Training physical state abnormity monitoring method and system based on clustering processing

Abstract

The invention discloses a training physical state abnormity monitoring method and system based on clustering processing. The method comprises the steps of data collection, clustering data set construction, clustering processing and abnormity judgment. The invention relates to the technical field of training intelligent monitoring, in particular to a training physical state abnormity monitoring method and system based on clustering processing, according to the scheme, window features are embedded in a low-dimensional mode, and local density and linear reconstruction residual errors are evaluated at the same time in a multi-scale neighborhood to select clustering seeds; the clustering structure is dynamically maintained by innovatively utilizing an information criterion and a splitting and merging strategy based on transfer entropy; soft distribution, time consistency penalty and boundary confidence estimation are introduced, clustering robustness, adaptability and time continuity are improved, and boundary uncertainty is quantified; through weighted fusion of three types of complementary anomaly metrics and scaling by using clustering confidence, a continuous window triggering rule is generated, space-time anomaly changes are considered, the false alarm rate is remarkably reduced through multi-index mutual verification, and the reliability and interpretability of anomaly monitoring are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent training monitoring technology, specifically to a method and system for monitoring abnormal training physical fitness based on clustering processing. Background Technology

[0002] In recent years, with the development of sensor signal acquisition, edge computing and machine learning, continuous monitoring of training physical fitness has become an important research direction. Multimodal signals can provide rich information on exercise intensity, fatigue and physiological stress. At the same time, with the maturity of unsupervised and online learning algorithms, data-driven state modeling and anomaly detection have shown broad application prospects in personalized training monitoring, real-time risk warning and adaptive adjustment of training programs.

[0003] Existing technologies still have shortcomings in practical applications: many methods directly perform clustering or distance measurement on high-dimensional original features or simple statistical features, which are easily affected by dimensionality, noise and sensor drift, and are sensitive to initial parameters and neighborhood scale; most methods are mainly offline batch processing, which is difficult to meet the requirements of online updates and real-time response; at the same time, single spatial distance or threshold detection often ignores short-term cluster migration behavior and statistical mutations within the group, resulting in a high false alarm or false negative rate; in addition, the lack of quantification and time stability constraints on the uncertainty of cluster boundaries makes the reliability and interpretability of anomaly judgment insufficient, which is not conducive to reliable alarm and intervention for risks in actual training and monitoring. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method and system for monitoring abnormal training physical condition based on clustering. Addressing the problems of traditional clustering methods being susceptible to initialization, scale selection, and noise in high-dimensional physiological or exercise data, struggling to capture local low-dimensional structures and temporal dynamics, and typically unable to update online and quantify boundary uncertainties, this solution robustly selects clustering seeds by first performing low-dimensional embedding on window features and simultaneously evaluating local density and local linear reconstruction residuals at multiple neighborhood scales. Online updates are achieved using an incremental Gaussian mixture model, and information criteria and a splitting and merging strategy based on transfer entropy are used for dynamic... The system maintains the cluster structure and employs soft allocation with temporal consistency penalties and cluster boundary confidence estimation to improve the robustness, adaptability, and temporal continuity modeling capabilities of the clusters. Addressing the issues of traditional anomaly detection methods that often rely on single spatial distances or thresholds, are prone to false positives or false negatives, ignore short-term cluster migrations and intra-cluster statistical mutations, and lack confidence-based scaling and temporal continuity judgment, this scheme constructs three complementary anomaly measures simultaneously, fuses them by weight, uses cluster confidence to scale scores, and incorporates statistical thresholds from historical normal data and continuous window triggering rules, thus balancing spatial anomalies with temporal behavior changes.

[0005] The technical solution adopted in this invention is as follows: a training physical fitness anomaly monitoring method based on clustering processing, the method comprising the following steps:

[0006] Step S1: Data acquisition, collecting heart rate sequence, triaxial acceleration, respiratory rate, skin conductance signal, and skin temperature of historical trainees;

[0007] Step S2: Construct a clustering dataset. For each time window, extract statistical features from the original signal and construct a window feature vector to form a sample set for clustering.

[0008] Step S3: Clustering. First, the window features are mapped to low-dimensional encoding, and the local similarity and local linear reconstruction error of each sample point are evaluated at multiple neighborhood scales. Points that simultaneously satisfy high density and low residual are used as cluster seeds. Then, starting from these seeds, an incremental Gaussian mixture model is used to maintain the statistical view of each cluster. A merging strategy is designed based on information criteria, and a splitting strategy is designed based on the multimodal and time series transition characteristics within the cluster. At the same time, soft assignment is used for boundary points, introducing a time stability penalty, and giving a confidence estimate of the cluster boundary.

