Oxygen uptake telemetering system suitable for overwater training of boat racing project

Abstract

The invention relates to the technical field of boat racing training physiological monitoring, and discloses an oxygen uptake telemetering system suitable for boat racing project water training. A data acquisition module of the system obtains multi-dimensional physiological data of an athlete through a mouthpiece type gas collection device, and the multi-dimensional physiological data are integrated to form a physiological data warehouse; the feature construction module extracts features from the data warehouse to obtain a multi-dimensional physiological feature matrix, and constructs a dimension association graph and a physiological state map; the cloud fusion module transmits the feature matrix to a cloud server, and a local physiological state clustering center and a training scene preliminary discrimination clustering center are obtained through clustering analysis; the sequence generation module fuses the physiological state atlas and the two clustering centers to form training scene multi-source fusion feature representation, and generates a real-time core physiological feature sequence; and the feedback module realizes real-time dynamic feedback of the core physiological feature sequence through the mobile client. The system can accurately and comprehensively monitor the water training physiological state of the boat racing athlete, and assists in improving the training effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of physiological monitoring for rowing training, in particular to an oxygen uptake telemetry system suitable for water training of rowing events. BACKGROUND

[0002] As a water event requiring high physical fitness, endurance and coordination of athletes, the physiological state of athletes during rowing training directly affects the training effect and the improvement of sports performance. In the water training scene of rowing, accurately grasping the core physiological data such as oxygen uptake of athletes is an important basis for coaches to develop scientific training programs, adjust training intensity and prevent sports injuries. However, there are still many problems to be solved in the current physiological monitoring technology for water training of rowing events. Traditional physiological data collection methods rely on wired monitoring devices. Such devices not only affect the coherence of the rowing action of athletes due to cable constraints in water training environment, but also may cause device failure due to water shaking and humid environment, making it difficult to achieve long-term stable data collection. Even if some wireless monitoring devices can get rid of the cable constraints, there are obvious shortcomings in the dimension and accuracy of data collection. Most of them can only obtain single heart rate data and cannot fully capture multi-dimensional physiological indicators such as oxygen uptake, breathing rate and energy consumption closely related to rowing training, making it difficult to form a complete system of physiological data of athletes. In the data processing and analysis link, the existing technology often stays at the level of simple data recording and display, lacking the ability to deeply integrate and feature mine the collected physiological data. Since the physiological state of athletes changes dynamically during rowing training, such as training time, rowing rhythm, water flow environment and other factors, single-dimensional data cannot reflect the internal relationship between physiological indicators, nor can it build a physiological model that can accurately represent the training state of athletes. In addition, the data processing link of most monitoring systems is limited to local devices with limited computing power, which cannot efficiently cluster analyze and multi-source information fuse a large amount of physiological data, making it difficult to quickly identify the physiological state change rule of athletes in different training scenarios and provide targeted training guidance for coaches. The existing monitoring technology generally lacks real-time dynamic feedback mechanism. The collected physiological data usually needs to be analyzed offline after the training ends, so the coach and the athlete cannot grasp the changes of the physiological state in time during the training process, and it is difficult to adjust the training strategy according to the real-time data. For example, when the athlete has abnormal fluctuation of oxygen uptake or the physiological load exceeds the tolerance range, if the training intensity cannot be adjusted in time, not only the training effect will be affected, but also the risk of sports injury will be increased. At the same time, the existing system cannot realize the effective association of physiological data and training scene, cannot accurately judge the physiological adaptation of the athlete in different training scenes, and leads to the lack of scene pertinence of training guidance, which is difficult to meet the needs of fine training of rowing events. SUMMARY

[0003] The purpose of the present application is to provide an oxygen uptake telemetry system suitable for water training of rowing events to solve the problems raised in the background art.

[0004] To achieve the above purpose, the present application provides an oxygen uptake telemetry system suitable for water training of rowing events, which comprises: A data acquisition module for acquiring multi-dimensional physiological data of rowing athletes through a mouthpiece gas collection device, integrating the multi-dimensional physiological data to obtain an athlete physiological data warehouse; A feature construction module for extracting features from the athlete physiological data warehouse to obtain a multi-dimensional physiological feature matrix, determining a dimension correlation graph of each dimension in the multi-dimensional physiological feature matrix, and constructing an athlete training-related physiological state atlas through the dimension correlation graph of each dimension; A cloud fusion module for transmitting the multi-dimensional physiological feature matrix to a cloud server, performing clustering analysis on the multi-dimensional physiological feature matrix to obtain a local physiological state clustering center and a training scene preliminary discrimination clustering center; A sequence generation module for fusing the local physiological state clustering center and the training scene preliminary discrimination clustering center based on the physiological state atlas to obtain a training scene multi-source fusion feature representation, and generating a real-time core physiological feature sequence of the athlete according to the training scene multi-source fusion feature representation; A feedback module for real-time dynamic feedback of the core physiological feature sequence through a mobile client.

