Data anomaly detection system and method based on wearable device

Abstract

The invention relates to the technical field of abnormal data detection, in particular to a data anomaly detection system and method based on a wearable device, and the method comprises the steps: obtaining multi-modal data of a plurality of sensors, and building time series data; a sliding window is adopted to extract window features of the time series data, short-period fluctuation is extracted according to the window features, seasonal components and trend components of the time series data are extracted, a long-period rhythm is obtained, and the short-period fluctuation and the long-period rhythm are spliced into multi-granularity features; establishing an individual health baseline and predicting an index trend; constructing a dynamic association graph according to the multi-modal data, and calculating an association weight and an association causality between nodes through an attention mechanism and a transfer entropy edge weight; and performing convolution operation on the dynamic association graph by using a graph convolution network, and positioning an abnormal node and an abnormal propagation path after neighborhood information is aggregated. The local physiological fluctuation and global state evolution trend can be accurately analyzed, and the sensitivity and robustness of anomaly detection in health monitoring are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of abnormal data detection, and particularly relates to a data abnormality detection system and method based on a wearable device. BACKGROUND

[0002] With the continuous progress of technology, smart wearable devices have gradually become popular in the field of health monitoring. These devices can monitor users' physiological parameters in real time by integrating various sensors, and transmit data to smartphones or cloud platforms, thereby providing health management services. The widespread use of smart wearable devices enables users to more actively manage their own health conditions and detect potential health problems early.

[0003] In the past few years, wearable devices have mainly focused on simple health monitoring, such as step counting, heart rate monitoring, and sleep monitoring. With the continuous progress of technology, the functions of wearable devices have gradually developed towards more comprehensive health management, especially multi-modal health data collection and abnormality detection have become new research hotspots. These devices can not only collect single physiological parameters, but also integrate multiple physiological signals to more comprehensively reflect users' health conditions.

[0004] However, existing wearable devices, although they can collect health data, have some problems: Single data monitoring: Most wearable devices mainly focus on the monitoring of single physiological indicators, such as heart rate and step count, and cannot achieve comprehensive health monitoring. Health conditions are often a comprehensive reflection of multiple physiological parameters, and single data sources cannot provide comprehensive health assessment.

[0005] Data accuracy and reliability problems: Due to the limitations of sensors and the imperfections of algorithms, the data collected by existing wearable devices often has noise or errors, especially in some special situations, which may lead to data abnormalities or inaccurate measurement results.

[0006] Insufficient data abnormality detection: Traditional health monitoring systems generally only focus on data collection and display, and lack intelligent detection and analysis of data abnormalities. Once the user's health data is abnormal, it is often difficult to detect and intervene in time, which may delay the best medical opportunity.

[0007] Difficulty in integrating multi-modal data: Modern wearable devices can collect multiple physiological indicator data, but different types of data often have different collection methods, frequencies, and formats. How to effectively integrate these data and analyze and warn through intelligent algorithms is a challenge.

[0008] Insufficient personalization of user health management: most existing wearable devices are general user-oriented health data collection tools, lacking personalized health recommendations and intervention programs, and it is also difficult to adjust the monitoring strategy according to the specific needs and health status of the user. SUMMARY

[0009] Embodiments of the present application provide a data anomaly detection method based on a wearable device, which solves the problems of difficulty in multi-modal data integration and insufficient data anomaly detection.

[0010] In another aspect, embodiments of the present application provide a data anomaly detection system based on a wearable device.

[0011] The present application is implemented as follows: The present application provides a data anomaly detection method based on a wearable device, the method comprising: S1 acquiring multi-modal data of a plurality of sensors, the multi-modal data comprising indicators of heart rate, electromyographic signal, body temperature and gait information at different time points, and establishing time series data; S2 extracting window features of the time series data using a sliding window, extracting short-period fluctuations, seasonal components and trend components of the time series data according to the window features, obtaining long-period rhythms, and concatenating the short-period fluctuations and the long-period rhythms into multi-granularity features; S3 establishing an individual health baseline and predicting an indicator trend according to the multi-granularity features; S4 constructing a dynamic association graph according to the multi-modal data, each node in the dynamic association graph being a multi-granularity feature of a time series slice, and using attention mechanisms and transfer entropy edge weights as the association weights and association causes between nodes; S5 performing convolution operations on the dynamic association graph using a graph convolution network, and positioning abnormal nodes and abnormal propagation paths after aggregating neighborhood information.