[0009] Step S4: Anomaly detection. After clustering and assigning the new windows of the monitored personnel, three independent anomaly measures are constructed: spatial anomaly degree relative to the optimal cluster center, anomaly degree of short-term cluster transition probability, and statistical mutation degree within the cluster. These three items are then combined into a final anomaly score according to their weights. The score is scaled using the cluster confidence score and detected based on a threshold obtained from historical normal data statistics. When the anomaly score of a window exceeds the threshold, an anomaly alarm is triggered, indicating that the training physical condition is abnormal.

[0010] Further, in step S1, the data acquisition involves collecting the heart rate sequence, triaxial acceleration, respiratory rate, skin conductance signal, and skin temperature of historical trainees; segmenting the data according to the sliding window and step size, and generating an observation sample vector for each window.

[0011] Further, in step S2, constructing the clustering dataset specifically includes the following steps:

[0012] Step S21: Define the feature window, set the sliding window length and step size, and construct a window feature vector for each window;

[0013] Step S22: Calculate the original domain statistical features. The feature vector includes the original domain statistical features. The original domain statistical features are the mean and standard deviation of any time series calculated over a window.

[0014] Step S23: Construct a window vector based on the original domain statistical features.

[0015] Further, in step S3, the clustering process specifically includes the following steps:

[0016] Step S31: Select clustering seeds. For each window vector, first map it to a low-dimensional encoding, construct a local similarity map at multiple scales, and simultaneously calculate two complementary indices: local density peak score and projection residual from the point to the local area. Set the points that simultaneously satisfy high density and low residual as clustering seeds. Specifically, this includes the following steps:

[0017] Step S311: Window vector embedding, mapping each window vector to a low-dimensional code through linear mapping;

[0018] Step S312: Construct a multi-scale neighborhood set and define the scale set; utilize The nearest neighbor algorithm constructs a set of neighbors for each scale;

[0019] Step S313: Local density scoring, define local density scores for sample points at each scale;

[0020] Step S314: Calculate the projection residuals, perform weighted local principal component analysis in the neighborhood, and calculate the r-dimensional principal component subspace;

[0021] Step S315: Cluster seed selection, synthesize scores for all scales, set a fixed threshold; select points with scores higher than the fixed threshold as initial cluster seeds;

[0022] Step S32: Cluster generation. Starting with the cluster seed, an incremental Gaussian mixture model is used to maintain the statistical view of each cluster: mean, covariance matrix, and sample count. Merging decisions are designed based on information gain according to the Bayesian information criterion, and splitting decisions are designed based on multimodal attitude quantities and transfer entropy within the cluster. Specifically, this includes the following steps:

[0023] Step S321: Cluster parameter statistics, maintain statistics for each cluster: number of samples, mean, covariance matrix;

[0024] Step S322: Merge decision, based on log-likelihood value and model complexity;

[0025] Step S323: Split determination, record the time index set of each cluster in the time window sequence; perform local splitting within the cluster. Clustering is performed to obtain candidate sub-clusters, and the splitting bonus is calculated.

[0026] Step S33: Calculate the cluster boundary confidence score, apply soft assignment to the boundary points and introduce a time stability penalty; simultaneously, model the uncertainty of the number of clusters using an approximate posterior, and give the cluster boundary confidence interval, specifically including the following steps:

[0027] Step S331: Soft assignment, calculate unnormalized similarity weights based on Gaussian negative logarithmic distance, and apply time stability penalty at the same time;

[0028] Step S332: Calculate the cluster boundary confidence score, define the activity score for each cluster, and calculate the cluster boundary confidence score.

[0029] Furthermore, in step S4, the anomaly determination specifically includes the following steps:

[0030] Step S41: Integrate anomaly scoring. Real-time collection of training data from personnel requiring monitoring. A data vector is generated using step S2. This new data to be monitored is recorded as the data vector for a new window. For each new window, three independent anomaly metrics are constructed based on the clustering assignment results and inter-cluster transition statistics: Mahalanobis distance anomaly to the optimal cluster; short-term cluster transition probability anomaly; and intra-cluster statistical mutation anomaly. The final anomaly score is obtained, and the score is scaled using cluster confidence. The specific steps include:

[0031] Step S411: Calculate spatial anomaly, take the cluster center of the best cluster, and calculate the spatial anomaly.