[0005] Preferably, the data acquisition module specifically includes the following steps in data integration of the multi-dimensional physiological data: Pretreating the multi-dimensional physiological data acquired through the mouthpiece gas collection device to obtain pretreated multi-dimensional physiological data; Performing semantic mapping on the pretreated multi-dimensional physiological data to obtain standardized physiological data of each dimension; The athlete physiological data warehouse is constructed by the standardized physiological data of all dimensions.

[0006] Preferably, the feature construction module performing feature extraction on the athlete physiological data warehouse specifically includes: The feature extraction is performed on the standardized physiological data of each dimension in the athlete physiological data warehouse respectively, and then the physiological feature vectors of each dimension are obtained; The multidimensional physiological feature matrix is constructed according to the physiological feature vectors of all dimensions.

[0007] Preferably, the feature construction module determining the dimension correlation graph of each dimension in the multidimensional physiological feature matrix specifically includes: A dimension is selected from all dimensions of the multidimensional physiological feature matrix, and the physiological feature vector of the selected dimension is obtained; The feature correlation degrees between the physiological features in the physiological feature vector of the selected dimension are determined; The dimension correlation graph of the selected dimension is constructed according to the feature correlation degrees between the physiological features, and then the dimension correlation graph of each dimension in the multidimensional physiological feature matrix is obtained.

[0008] Preferably, the feature construction module constructing the physiological state atlas related to athlete training through the dimension correlation graphs of each dimension specifically includes: The key physiological features of each dimension are determined, and then the feature correlation degrees between the key physiological features are determined; The dimension correlation graphs of each dimension are connected according to the feature correlation degrees between the key physiological features, and then the physiological state atlas related to athlete training is obtained.

[0009] Preferably, the cloud fusion module performing clustering analysis on the multidimensional physiological feature matrix specifically includes: The physiological feature vectors of each dimension in the multidimensional physiological feature matrix are fully connected coding to obtain a set of physiological feature fully connected embedding coding vectors; The set of physiological feature fully connected embedding coding vectors are density clustering to obtain the local physiological state clustering centers.

[0010] Preferably, the cloud fusion module performing clustering analysis on the multidimensional physiological feature matrix further includes: The training scene preliminary discrimination types of each dimension in the multidimensional physiological feature matrix are one-hot coding to obtain a set of training scene discrimination one-hot coding vectors; The set of training scene discrimination one-hot coding vectors are partition clustering to obtain a plurality of local training scene discrimination clustering centers; Calculate a weighted sum of the plurality of training scene local discriminative clustering centers to obtain the training scene preliminary discriminative clustering center.

[0011] Preferably, the sequence generation module fuses the local physiological state clustering center and the training scene preliminary discriminative clustering center based on the physiological state atlas, and the fusion specifically comprises: Cross-dimension isomorphism is performed on the training scene preliminary discriminative clustering center and the local physiological state clustering center to obtain an optimized training scene discriminative clustering center. Cascade is performed on the optimized training scene discriminative clustering center and the local physiological state clustering center to obtain a training scene multi-source fusion feature representation.

[0012] Preferably, the sequence generation module generates a real-time core physiological feature sequence of the athlete according to the training scene multi-source fusion feature representation, and the generation specifically comprises: For each physiological feature in the multi-dimensional physiological feature matrix, a physiological dimension entropy of a dimension in which the physiological feature is located is determined. All feature correlation degrees corresponding to the physiological feature are extracted in the physiological state atlas. The physiological core entropy of the physiological feature is determined through the physiological dimension entropy of the dimension in which the physiological feature is located and the corresponding all feature correlation degrees, and the physiological core entropy of each physiological feature in the multi-dimensional physiological feature matrix is obtained. All core physiological features are screened according to the physiological core entropy of each physiological feature. The core physiological feature sequence is constructed according to all the core physiological features.

[0013] Preferably, the feedback module performs real-time dynamic feedback on the core physiological feature sequence through the mobile client by transmitting the core physiological feature sequence to the mobile client through the long-distance data transmission module.

