Intelligent monitoring method and system for intelligent jewelry

Abstract

The invention discloses an intelligent monitoring method and system for intelligent jewelry, and relates to the technical field of intelligent monitoring. Performing time alignment on the target traffic data through a dynamic time warping algorithm to obtain an original data set; preprocessing all types of traffic data in the original data set to obtain an initial data set; for each type of traffic data in the initial data set, extracting a time-frequency feature corresponding to the type of traffic data based on a preset sliding window; substituting the time-frequency features corresponding to all types of traffic data into a preset lightweight classification model to obtain a user state; and dynamically configuring a sensor sampling strategy according to the user state. According to the intelligent jewelry monitoring method and device, preprocessing and time frequency feature extraction are carried out after time synchronization is carried out on various sensor data, the accuracy of the data is improved, then the extracted time frequency features are substituted into the preset lightweight classification model to determine the user state, low-power-consumption detection is carried out on the premise of accurate measurement, and the monitoring efficiency of the intelligent jewelry is improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent monitoring technology, and particularly relates to an intelligent monitoring method and system for smart jewelry. Background Art

[0002] With the rapid development of wearable technology, smart jewelry, thanks to its portability and decorative properties, has shown great potential in the field of health monitoring. By integrating multiple sensors (such as electrocardiogram (ECG) sensors, inertial measurement units (IMUs), and photoplethysmography (PPG) sensors), smart jewelry can collect real-time physiological and motion data from users, providing data support for health management and sports analysis.

[0003] However, current health monitoring technology based on smart jewelry still faces numerous challenges. First, different types of sensors collect data at different frequencies and triggering mechanisms, resulting in time asynchrony between multi-source data such as ECG, IMU, and PPG, making direct fusion analysis difficult. Second, during user movement, acceleration changes in IMU data can interfere with physiological signals such as ECG, causing noise issues such as baseline drift, seriously impacting data quality and the accuracy of subsequent analysis. Furthermore, raw sensor data contains a large amount of redundant information, making it difficult for traditional feature extraction methods to efficiently identify key features. Complex models also increase computing load and device power consumption, making them unable to meet the low-power, real-time application requirements of smart jewelry, resulting in low monitoring efficiency. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem of low monitoring efficiency of smart jewelry and to propose an intelligent monitoring method and system for smart jewelry.

[0005] In a first aspect of the present invention, a smart monitoring method for smart jewelry is first proposed, the method comprising:

[0006] The target type flow data is collected in real time by multi-source sensors, and the target type flow data is time-aligned by a dynamic time warping algorithm to obtain an original data set; the target type flow data includes ECG data, IMU data and PPG data;

[0007] Preprocessing all types of traffic data in the original data set to obtain an initial data set;

[0008] For each type of traffic data in the initial data set, the time-frequency features corresponding to the type of traffic data are extracted based on the preset sliding window;

[0009] Substitute the time-frequency features corresponding to all types of traffic data into the preset lightweight classification model to obtain the user status;

[0010] The sensor sampling strategy is dynamically configured according to the user status.

[0011] Optionally, preprocessing all types of traffic data in the original data set to obtain an initial data set includes:

[0012] Performing median filtering on the IMU data to obtain first IMU data, and performing feature extraction on the first IMU data to obtain a first feature set; the first feature set includes a mean, a variance, a root mean square, and a frequency domain feature of acceleration;

[0013] Extracting waveform features of the ECG data to obtain a second feature, concatenating the first feature set and the second feature to obtain a fusion feature, and substituting the fusion feature into a preset regression model to obtain interference data;

[0014] performing bandpass filtering on the ECG data to obtain first ECG data, and correcting the first ECG data according to the interference data to obtain second ECG data;

[0015] Performing low-pass filtering on the first IMU data to obtain second IMU data;

[0016] The PPG data is processed by a least mean square algorithm to obtain first PPG data, and an initial data set is obtained according to the second ECG data, the second IMU data and the first PPG data.

