A health monitoring system based on human feature data

Abstract

The application discloses a health monitoring system based on human feature data and belongs to the technical field of health monitoring, and comprises a sensing terminal, an edge node and a cloud platform. The sensing terminal is a wearable patch device integrated with a flexible temperature sensor, an environmental temperature sensor and a low-power MCU, which collects original human body surface temperature data at a sampling frequency of 1 Hz. Environmental temperature data are synchronously collected to provide environmental parameter input for drift compensation. Preliminary hardware-level signal conditioning and data packaging are performed. The original data are transmitted to the edge node through Bluetooth. The cloud platform generates personalized health reports and intervention suggestions. User feedback data are received to continuously optimize evaluation rules and model parameters. The whole-link optimization innovation is carried out from the sensing layer, the algorithm layer and the architecture layer for the core bottleneck of the existing flexible body temperature health monitoring system, and five problems of sensor drift, poor data quality, single feature dimension, weak model generalization and poor module cooperation are effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of health monitoring equipment technology, and specifically relates to a health monitoring system based on human characteristic data. Background Technology

[0002] Body temperature, as a core physiological indicator reflecting human health status, is crucial for accurate collection and dynamic prediction of health risks, enabling timely intervention and treatment. Currently, mainstream body temperature monitoring methods are mainly divided into traditional contact and non-contact methods, such as mercury thermometers and infrared thermometers, both of which have significant limitations.

[0003] While traditional contact temperature measurement devices offer high accuracy, they generally suffer from poor portability and the inability to perform continuous monitoring, making it difficult to capture dynamic trends in body temperature. Non-contact temperature measurement devices, on the other hand, are susceptible to interference from factors such as ambient temperature and measurement distance, resulting in larger measurement errors. Furthermore, they also cannot meet the needs of long-term continuous monitoring.

[0004] To address the issue of continuous monitoring, existing wearable body temperature monitoring devices have begun to adopt flexible sensor technology. However, several technical bottlenecks remain: First, flexible sensors are susceptible to drift due to their own material properties and environmental factors, leading to a decrease in data accuracy over long-term monitoring. Second, the simple data preprocessing methods are ineffective in handling outliers and noise, affecting the accuracy of subsequent feature extraction. Third, feature extraction often focuses on static statistical features, ignoring the dynamic changes in body temperature over time, resulting in insufficient comprehensiveness in health status assessment. Fourth, models used for body temperature prediction often rely on small sample data from a single user for training, exhibiting poor generalization ability, low training efficiency, difficulty in adapting to individual differences among users, and difficulty in guaranteeing prediction accuracy.

[0005] In addition, the existing body temperature monitoring system has many functional gaps in its module design, and the coordination between data collection, processing, prediction and evaluation is poor, making it impossible to form a complete technical loop. This results in insufficient timeliness and practicality of health assessment reports, making it difficult to effectively support users' health management decisions. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a health monitoring system based on human body feature data, comprising: a sensing terminal, an edge node, and a cloud platform; the sensing terminal is a wearable patch device integrating a flexible temperature sensor, an ambient temperature sensor, and a low-power MCU, which collects raw human body surface temperature data at a sampling frequency of 1Hz; it simultaneously collects ambient temperature data to provide environmental parameter input for drift compensation; it performs preliminary hardware-level signal conditioning and data packaging; and it transmits the raw data to the edge node via Bluetooth. Edge nodes, carried by users' smartphones or home gateways, undertake processing tasks with high real-time requirements and moderate computational load: receiving raw data from the terminal, executing drift compensation algorithms and data preprocessing; extracting lightweight time-series features, running local lightweight prediction models and anomaly warning logic; triggering local audible and visual warnings when abnormal body temperature is detected; aggregating data by time granularity, compressing it, and uploading it to the cloud platform to reduce bandwidth consumption. The cloud platform is responsible for big data processing and model iteration: storing all users' historical body temperature data and health records; performing batch training and iterative updates of meta-learning models, and regularly sending optimized model parameters to the edge; performing in-depth health status assessments based on multi-dimensional features, generating personalized health reports and intervention suggestions; and receiving user feedback data to continuously optimize assessment rules and model parameters.