[0014] Compared with the prior art, the present application has the following beneficial effects: In the aspect of data acquisition, the system adopts a mouthpiece type gas collection device to obtain multi-dimensional physiological data of the athlete, the device conforms to the breathing habit of the athlete during training, does not interfere with the rowing action, and can stably collect various core physiological indexes such as oxygen uptake, breathing rate and gas exchange rate on the premise of ensuring the coherence of the athlete's training. At the same time, the athlete physiological data warehouse formed through data integration breaks the limitation of scattered data and single dimension in traditional monitoring, and provides a comprehensive and complete data basis for subsequent physiological state analysis. The feature construction module extracts features from the physiological data warehouse to form a multi-dimensional physiological feature matrix, and further constructs a dimensional correlation graph and a physiological state map, achieving in-depth mining and integration of physiological data. The multi-dimensional physiological feature matrix covers the changing characteristics of various key physiological indicators during athlete training, while the dimensional correlation graph clearly presents the intrinsic relationships between different physiological indicators, such as the dynamic correlation between oxygen uptake and heart rate, and energy consumption. The physiological state map constructed on this basis can intuitively and systematically reflect the changing trends of athletes' physiological states during training, allowing coaches to clearly grasp the correlation between athletes' physiological states and factors such as training intensity and duration, providing an intuitive reference for understanding athletes' energy consumption patterns and endurance limits. The cloud-based fusion module transmits a multi-dimensional physiological feature matrix to a cloud server for cluster analysis. Leveraging the powerful computing capabilities of the cloud, it can quickly and efficiently process large amounts of physiological data, forming local physiological state cluster centers and preliminary training scenario discrimination cluster centers. The local physiological state cluster centers can accurately locate the physiological state characteristics of athletes at different training stages, such as the clustering patterns of physiological indicator changes in the early, middle, and late stages of training. The preliminary training scenario discrimination cluster centers can preliminarily distinguish different training scenario types based on physiological data characteristics, achieving an initial correlation between physiological data and training scenarios. This cloud-based cluster analysis method overcomes the limitations of local device computing power, improves the efficiency and accuracy of data processing, and lays the foundation for subsequent multi-source feature fusion. The sequence generation module integrates physiological state maps and two types of cluster centers to form a multi-source fusion feature representation of the training scenario, thereby generating a real-time core physiological feature sequence. This enables precise characterization and dynamic tracking of the athlete's physiological state. The multi-source fusion feature representation integrates physiological data features, indicator correlation patterns, and scenario discrimination information, comprehensively reflecting the athlete's overall physiological state in a specific training scenario. The core physiological feature sequence, in the form of a dynamic sequence, presents the changes in key physiological indicators such as the athlete's oxygen uptake in real time, allowing coaches to promptly grasp the athlete's physical fitness changes and accurately determine whether the athlete is in optimal training condition or at risk of physical exhaustion. The feedback module realizes real-time dynamic feedback of the core physiological feature sequence through the mobile client, so that the coach and the athlete can view the physiological state data at any time during the training process, and breaks the time limit of traditional offline analysis. The coach can adjust the training intensity, rowing rhythm and other training strategies in time according to the real-time feedback physiological data, for example, when it is found that the oxygen uptake of the athlete is continuously too high and the physiological load is close to the tolerance limit, the training intensity can be immediately reduced to avoid sports injuries; the athlete can also understand the physical state of himself through real-time data, and actively adjust the exercise posture and breathing rhythm to improve the autonomy and effectiveness of training. In addition, real-time dynamic feedback also provides convenience for instant communication and guidance during training, and the coach can provide targeted on-site guidance for the athlete based on real-time physiological data, further improving the training effect. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 Timing diagram of the oxygen uptake telemetry system for water training of the applicable racing boat project according to the present application; Figure 2 Flow chart of multi-dimensional physiological data integration of the data acquisition module of the oxygen uptake telemetry system; Figure 3 Flow chart of dimension association graph determination of the feature construction module of the oxygen uptake telemetry system. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0017] Please refer to Figure 1 The present application provides an oxygen uptake telemetry system for water training of the applicable racing boat project, which comprises a data acquisition module, a feature construction module, a cloud fusion module, a sequence generation module and a feedback module.

[0018] The data acquisition module obtains multi-dimensional physiological data of the rowing athlete through the mouthpiece gas collection device, and integrates the multi-dimensional physiological data to obtain an athlete physiological data warehouse; the feature construction module extracts features from the athlete physiological data warehouse to obtain a multi-dimensional physiological feature matrix, determines a dimension correlation graph for each dimension in the multi-dimensional physiological feature matrix, and constructs an athlete training-related physiological state graph through the dimension correlation graph of each dimension; the cloud fusion module transmits the multi-dimensional physiological feature matrix to a cloud server, performs clustering analysis on the multi-dimensional physiological feature matrix to obtain a local physiological state clustering center and a training scene preliminary discrimination clustering center; the sequence generation module fuses the local physiological state clustering center and the training scene preliminary discrimination clustering center based on the physiological state graph to obtain a training scene multi-source fusion feature representation, generates a real-time core physiological feature sequence of the athlete according to the training scene multi-source fusion feature representation; and the feedback module performs real-time dynamic feedback on the core physiological feature sequence through a mobile client.

[0019] Embodiment 1: see Figure 2 The original multi-dimensional physiological data obtained by the mouthpiece gas collection device is preprocessed. In the preprocessing stage, a multi-stage filtering technique is used to eliminate mixed environmental electromagnetic noise and motion artifacts in the collected signal. At the same time, a sliding window algorithm is used to detect and eliminate abnormal peak data caused by temporary unstable connection of the equipment. Subsequently, the minimum-maximum normalization method is used to unify the data scales of each dimension to the same numerical range, thereby obtaining the preprocessed multi-dimensional physiological data. Next, the preprocessed multi-dimensional physiological data is subjected to semantic mapping. This process establishes a data annotation system based on ontology, assigns a unique semantic identifier to each data dimension, for example, maps the airflow rate data to "flow_rate" and associates its unit of measurement "L / min", and maps the blood oxygen saturation data to "spo2" and marks its percentage attribute. At the same time, a uniform timestamp and athlete identity code are added to all time series data, thereby obtaining standardized physiological data for each dimension. These data not only have a standard format but also have clear physical meaning. Finally, an athlete physiological data warehouse is constructed through all the standardized physiological data of each dimension. The warehouse is stored using a time series database architecture, and a multi-level indexing mechanism is designed with athlete ID as the primary key and timestamp as the auxiliary index. This supports efficient range queries and streaming data access. Each data point retains its original collection value, standardized value, and semantic label triple structure.

