster, which will not be elaborated here.

[0028] S33. Perform soft classification optimization based on state category probability distribution and temporal feature vectors to obtain the optimized classification model and model parameters; S331. Obtain the physiological index sequence from the raw data, and obtain the load index value based on the physiological index sequence through trend fitting. In this step, a time window with a sliding step size of C and a length of M is first constructed. Within this time window, the corresponding RR interval sequence is obtained based on the heart rate data of the raw data. The RR interval sequence is a wave signal that records the heartbeat in an electrocardiogram and can also record the heart rate. Then, the standard deviation of the RR interval sequence is calculated. This standard deviation value represents the heart rate variability within this time window. Subsequently, in order to obtain the physiological load index, the heart rate variability calculated from different time windows can be further processed, for example, multiple... The physiological load index is obtained by averaging the heart rate variability over a continuous time window or by adding the heart rate variability of different time windows according to a certain weight. For example, when monitoring the heart rate of an athlete during training, a time window of 5 minutes is selected. In the first 5-minute time window, the standard deviation of the RR interval sequence is 20, and in the second 5-minute time window, the standard deviation of the RR interval sequence is 25. If the average method is used, the two standard deviations are added together and then divided by 2, which equals 22.5. This 22.5 is the physiological load index calculated based on the data from these two time windows.

[0029] S332. Based on the state category probability distribution and load index values, a personalized state vector is obtained through weighted fusion. This step first involves collecting, organizing, and analyzing user performance data over a period of time in similar scenarios or tasks. For example, this includes statistically analyzing average speed and task completion accuracy across multiple training sessions. A representative benchmark value is then calculated to measure the user's current performance level. Next, the state with the highest probability is identified from the state probability distribution and multiplied by a learnable fusion weight. The learnable fusion weight, along with initial values ​​obtained through the analytic hierarchy process (AHP), is then iteratively optimized using Bayesian optimization or particle swarm optimization, which are commonly used parameter optimization methods. The optimization method, without using an exponent here, involves dividing the current physiological load index by a pre-set maximum physiological load index. The maximum physiological load index can be the maximum value of the corresponding physiological load index in the database, thereby normalizing the physiological load index to a certain range. The normalized result is then multiplied by the corresponding fusion weight. Subsequently, the current performance metric is divided by the historical performance baseline. The current performance metric can be the time taken for a user's fitness task at the current moment, using time as an example, but other physical quantities can also be used, to obtain the relative level of the current performance compared to the historical baseline. This is also multiplied by the corresponding fusion weight. Finally, the results obtained from the above three steps are combined to form the user's personalized state vector.

[0030] S333. Based on temporal feature vectors and personalized state vectors, parameter optimization is performed on multi-level LSTM encoding layers and Gaussian mixture clustering layers to obtain optimized classification models and model parameters. In this step, for the multi-level LSTM encoding layer, based on the two newly input vectors, the key information in the temporal feature vectors and state vectors is gradually passed and integrated into the hidden state representation according to the LSTM processing method. This process is similar to the layer-by-layer filtering and refinement of information in a multi-level network structure. Each layer combines the information from the previous layer and the current input, and updates the hidden state through specific weighted operations, such as the multiplication of the connection weights between different neurons with the input data, plus a bias term. This allows the hidden state to more accurately reflect the temporal and state features contained in the current data. For the Gaussian mixture clustering layer, the parameters are optimized based on the temporal feature vectors and personalized state vectors. Based on the updated hidden state representation, the parameters of each Gaussian distribution, such as mean and variance, are recalculated. For example, the shape and position of the Gaussian distribution are adjusted by using the data in the hidden state representation and the current parameters to more accurately fit the distribution of the data. Finally, the updated hidden state and the Gaussian mixture clustering layer with optimized parameters are merged to obtain an optimized classification model and model parameters. This invention uses LSTM to accurately capture the dynamic changes in motion state, effectively handles long-term dependencies in long-sequence data, and combines Gaussian mixture clustering layers to quantify uncertainty, such as accurately assessing fatigue probability. This significantly improves the accuracy and interpretability of action completion quality and real-time fatigue assessment. Furthermore, through incremental online learning, it can dynamically adapt to the long-term evolution of user capabilities, providing users with a more scientific and personalized motion state assessment.