[0007] Furthermore, the overall system workflow is as follows: the sensing terminal collects raw data, the edge terminal performs drift compensation and preprocessing, the edge terminal extracts features and completes real-time prediction and early warning, the data is synchronized to the cloud, the cloud performs in-depth evaluation and model iteration, the evaluation results and optimization parameters are sent back to the edge terminal, and the user obtains health advice and feedback on the effect, forming a closed loop of the entire chain.

[0008] Furthermore, to address the coupling issue of temperature drift and long-term creep drift in flexible sensors, a two-stage drift compensation mechanism is proposed, consisting of polynomial preliminary compensation and adaptive Kalman filter real-time estimation. This mechanism simultaneously covers both the systematic temperature drift caused by ambient temperature and the random drift caused by material aging.

[0009] Furthermore, temperature is the core cause of drift in flexible sensors. The temperature coefficient of resistance of the sensitive material and the thermal expansion effect of the substrate jointly lead to output value deviation. A second-order polynomial temperature drift model is established to describe the nonlinear relationship between the sensor's zero-point output and the ambient temperature. At the same time, a sensitivity temperature compensation model is established to correct the sensor's response sensitivity at different temperatures. After preliminary temperature drift compensation, the sensor output results show that polynomial compensation can only eliminate systematic temperature drift and cannot cover random slow drift caused by material creep and mechanical deformation. The drift amount is introduced as a state variable into the Kalman filter framework to achieve real-time estimation and compensation of the remaining drift.

[0010] Furthermore, to address the issues of low outlier removal accuracy and loss of effective information during noise smoothing in existing preprocessing methods, a two-stage joint preprocessing algorithm combining outlier detection and noise smoothing is adopted to improve data quality while preserving the dynamic characteristics of body temperature. The traditional 3σ criterion, which identifies outliers based on the mean and standard deviation, is inherently susceptible to extreme values. Hampel filtering, which uses the median and median absolute deviation (MAD) as statistics, exhibits strong robustness. The traditional Hampel filtering is improved with window adaptive adjustment, calibrating MAD to an equivalent quantity of the standard deviation under a normal distribution. If the data at the center of the window exceeds the threshold range, it is identified as an outlier and replaced with the median of the window.

[0011] The beneficial effects of this invention include: This invention provides a health monitoring system based on human body feature data, including a sensing terminal, edge nodes, and a cloud platform. The sensing terminal is a wearable patch device integrating a flexible temperature sensor, an ambient temperature sensor, and a low-power MCU. Its core functions include: acquiring raw human body surface temperature data at a sampling frequency of 1Hz; and simultaneously acquiring ambient temperature data to provide environmental parameter input for drift compensation. This solution addresses the core bottlenecks of existing flexible body temperature health monitoring systems by optimizing and innovating the entire chain from the sensing layer, algorithm layer, and architecture layer, effectively solving five major problems: sensor drift, poor data quality, single feature dimension, weak model generalization, and poor module coordination. Experimental results show that this solution significantly outperforms traditional solutions in monitoring accuracy, prediction performance, and assessment accuracy, demonstrating high clinical application value and feasibility. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of a health monitoring system based on human characteristic data according to an embodiment of the present invention. Detailed Implementation

[0014] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0015] like Figure 1As shown, this invention discloses a health monitoring system based on human body feature data, including a sensing terminal, an edge node, and a cloud platform. The sensing terminal is a wearable patch device integrating a flexible temperature sensor, an ambient temperature sensor, and a low-power MCU. Its core functions include: acquiring raw human body surface temperature data at a sampling frequency of 1Hz; simultaneously acquiring ambient temperature data to provide environmental parameter input for drift compensation; performing preliminary hardware-level signal conditioning and data packaging; and transmitting raw data to the edge node via Bluetooth. This layer adopts a low-power design, supporting continuous monitoring for 14 days on a single charge.

[0016] Edge nodes, carried by users' smartphones or home gateways, undertake processing tasks with high real-time requirements and moderate computational load: receiving raw data from terminals, executing drift compensation algorithms and data preprocessing; extracting lightweight time-series features, running local lightweight prediction models and anomaly warning logic; triggering local audible and visual warnings immediately when abnormal body temperature is detected, without waiting for cloud response; aggregating data by time granularity, compressing it, and uploading it to the cloud platform to reduce bandwidth consumption.