[0020] When the feature construction module extracts features from the athlete physiological data warehouse, it first calculates features for each dimension of the standardized physiological data in the warehouse. For the respiratory rate dimension, it calculates the moving average and coefficient of variation of breaths per minute using a sliding time window. For the oxygen uptake dimension, it extracts the rising slope feature and steady-state value. It also extracts statistical features for all numerical dimensions, including the root mean square value, peak-to-peak value, and sample entropy. This results in a physiological feature vector for each dimension, which contains multiple feature indicators for that dimension within a specific time window. A multidimensional physiological feature matrix is constructed based on the physiological feature vectors for all dimensions. During matrix construction, a time alignment mechanism is used to ensure that feature vectors for different dimensions remain synchronized within the same time segment. The row dimension of the matrix represents a sequence of consecutive time segments, and the column dimension arranges groups of feature vectors for different physiological parameters. Each matrix element stores a feature value for a specific dimension within a certain time segment, ultimately forming a two-dimensional data structure that comprehensively reflects changes in the athlete's physiological state. The matrix also introduces a missing value processing mechanism. When there is a data interruption during data collection for a certain time segment, an interpolation method is used to fill in the data from adjacent segments, ensuring the time continuity of the matrix and providing complete data input for subsequent correlation graph construction. The storage format of the matrix uses sparse matrix compression technology, storing only non-zero feature values and their time position indexes to optimize data transmission and processing efficiency.

[0021] Example 2: see Figure 3 Among all dimensions of the matrix, a single dimension is selected for analysis in sequence. When a specific dimension is selected, the system extracts the physiological feature vector corresponding to that dimension from the matrix, which contains all feature values for that dimension over a continuous time sequence. To determine the feature correlation between each physiological feature in the selected dimension's physiological feature vector, a statistical correlation analysis method is used to calculate the Pearson correlation coefficient between each pair of features to quantify their linear correlation strength. Mutual information calculation is also used to capture non-linear correlation relationships, resulting in a comprehensive correlation score for each feature pair. To construct the dimension correlation graph for the selected dimension based on the feature correlation between each physiological feature, each physiological feature is treated as a node in the graph structure. Whether to establish a connection edge between nodes is determined based on the calculated feature correlation value. A correlation threshold is set to automatically filter out weak correlations and retain significant feature correlations, forming a directed and weighted graph representation of the dimension. The weight of the edge in the graph directly corresponds to the numerical value of the feature correlation. In this way, the dimension correlation graph for each dimension in the multidimensional physiological feature matrix is gradually constructed.

[0022] In the process of constructing the physiological state atlas related to the training of athletes by the dimension correlation graph of each dimension, it is necessary to determine the key physiological characteristics of each dimension, to identify the most important nodes in each dimension correlation graph by using the algorithm based on the centrality of the graph, to comprehensively evaluate the importance of each physiological characteristic in the dimension by calculating the degree centrality, closeness centrality and eigenvector centrality of the characteristic nodes, and to select the top-ranked characteristics as the key physiological characteristics. Then, the feature correlation degree between each key physiological characteristic is determined, which requires cross-dimension calculation of the correlation between different key characteristics of different dimensions, the use of time series alignment technology to ensure that the characteristic values of different dimensions are on the same time reference, and then the use of the same correlation calculation method to establish the cross-dimension feature correlation.

[0023] The dimension correlation graphs of each dimension are connected according to the feature correlation degree between each key physiological characteristic, cross-dimension connection edges are established between the key characteristic nodes of different dimensions, the weight of the edge is determined by the cross-dimension feature correlation degree, and finally a unified and cross-dimension graph network structure is formed, which is the physiological state atlas related to the training of athletes, which fully reflects the complex correlation between different physiological dimensions and the characteristics within the dimensions, and provides a structured data basis for subsequent physiological state analysis. In the process of constructing the atlas, a dynamic updating mechanism is introduced, which can adjust the feature correlation degree and the atlas structure in real time according to new training data, ensuring that the atlas always reflects the latest physiological state characteristics of athletes.