[0031] S4. Based on the optimized classification model and model parameters, personalized training plans and plan generation models are generated through adaptive processing. Existing technologies, in determining fitness plans, use fixed thresholds for plan generation, making it difficult to cover complex scenarios and unable to dynamically optimize based on user feedback. This results in rigid plans that fail to perceive the user's real-time state, such as fatigue and heart rate fluctuations, and easily generate plans that do not match the user's current ability. To solve these problems, the specific implementation steps are as follows: S41. Based on the optimized classification model and model parameters, user classification identifiers are obtained through clustering. Key features are extracted from the dynamic model parameters of users, using the center points of the Gaussian mixture layer as representatives of key features. Then, the K-means clustering algorithm is used to group users with similar key features into the same category. When calculating the cluster center, for each cluster, the dynamic model parameters of all users in that cluster are summed and then divided by the number of users in that cluster. The final result is the center point data of that cluster. In this way, the range of each cluster is continuously adjusted until all users can be reasonably classified. Finally, the similar user group to which the current user belongs is determined, and the corresponding user classification identifier is output. For example, in a sports and fitness application, there is a set of dynamic model parameters for 100 users. From the parameters of each user, the key features are extracted... Using the Gaussian mixture clustering center points as key features, assuming the initial goal is to divide users into three similar user groups, we begin processing using the K-means clustering algorithm. For the first cluster with 30 users, we sum the data corresponding to the Gaussian mixture clustering center points of these 30 users, assuming the sum is 1500. Then, we divide 1500 by 30, obtaining 50, which is the center point data of this cluster. We calculate the center point data of the other two clusters in the same way. Then, based on the distance between each user's key features and these center point data, the user is assigned to the nearest cluster, thus classifying the user and finally obtaining the user classification label. The center points of the Gaussian mixture clustering layer are determined through iterative optimization using the expectation-maximization algorithm. The K-means clustering algorithm is a commonly used data classification method, which will not be elaborated here.

[0032] S42. Collect candidate solutions based on user classification identifiers to obtain a candidate solution set. In this step, a historical solution library is first built based on the user's historical fitness plans. Historical fitness plans include information such as user name, height, weight, health status, fitness goals, time period, training schedule, and training content. Since this information belongs to the user's personal information, it is all encrypted information and processed. It is decrypted before processing to ensure the user's information security. The solutions used by the group in the historical solution library are selected, and then those solutions that have successfully achieved the goal and have high user experience scores are further selected. These solutions represent the group's high-quality experience in achieving good results in past practice. Then, these selected successful solutions are sampled. The number of samples can be set to a base value based on actual needs and experience, such as 10. Then, combined with the importance or urgency of the current task, the base value is multiplied by a corresponding adjustment coefficient. This adjustment coefficient can be greater than 1, indicating that more candidate solutions are needed to increase the diversity of choices, or less than 1, indicating that the number of candidate solutions is appropriately reduced to improve the screening efficiency. Finally, the number of candidate seed solutions is obtained through such calculation, and a corresponding number of solutions are randomly selected from the successful solutions to form a candidate solution set.