[0017] The cloud platform is responsible for big data processing and model iteration: storing all users' historical body temperature data and health records; performing batch training and iterative updates of meta-learning models, and regularly sending optimized model parameters to the edge; performing in-depth health status assessments based on multi-dimensional features, generating personalized health reports and intervention suggestions; and receiving user feedback data to continuously optimize assessment rules and model parameters.

[0018] The overall workflow of the system is as follows: the sensing terminal collects raw data → the edge terminal performs drift compensation and preprocessing → the edge terminal extracts features and completes real-time prediction and early warning → the data is synchronized to the cloud → the cloud performs in-depth evaluation and model iteration → the evaluation results and optimization parameters are sent back to the edge terminal → users obtain health advice and provide feedback on the effects, forming a closed loop of the entire chain.

[0019] To address the coupling issue of temperature drift and long-term creep drift in flexible sensors, this solution proposes a two-stage drift compensation mechanism consisting of polynomial preliminary compensation and adaptive Kalman filter (AKF) real-time estimation, which simultaneously covers both systematic temperature drift caused by ambient temperature and random drift caused by material aging.

[0020] Temperature is the core cause of drift in flexible sensors; the temperature coefficient of resistance of the sensitive material and the thermal expansion effect of the substrate both contribute to output value shifts. This scheme establishes a second-order polynomial temperature drift model to describe the nonlinear relationship between the sensor's zero-point output and ambient temperature:

[0021] Where V0 is the sensor zero-point output value at the reference temperature T0 (usually taken as 25℃); α is the first-order temperature coefficient, characterizing the linear temperature drift intensity; and β is the second-order temperature coefficient, characterizing the nonlinear temperature drift intensity. The model parameters were obtained through high and low temperature chamber calibration experiments and fitted using the least squares method.

[0022] Simultaneously, a sensitivity temperature compensation model is established to correct the sensor response sensitivity at different temperatures:

[0023] In the formula, S0 is the sensitivity at the reference temperature, and γ and δ are the first and second temperature coefficients of the sensitivity, respectively.

[0024] The sensor output after preliminary temperature drift compensation is:

[0025] in This is the sensor's original output voltage.

[0026] Polynomial compensation can only eliminate systematic temperature drift and cannot cover random slow drift caused by material creep and mechanical deformation. This scheme introduces the drift amount as a state variable into the Kalman filter framework to achieve real-time estimation and compensation of the remaining drift.

[0027] Define the system state vector as ,dk]T, where The actual body temperature value at time k. Let k be the remaining drift at time k. Establish the state equation:

[0028] The state transition matrix in the formula It is assumed that body temperature and drift remain approximately constant within short sampling intervals; The noise is the process noise, which follows a mean of 0 and a covariance of . The Gaussian distribution.

[0029] To address the issues of fixed noise covariance and poor adaptability in traditional Kalman filters, this scheme introduces an adaptive noise estimation mechanism: real-time updating of variance based on a sliding window of the innovation sequence. and New information is defined as The formula for updating the observation noise variance is:

[0030] in The length of the sliding window. This is the prior estimate of the error covariance.

[0031] The final output, after compensation, is the true body temperature component estimated by the Kalman filter. This achieves dual suppression of systematic temperature drift and random drift.

[0032] To address the issues of low accuracy in outlier removal and loss of useful information during noise smoothing in existing preprocessing methods, this solution proposes a two-stage joint preprocessing algorithm of "outlier detection - noise smoothing" to improve data quality while preserving the dynamic characteristics of body temperature.

[0033] The traditional 3σ criterion identifies outliers based on the mean and standard deviation, but it is inherently susceptible to extreme values. The Hampel filter, using the median and median absolute deviation (MAD) as statistics, exhibits strong robustness. This scheme improves upon the traditional Hampel filter by implementing an adaptive windowing mechanism: For length of Data sequence within the sliding window Calculate the number of pixels in the window. Absolute deviation from the median:

[0034] The anomaly detection threshold is set as follows:

[0035] The coefficient 0.6745 is used to calibrate MAD to the equivalent of the standard deviation under a normal distribution, and a k value of 3 corresponds to a 99.7% confidence interval. If the data in the center of the window exceeds the threshold range, it is judged as an outlier and replaced with the window center value.