[0012] Further, the individual health baseline, the indicator trend, the abnormal nodes and the abnormal propagation paths are further generated into spatio-temporal unified representation data through a gated attention mechanism, and a hierarchical early warning is triggered according to the spatio-temporal unified representation data.

[0013] Further, in S3, an LSTM network is used to train the time series data of at least 30 days in the past, and a dilated convolution is used to expand the time receptive field to 48 hours to capture long-term health baseline and identify chronic drift or acute fluctuations.

[0014] Further, the abnormal amplitude and frequency in the spatio-temporal unified representation data are extracted and used as threat data; The abnormal nodes and the abnormal propagation paths in the spatio-temporal unified representation data are extracted and used as vulnerability data; The threat data and the vulnerability data are combined to form a unified health risk score: , wherein, is a threat index score, is a vulnerability index score, is a comprehensive risk assessment result, is a weight.

[0015] Further, according to the spatio-temporal unified representation data, a hierarchical early warning is triggered, including: a prediction residual of the index trend is calculated by dynamic time warping, and when the prediction residual exceeds 90% of the historical quantile, a wearable device is started to issue a vibration reminder; a short-period fluctuation that is continuously sustained for more than three sliding windows is taken as an abnormal signal, a causal report with a confidence of >80% is generated, and an SMS push is performed.

[0016] Further, a residual between a current observation value and a predicted value of an abnormal node is calculated as an abnormal score, and an alarm is triggered according to the level of the abnormal score.

[0017] In another aspect of the present application, a data anomaly detection system based on a wearable device, the system comprising: an acquisition module configured to acquire multi-modal data of a plurality of sensors, the multi-modal data comprising indexes of heart rate, electromyographic signal, body temperature and gait information at different time points, and to establish time series data; a feature extraction module configured to extract window features of the time series data using a sliding window, to extract short-period fluctuations according to the window features, to extract seasonal components and trend components of the time series data, and to obtain long-period rhythms, and to splice the short-period fluctuations and the long-period rhythms into multi-granularity features; a trend prediction module configured to establish an individual health baseline and predict an index trend according to the multi-granularity features; a dynamic correlation graph construction module configured to construct a dynamic correlation graph according to the multi-modal data, each node in the dynamic correlation graph being a multi-granularity feature of a time series slice, and the correlation weight and the correlation cause between nodes being calculated by an attention mechanism and a transfer entropy edge weight; an anomaly detection module configured to perform convolution operation on the dynamic correlation graph by a graph convolution network, to locate abnormal nodes and abnormal propagation paths after aggregating neighborhood information.

[0018] Further, it further comprises an early warning module configured to generate spatio-temporal unified representation data by a gated attention mechanism according to the individual health baseline, the index trend, the abnormal nodes and the abnormal propagation paths, and to trigger a hierarchical early warning according to the spatio-temporal unified representation data.

[0019] Compared with the prior art, the present application has the beneficial effects of: Compared with traditional health monitoring devices, the system of the application is a non-invasive health monitoring method, which can model the complex spatial dependence and correlation between nodes, accurately analyze local physiological fluctuations and global state evolution trend through the introduction of spatiotemporal feature interaction modeling mechanism, and improve the sensitivity and robustness of abnormal detection in health monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A structural schematic diagram of a wearable device provided for an embodiment of the application is shown in FIG. 1. Figure 2 A flowchart of a data anomaly detection method based on a wearable device provided for an embodiment of the application is shown in FIG. 2. Figure 3 A software framework diagram of a data anomaly detection system based on a wearable device provided for an embodiment of the application is shown in FIG. 3. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the application clearer and more apparent, the application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0022] One of the cores of the wearable device based on the embodiments of the application is to collect multi-dimensional physiological data of the wearer in real time through various sensors. These sensors include optical sensors, gas sensors, pressure sensors, etc., and each sensor is responsible for monitoring different physiological indicators.