[0024] The implementation process of the feature construction module is illustrated by taking a water training of a rowing athlete as an example. The athlete wears a mouthpiece gas collection device to perform continuous rowing training, and the device collects four-dimensional physiological data of breathing rate, oxygen uptake, heart rate variability and blood oxygen saturation in real time. The data collection module has completed preprocessing and standardization, forming a four-dimensional physiological data warehouse containing 120 time points. The feature construction module first extracts the four-dimensional data from the data warehouse for feature calculation, extracts the moving average, standard deviation and sample entropy features of the breathing rate dimension to form a 3-dimensional feature vector, calculates the slope feature, peak maintenance and fluctuation coefficient of the oxygen uptake dimension to form a 3-dimensional feature vector, analyzes the time domain SDNN feature and frequency domain LF / HF ratio of the heart rate variability to form a 2-dimensional feature vector, and extracts the falling rate and minimum value maintenance time features of the blood oxygen saturation to form a 2-dimensional feature vector. After aligning all the feature vectors by time, they are combined into a 10x120-dimensional multi-dimensional physiological feature matrix, each row of which represents a time point and each column corresponds to a characteristic dimension.

[0025] In the construction of the dimension correlation graph of the respiratory frequency dimension, the system calculated the correlation coefficient matrix between the three characteristics of this dimension, and found that the negative correlation coefficient between the moving average and the sample entropy reached-0.82, the correlation coefficient between the moving average and the standard deviation was 0.75, and the correlation coefficient between the standard deviation and the sample entropy was-0.68. According to the set threshold of 0.7, a strong negative correlation edge (weight 0.82) was established between the moving average and the sample entropy, a positive correlation edge (weight 0.75) was established between the moving average and the standard deviation, and a dimension correlation graph containing 3 nodes and 2 edges was formed. In the same way, the correlation graphs of the remaining dimensions were constructed: in the oxygen uptake dimension, the correlation coefficient between the slope characteristic and the peak value maintenance was 0.79, and the correlation coefficient with the fluctuation coefficient was-0.83, two strong correlation edges were established; in the heart rate variability dimension, the correlation coefficient between SDNN and LF / HF ratio was-0.71, and one correlation edge was established; in the oxygen saturation dimension, the correlation coefficient between the decline rate and the minimum value maintenance time was-0.86, and one strong correlation edge was established.

[0026] In the determination of the key physiological characteristics of each dimension, it was found through the calculation of the characteristic vector centrality that the moving average node centrality of the respiratory frequency dimension was the highest (0.62), the fluctuation coefficient node centrality of the oxygen uptake dimension was the highest (0.58), the LF / HF ratio node centrality of the heart rate variability dimension was the highest (0.61), and the decline rate node centrality of the oxygen saturation dimension was the highest (0.59). These key characteristics were used as anchor points for cross-dimension connection. The cross-dimension correlation degree between the key characteristics was calculated: the correlation coefficient between the respiratory frequency moving average and the oxygen uptake fluctuation coefficient was 0.73, the correlation coefficient with the heart rate variability LF / HF ratio was-0.69, and the correlation coefficient with the oxygen saturation decline rate was 0.71; the correlation coefficient between the oxygen uptake fluctuation coefficient and the heart rate variability LF / HF ratio was-0.76, and the correlation coefficient with the oxygen saturation decline rate was 0.82. According to these correlation degree values, cross-dimension connection edges were established between the correlation graphs of the four dimensions, and finally a physiological state graph containing 10 characteristic nodes, 9 intra-dimension connection edges and 6 cross-dimension connection edges was formed. This graph fully reveals the internal relationship between various physiological parameters during the training of athletes: the stability of respiratory frequency is strongly positively correlated with oxygen uptake fluctuation, negatively correlated with heart rate regulation ability, and positively correlated with oxygen decline rate; the fluctuation characteristics of oxygen uptake are highly correlated with autonomic nervous regulation function and closely linked with oxygen change rate.

[0027] In the process of clustering analysis of the multi-dimensional physiological feature matrix by the cloud fusion module to obtain the local physiological state clustering center, first, the physiological feature vectors of each dimension in the matrix are subjected to full connection coding processing. The coding process adopts a neural network layer with a nonlinear activation function to map the feature vectors of each dimension to a high-dimensional embedding space. This mapping can capture the complex nonlinear relationship between the features to generate high-dimensional feature representation vectors. The embedding vectors of all dimensions collectively constitute a set of physiological feature full connection embedding coding vectors. When performing density clustering analysis on the vector set, a density-based spatial clustering algorithm is used. This algorithm automatically determines the number of clusters and identifies noise points by calculating the density distribution of sample points in the vector space. The algorithm first finds core sample points and then expands the clustering area to finally form multiple clusters of density-connected sample points. The center point of each cluster is determined as the local physiological state clustering center. These center points represent the typical patterns of different physiological states.

[0028] In further clustering analysis of the multi-dimensional physiological feature matrix to obtain the training scene preliminary discrimination clustering center, the training scene preliminary discrimination types of each dimension need to be converted by one-hot encoding. This conversion process converts discrete scene category labels into binary vector form. In each vector, only the element at the corresponding category position is 1 and the rest are 0. These encoding vectors of all dimensions constitute a set of training scene discrimination one-hot encoding vectors. The vector set is subjected to partition clustering processing. A partition-based clustering algorithm is used to divide the vector space into multiple regions. The algorithm continuously adjusts the clustering center position through iterative optimization so that the vectors within the same region have a high degree of similarity. Finally, multiple local discrimination clustering centers of the training scene are obtained, each center representing the scene feature pattern of a local region.