[0036] Based on this, a window adaptation mechanism is introduced: calculating the volatility of data within the window. When volatility is higher than the threshold, the window is narrowed to improve the sensitivity of anomaly detection; when volatility is lower than the threshold, the window is widened to reduce the false detection rate, thus achieving differentiated processing of stable and volatile segments.

[0037] Even after outlier removal, the data still contains high-frequency noise caused by human movement and electromagnetic interference. This solution employs an improved wavelet thresholding denoising algorithm to address the shortcomings of traditional soft thresholding functions, which suffer from constant bias, and hard thresholding functions, which are discontinuous.

[0038] First, the preprocessed data is subjected to J-level wavelet decomposition to obtain the wavelet coefficients of each level. (j is the decomposition level number, k is the coefficient index). For each wavelet coefficient level, calculate the noise standard deviation estimate for that level:

[0039] The corresponding adaptive threshold is:

[0040] in denoted as the number of wavelet coefficients at the j-th layer.

[0041] An improved semi-soft thresholding function is used to shrink the wavelet coefficients:

[0042] Where α∈[0,1] is the adjustment factor, which adaptively adjusts with the number of decomposition layers: α takes a smaller value in low-frequency layers to retain effective signals, and α takes a larger value in high-frequency layers to enhance the denoising effect. This function is continuous at the threshold, and at the same time, it balances bias suppression and feature preservation through the adjustment factor, which is superior to traditional soft and hard threshold functions.

[0043] Finally, the processed wavelet coefficients are reconstructed to obtain the denoised and clean body temperature time series data.

[0044] To address the problem that existing feature extraction methods neglect temporal dynamic patterns, this scheme constructs a three-dimensional feature system comprising static statistical features, temporal dynamic features, and time-frequency domain features. By employing Empirical Mode Decomposition (EMD) and Hilbert-Huang Transform (HHT), the implicit time-frequency information of body temperature signals is mined, comprehensively characterizing the patterns of body temperature changes.

[0045] Static statistical features: Eight dimensions of body temperature are selected within a fixed time window (e.g., 30 minutes): mean, median, maximum, minimum, standard deviation, interquartile range, skewness, and kurtosis, to reflect the overall distribution and dispersion of body temperature.

[0046] Temporal dynamic features: The calculation of body temperature within the window includes five dimensions: the mean of the first-order difference (average rate of change), the sum of the absolute values ​​of the first-order differences (cumulative fluctuation), the proportion of rising time, the proportion of falling time, and the sample entropy, which characterize the dynamic trend and complexity of body temperature changes.

[0047] Body temperature signals are nonlinear and non-stationary time series. Both Fourier transform and wavelet transform rely on prior basis functions and cannot adaptively characterize their local time-frequency properties. This scheme uses the EMD-HHT method to extract time-frequency domain features: The first step is to adaptively decompose the body temperature time series x(t) into n intrinsic mode functions (IMFs) and 1 residual component using EMD:

[0048] in For the i-th IMF component, the following conditions must be met: "the number of extreme points and zero-crossing points differs by no more than 1, and the local upper and lower envelopes are symmetric about the time axis at any time"; The residual component represents the long-term trend of the signal.

[0049] The second step is to process each IMF component. Perform a Hilbert transform to obtain its analytic signal:

[0050] in for Hilbert transform; Instantaneous amplitude; instantaneous phase The third step is to calculate the instantaneous frequency and marginal spectral characteristics. The formula for instantaneous frequency is:

[0051] The marginal spectrum reflects the body temperature signal at various frequencies by integrating the instantaneous amplitudes of all IMFs, yielding the marginal spectrum:

[0052] Energy cumulative distribution. Finally, six-dimensional time-frequency domain features are extracted for each IMF: average instantaneous frequency, instantaneous amplitude standard deviation, marginal spectrum main peak frequency, and marginal spectrum energy entropy.

[0053] The maximum mutual information coefficient (MIC) is used to calculate the correlation between all 19 features and the health status label. Redundant features with MIC values ​​below 0.1 are removed, and high-discrimination features are retained to form the final feature set, thereby reducing the computational complexity of subsequent models.