[0017] Optionally, the preset regression model construction includes:

[0018] Synchronously collect IMU data and ECG data under different motion states through preset protocols;

[0019] For each set of IMU data and ECG data, feature extraction is performed on the IMU data and ECG data to obtain IMU features and ECG features;

[0020] Subtracting the ECG data from the preset ECG data to obtain noisy ECG data; the preset ECG data is ECG data collected in a non-motion scene;

[0021] The IMU features and ECG features are fused to obtain fused features, and the noisy ECG data is used as the output label of the fused features;

[0022] Initialize the support vector regression model, substitute the fusion features of all groups and the corresponding noisy ECG data into the support vector regression model for training to obtain training parameters, and update the support vector regression model according to the training parameters to obtain a preset regression model.

[0023] Optionally, for each type of traffic data in the initial data set, extracting the time-frequency features corresponding to the type of traffic data based on a preset sliding window includes:

[0024] Acquire data from the initial data set according to a preset sliding window to obtain an identification data set; the identification data set includes identification ECG data, identification IMU data, and identification PPG data;

[0025] detecting R waves of the identified ECG data using a Pan-Tompkins algorithm, and calculating adjacent intervals between adjacent R waves to obtain a target interval;

[0026] Calculating the corresponding pulse transit time according to the identified PPG data to obtain a target transit time;

[0027] If the target interval and the target conduction time are both within the corresponding preset ranges, the identified data set is recorded as a valid data set;

[0028] For all types of traffic data in the valid data set, the time domain features of the traffic data of this type are calculated, and the time domain signals of the traffic data of this type are converted into frequency domain signals for frequency domain feature extraction;

[0029] The extracted time domain features and frequency domain features are fused to form time-frequency features.

[0030] Optionally, the preset lightweight classification model construction includes:

[0031] Obtaining a time-frequency feature set corresponding to different motion states; the time-frequency feature set includes ECG time-frequency features, IMU time-frequency features, and PPG time-frequency features;

[0032] For each motion state, the corresponding ECG time-frequency features are substituted into the ECG input branch, the IMU time-frequency features are substituted into the IMU input branch, and the PPG time-frequency features are substituted into the PPG input branch;

[0033] All input features are convolved across channels and then weighted fused to obtain the target fusion features;

[0034] Substituting the target fusion feature into the gated recurrent unit to obtain a temporal dependency feature;

[0035] The temporal dependency features are globally averaged pooled and fed into the fully connected layer, and the probability distribution of each motion state is obtained through the Softmax activation function.

[0036] In a second aspect of the present invention, an intelligent monitoring system for smart jewelry is provided, comprising:

[0037] A data time alignment module is used to collect target type flow data in real time through multi-source sensors and perform time alignment on the target flow data using a dynamic time warping algorithm to obtain an original data set; the target type flow data includes ECG data, IMU data, and PPG data;

[0038] A data preprocessing module, configured to preprocess all types of traffic data in the original data set to obtain an initial data set;

[0039] The first feature extraction module is used to extract the time-frequency features corresponding to each type of traffic data in the initial data set based on a preset sliding window;

[0040] The user status recognition module is used to substitute the time-frequency features corresponding to all types of traffic data into the preset lightweight classification model to obtain the user status;

[0041] The sampling strategy determination module is used to dynamically configure the sensor sampling strategy according to the user status.

[0042] Optionally, the data preprocessing module includes:

[0043] a first feature set extraction module, configured to perform median filtering on the IMU data to obtain first IMU data, and perform feature extraction on the first IMU data to obtain a first feature set; the first feature set includes the mean, variance, root mean square, and frequency domain features of acceleration;

[0044] A first feature splicing module is used to extract waveform features of the ECG data to obtain a second feature, splice the first feature set and the second feature to obtain a fusion feature, and substitute the fusion feature into a preset regression model to obtain interference data;

[0045] An ECG data correction module, configured to perform bandpass filtering on the ECG data to obtain first ECG data, and correct the first ECG data according to the interference data to obtain second ECG data;

[0046] A low-pass filtering processing module, configured to perform low-pass filtering on the first IMU data to obtain second IMU data;

[0047] An initial data set determination module is configured to process the PPG data using a least mean square algorithm to obtain first PPG data, and to obtain an initial data set based on the second ECG data, the second IMU data, and the first PPG data.