[0023] In order to respond to the health needs of the wearer in a timely manner, the wearable device will provide real-time feedback and alarm mechanisms. The health feedback function analyzes real-time data, and the system will provide health advice or intervention measures to the user according to the evaluation results. When the system detects abnormal data (such as rapid heart rate, low blood oxygen, abnormal body temperature, etc.), the wearable device will remind the wearer in time through vibration, voice or push notification. These feedbacks not only provide health warnings, but also notify family members through emergency contacts, or directly contact medical service platforms for remote intervention, thereby providing comprehensive protection for the health of the wearer.

[0024] The analysis principle of the wearable device based on the embodiments of the application is as follows: Physiological data such as heart rate, blood pressure, blood oxygen, body temperature, and respiratory rate are monitored in real time by integrated sensors. The data collected by each sensor is preliminarily preprocessed (such as filtering, denoising, etc.) by the real-time acquisition module to ensure accuracy. Acceleration / gyroscope sensors are used to monitor the activity state of the wearer, helping to judge the intensity of exercise, posture or accidental falling, etc.

[0025] Data is transmitted in real time to the local processing unit through wireless communication technology or directly uploaded to the cloud platform for further processing. The cloud platform aggregates multi-device data and performs deep analysis and fusion through a centralized computing platform. After introducing edge computing functions, some preliminary data processing is completed on the device side, reducing delay and improving real-time performance.

[0026] Since the health data of the elderly is usually time-series data, that is, each physiological indicator changes over time. The regularity and dynamics of time-series data are important characteristics of health monitoring. Through time-series data analysis and graph neural network (GNN), the time correlation of the data can be mined by using an LSTM network model, the potential pattern of health changes of the elderly can be identified, and abnormal conditions can be discovered in a timely manner.

[0027] A hybrid model using LSTM network and graph neural network (GNN) realizes deep analysis of health data through "time-series-graph dual-channel cooperation", including: real-time collection of physiological signals by multi-modal sensors, alignment of time stamps by neural differential equation and extraction of multi-granularity features (sliding window captures short-period fluctuations, STL decomposes long-period rhythms); LSTM network uses dilated convolution to expand the time horizon, models individual health baseline and predicts index trend, and calculates prediction residual by dynamic time warping (DTW); the graph neural network constructs a dynamic correlation graph, the nodes are multi-dimensional index data of time-series slices, and the edge weights are screened for causal association (such as "blood sugar → heart rate") through attention mechanism and transfer entropy, and the graph convolution is used to aggregate neighborhood information and locate abnormal propagation path (such as sleep deficiency causing abnormal blood pressure). After the dual-channel output is fused by the gated attention, the hierarchical early warning is triggered, and the causal chain is verified by counterfactual reasoning to improve the explainability. The present application is significantly superior to traditional methods in chronic disease prediction, acute event response and false positive rate, and realizes closed-loop health management from data perception to intelligent decision-making.