[0054] To address the issues of poor generalization ability and slow individual adaptation of existing models with small sample sizes, this solution proposes a body temperature prediction framework based on model-independent meta-learning (MAML). Using a lightweight LSTM as the backbone network, it achieves high-precision personalized prediction with a small amount of user data through a two-stage process of "meta-training - rapid adaptation".

[0055] A two-layer LSTM network is used as the backbone for time-series prediction. The input is the historical T-step body temperature feature sequence, and the output is the predicted body temperature value for the next H steps. The network structure includes: an input layer, two LSTM hidden layers, and a fully connected output layer, with a total of approximately 12,000 parameters. It can be deployed on edge devices.

[0056] The core idea of ​​meta-learning is "learning how to learn". Through extensive training across user tasks, a set of initial parameters that are sensitive to new tasks are obtained. New users only need a small amount of data to perform a few gradient updates to achieve good prediction performance.

[0057] Define task distribution Each task The corresponding user's body temperature prediction task includes a support set. With query set Support sets are used for in-task adaptation, while query sets are used to evaluate the adaptation effect.

[0058] Inner loop (task adaptation phase): For each sampled task Starting with the initial parameters θ, the loss is calculated using the support set data, and a single-step gradient update is performed to obtain the task-specific parameters. :

[0059] Where α is the inner loop learning rate. Let the mean squared error (MSE) loss function be used. For parameters The corresponding prediction model.

[0060] Outer loop (meta-optimization phase): Based on the loss of all tasks on the query set, update the meta-initial parameters θ, with the goal of minimizing the expected loss after all tasks have been adapted.

[0061] Parameter updates are performed using batch gradient descent:

[0062] Where β is the outer loop learning rate and B is the number of tasks in the meta-training batch.

[0063] The meta-training phase uses a large-scale public body temperature dataset and multi-user data accumulated in the cloud to construct hundreds of prediction tasks for iteration, ultimately obtaining meta-initial parameters with strong generalization.

[0064] For new users, only 24 hours of body temperature data need to be collected as a support set. The initial parameters are then fine-tuned in 1-3 steps using an inner loop update rule to obtain a personalized prediction model that fits the user's body temperature baseline and rhythm. The adaptation time is shortened by more than 90% compared to traditional training from scratch, and the prediction accuracy is significantly higher than that of the direct training model with a small sample size.

[0065] To address the issue of module fragmentation, this solution clearly defines the functional boundaries and data flow rules for each level, achieving end-to-end collaboration: End-side: Responsible for raw data acquisition and hardware-level signal conditioning to ensure data source stability; Side-side: Handles drift compensation, preprocessing, feature extraction, local prediction and anomaly warning with high real-time requirements, with anomaly response latency of less than 1 second; simultaneously aggregates data at the hourly granularity, encrypts it and uploads it to the cloud; In the cloud: it undertakes non-real-time heavy computation tasks, including batch training of meta-models, deep health assessment, and historical data backtracking analysis; it synchronizes optimized model parameters to the edge every 7 days to achieve silent iteration of the model.

[0066] The data flow adopts a model of real-time processing at the edge and periodic optimization in the cloud, which balances real-time performance and model performance, avoiding the latency problem of traditional pure cloud architecture and the insufficient computing power problem of pure edge architecture.

[0067] Based on the extracted multi-dimensional body temperature features and predicted trends, combined with the user's basic health information (age, gender, and basic medical history), a fuzzy comprehensive assessment model is constructed to output the health status level and intervention suggestions.

[0068] The assessment indicator system includes: current body temperature level (weight 0.3), degree of body temperature fluctuation (weight 0.2), trend of body temperature change (weight 0.25), integrity of circadian rhythm (weight 0.15), and frequency of historical abnormalities (weight 0.1). The Analytic Hierarchy Process (AHP) is used to determine the indicator weights. Membership functions are used to calculate the membership degree of each indicator to the four levels of "normal, mild abnormality, moderate abnormality, and severe abnormality." Finally, the assessment results are output based on the principle of maximum membership degree.

[0069] The system generates differentiated recommendations based on the assessment level: for normal conditions, it outputs health maintenance recommendations; for mild abnormalities, it prompts users to pay attention to rest and hydration; for moderate abnormalities, it recommends physical cooling and continuous monitoring; for severe abnormalities, it immediately triggers an alert, recommends seeking medical attention promptly, and can authorize push notifications to the bound emergency contacts.