[0032] Step S412: Calculate the transition probability anomaly degree, define the empirical transition probability matrix, and calculate the transition probability anomaly degree;

[0033] Step S413: Calculate the statistical mutation rate within the cluster. Compare the current sample window vector with the mean of the samples most recently assigned to the optimal cluster, and calculate the statistical mutation rate within the cluster.

[0034] Step S414: Merge abnormal scores;

[0035] Step S42: Anomaly Alarm, based on historical normal training data The criteria set an abnormal threshold; if the abnormal scores of three consecutive windows exceed the threshold, an abnormal alarm is triggered.

[0036] The training physical fitness status abnormality monitoring system based on clustering processing provided by the present invention includes a data acquisition module, a clustering dataset construction module, a clustering processing module, and an anomaly determination module;

[0037] The data acquisition module collects heart rate sequences, triaxial acceleration, respiratory rate, skin conductance signals, and skin temperature of historical trainees, and sends the data to the clustering dataset construction module.

[0038] The clustering dataset construction module receives data sent by the data acquisition module, extracts statistical features from the original signal for each time window and constructs a window feature vector to form a sample set for clustering, and sends the data to the clustering processing module and the anomaly detection module.

[0039] The clustering processing module receives data sent by the clustering dataset construction module. First, it maps window features to low-dimensional encoding and evaluates the local similarity and local linear reconstruction error of each sample point at multiple neighborhood scales. Points that simultaneously satisfy high density and low residual are used as clustering seeds. Then, starting from these seeds, an incremental Gaussian mixture model is used to maintain the statistical view of each cluster. Merging strategies are designed based on information criteria, and splitting strategies are designed based on multimodal and time series transition characteristics within clusters. At the same time, soft assignment is used for boundary points, introducing time stability penalties, giving a confidence estimate of the cluster boundaries, and sending the data to the anomaly detection module.

[0040] The anomaly detection module receives data sent by the clustering dataset construction module and the clustering processing module. After clustering and assigning new windows of the monitored personnel, it constructs three independent anomaly measures: spatial anomaly degree relative to the optimal cluster center, anomaly degree of short-term cluster transition probability, and statistical mutation degree within the cluster. These three items are then combined into a final anomaly score according to weights. The score is scaled using clustering confidence and detected based on a threshold obtained from historical normal data statistics. When the anomaly score of a window exceeds the threshold, an anomaly alarm is triggered, indicating that the training physical condition is abnormal.

[0041] The beneficial effects achieved by the present invention using the above solution are as follows:

[0042] (1) To address the problems of traditional clustering methods being susceptible to initialization, scale selection and noise in high-dimensional physiological or motion data, having difficulty capturing local low-dimensional structures and temporal dynamics, and usually being unable to update online and quantify boundary uncertainties, this scheme first performs low-dimensional embedding on window features, and simultaneously evaluates local density and local linear reconstruction residuals at multiple neighborhood scales to robustly select clustering seeds; uses an incremental Gaussian mixture model to achieve online updates and uses information criteria and splitting and merging strategies based on transfer entropy to dynamically maintain the clustering structure; at the same time, it adopts soft allocation and introduces time consistency penalty and clustering boundary confidence estimation to improve the robustness, adaptability and temporal continuity modeling ability of clustering, and can provide a quantitative explanation of boundary uncertainties.

[0043] (2) In view of the problems that traditional anomaly detection often relies on a single spatial distance or threshold, is prone to false alarms or false alarms, ignores short-term cluster migration and statistical mutation within the group, and lacks confidence-based scaling and time continuity judgment, this scheme constructs three complementary anomaly measures at the same time, integrates them by weight and uses cluster confidence to scale the score, and combines the statistical threshold of historical normal data and continuous window triggering rules. It takes into account both spatial anomalies and changes in temporal behavior, and significantly reduces the false alarm rate and improves the reliability and interpretability of anomaly detection through multi-indicator mutual verification and continuity requirements. Attached Figure Description

[0044] Figure 1 A schematic diagram of the training physical condition abnormality monitoring method based on clustering processing provided by the present invention;

[0045] Figure 2 A schematic diagram of the training physical fitness abnormality monitoring system based on clustering processing provided by the present invention;

[0046] Figure 3 This is a schematic diagram of step S2;

[0047] Figure 4 This is a schematic diagram of step S3;

[0048] Figure 5 This is a schematic diagram of step S4.

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0051] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0052] Example 1, see Figure 1 The present invention provides a method for monitoring abnormal training physical condition based on clustering processing, which includes the following steps:

[0053] Step S1: Data acquisition, collecting heart rate sequence, triaxial acceleration, respiratory rate, skin conductance signal, and skin temperature of historical trainees;

[0054] Step S2: Construct a clustering dataset. For each time window, extract statistical features from the original signal and construct a window feature vector to form a sample set for clustering.