[0048] Optionally, the preset regression model construction includes:

[0049] Data synchronization acquisition module, used to synchronously collect IMU data and ECG data under different motion states through preset protocols;

[0050] The second feature extraction module is used to extract features from each set of IMU data and ECG data to obtain IMU features and ECG features;

[0051] A noise ECG data extraction module is used to subtract ECG data from preset ECG data to obtain noisy ECG data; the preset ECG data is ECG data collected in a non-motion scene;

[0052] The feature fusion module is used to fuse the IMU features and ECG features to obtain a fused feature, and use the noisy ECG data as the output label of the fused feature;

[0053] The preset regression model construction module is used to initialize the support vector regression model, substitute the fusion features of all groups and the corresponding noisy ECG data into the support vector regression model for training to obtain training parameters, and update the support vector regression model according to the training parameters to obtain the preset regression model.

[0054] Optionally, the first feature extraction module includes:

[0055] an identification data set determination module, configured to obtain data from the initial data set according to a preset sliding window to obtain an identification data set; the identification data set includes identification ECG data, identification IMU data, and identification PPG data;

[0056] a target interval determination module, configured to detect the R wave of the identified ECG data using a Pan-Tompkins algorithm and calculate adjacent intervals between adjacent R waves to obtain a target interval;

[0057] a target conduction time determination module, configured to calculate the corresponding pulse conduction time according to the identified PPG data to obtain a target conduction time;

[0058] a valid data set determining module, configured to record the identified data set as a valid data set if the target interval and the target conduction time are both within corresponding preset ranges;

[0059] The time-frequency feature extraction module is used to calculate the time domain features of all types of traffic data in the valid data set, convert the time domain signals of the traffic data into frequency domain signals, and perform frequency domain feature extraction;

[0060] The time-frequency feature fusion module is used to fuse the extracted time domain features and frequency domain features to form time-frequency features.

[0061] Optionally, the preset lightweight classification model construction includes:

[0062] A time-frequency feature set determination module is used to obtain the time-frequency feature set corresponding to different motion states; the time-frequency feature set includes ECG time-frequency features, IMU time-frequency features and PPG time-frequency features;

[0063] The feature input module is used to substitute the corresponding ECG time-frequency features into the ECG input branch, the IMU time-frequency features into the IMU input branch, and the PPG time-frequency features into the PPG input branch for each motion state;

[0064] The target fusion feature determination module is used to perform cross-channel convolution on all input features and then perform weighted fusion to obtain the target fusion feature;

[0065] A time-dependent feature extraction module, configured to substitute the target fusion feature into the gated loop unit to obtain a time-dependent feature;

[0066] The probability distribution determination module is used to perform global average pooling on the temporal dependency features and then input them into the fully connected layer, and obtain the probability distribution of each motion state through the Softmax activation function.

[0067] Beneficial effects of the present invention:

[0068] This invention proposes an intelligent monitoring method for smart jewelry. This method uses multiple sensors to collect target flow data in real time. The target flow data is time-aligned using a dynamic time warping algorithm to obtain an original data set. The target flow data includes ECG data, IMU data, and PPG data. All flow data types in the original data set are preprocessed to obtain an initial data set. For each type of flow data in the initial data set, the corresponding time-frequency features are extracted based on a preset sliding window. The corresponding time-frequency features of all flow data types are substituted into a preset lightweight classification model to obtain the user status. The sensor sampling strategy is dynamically configured based on the user status. The dynamic time warping (DTW) algorithm is used to synchronize the time of multiple sensor data (ECG, IMU, and PPG), followed by preprocessing and time-frequency feature extraction, improving data accuracy. The extracted time-frequency features are then substituted into a preset lightweight classification model to determine the user status, thereby determining the sensor sampling strategy. This method achieves low-power detection while maintaining accurate measurement, improving the monitoring efficiency of smart jewelry. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The present invention will be further described below with reference to the accompanying drawings.

[0070] Figure 1 A flow chart of an intelligent monitoring method for smart jewelry provided by an embodiment of the present invention;

[0071] Figure 2 A framework diagram of an intelligent monitoring system for smart jewelry provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0073] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0074] The embodiment of the present invention provides an intelligent monitoring method for smart jewelry. Figure 1 , Figure 1 A flowchart of an intelligent monitoring method for smart jewelry provided in an embodiment of the present invention. The method includes the following steps:

[0075] S101, collecting target type traffic data in real time through multi-source sensors, and performing time alignment on the target traffic data using a dynamic time warping algorithm to obtain an original data set;

[0076] S102, preprocessing all types of traffic data in the original data set to obtain an initial data set;

[0077] S103, for each type of traffic data in the initial data set, extracting the time-frequency features corresponding to the type of traffic data based on a preset sliding window;

[0078] S104: Substitute the time-frequency features corresponding to all types of traffic data into a preset lightweight classification model to obtain user status;

[0079] S105: Dynamically configure the sensor sampling strategy according to the user status.