[0028] Referring to Figure 1The wearable device based on the present application shown adopts a flexible vest design, the base material adopts a three-layer composite structure, the outer layer is an antibacterial nylon blended fabric, the middle layer is embedded with a flexible circuit and a graphene heating film (temperature uniformity ±0.5℃), and the inner layer is a medical-grade silicone sensor fitting layer. The electromyography sensor 3, the heart rate sensor 1, the body temperature sensor 2 and the acceleration / gyroscope sensor are integrated in the inner layer. The electromyography sensor 3, the heart rate sensor 1 and the body temperature sensor 2 are connected to the data processing transmission module 5 through the data transmission line 4. The flexible vest is provided with a waistband 6, and flexible buffer cotton 7 is arranged on the flexible vest and fixed on both sides through a zipper 8. The electromyography sensor 3 (Electromyography Sensor) is used to collect weak electrical signals generated by muscle activity, monitor the muscle movement state of the user, and is commonly used in rehabilitation training, sports evaluation and other scenes. The heart rate sensor 1 (Heart Rate Sensor) monitors the heart rate data of the user in real time, supports the measurement of resting heart rate and dynamic heart rate, and is used for health management and sports safety evaluation. The body temperature sensor 2 (Body Temperature Sensor) is attached to the skin surface and senses the body temperature change of the user, which can be used for early warning of fever, monitoring of body temperature trend, etc. The data transmission line 4 is a connection channel between each sensor and the main control module, which is laid in the device using flexible conductive material to ensure signal stability and human comfort. The data processing transmission module 5 integrates Bluetooth, Wi-Fi or LoRa communication protocols, which is used to transmit the collected data to the mobile terminal or cloud platform in real time to realize remote health monitoring. The waistband 6 is used to firmly fix the sensor module on the waist of the human body, which adopts an elastic adjustable design to adapt to different body types, and ensures the stability and accuracy of data acquisition.

[0029] The embodiment of the present application provides a data anomaly detection method based on a wearable device. Figure 2 As shown in the figure, a data anomaly detection method based on a wearable device comprises the following steps. S1, acquiring multi-modal data of a plurality of sensors, the multi-modal data comprising indexes of heart rate, electromyography signal, body temperature and gait information at different time points, and establishing time series data; The plurality of sensors comprise an electromyography sensor 3, a heart rate sensor 1, a body temperature sensor 2 and an acceleration / gyroscope sensor. The original data has the problem of inconsistent sampling frequency, so a neural differential equation (Neural ODE) is introduced to uniformly align the non-uniform data and generate time series data streams with equal time intervals.

[0030] Suppose there are N nodes of wearable devices, each node collects a d-dimensional feature vector (such as heart rate, body temperature, gait, etc.) at time t, denoted as: .

[0031] A graph structure G=(V, E) is constructed, where V represents a set of device nodes, and E represents the relationship edges between channels. The edges can be defined according to geographical distance, physiological signal correlation, etc. to obtain an adjacency matrix A ∈ R N×N .

[0032] S2 extracts window features of the time series data using a sliding window, extracts short-period fluctuations based on the window features, extracts seasonal components and trend components of the time series data, and obtains long-period rhythms by splicing the short-period fluctuations and the long-period rhythms into multi-granularity features. For example, short-period fluctuation statistics and time-frequency features are extracted within a sliding window of every n minutes, and the STL (season-trend-residual) decomposition is used to strip off slow variable trends such as diurnal / weekly rhythms; thus, it is helpful to focus on short-period fluctuations or abnormalities. Long-period rhythms refer to periodic changes with a long duration in time series, such as seasonal changes, annual cycles, or regular fluctuations with a longer time span. Compared with short-period fluctuations (such as diurnal rhythms and weekly rhythms), long-period rhythms can better reflect the long-term trends and periodic characteristics of time series. They are used to identify long-term health trends, such as seasonal changes in heart rate.

[0033] S3 establishes an individual health baseline and predicts index trends based on multi-granularity features; LSTM paths train the individual time series of the past 30 days, and use dilation convolution (dilation=4) to expand the time receptive field to 48 hours to capture long-term health baselines such as resting heart rate and blood pressure fluctuations. The system compares the predicted values with the measured values based on dynamic time warping (DTW) to identify chronic drifts (such as a 0.5% weekly decrease in blood oxygen) or acute fluctuations.

[0034] S4 constructs a dynamic association graph based on multi-modal data, where each node in the dynamic association graph is a multi-granularity feature of a time series slice, and the association weight and causal relationship between nodes are calculated through an attention mechanism and transfer entropy edge weight; The construction of the dynamic association graph requires the definition of nodes and edges, as well as the calculation of the association weight and causal relationship between nodes. Each node represents a multi-granularity feature within a time window. The attention mechanism can dynamically calculate the association weight between nodes. Transfer entropy is a method for measuring the causal relationship between two time series. The attention weight and transfer entropy are combined to construct the adjacency matrix of the dynamic association graph, which needs to be dynamically adjusted according to the update of the time series data. For example, the graph structure is updated every 10 minutes.