[0029] When calculating the weighted sum of the multiple local discrimination clustering centers of the training scene to obtain the training scene preliminary discrimination clustering center, the following calculation method is adopted:

[0030] Wherein: represents the final obtained training scene preliminary discrimination clustering center vector, represents the total number of local discrimination clustering centers, is the weight coefficient of the i-th local clustering center, which is determined according to the data density and quality of the corresponding clustering region, The vector represents the local discriminant clustering center of the ith training scene. The weighted calculation process ensures that important clustering centers have a greater contribution to the final result. During the entire clustering analysis process, a clustering quality evaluation mechanism is also introduced. The clustering effect is evaluated by calculating the tightness within the cluster and the separation between clusters. According to the evaluation results, the clustering parameters are dynamically adjusted to ensure that the clustering centers obtained are representative and discriminant. All clustering operations are completed on the cloud server. Distributed computing framework is used to process large-scale data, improve computing efficiency and scalability, and the final clustering center results will be used for subsequent feature fusion and sequence generation steps.

[0031] Taking the continuous training monitoring of a rowing athlete as an example, the cloud fusion module needs to process a set of multi-dimensional physiological feature matrix, which contains four-dimensional feature data of respiratory rate, oxygen uptake, heart rate and blood oxygen saturation collected in the past 30 minutes of training. Each dimension contains 10 feature indicators, forming a 40x180 feature matrix (40 feature dimensions, 180 time points). When performing full connection encoding processing on the feature matrix, a three-layer neural network structure is used. The number of input layer nodes corresponds to the number of features in each dimension. The hidden layer is set to 128 nodes and uses the ReLU activation function. The output layer generates a 32-dimensional embedding vector. The 10 features of the respiratory rate dimension are encoded to obtain a 32-dimensional embedding vector. The oxygen uptake, heart rate and blood oxygen saturation dimensions also generate their own 32-dimensional embedding vectors. All embedding vectors at different time points form an 128x180 embedding vector set (128-dimensional features, 180 time points). When performing density clustering on the embedding vector set, an improved DBSCAN algorithm is used, with a neighborhood radius of 1.2 and a minimum sample size of 5. The algorithm first calculates the distance distribution of each vector and its k-nearest neighbors to determine the boundary of the core sample point. Then it expands the clustering area from any core point. Finally, the 180 time point vectors are divided into 3 clustering clusters and 1 noise point set. The mean of all vectors in each cluster is calculated to obtain 3 local physiological state clustering centers, representing three different physiological state patterns.

[0032] In the one-hot encoding of the training scene, the training scene is divided into three types of "aerobic endurance", "anaerobic strength" and "recovery adjustment", and the scene type is labeled according to the real-time training intensity at each time point. "Aerobic endurance" is encoded as [1, 0, 0], "anaerobic strength" is encoded as [0, 1, 0], and "recovery adjustment" is encoded as [0, 0, 1]. The one-hot encoding vector of 180 time points forms a 3x180 encoding matrix. When the one-hot encoding matrix is partitioned and clustered, the K-means algorithm is used to divide the 180 time points into 4 regions, each region containing 45 time points. The mean of the one-hot encoding vector in each region is calculated to obtain 4 local discriminant clustering centers of the training scene, and the values thereof reflect the proportion distribution of different types of training scenes in different regions.

[0033] When calculating the weighted sum of the 4 local discriminant clustering centers, the weight coefficients are set according to the number of time points contained in each region. The first region contains 50 time points with a weight of 0.28, the second region contains 40 time points with a weight of 0.22, the third region contains 45 time points with a weight of 0.25, and the fourth region contains 45 time points with a weight of 0.25. The vector representation of the 4 clustering centers is weighted and averaged according to the weights, and finally a 3-dimensional preliminary discriminant clustering center of the training scene is obtained, and the values thereof reflect the comprehensive distribution of different types of scenes in the whole training process. The entire clustering analysis process runs on a cloud distributed computing framework, and parallel computing is realized using the SparkML library. The processing tasks of each dimension are allocated to different computing nodes for execution, and the final generated clustering center result is stored in a distributed file system to provide input data for subsequent feature fusion. The system also records quality evaluation indicators during the clustering process, including clustering silhouette coefficients and intra-class distances, which are used to monitor the clustering effect and adjust algorithm parameters.

[0034] In the process of fusing the local physiological state clustering center and the preliminary discriminant clustering center of the training scene based on the physiological state atlas in the sequence generation module, the preliminary discriminant clustering center of the training scene and the local physiological state clustering center need to be processed in a cross-dimension isomorphism manner in the specific implementation. This processing process projects the clustering centers in different feature spaces into a unified feature space through a dimension mapping function. For example, when the training scene discriminant clustering center contains 5-dimensional features and the physiological state clustering center contains 8-dimensional features, the dimension of the two centers is expanded to the same 12-dimensional feature space through a feature alignment algorithm. This expansion process uses feature padding and dimension transformation techniques to ensure the comparability and operability of clustering centers from different sources. After such cross-dimension isomorphism processing, an optimized training scene discriminant clustering center is finally obtained.