[0055] Step S3: Clustering. First, the window features are mapped to low-dimensional encoding, and the local similarity and local linear reconstruction error of each sample point are evaluated at multiple neighborhood scales. Points that simultaneously satisfy high density and low residual are used as cluster seeds. Then, starting from these seeds, an incremental Gaussian mixture model is used to maintain the statistical view of each cluster. A merging strategy is designed based on information criteria, and a splitting strategy is designed based on the multimodal and time series transition characteristics within the cluster. At the same time, soft assignment is used for boundary points, introducing a time stability penalty, and giving a confidence estimate of the cluster boundary.

[0056] Step S4: Anomaly detection. After clustering and assigning the new windows of the monitored personnel, three independent anomaly measures are constructed: spatial anomaly degree relative to the optimal cluster center, anomaly degree of short-term cluster transition probability, and statistical mutation degree within the cluster. These three items are then combined into a final anomaly score according to their weights. The score is scaled using the cluster confidence score and detected based on a threshold obtained from historical normal data statistics. When the anomaly score of a window exceeds the threshold, an anomaly alarm is triggered, indicating that the training physical condition is abnormal.

[0057] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the data acquisition involves collecting the heart rate sequences of historical trainees. Triaxial acceleration respiratory rate Electrodermal signals Skin temperature ; Click the sliding window With step size Segmentation, denoted as the nth window. Each window generates an observation sample vector: , Indicates the first The original signal vector of each window, Indicates the start time of the window.

[0058] Example 3, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S2, the construction of the clustering dataset specifically includes the following steps:

[0059] Step S21: Define the feature window and set the sliding window length to... seconds, with a value range of Step size is seconds, with a value of For each window index Constructing window feature vectors ;

[0060] Step S22: Calculate the statistical features of the original domain, and the feature vector. This includes the original domain statistical features; the original domain statistical features are for any time series. The mean and standard deviation are calculated on the window, as shown below:

[0061] ;

[0062] in, and Representing separate time series In the The mean and standard deviation signals over a sliding window, This indicates the start time of the nth sliding window;

[0063] Step S23: Construct a window vector. Based on the original domain statistical features, construct a window vector as follows:

[0064] ;

[0065] in, Indicates the first Window vectors on a sliding window, and Representing heart rate sequences In the The mean and standard deviation signals over a sliding window, This represents the original signal of the heart rate sequence at the start of the nth sliding window. The window vector is formed by concatenating these three signals from all time series, with a total dimension of [missing value]. .

[0066] Example 4, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S3, the clustering process specifically includes the following steps:

[0067] Step S31: Select a clustering seed for each window vector. First, it is mapped to low-dimensional encoding. The method involves constructing local similarity maps at multiple scales and simultaneously calculating two complementary metrics: local density peak score and projection residual from a point to a local area. Points that simultaneously satisfy high density and low residual are set as clustering seeds. The specific steps include:

[0068] Step S311: Window vector embedding, which involves embedding each window vector through a linear mapping. Mapping to low-dimensional encoding , means as follows:

[0069] ;

[0070] in, Indicates the first A low-dimensional representation of a window, with dimension . , Representation from low dimension The window vector obtained from decoding; , , and This represents the trainable mapping parameters, which are optimized through unsupervised training. The training objective is to minimize the reconstruction error. ,use Optimizer, learning rate set to The maximum number of iterations is Round, when the reduction in reconstruction error is less than 20 consecutive rounds Training is stopped when the time is right; the training data is a window vector set of at least 100 historical normal trainees, with each person's data collection time exceeding 60 minutes;

[0071] Step S312: Construct a multi-scale neighborhood set and define the scale set: ;use The nearest neighbor algorithm constructs a neighbor set for each scale k. ;in, Represents a predefined set of integers representing neighborhood sizes. Representing space and The set of indices of the k nearest points;

[0072] Step S313: Local density scoring, defining local density scoring for scale k: ;in, express The square of the norm, This represents the local density score at scale k. This represents the kernel bandwidth at scale k, specifically the value of the k-th kernel. At each scale, it is half the median of the pairwise Euclidean distance between the low-dimensional codes of all historically trained individuals.