[0080] Among them, target type flow data includes ECG data, IMU data and PPG data.

[0081] An intelligent monitoring method for smart jewelry provided by an embodiment of the present invention uses a dynamic time warping (DTW) algorithm to synchronize the time of multiple sensor data (ECG, IMU, PPG), and then performs preprocessing and time-frequency feature extraction to improve data accuracy. The extracted time-frequency features are then substituted into a preset lightweight classification model to determine the user status, thereby determining the sensor sampling strategy. This achieves low-power detection while maintaining accurate measurement, thereby improving the monitoring efficiency of smart jewelry.

[0082] In one implementation method, the dynamic time warping algorithm (DTW) is used to time-align multi-source heterogeneous data such as ECG (electrocardiogram), IMU (inertial measurement unit), and PPG (photoplethysmography) to solve the timing misalignment problem caused by sampling frequency differences and signal transmission delays of different sensors, and ensure the spatiotemporal consistency of subsequent analysis.

[0083] In one implementation, ECG (electrocardiogram) provides cardiac electrophysiological signals, PPG (photoplethysmography) reflects blood flow changes, and IMU (inertial unit) captures motion posture. The combination of the three can cross-verify data reliability (such as the elimination of motion artifacts). A single sensor is susceptible to noise (such as PPG is sensitive to motion), and multi-source data can reduce the misjudgment rate through time-frequency feature fusion, which is particularly suitable for complex scenarios (such as heart rate monitoring during exercise).

[0084] In one implementation, different sensors have different sampling frequencies (e.g., ECG is typically 100 Hz, while IMU may be 50 Hz). DTW eliminates time offsets through nonlinear alignment, ensuring the consistency of multimodal data on the time axis. Compared with simple interpolation or downsampling, DTW maintains the dynamic changes of the original signal, providing more accurate input for time-frequency analysis.

[0085] In one implementation, dynamically adjusting the sensor sampling rate according to the user's state (stationary, moving, sleeping, abnormal) (such as reducing the IMU frequency during sleep) can increase the usage time of smart jewelry; preprocessing the original data set can improve the signal-to-noise ratio and reduce the complexity of subsequent feature extraction.

[0086] In one implementation, the multi-source sensors include an electrocardiogram sensor, an inertial measurement unit, a photoplethysmography sensor, etc.; the preset sliding window is pre-set by a technician; when the user is in a stationary state, the ECG sampling rate is 100Hz, the IMU sampling rate is 20Hz, and the PPG sampling rate is 50Hz; when the user is in motion, the ECG sampling rate is 250Hz, the IMU sampling rate is 100Hz, and the PPG sampling rate is 100Hz; when the user is in sleep, the ECG sampling rate is 50Hz, the IMU sampling rate is 10Hz, and the PPG sampling rate is 25Hz; when the user is in sleep, the ECG sampling rate is 500Hz, the IMU sampling rate is 50Hz, and the PPG sampling rate is 200Hz.

[0087] In one embodiment, preprocessing all types of traffic data in the original data set to obtain an initial data set includes:

[0088] Performing median filtering on the IMU data to obtain first IMU data, and performing feature extraction on the first IMU data to obtain a first feature set; the first feature set includes a mean, a variance, a root mean square, and a frequency domain feature of the acceleration;

[0089] Extract the waveform features of the ECG data to obtain the second feature, concatenate the first feature set and the second feature set to obtain the fused feature, and substitute the fused feature into the preset regression model to obtain the interference data;

[0090] performing bandpass filtering on the ECG data to obtain first ECG data, and correcting the first ECG data according to the interference data to obtain second ECG data;

[0091] Performing low-pass filtering on the first IMU data to obtain second IMU data;

[0092] The PPG data is processed by a least mean square algorithm to obtain first PPG data, and an initial data set is obtained according to the second ECG data, the second IMU data and the first PPG data.