[0035] When cascading the optimized training scene discriminative clustering center and the local physiological state clustering center, the feature vector splicing method is used to combine the feature representations of the two centers into a composite vector. For example, the optimized training scene discriminative clustering center is a 12-dimensional vector, and the local physiological state clustering center is a 10-dimensional vector. After cascading, a 22-dimensional comprehensive feature vector is generated. This cascading process maintains the spatial order and numerical accuracy of the original features. The final training scene multi-source fusion feature representation contains dual information features of scene discrimination and physiological state. Referring to Table 1, the feature vector changes before and after cross-dimension isomorphic processing are shown.

[0036] Table 1: Cross-dimension isomorphic processing data

[0037] In the specific implementation process, cross-dimension isomorphic processing uses a feature projection matrix to achieve dimension expansion. This projection matrix is learned from training data and can maintain the semantic information and numerical relationship of the original features. For example, for each original feature dimension of the training scene discriminative clustering center, the projection matrix generates a corresponding new dimension mapping value while filling in the default feature value for the missing dimension. During the cascading operation, feature standardization processing is implemented to normalize all feature values to the same numerical range, eliminating the bias caused by different feature scales and ensuring that the fused feature representation has consistent numerical characteristics.

[0038] The entire fusion process implements a dynamic weight adjustment mechanism. According to the real-time physiological data quality and the confidence level of the training scene judgment, the contribution weights of the two clustering centers in the fusion process are automatically adjusted. When the training scene discriminative confidence is high, the weight of the scene clustering center is appropriately increased. When the physiological state data quality is good, the weight of the state clustering center is increased. This dynamic adjustment enables the fusion feature representation to adapt to different training environments and data conditions. The final generated training scene multi-source fusion feature representation will serve as the basis for the generation of core physiological feature sequences, providing feature support for subsequent real-time feedback.

[0039] In the physiological state atlas, all feature correlation degrees corresponding to the physiological features need to be extracted. The edge weights connected to the feature node in the atlas represent the correlation strength between the feature and other features. For example, the oxygen uptake feature may be connected to multiple features such as heart rate and vital capacity. The weight value of each edge reflects the cooperative change relationship between these physiological parameters. The extraction process uses a graph traversal algorithm to collect the weight values of all adjacent edges to form the feature correlation degree set of the feature.

[0040] In the physiological state atlas, all feature correlation degrees corresponding to the physiological features need to be extracted. The edge weights connected to the feature node in the atlas represent the correlation strength between the feature and other features. For example, the oxygen uptake feature may be connected to multiple features such as heart rate and vital capacity. The weight value of each edge reflects the cooperative change relationship between these physiological parameters. The extraction process uses a graph traversal algorithm to collect the weight values of all adjacent edges to form the feature correlation degree set of the feature.

[0041] When constructing the core physiological feature sequence according to all core physiological features, the screened feature values are arranged in time sequence, and each time point in the sequence contains the values of all core features at that moment. The sequence construction uses a ring buffer data structure to support real-time streaming data processing. The latest generated feature values are added to the sequence in time, while the outdated historical data is removed, keeping the constant length of the sequence and the real-time nature of the data. When the feedback module provides real-time dynamic feedback of the core physiological feature sequence through the mobile client, the remote data transmission module uses 4G / 5G wireless communication technology to package and transmit the sequence data. The data packet is encoded in a compact binary format to reduce transmission delay. After receiving the data, the mobile client parses and decodes it, and displays the trend curve of the core physiological features in real time through the graphical interface. At the same time, a threshold alarm mechanism is set, which automatically sends visual and audible warnings when some feature values exceed the normal range. Coaches can view detailed feature data for a specific time period through the interface interaction function, and also adjust the display parameters and alarm thresholds to adapt to different training monitoring needs.

[0042] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0043] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A remote oxygen uptake measurement system suitable for on-water training in rowing events, characterized in that, The system includes: The data acquisition module is used to acquire multidimensional physiological data of rowers through a mouthpiece gas collection device, and to integrate the multidimensional physiological data to obtain an athlete physiological data warehouse. The feature construction module is used to extract features from the athlete physiological data warehouse to obtain a multi-dimensional physiological feature matrix, determine the dimensional correlation graph of each dimension in the multi-dimensional physiological feature matrix, and construct a physiological state map related to athlete training through the dimensional correlation graphs of each dimension. The cloud fusion module is used to transmit the multi-dimensional physiological feature matrix to the cloud server and perform cluster analysis on the multi-dimensional physiological feature matrix to obtain local physiological state cluster centers and preliminary discrimination cluster centers of training scenarios. The sequence generation module is used to fuse the local physiological state clustering center and the preliminary discrimination clustering center of the training scene based on the physiological state map to obtain the multi-source fusion feature representation of the training scene, and to generate the real-time core physiological feature sequence of the athlete based on the multi-source fusion feature representation of the training scene. The feedback module is used to provide real-time dynamic feedback of the core physiological feature sequence via a mobile client.