[0073] Step S314: Calculate the projection residual in the neighborhood. Weighted local principal component analysis is performed on the above to calculate the r-dimensional principal component subspace. ,in Selection based on the cumulative explained variance ratio: choose the smallest one. Make the former The proportion of the sum of variances explained by each principal component to the total variance of the neighborhood reaches a threshold. The residual is defined as: ;in, Indicates in the neighborhood The local principal component basis calculated above has the following column number: ; This represents the nth sliding window. Projected residuals at various scales Indicates the transpose symbol;

[0074] Step S315: Cluster seed screening, synthesizing scores for all scales, as shown below: ;in, This represents the scale composition weights, with values ​​of 0.3, 0.4, and 0.3 for the three scales, respectively. , This represents the residual penalty coefficient, with a value range of [value missing]. The initial value is set to 3; This represents the seed score for synthetic clustering, with a fixed threshold set. The value is taken from all historical normal samples. of Quantiles; Selection Above a fixed threshold The points are used as initial clustering seeds;

[0075] Step S32: Cluster generation. Starting with the cluster seed, an incremental Gaussian mixture model is used to maintain the statistical view of each cluster: mean, covariance matrix, and sample count. Merging decisions are designed based on information gain according to the Bayesian information criterion, and splitting decisions are designed based on multimodal attitude quantities and transfer entropy within the cluster. Specifically, this includes the following steps:

[0076] Step S321: Cluster parameter statistics, maintain statistics for each cluster c: number of samples mean Covariance matrix When new samples When assigned to c, updates are performed incrementally, as shown below:

[0077] ;

[0078] in, , and Let represent the number of samples, mean, and cumulative second moment term of cluster c after the update, respectively. This represents updating the second-order moment cumulative term of the previous cluster c, and updating the covariance matrix: ;

[0079] Step S322: Merge decision. Merge decision is made based on log-likelihood value and model complexity. There are two clusters. , Fit their respective Gaussian models , ; Clustering after merging From the combined statistics , and Define the consolidated revenue as follows:

[0080] ;

[0081] in, Clustering The log-likelihood function value, the log-likelihood function value after cluster merging. according to and calculate: ; This represents the difference in the number of free parameters before and after the merge. Low-dimensional representation Dimensions Indicates complexity weights. Clustering and The combined gains, if Merging is then permitted. This indicates taking the determinant. and Representing clustering and The log-likelihood function value, Represents the natural logarithm function. Clustering The number of samples, Clustering The number of samples, Clustering and The covariance matrix of the combined clusters;

[0082] Step S323: Split determination, for clustering Record a set of time series indices within the time window series. Perform local operations within cluster c. Clustering to obtain candidate sub-clusters , The splitting payoff is calculated and expressed as follows:

[0083] ;

[0084] in, Indicates the benefits of splitting. This represents the time series transfer entropy from samples within a cluster to samples in other clusters. Clustering The transfer entropy, Indicates based on the most recent The clustering assignment results for a continuous sample window represent the statistical distribution of clusters on a time-continuous sample sequence. To clustering The empirical transition probability, The range of values ​​is ,default ; Represents statistical clustering To clustering Number of transfers , , Represents a clustering index. Indicates pseudo-counting; and Representing clustering and The log-likelihood function value;

[0085] Step S33: Calculate the cluster boundary confidence score, apply soft assignment to the boundary points and introduce a time stability penalty; simultaneously, model the uncertainty of the number of clusters using an approximate posterior, and give the cluster boundary confidence interval, specifically including the following steps:

[0086] Step S331: Soft assignment, calculating unnormalized similarity weights based on Gaussian negative logarithmic distance, while applying a time stability penalty; for the current cluster set Define the sample Assigned to cluster Unnormalized weights: Define time consistency items: Finally, perform normalized soft allocation: ;in, This represents the currently existing cluster set. Indicates sample Assigned to cluster Unnormalized weights Indicates sample Assigned to cluster Time consistency items, Represents cluster set The index of any cluster in the dataset. Indicates sample Assigned to cluster Unnormalized weights Indicates sample Assigned to cluster Time consistency items, Display window The normalized soft-assigned vector, This indicates the clustering relationship between the current window and the previous window. The smaller the difference in allocation weights, the better. The closer to 1, the smaller the penalty; This represents the time stability penalty coefficient. Indicates taking Norm;

[0087] Step S332: Calculate the cluster boundary confidence score for the clusters. Define activity level: Cluster boundary confidence: The maximum normalized similarity weight for sample points. Corresponding cluster c and Then the sample points are marked as highly uncertain boundary points; where, Indicates fixed Prior constants, Clustering activity level Clustering activity level This represents a confidence index for cluster boundaries. This represents the confidence threshold.