[0033] S43. Perform scheme generation processing on the candidate scheme set and personalized state vector to obtain personalized training scheme and scheme generation model; S431. Based on the candidate solution set and personalized state vector, the original decision action is obtained through a policy network. First, a policy network for the TD3 algorithm is constructed, typically using a multi-layer fully connected structure. The input layer receives state information, which is then abstracted layer by layer through the hidden layers. Finally, the output layer provides a deterministic action to achieve the mapping from state to action. Then, the personalized state vector is input into the policy network, which analyzes and processes this state. The specific steps are as follows: the input layer receives the personalized state vector and passes its value to the hidden layers. In the hidden layers, numerous neurons perform linear operations with the input value through the weight matrix. After adding a bias term, the data is then processed by an activation function, such as the ReLU function, to introduce a non-linear transformation, allowing the network to learn complex patterns. This process is passed layer by layer through the hidden layers, continuously extracting and abstracting state features, and mining the information hidden in the state that is related to the optimal action. As data flows between the hidden layers, the network gradually transforms the original state information. To provide a more representative feature representation, the output layer ultimately outputs a deterministic action value through a linear transformation based on these features. To encourage the agent to explore during the decision-making process and avoid getting trapped in local optima, exploration noise is added to this basic action value. The action value output by the policy network is then added to the exploration noise to obtain the original decision action. Linear transformation is a commonly used mathematical mapping method, which will not be elaborated here. For example, suppose a user's personalized state vector describes the user's physical fatigue level, exercise ability, and other state information after continuous exercise for a period of time. The policy network calculates and provides a basic action value, indicating that the action complexity of the next beat is 5 and the rest interval is 10 seconds. Meanwhile, the exploration noise is a random value, such as an increase in action complexity of 0.5 and a decrease in rest interval of 2 seconds. Then, after adding the basic action value and the exploration noise, the original decision action obtained is that the action complexity of the next beat is 5.5 and the rest interval is 8 seconds.

[0034] S432. Perform safety verification on the original decision action to obtain a safe action. In this step, the user's pre-set fitness safety limits are first obtained. Using a constraint satisfaction problem-solving approach, the original decision action is regarded as an initial value to be adjusted, with the fitness safety limits as constraints. In specific calculations, an action with the smallest difference from the original decision action is found. For example, the difference can be measured by the sum of the squares of the two subtractions, and the action is within the fitness safety limits. The final result is the safety-adjusted action. For example, if the original decision action is to make the user's heart rate reach 130 beats per minute in the next beat of exercise, while the fitness safety limit stipulates a maximum heart rate of 120 beats per minute, then the original heart rate of 130 is used as the initial value. The closest value to 130 is found within the fitness safety limit of no more than 120. After calculation and adjustment, the heart rate corresponding to the safety-adjusted action is 120 beats per minute, thus ensuring that the action is within the safe range.

[0035] S433. Based on safety movements, a structured transformation is used to obtain personalized training plans and plan generation models. This step first constructs a cloud resource library, including standard movement video clips, music and beat audio, voice clips, motivational sounds, image-guided animations, and other related data, resulting in a multimedia database. Based on the movement parameters of the safety movements, such as movement complexity, exercise intensity, and rest intervals, precise matching is performed in the multimedia database. Taking movement complexity as an example, a complexity threshold is set. If the movement complexity is greater than or equal to the threshold, a more difficult and varied segment is selected; otherwise, a simple and easy-to-learn segment is selected. The beat speed is then determined based on exercise intensity; higher intensity results in a faster beat, and lower intensity results in a slower beat. Simultaneously, appropriate instructions are inserted according to the rest intervals to remind users to rest, such as "You did great! You held on for a minute, please take a break." Finally, these selected dance video clips, set beats, and instructions are integrated into a personalized training plan according to a certain logical order. The integration method can be achieved through MPEG. The DASH dynamic adaptive stream encapsulation technology is used to generate a solution model that takes S1 to S4 as a whole. This invention utilizes clustering to quickly locate similar user groups, combines the sampling of successful solutions from the group's history to provide an empirical basis, avoids the problem of solution mismatch, dynamically adjusts training parameters based on the user's real-time state vector through a policy network, and introduces fitness safety restrictions to ensure the compliance of the solution with the user. This overcomes the limitations of existing technologies that rely on fixed presets and cannot be optimized in the long term, and greatly improves the effectiveness and sustainability of the solution.