[0035] S5 uses a graph convolution network to perform convolution operations on the dynamic association graph, aggregates neighborhood information, and locates abnormal nodes and abnormal propagation paths.

[0036] In an embodiment, the abnormal amplitude and frequency in the unified spatio-temporal representation data are extracted and used as threat data. Extract anomalous nodes and anomalous propagation paths from spatiotemporal unified representation data for use as vulnerability data; Combine threat data with vulnerability data to form a unified health risk score: , in, The threat index score, Score the vulnerability index. Based on the comprehensive risk assessment results, As weight.

[0037] Threat assessment extracts "threat data" based on the magnitude and frequency of abnormalities, such as persistent abnormal heart rate fluctuations; Vulnerability assessment focuses on the combined judgment of anomaly sources (such as inter-node correlations) and individual state context, utilizing dynamic causal graph reasoning (GNN path) to construct a spatiotemporal graph: The nodes are time slices of each indicator, and the edge weights are learned through the graph attention mechanism (GAT). Significant causal paths (p<0.01) were screened using transfer entropy, and then abnormal propagation paths (such as "insufficient sleep → increased heart rate → increased blood pressure") were visualized using GNNExplainer. By combining counterfactual reasoning methods, we can verify whether controlling the root cause variable can suppress overall anomalies.

[0038] A unified health risk score is formed by comprehensively assessing "threat" and "vulnerability".

[0039] In one embodiment, triggering a tiered early warning based on spatiotemporal unified characterization data includes: The predictive residuals of the indicator trend are calculated by dynamic time warping. When the predictive residuals exceed the historical 90th percentile, the wearable device is activated to issue a vibration reminder. Short-cycle fluctuations that last for more than three consecutive sliding windows are considered abnormal signals. A causal report with a confidence level of >80% is generated and sent via SMS.

[0040] It may also include: multi-indicator coordinated anomaly triggering device buzzers, cloud alarms, and connection to hospital emergency systems via the HL7 protocol. The memory LSTM-GNN anomaly warning module achieves accurate and interpretable health risk warnings by deeply integrating temporal memory and graph structure reasoning capabilities.

[0041] Anomaly detection and early warning utilizes the output of an LSTM network to predict the feature values ​​at the next time step. .

[0042] Calculate the residual between the current observation and the predicted value as the anomaly score: , Set a threshold θ, if ϵ t If the value is greater than θ, it is considered an abnormal state, triggering an alert.

[0043] The method in this application uses an LSTM network to analyze minute-level fluctuations (changes in exercise heart rate) and a GNN network to reveal daily / weekly cross-indicator correlations (long-term correlation between blood pressure and kidney function). Non-explicit medical correlations (non-linear coupling between sleep quality and blood glucose fluctuations) are automatically discovered through the edge weights of GNN networks. Graph neighborhood aggregation suppresses single-sensor noise (such as motion artifacts in ECG signals), and combined with pattern deviation detection (healthy embedding spatial comparison based on Mahalanobis distance) and causal anomaly localization (GNNExplainer + counterfactual reasoning to analyze the “sleep deprivation → heart rate ↑ → blood pressure ↑” transmission chain), it achieves interpretable analysis of the entire link from microscopic anomalies to macroscopic pathological transmission.