[0088] By performing the above operations, this scheme addresses the problems of traditional clustering methods being susceptible to initialization, scale selection, and noise in high-dimensional physiological or motion data, struggling to capture local low-dimensional structures and temporal dynamics, and typically failing to update online and quantify boundary uncertainties. It robustly selects clustering seeds by first embedding window features in a low dimension and simultaneously evaluating local density and local linear reconstruction residuals at multiple neighborhood scales. An incremental Gaussian mixture model is used for online updates, and information criteria and a splitting and merging strategy based on transfer entropy are used to dynamically maintain the clustering structure. Furthermore, soft allocation is employed, and time consistency penalties and cluster boundary confidence estimation are introduced, thereby improving the robustness, adaptability, and temporal continuity modeling capabilities of the clustering, and providing a quantitative explanation of boundary uncertainties.

[0089] Example 5, see Figure 1 and Figure 5 This embodiment is based on the above embodiment. In step S4, the anomaly determination specifically includes the following steps:

[0090] Step S41: Based on the comprehensive anomaly score, collect training data from personnel requiring monitoring in real time. Generate a data vector using the data vector generated in step S2, and record this new data to be monitored as a new window. The data vector for each new window Three independent anomaly measures are constructed based on the clustering assignment results and inter-cluster transition statistics: Mahalanobis distance anomaly to the optimal cluster; short-term cluster transition probability anomaly; and intra-cluster statistical mutation anomaly. The final anomaly score is then obtained and compared with the cluster confidence score. Scaling fractions involves the following steps:

[0091] Step S411: Calculate spatial anomalies and select the optimal cluster. Calculate spatial anomaly: ;in, This represents the optimal cluster center for the current window. Indicates a new window Data vectors and clustering Normalized similarity weights, Indicates a new window Spatial anomaly of the data vector, Clustering The inverse covariance matrix, Clustering The mean, Indicates a new window The low-dimensional representation;

[0092] Step S412: Calculate the transition probability anomaly degree and define the empirical transition probability matrix. The row element To cluster By applying empirical probabilities to other clusters and defining the elements of each row sequentially, we obtain the transition probability matrix. The previous window mainly clustered as Calculate the transition probability anomaly degree: ;in, Indicates a new window The anomaly degree of the transition probability of the data vector. Represents the empirical transition probability matrix The row element, Indicates from clustering To clustering The empirical transition probability;

[0093] Step S413: Calculate the statistical mutation rate within clusters using the current window vector. The most recently assigned cluster in the time series of Low-dimensional encoding of consecutive windows Compare and calculate the statistical mutation rate within each cluster, expressed as: ;in, Indicates a new window The instantaneous statistical abrupt change of the data vector;

[0094] Step S414: Fuse the anomaly score, as shown below:

[0095] ;

[0096] in, This represents the final fusion anomaly score. , , Indicates the weight of abnormal fusion. Represents optimal clustering Cluster boundary confidence;

[0097] Step S42: Anomaly alarm, set anomaly threshold , Based on historical normal training data The criteria are established. and These are the mean and standard deviation of outlier scores for the normal data; if the outlier scores are from three consecutive windows... Exceeding the threshold Issue an anomaly alert in a new window. The training performance of the data vectors is abnormal.

[0098] By performing the above operations, this solution addresses the problems of traditional anomaly detection, which often relies on a single spatial distance or threshold, is prone to false alarms or false negatives, ignores short-term cluster migration and intra-group statistical mutations, and lacks confidence-based scaling and temporal continuity determination. This solution constructs three complementary anomaly measures simultaneously, merges them by weight and uses cluster confidence to scale the scores, and combines statistical thresholds of historical normal data with continuous window triggering rules. This approach takes into account both spatial anomalies and temporal behavior changes, and significantly reduces the false alarm rate and improves the reliability and interpretability of anomaly detection through multi-indicator mutual verification and continuity requirements.

[0099] Example 6, see Figure 1 and Figure 2 Based on the above embodiments, the training physical condition abnormality monitoring system based on clustering processing provided by the present invention includes a data acquisition module, a clustering dataset construction module, a clustering processing module, and an anomaly determination module.

[0100] The data acquisition module collects heart rate sequences, triaxial acceleration, respiratory rate, skin conductance signals, and skin temperature of historical trainees, and sends the data to the clustering dataset construction module.

[0101] The clustering dataset construction module receives data sent by the data acquisition module, extracts statistical features from the original signal for each time window and constructs a window feature vector to form a sample set for clustering, and sends the data to the clustering processing module and the anomaly detection module.