[0036] S5. Implement a personalized training plan and acquire real-time fitness data. Based on the real-time fitness data, dynamically optimize the personalized training plan to obtain a new training plan. Existing technologies lack efficient feature extraction methods in fitness plan optimization, resulting in insufficient key information mining, potentially inaccurate state tracking, difficulty in effectively handling data noise, and possibly single-indicator performance evaluation that cannot comprehensively reflect training effectiveness. To solve the above problems, the specific implementation steps are as follows: S51. Based on real-time fitness data, obtain real-time feature data through lightweight LSTM feature processing. In this step, first collect real-time fitness data of users, including various data in the original data of personalized fitness and perform standardization processing. Use a lightweight LSTM model, which reads data points in the real-time fitness data one by one in time order. For each data point, it will calculate according to its correlation with the data before and after it and the preset rules. The specific steps include: the model continuously updates the hidden state through weight matrix operations, such as multiplying the current data point by a weight matrix and adding the output of the hidden state of the previous time step multiplied by another weight matrix, thereby capturing the long-term dependence and dynamic change pattern of the data in the time dimension. After such calculation and processing, more representative and key information is finally extracted from the original real-time fitness data and integrated into real-time feature data.

[0037] S52. Obtain the feature data from the previous time step, and perform Kalman filtering on the real-time feature data and the feature data from the previous time step to obtain optimized real-time features. In this step, a key Kalman gain value is first calculated based on the feature data from the previous time step. The specific process is as follows: first calculate the product of the prediction error covariance and the transpose of the observation matrix, and then multiply it by the inverse of the sum of the observation matrix, the prediction error covariance, and the observation noise covariance to obtain the Kalman gain. After obtaining the Kalman gain, we use it to correct the prediction state at the current time step. The correction method is to multiply the difference between the real-time feature data and the product of the observation matrix and the prediction state by... The Kalman gain is calculated, and then the result is added to the predicted state to obtain the optimized state estimate for the current time step. Simultaneously, the estimation error covariance is updated based on the Kalman gain and the observation matrix to provide more accurate prior information for the state estimate at the next time step. The observation matrix can be obtained using the least squares method, the prediction error covariance can be calculated using the formula for a diagonal matrix based on real-time feature data and the feature data from the previous time step, and the observation noise covariance can be obtained through maximum likelihood estimation. Updating the estimation error covariance, also known as posterior estimation error covariance calculation, is a commonly used method for updating the error covariance and will not be elaborated upon here.

[0038] S53. Based on optimized real-time features, generate real-time guidance instructions through error mapping. In this step, the optimized real-time features are first compared in detail with the expected state in the plan. The expected state is, for example, the heart rate should reach 90 beats per minute or more. The difference between the two is analyzed through the error mapping relationship of the pre-set rules. The pre-set rules include the threshold range of error for different state parameters and the corresponding guidance instruction triggering conditions. For example, for the parameter of spinal lateral tilt angle, it is set that when its error continues to exceed 5 degrees, a voice prompt instruction for core tightening will be triggered. The specific processing process is to check the error between the real-time estimated value and the expected value of each key state parameter one by one. Once the error of a certain parameter reaches or exceeds the threshold set in the rule, the corresponding real-time guidance instruction is generated according to the rule.

[0039] S54. Based on all optimized real-time features and corresponding real-time guidance instructions, a quantitative performance score is obtained through quantitative evaluation. In this step, all optimized real-time features and personalized training schemes are input. The schemes include the expected state to be achieved at each moment. The processing method revolves around calculations based on multiple key indicators, including goal achievement, consistency, and fatigue index, to comprehensively measure the effectiveness of real-time guidance instructions. Taking the consistency index calculation as an example, firstly, for each moment in the real-time guidance instruction, the absolute value of the difference between the optimized real-time feature and the expected state value at that moment is calculated, which measures the degree of deviation between the actual state and the expected state. Then, the average value of these absolute values ​​of difference over all moments is calculated. This average value reflects the overall situation of the degree of deviation between the actual state and the expected state throughout the entire session. Next, a standard range is determined, which can be preset by the user. It represents the maximum degree of state deviation within a normal or acceptable range. Finally, the average degree of deviation obtained above is divided by this standard range, and then 1 is subtracted from this ratio to obtain the consistency index. The calculation formulas for goal achievement and fatigue index are commonly used quantitative methods and will not be elaborated here.