[0070] The system establishes a closed-loop data feedback mechanism: users' adoption of health advice and changes in their actual health status are synchronized to the cloud. The cloud continuously optimizes the weights of evaluation indicators and model parameters based on the feedback data. At the same time, de-identified data of new users is added to the meta-training set to further improve the generalization ability of the meta-model, forming a positive cycle of "data accumulation - model optimization - experience improvement - data growth".

[0071] To verify the technical performance of this solution, an experimental testing environment was set up. A comparative experiment was conducted using the publicly available body temperature dataset PHYSIONET and a self-built dataset of 20 volunteers with 7 days of continuous monitoring. The verification was carried out from four dimensions: drift compensation, preprocessing, feature effectiveness, and prediction performance.

[0072] Comparing the compensation effects of traditional first-order linear compensation, second-order polynomial compensation, and this scheme, the mean absolute error (MAE) of 72 hours of continuous monitoring was calculated using standard medical body temperature probe readings as a benchmark: First-order linear compensation: MAE is 0.31℃, and the drift error increases to 0.42℃ after 72 hours; Second-order polynomial compensation: MAE is 0.18℃, which improves the nonlinear temperature drift suppression effect, but random drift still exists; This method achieves a MAE of 0.06℃, with a long-term monitoring error of less than 0.08℃ over 72 hours. The drift suppression effect is more than 70% better than traditional methods, meeting the accuracy requirements for clinical-grade body temperature monitoring.

[0073] The denoising effects of the moving average method, traditional wavelet denoising, and our proposed solution are compared, using signal-to-noise ratio (SNR) and mean square error (MSE) as indicators: Moving average method: SNR improved by 3.2dB, MSE was 0.021, which easily smoothed out the peak characteristics of body temperature; Traditional wavelet hard thresholding: SNR improved by 6.8dB, MSE is 0.012, and pseudo-Gibbs phenomenon exists; This solution improves SNR by 9.5dB and MSE by 0.006. While effectively suppressing noise, it fully preserves the instantaneous change characteristics of body temperature, and the outlier detection accuracy reaches 98.7%.

[0074] The accuracy of health status assessment using traditional methods that only use static features is compared with that of our multi-dimensional feature approach. Furthermore, the small-sample prediction performance of ordinary LSTM and MAML-LSTM is compared. Health assessment: The accuracy rate of traditional static features is 76.2%, while the accuracy rate of this multi-dimensional feature approach is 92.5%, an improvement of 16.3 percentage points, which can effectively identify early abnormal body temperature trends; Body temperature prediction: In a small sample scenario using only 24 hours of training data, the MAE of ordinary LSTM is 0.28℃, while the MAE of this solution MAML-LSTM is 0.16℃, improving the prediction accuracy by 42.9%, and reducing the adaptation time for new users from an average of 2 hours to less than 5 minutes.

[0075] The core innovations of this solution can be summarized in the following five points: Drift compensation innovation: A two-stage compensation scheme of second-order polynomial coupled adaptive Kalman filter is proposed, which simultaneously suppresses systematic temperature drift and random slow drift, and significantly improves long-term monitoring accuracy; Preprocessing innovation: Design a joint denoising algorithm combining adaptive Hampel filtering and improved wavelet thresholding, which balances the robustness of outlier removal with the preservation of effective features; Innovative Feature Extraction: Constructing a three-dimensional feature system and introducing EMD-HHT to extract time-frequency domain features, comprehensively exploring the dynamic changes in body temperature over time, and improving the comprehensiveness of health assessment; Predictive model innovation: Introducing MAML meta-learning into the body temperature prediction scenario solves the problems of poor model generalization and slow individual adaptation under small sample conditions, while taking into account both large-scale deployment and personalized needs; Architectural innovation: Adopting an edge-cloud collaborative architecture, connecting the data links of each module to form a complete technical closed loop, taking into account both real-time early warning and in-depth evaluation, and improving the system's usability.

[0076] This solution addresses the core bottlenecks of existing flexible body temperature health monitoring systems by implementing end-to-end optimization and innovation across the perception, algorithm, and architecture layers. It effectively solves five major problems: sensor drift, poor data quality, limited feature dimensions, weak model generalization, and poor module coordination. Experimental results show that this solution significantly outperforms traditional solutions in monitoring accuracy, prediction performance, and assessment accuracy, demonstrating high clinical application value and feasibility.