[0102] The clustering processing module receives data sent by the clustering dataset construction module. First, it maps window features to low-dimensional encoding and evaluates the local similarity and local linear reconstruction error of each sample point at multiple neighborhood scales. Points that simultaneously satisfy high density and low residual are used as clustering seeds. Then, starting from these seeds, an incremental Gaussian mixture model is used to maintain the statistical view of each cluster. Merging strategies are designed based on information criteria, and splitting strategies are designed based on multimodal and time series transition characteristics within clusters. At the same time, soft assignment is used for boundary points, introducing time stability penalties, giving a confidence estimate of the cluster boundaries, and sending the data to the anomaly detection module.

[0103] The anomaly detection module receives data sent by the clustering dataset construction module and the clustering processing module. After clustering and assigning new windows of the monitored personnel, it constructs three independent anomaly measures: spatial anomaly degree relative to the optimal cluster center, anomaly degree of short-term cluster transition probability, and statistical mutation degree within the cluster. These three items are then combined into a final anomaly score according to weights. The score is scaled using clustering confidence and detected based on a threshold obtained from historical normal data statistics. When the anomaly score of a window exceeds the threshold, an anomaly alarm is triggered, indicating that the training physical condition is abnormal.

[0104] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0106] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for monitoring abnormal training physical condition based on clustering processing, characterized in that, The method includes the following steps: Step S1: Data acquisition, collecting heart rate sequence, triaxial acceleration, respiratory rate, skin conductance signal, and skin temperature of historical trainees; Step S2: Construct a clustering dataset. For each time window, extract statistical features from the original signal and construct a window feature vector to form a sample set for clustering. Step S3: Clustering. First, the window features are mapped to low-dimensional encoding, and the local similarity and local linear reconstruction error of each sample point are evaluated at multiple neighborhood scales. Points that simultaneously satisfy high density and low residual are used as cluster seeds. Then, starting from these seeds, an incremental Gaussian mixture model is used to maintain the statistical view of each cluster. A merging strategy is designed based on information criteria, and a splitting strategy is designed based on the multimodal and time series transition characteristics within the cluster. At the same time, soft assignment is used for boundary points, introducing a time stability penalty, and giving a confidence estimate of the cluster boundary. Step S4: Anomaly detection. After clustering and assigning the new windows of the monitored personnel, three independent anomaly measures are constructed: spatial anomaly degree relative to the optimal cluster center, anomaly degree of short-term cluster transition probability, and statistical mutation degree within the cluster. These three items are then combined into a final anomaly score according to their weights. The score is scaled using the cluster confidence score and detected based on a threshold obtained from historical normal data statistics. When the anomaly score of a window exceeds the threshold, an anomaly alarm is triggered, indicating that the training physical condition is abnormal.

2. The training physical condition abnormality monitoring method based on clustering processing according to claim 1, characterized in that: In step S2, constructing the clustering dataset specifically includes the following steps: Step S21: Define the feature window, set the sliding window length and step size, and construct a window feature vector for each window; Step S22: Calculate the original domain statistical features. The feature vector includes the original domain statistical features. The original domain statistical features are the mean and standard deviation of any time series calculated over a window. Step S23: Construct a window vector based on the original domain statistical features.

3. The training physical condition abnormality monitoring method based on clustering processing according to claim 1, characterized in that: In step S3, the clustering process specifically includes the following steps: Step S31: Select clustering seeds. For each window vector, first map it to a low-dimensional encoding, construct a local similarity map at multiple scales, and simultaneously calculate two complementary indices: local density peak score and projection residual from the point to the local area. Set the points that simultaneously satisfy high density and low residual as clustering seeds. Specifically, this includes the following steps: Step S311: Window vector embedding, mapping each window vector to a low-dimensional code through linear mapping; Step S312: Construct a multi-scale neighborhood set and define the scale set; utilize The nearest neighbor algorithm constructs a set of neighbors for each scale; Step S313: Local density scoring, define local density scores for sample points at each scale; Step S314: Calculate the projection residuals, perform weighted local principal component analysis in the neighborhood, and calculate the r-dimensional principal component subspace; Step S315: Cluster seed selection, synthesize scores for all scales, set a fixed threshold; select points with scores higher than the fixed threshold as initial cluster seeds; Step S32: Cluster generation. Starting with the cluster seed, an incremental Gaussian mixture model is used to maintain the statistical view of each cluster: mean, covariance matrix, and sample count. Merging decisions are designed based on information gain according to the Bayesian information criterion, and splitting decisions are designed based on multimodal attitude quantities and transfer entropy within the cluster. Specifically, this includes the following steps: Step S321: Cluster parameter statistics, maintain statistics for each cluster: number of samples, mean, covariance matrix; Step S322: Merge decision, based on log-likelihood value and model complexity; Step S323: Split determination, record the time index set of each cluster in the time window sequence; perform local splitting within the cluster. Clustering is performed to obtain candidate sub-clusters, and the splitting bonus is calculated. Step S33: Calculate the cluster boundary confidence score, apply soft assignment to the boundary points and introduce a time stability penalty; simultaneously, model the uncertainty of the number of clusters using an approximate posterior, and give the cluster boundary confidence interval, specifically including the following steps: Step S331: Soft assignment, calculate unnormalized similarity weights based on Gaussian negative logarithmic distance, and apply time stability penalty at the same time; Step S332: Calculate the cluster boundary confidence score, define the activity score for each cluster, and calculate the cluster boundary confidence score.