[0044] On the other hand, this application provides a data anomaly detection system based on wearable devices. Includes: an acquisition module for acquiring multimodal data from multiple sensors, the multimodal data including indicators of heart rate, electromyography signals, body temperature and gait information at different time points, and establishing time-series data; The feature extraction module is used to extract window features of time series data using a sliding window, extract short-period fluctuations based on window features, extract seasonal and trend components of time series data to obtain long-period rhythms, and concatenate short-period fluctuations and long-period rhythms into multi-granular features. The trend prediction module establishes an individual health baseline and predicts indicator trends based on multi-granularity features. The dynamic association graph construction module is used to construct dynamic association graphs based on multimodal data. Each node in the dynamic association graph is a multi-granular feature of a time-series slice. Attention mechanism and transfer entropy edge weights are used to calculate the association weights and association causality between nodes. The anomaly detection module is used to perform convolution operations on a dynamic relational graph using a graph convolutional network, and locate abnormal nodes and anomaly propagation paths after aggregating neighborhood information. The early warning module is used to generate spatiotemporally unified representation data of individual health baselines, indicator trends, abnormal nodes, and abnormal propagation paths through a gating attention mechanism, and to trigger graded early warnings based on the spatiotemporally unified representation data.

[0045] During operation, the electromyography (EMG) sensor collects muscle electrical signals, the heart rate sensor monitors pulse waves, and the body temperature sensor records changes in body surface temperature. All data is transmitted to the central control module or data transmission module via a flexible data transmission line, ensuring stable signal transmission. Subsequently, the system performs preliminary processing such as filtering and noise reduction on the collected signals and uploads the data to the cloud via Bluetooth wireless communication. The wearable device-based data anomaly detection system can be set up in the cloud for remote monitoring and health data management. An intelligent feedback mechanism is also set up; when conditions such as abnormal heart rate, sudden EMG changes, or elevated body temperature are detected, closed-loop regulation can be implemented through local device vibration or remote alarms, providing timely intervention and safety assurance.

[0046] The framework implemented by this application through software, as shown in Figure 3, is divided into four layers: multi-source data layer, situation indicator extraction layer, element layer, and overall layer.

[0047] Multi-source data layer: Real-time collection of multi-source physiological signals, including heart rate, body temperature, and gait, via wearable devices, supplemented by accelerometer events (such as falls) as anomaly anchors. Since the raw data suffers from inconsistent sampling frequencies, Neural ODEs are introduced to uniformly align the non-uniform data, generating time-series data streams with equal time intervals.

[0048] The situation indicator extraction layer extracts short-cycle statistics and time-frequency features within a 5-minute sliding window, and uses STL (Seasonal-Trend-Residual) decomposition to remove slow-moving variables such as diurnal / weekly rhythms. An LSTM network is trained on individual time series data from the past 30 days, and dilated convolution (dilation=4) is used to expand the temporal receptive field to 48 hours to capture long-term health baselines such as resting heart rate and blood pressure fluctuations. The system uses Dynamic Time Warping (DTW) to compare the morphological differences between predicted and measured values, thereby identifying chronic drift (e.g., a 0.5% weekly decrease in blood oxygen saturation) or acute fluctuations.

[0049] Element Layer: Based on the extraction of situation indicators, further distinction is made between two categories of indicators: "threat" and "vulnerability". The threat assessment module extracts "threat data" based on the magnitude and frequency of anomalies, such as continuous abnormal heart rate fluctuations. The vulnerability assessment module focuses on the combined judgment of the anomaly source (such as the correlation between nodes) and the individual state background, and uses dynamic causal graph reasoning (GNN network) to construct a spatiotemporal graph: nodes are time slices of each indicator, and edge weights are learned through graph attention mechanism (GAT). Significant causal paths (p<0.01) are screened using transfer entropy, and then the anomaly propagation path (such as "insufficient sleep → increased heart rate → increased blood pressure") is visualized by GNNExplainer. Combined with counterfactual reasoning methods, it is verified whether overall anomalies can be suppressed if the root cause variable is controlled.

[0050] Overall level: Comprehensive assessment of "threat" and "vulnerability" to form a unified health risk score. , of which M w For the threat index score, M v M represents the vulnerability index score, and M represents the overall risk assessment result.

[0051] Those skilled in the art should understand that the image processing method of this application is not limited to specific lung image analysis scenarios, and can be extended to other applications that require comprehensive image information extraction and analysis by combining multiple modal medical images as needed.