2. The oxygen uptake remote sensing system for rowing water training as described in claim 1, characterized in that, The data acquisition module integrates the multidimensional physiological data, specifically including: The multidimensional physiological data acquired through the mouthpiece gas collection device is preprocessed to obtain preprocessed multidimensional physiological data. Semantic mapping is performed on the preprocessed multidimensional physiological data to obtain standardized physiological data for each dimension. A physiological data warehouse for athletes is built using standardized physiological data across all dimensions.

3. The oxygen uptake remote sensing system for rowing water training as described in claim 1, characterized in that, The feature construction module performs feature extraction on the athlete physiological data warehouse, specifically including: Feature extraction is performed on the standardized physiological data of each dimension in the athlete physiological data warehouse to obtain physiological feature vectors for each dimension. A multidimensional physiological feature matrix is ​​constructed based on the physiological feature vectors of all dimensions.

4. The oxygen uptake remote sensing system for on-water training in rowing events as described in claim 1, characterized in that, The feature construction module determines the dimensional correlation graph of each dimension in the multi-dimensional physiological feature matrix, specifically including: Select one dimension from all dimensions of the multi-dimensional physiological feature matrix and obtain the physiological feature vector of the selected dimension; Determine the feature correlation degree between each physiological feature in the selected dimension of the physiological feature vector; Based on the feature correlation degree between various physiological features, a dimension correlation graph of the selected dimension is constructed, thereby obtaining the dimension correlation graph of each dimension in the multi-dimensional physiological feature matrix.

5. The oxygen uptake remote sensing system for on-water training in rowing events as described in claim 1, characterized in that, The feature construction module constructs a physiological state map related to athlete training through dimensional correlation graphs of various dimensions, specifically including: Identify the key physiological characteristics of each dimension, and then determine the feature correlation between each key physiological characteristic; By connecting the dimensional correlation diagrams of each dimension based on the feature correlation degree between each key physiological characteristic, a physiological state map related to athlete training can be obtained.

6. The oxygen uptake remote sensing system for on-water training in rowing events as described in claim 1, characterized in that, The cloud-based fusion module performs cluster analysis on the multi-dimensional physiological feature matrix, specifically including: Fully connected encoding is performed on the physiological feature vectors of each dimension in the multi-dimensional physiological feature matrix to obtain a set of fully connected embedded encoding vectors of physiological features; Density clustering is performed on the set of fully connected embedded encoding vectors of the physiological features to obtain the cluster centers of the local physiological states.

7. The oxygen uptake remote sensing system for rowing water training as described in claim 1, characterized in that, The cloud-based fusion module also performs cluster analysis on the multi-dimensional physiological feature matrix, including: One-hot encoding is performed on the preliminary discrimination type of each dimension of the training scene in the multi-dimensional physiological feature matrix to obtain a set of training scene discrimination one-hot encoded vectors; The set of discriminative one-hot encoded vectors for the training scene is partitioned and clustered to obtain multiple local discriminative cluster centers for the training scene; Calculate the weighted sum of the local discriminative cluster centers of the multiple training scenarios to obtain the preliminary discriminative cluster centers of the training scenarios.

8. The oxygen uptake remote sensing system for on-water training in rowing events as described in claim 1, characterized in that, The sequence generation module, based on the physiological state map, fuses the local physiological state cluster centers and the preliminary cluster center determination of the training scenario, specifically includes: The training scene preliminary discrimination cluster center and the local physiological state cluster center are subjected to cross-dimensional isomorphism to obtain the optimized training scene discrimination cluster center; The optimized training scene discrimination cluster center and the local physiological state cluster center are cascaded to obtain a multi-source fusion feature representation of the training scene.

9. The oxygen uptake remote sensing system for on-water training in rowing events as described in claim 1, characterized in that, The sequence generation module generates a real-time core physiological feature sequence of the athlete based on the multi-source fusion feature representation of the training scenario, specifically including: For each physiological feature in the multidimensional physiological feature matrix, determine the physiological dimension entropy of the dimension in which the physiological feature is located; Extract all feature correlation degrees corresponding to the physiological features from the physiological state map; The physiological core entropy of the physiological feature is determined by the physiological dimension entropy of the dimension in which the physiological feature is located and the correlation degree of all corresponding features, thereby obtaining the physiological core entropy of each physiological feature in the multi-dimensional physiological feature matrix. All core physiological features are selected based on the physiological core entropy of each physiological feature. Construct a core physiological feature sequence based on all core physiological features.

10. The oxygen uptake remote sensing system for on-water training in rowing events as described in claim 1, characterized in that, The feedback module provides real-time dynamic feedback of the core physiological feature sequence through a mobile client by transmitting the core physiological feature sequence to the mobile client via a long-distance data transmission module.

Images