4. The training physical condition abnormality monitoring method based on clustering processing according to claim 1, characterized in that: In step S4, the anomaly determination specifically includes the following steps: Step S41: Integrate anomaly scoring. Real-time collection of training data from personnel requiring monitoring. A data vector is generated using step S2. This new data to be monitored is recorded as the data vector for a new window. For each new window, three independent anomaly metrics are constructed based on the clustering assignment results and inter-cluster transition statistics: Mahalanobis distance anomaly to the optimal cluster; short-term cluster transition probability anomaly; and intra-cluster statistical mutation anomaly. The final anomaly score is obtained, and the score is scaled using cluster confidence. The specific steps include: Step S411: Calculate spatial anomaly, take the cluster center of the best cluster, and calculate the spatial anomaly. Step S412: Calculate the transition probability anomaly degree, define the empirical transition probability matrix, and calculate the transition probability anomaly degree; Step S413: Calculate the statistical mutation rate within the cluster. Compare the current sample window vector with the mean of the samples most recently assigned to the optimal cluster, and calculate the statistical mutation rate within the cluster. Step S414: Merge abnormal scores; Step S42: Anomaly Alarm, based on historical normal training data The criteria set an abnormal threshold; if the abnormal scores of three consecutive windows exceed the threshold, an abnormal alarm is triggered.

5. The training physical condition abnormality monitoring method based on clustering processing according to claim 1, characterized in that: In step S1, the data acquisition involves collecting the heart rate sequence, triaxial acceleration, respiratory rate, skin conductance signal, and skin temperature of historical trainees; the data is segmented according to the sliding window and step size, and each window generates an observation sample vector.

6. A training physical fitness status anomaly monitoring system based on clustering processing, used to implement the training physical fitness status anomaly monitoring method based on clustering processing as described in any one of claims 1-5, characterized in that: It includes a data acquisition module, a clustering dataset construction module, a clustering processing module, and an anomaly detection module.

7. The training physical fitness status abnormality monitoring system based on clustering processing according to claim 6, characterized in that: The data acquisition module collects heart rate sequences, triaxial acceleration, respiratory rate, skin conductance signals, and skin temperature of historical trainees, and sends the data to the clustering dataset construction module. The clustering dataset construction module receives data sent by the data acquisition module, extracts statistical features from the original signal for each time window and constructs a window feature vector to form a sample set for clustering, and sends the data to the clustering processing module and the anomaly detection module. The clustering processing module receives data sent by the clustering dataset construction module. First, it maps window features to low-dimensional encoding and evaluates the local similarity and local linear reconstruction error of each sample point at multiple neighborhood scales. Points that simultaneously satisfy high density and low residual are used as clustering seeds. Then, starting from these seeds, an incremental Gaussian mixture model is used to maintain the statistical view of each cluster. Merging strategies are designed based on information criteria, and splitting strategies are designed based on multimodal and time series transition characteristics within clusters. Simultaneously, soft assignment is applied to boundary points, a time stability penalty is introduced, a confidence estimate of the cluster boundary is given, and the data is sent to the anomaly detection module; The anomaly detection module receives data sent by the clustering dataset construction module and the clustering processing module. After clustering and assigning new windows of the monitored personnel, it constructs three independent anomaly measures: spatial anomaly degree relative to the optimal cluster center, anomaly degree of short-term cluster transition probability, and statistical mutation degree within the cluster. These three items are then combined into a final anomaly score according to weights. The score is scaled using clustering confidence and detected based on a threshold obtained from historical normal data statistics. When the anomaly score of a window exceeds the threshold, an anomaly alarm is triggered, indicating that the training physical condition is abnormal.

Images