[0040] S55. Adjust the program generation model based on the quantitative effect score and real-time fitness data to obtain an optimized new training program. This step first fine-tunes the program generation model, using real-time fitness data as a key basis. It deeply analyzes the user's actual training performance, physical responses, and other characteristics reflected in the data. Based on this detailed information, the program generation model is meticulously adjusted. This includes adjusting model parameters with a small learning rate or using the backpropagation algorithm to more accurately match the user's current physical state and training ability. Next, a simple judgment is made based on the quantitative effect score: if the quantitative effect score is greater than or equal to the passing threshold, a good evaluation result is generated, indicating that the overall effect of this training program is good; if the quantitative effect score is less than the passing threshold, the effect is poor, and a new personalized training program is generated. The passing threshold can be obtained based on historical data through maximum likelihood estimation. Through this adjustment, the program generation model can continuously optimize its decision-making direction based on actual training effects. Subsequently, by comprehensively quantifying the effect score, real-time fitness data, and the adjusted plan generation model, and combining the user's historical training data, basic physical condition, and other multi-dimensional information, a specific algorithm is used to generate an optimized new training plan for the next cycle. This specific algorithm can use the decision tree algorithm to generate the new training plan. The decision tree algorithm is a decision-making algorithm based on a tree structure. It constructs a tree model by dividing data features layer by layer and is a commonly used machine learning algorithm, which will not be elaborated upon here. This invention uses lightweight LSTM feature extraction to extract real-time data streams, efficiently extracting key features and reducing computational resource consumption. Kalman filtering is used for state tracking, which can accurately optimize real-time state estimation, effectively cope with data noise interference, and quickly generate targeted guidance instructions to correct user actions in a timely manner. Simultaneously, the evaluation considers multiple indicators comprehensively and accurately, and finally, the model and plan are iteratively updated based on the evaluation results to achieve dynamic optimization of the training plan. This provides users with highly personalized, scientific, and effective training guidance, greatly improving training effectiveness and user experience.

[0041] Example 2: Due to the difficulty of conducting cross-sectional comparative analysis using existing technologies and the lack of a unified processing method, the accuracy and validity of fitness data analysis results are relatively low. Please refer to [link to relevant documentation]. Figure 2 The diagram shown is a structural block diagram of the cloud-based personalized fitness data management system provided in this embodiment. It can ensure the uniformity of multiple batches of data, construct a unified analysis method and process, effectively solve the problem that data is difficult to compare horizontally due to equipment differences, and provide a unified data processing method. The system includes: a preprocessing module, a cloud synchronization module, a dynamic construction module, a scheme generation module, and a scheme fine-tuning module. The preprocessing module is used to collect raw data for personalized fitness, and obtain preprocessed data based on the raw data through preprocessing. The cloud synchronization module is used to obtain a three-dimensional data cube based on preprocessed data through cloud synchronization and data integration; The dynamic building module is used to obtain an optimized classification model and model parameters based on a 3D data cube through personalized state evaluation and dynamic modeling. The scheme generation module is used to generate personalized training schemes and scheme generation models based on the optimized classification model and model parameters through adaptive scheme processing. The program fine-tuning module is used to implement personalized training programs and acquire real-time fitness data. Based on the real-time fitness data, the personalized training programs are dynamically optimized to obtain new training programs.

[0042] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0043] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A cloud-based personalized fitness data management method, characterized in that, The steps of this method are as follows: collect raw data of personalized fitness, and obtain preprocessed data based on the raw data through preprocessing; A three-dimensional data cube is obtained by cloud synchronization and data integration based on the preprocessed data; Based on the three-dimensional data cube, an optimized classification model and model parameters are obtained through personalized state assessment and dynamic modeling. Based on the optimized classification model and model parameters, personalized training schemes and scheme generation models are generated through adaptive processing. Implement personalized training programs and acquire real-time fitness data. Based on the real-time fitness data, dynamically optimize the personalized training programs to obtain new training programs.