[0052] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A data anomaly detection method based on wearable devices, characterized in that, The method includes: S1 acquires multimodal data from multiple sensors, including indicators of heart rate, electromyography signals, body temperature, and gait information at different time points, and establishes time-series data; S2 uses a sliding window to extract window features from time series data, and extracts short-cycle fluctuations based on window features. It also extracts seasonal and trend components from time series data to obtain long-cycle rhythms, and then concatenates short-cycle fluctuations and long-cycle rhythms into multi-granular features. S3 establishes individual health baselines and predicts indicator trends based on multi-granularity features; S4 constructs a dynamic association graph based on multimodal data. Each node in the dynamic association graph is a multi-granular feature of a time-series slice. The attention mechanism and the transfer entropy edge weights are used to calculate the association weights and association causality between nodes. S5 uses graph convolutional networks to perform convolution operations on dynamic relational graphs, and after aggregating neighborhood information, it locates abnormal nodes and abnormal propagation paths.

2. The data anomaly detection method based on wearable devices according to claim 1, characterized in that, It also includes generating spatiotemporally unified representation data from individual health baselines, indicator trends, abnormal nodes, and abnormal propagation paths through a gating attention mechanism, and triggering graded early warnings based on the spatiotemporally unified representation data.

3. The data anomaly detection method based on wearable devices according to claim 1, characterized in that, S3 includes: training an LSTM network on time-series data from at least 30 days ago, using dilated convolutions to extend the temporal receptive field to 48 hours, capturing a long-term healthy baseline, and identifying chronic drift or acute fluctuations.

4. The data anomaly detection method based on wearable devices according to claim 1, characterized in that, Extract the magnitude and frequency of anomalies from spatiotemporal unified characterization data for use as threat data; Extract anomalous nodes and anomalous propagation paths from spatiotemporal unified representation data for use as vulnerability data; Combine threat data with vulnerability data to form a unified health risk score: ， in, The threat index score, Score the vulnerability index. Based on the comprehensive risk assessment results, As weight.

5. The data anomaly detection method based on wearable devices according to claim 1, characterized in that, Tiered early warnings are triggered based on unified spatiotemporal representation data, including: The predictive residuals of the indicator trend are calculated by dynamic time warping. When the predictive residuals exceed the historical 90th percentile, the wearable device is activated to issue a vibration reminder. Short-cycle fluctuations that last for more than three consecutive sliding windows are considered abnormal signals. A causal report with a confidence level of >80% is generated and sent via SMS.

6. The data anomaly detection method based on wearable devices according to claim 2, characterized in that, The residual between the current observed value and the predicted value of the abnormal node is calculated as the abnormal score, and an alarm is triggered based on the abnormal score.

7. A data anomaly detection system based on wearable devices, characterized in that, The system includes: The acquisition module is used to acquire multimodal data from multiple sensors, including indicators such as heart rate, electromyography signals, body temperature, and gait information at different time points, and to establish time-series data. The feature extraction module is used to extract window features of time series data using a sliding window, extract short-period fluctuations based on window features, extract seasonal and trend components of time series data to obtain long-period rhythms, and concatenate short-period fluctuations and long-period rhythms into multi-granular features. The trend prediction module establishes an individual health baseline and predicts indicator trends based on multi-granularity features. The dynamic association graph construction module is used to construct dynamic association graphs based on multimodal data. Each node in the dynamic association graph is a multi-granular feature of a time-series slice. Attention mechanism and transfer entropy edge weights are used to calculate the association weights and association causality between nodes. The anomaly detection module is used to perform convolution operations on a dynamic relational graph using a graph convolutional network, and locate abnormal nodes and anomaly propagation paths after aggregating neighborhood information.

8. The data anomaly detection system based on wearable devices according to claim 7, characterized in that, It also includes an early warning module, which generates spatiotemporally unified representation data of individual health baselines, indicator trends, abnormal nodes and abnormal propagation paths through a gating attention mechanism, and triggers graded early warnings based on the spatiotemporally unified representation data.

Images