[0077] In the future, multimodal human characteristic data such as heart rate and blood oxygen can be further integrated to expand the dimensions of health assessment; at the same time, a federated learning mechanism can be introduced to achieve cross-device model collaborative optimization while protecting user data privacy, thereby further improving the system's universality and security.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A health monitoring system based on human characteristic data, characterized in that, include: The system comprises a sensing terminal, edge nodes, and a cloud platform. The sensing terminal is a wearable patch device integrating a flexible temperature sensor, an ambient temperature sensor, and a low-power MCU. It collects raw human body surface temperature data at a sampling frequency of 1Hz; it simultaneously collects ambient temperature data to provide environmental parameter input for drift compensation; it performs preliminary hardware-level signal conditioning and data packaging; and it transmits the raw data to the edge nodes via Bluetooth. Edge nodes, carried by users' smartphones or home gateways, undertake processing tasks with high real-time requirements and moderate computational load: receiving raw data from the terminal and performing drift compensation algorithms and data preprocessing. Extract lightweight time-series features and run a local lightweight prediction model and anomaly warning logic; When an abnormal body temperature is detected, a local audible and visual warning is triggered. Data is aggregated by time granularity, compressed, and then uploaded to the cloud platform to reduce bandwidth consumption; The cloud platform is responsible for big data processing and model iteration: storing all users' historical body temperature data and health records; performing batch training and iterative updates of meta-learning models, and regularly sending optimized model parameters to the edge; performing in-depth health status assessments based on multi-dimensional features, and generating personalized health reports and intervention suggestions. We receive user feedback data and continuously optimize evaluation rules and model parameters.

2. The health monitoring system based on human characteristic data as described in claim 1, characterized in that, The overall workflow of the system is as follows: the sensing terminal collects raw data, the edge terminal performs drift compensation and preprocessing, the edge terminal extracts features and completes real-time prediction and early warning, the data is synchronized to the cloud, the cloud performs in-depth evaluation and model iteration, the evaluation results and optimization parameters are sent back to the edge terminal, and the user obtains health advice and feedback on the effect, forming a closed loop of the whole chain.

3. A health monitoring system based on human characteristic data as described in claim 1, characterized in that, To address the coupling issue of temperature drift and long-term creep drift in flexible sensors, a two-stage drift compensation mechanism is proposed, consisting of polynomial preliminary compensation and adaptive Kalman filter real-time estimation. This mechanism simultaneously covers both the systematic temperature drift caused by ambient temperature and the random drift caused by material aging.

4. A health monitoring system based on human characteristic data as described in claim 3, characterized in that, Temperature is the core cause of drift in flexible sensors. The temperature coefficient of resistance of the sensitive material and the thermal expansion effect of the substrate will jointly lead to the shift in output value. A second-order polynomial temperature drift model is established to describe the nonlinear relationship between the sensor's zero-point output and the ambient temperature. Simultaneously, a sensitivity temperature compensation model is established to correct the sensor response sensitivity at different temperatures. After preliminary temperature drift compensation, the sensor output results show that polynomial compensation can only eliminate systematic temperature drift and cannot cover random slow drift caused by material creep and mechanical deformation. The drift amount is introduced as a state variable into the Kalman filter framework to achieve real-time estimation and compensation of the remaining drift.

5. A health monitoring system based on human characteristic data as described in claim 4, characterized in that, To address the issues of low outlier removal accuracy and loss of effective information during noise smoothing in existing preprocessing methods, a two-stage joint preprocessing algorithm combining outlier detection and noise smoothing is adopted to improve data quality while preserving the dynamic characteristics of body temperature. The traditional 3σ criterion, which identifies outliers based on the mean and standard deviation, is inherently susceptible to extreme values. Hampel filtering, which uses the median and median absolute deviation (MAD) as statistics, exhibits strong robustness. The traditional Hampel filtering is improved by adaptive windowing, calibrating MAD to an equivalent quantity of the standard deviation under a normal distribution. If the data at the center of the window exceeds the threshold range, it is identified as an outlier and replaced with the median of the window.

Images