2. The personalized fitness data management method based on cloud computing according to claim 1, characterized in that, Preprocessed data is obtained from the original data through preprocessing, including: normalized and filtered data is obtained from the original data through window normalization and filtering. Cleaned data is obtained by removing outliers from normalized filtered data; The normalized filtered data and the cleaned data are aligned and differentially compressed to obtain preprocessed data.

3. The personalized fitness data management method based on cloud computing according to claim 1, characterized in that, Based on the preprocessed data, cloud synchronization and data integration are performed, including: performing transport layer encryption on the preprocessed data to obtain an encrypted data stream; The encrypted data stream is decrypted and processed in the cloud to obtain the decrypted dataset; A three-dimensional data cube is obtained by constructing a three-dimensional data cube based on the decrypted dataset.

4. The personalized fitness data management method based on cloud computing according to claim 1, characterized in that, Personalized state evaluation and dynamic modeling based on 3D data cubes include: constructing multi-level LSTM encoding layers based on 3D data cubes to obtain temporal feature vectors; A Gaussian mixture clustering layer is constructed to perform probabilistic clustering on the temporal feature vectors, thereby obtaining the probability distribution of state categories. Soft classification optimization is performed based on the state category probability distribution and temporal feature vectors to obtain the optimized classification model and model parameters.

5. The personalized fitness data management method based on cloud computing according to claim 4, characterized in that, Soft classification optimization based on state category probability distribution and time-series feature vectors includes: obtaining physiological index sequences from the original data, and obtaining load index values ​​by trend fitting based on the physiological index sequences; A personalized state vector is obtained by weighted fusion based on the probability distribution of state categories and load index values. Based on temporal feature vectors and personalized state vectors, the parameters of the multi-level LSTM encoding layer and Gaussian mixture clustering layer are optimized to obtain an optimized classification model and model parameters.

6. The personalized fitness data management method based on cloud computing according to claim 1, characterized in that, The process involves optimizing the classification model and model parameters, and then using an adaptive approach, including: obtaining user classification identifiers through clustering based on the optimized classification model and model parameters. Candidate solutions are collected based on user classification identifiers to obtain a candidate solution set. The candidate solution set and personalized state vector are processed to generate a solution, resulting in a personalized training solution and a solution generation model.

7. The personalized fitness data management method based on cloud computing according to claim 6, characterized in that, The process of generating a solution involves processing the candidate solution set and personalized state vectors through a policy network to obtain the original decision action. Perform security verification on the original decision-making actions to obtain the secure actions; Based on safety actions, a structured transformation is used to obtain personalized training schemes and scheme generation models.

8. The personalized fitness data management method based on cloud computing according to claim 1, characterized in that, Dynamic optimization of personalized training programs based on real-time fitness data includes: obtaining real-time feature data by using lightweight LSTM feature processing based on real-time fitness data; The feature data from the previous time step is obtained, and Kalman filtering is performed on the real-time feature data and the feature data from the previous time step to obtain optimized real-time features. Real-time guidance instructions are generated based on optimized real-time features and error mapping. Based on all optimized real-time features and corresponding real-time guidance instructions, a quantitative performance score is obtained through quantitative evaluation. The program generation model is adjusted based on quantitative performance scores and real-time fitness data to obtain an optimized new training program.

9. A system applied to the cloud-based personalized fitness data management method according to any one of claims 1-8, characterized in that, The system includes: The preprocessing module is used to collect raw data for personalized fitness, and obtain preprocessed data based on the raw data through preprocessing. The cloud synchronization module is used to obtain a three-dimensional data cube based on preprocessed data through cloud synchronization and data integration; The dynamic building module is used to obtain an optimized classification model and model parameters based on a 3D data cube through personalized state evaluation and dynamic modeling. The scheme generation module is used to generate personalized training schemes and scheme generation models based on the optimized classification model and model parameters through adaptive scheme processing. The program fine-tuning module is used to implement personalized training programs and acquire real-time fitness data. Based on the real-time fitness data, the personalized training programs are dynamically optimized to obtain new training programs.

Images