[0093] In one implementation, IMU median filtering is used to eliminate impulse noise (such as instantaneous collision interference) from the accelerometer / gyroscope, retain the true motion trend, and improve the reliability of subsequent feature extraction. By fusing IMU motion features (acceleration variance, frequency domain energy) with ECG waveform features (R wave amplitude, ST segment slope), the regression model can accurately quantify the interference intensity of motion on the ECG (such as baseline drift and electromyographic noise).

[0094] In one implementation, the IMU performs low-pass filtering (such as a 5Hz cutoff) to focus on the main frequency of human motion to prevent high-frequency vibration from interfering with the PPG; the PPG signal is adaptively corrected through the least mean square algorithm (LMS) to synchronize it with the heart rate trend of the second ECG data, solving the problem of multi-sensor sampling time offset; the frequency domain characteristics are the main frequency (pre-set frequencies, determined by technical personnel) components and their amplitudes obtained by fast Fourier transform FFT.

[0095] In one embodiment, the preset regression model construction includes:

[0096] Synchronously collect IMU data and ECG data under different motion states through preset protocols;

[0097] For each set of IMU data and ECG data, feature extraction is performed on the IMU data and ECG data to obtain IMU features and ECG features;

[0098] Noise ECG data is obtained by subtracting the ECG data from the preset ECG data; the preset ECG data is ECG data collected in a non-motion scene;

[0099] The IMU features and ECG features are fused to obtain fused features, and the noisy ECG data is used as the output label of the fused features;

[0100] Initialize the support vector regression model, substitute the fusion features of all groups and the corresponding noisy ECG data into the support vector regression model for training to obtain training parameters, and update the support vector regression model according to the training parameters to obtain a preset regression model.

[0101] In one implementation, the preset protocol is the IEEE1588 protocol; the motion noise is directly quantified by comparing the preset ECG (non-motion scene) with the measured ECG (motion scene), which is more physically meaningful than traditional methods (such as blind source separation) and converts the complex noise separation problem into a supervised learning regression problem.

[0102] In one implementation, the IMU (inertial measurement unit) can accurately capture motion states (such as acceleration and angular velocity), while the ECG records electrocardiogram signals. Fusion of IMU features and ECG features can more comprehensively characterize the impact of motion on the ECG, helping to improve the model's ability to perceive and model motion artifacts.

[0103] In one implementation, SVR excels at processing small samples, nonlinearity, and high-dimensional feature data, making it suitable for modeling complex IMU-ECG relationships. The trained regression model can accurately predict ECG noise under different motion states, helping to restore more realistic and clean ECG signals.

[0104] In one implementation, IMU features include acceleration mean, variance, spectral energy, etc., which reflect the motion amplitude and frequency characteristics; ECG features include waveform morphology parameters (such as R wave amplitude and QRS width), which characterize the inherent characteristics of the ECG signal; SVR (support vector regression) maps the features to a high-dimensional space through a kernel function (such as the RBF kernel) to capture the complex nonlinear relationship between the IMU and ECG noise.

[0105] In one embodiment, for each type of traffic data in the initial data set, extracting the time-frequency features corresponding to the type of traffic data based on a preset sliding window includes:

[0106] Acquire data from the initial data set according to a preset sliding window to obtain a recognition data set; the recognition data set includes recognition ECG data, recognition IMU data, and recognition PPG data;

[0107] The R wave of ECG data is detected and identified by the Pan-Tompkins algorithm, and the adjacent intervals between adjacent R waves are calculated to obtain the target interval;

[0108] Calculate the corresponding pulse transmission time according to the identified PPG data to obtain the target transmission time;

[0109] If the target interval and target conduction time are both within the corresponding preset range, the identified data set is recorded as a valid data set;

[0110] For all types of traffic data in the valid data set, the time domain features of the traffic data of this type are calculated, and the time domain signals of the traffic data of this type are converted into frequency domain signals for frequency domain feature extraction;

[0111] The extracted time domain features and frequency domain features are fused to form time-frequency features.

[0112] In one implementation, through dual determination of the R-wave interval (RR interval) and pulse transit time (PTT), abnormal or noise-prone data segments are effectively eliminated. Only when both key physiological parameters are within a reasonable range is the data retained as a "valid data set," improving the accuracy and credibility of subsequent analysis from the source.

[0113] In one implementation, the ECG data, the IMU data identification, and the PPG data identification are initial data sets obtained based on the second ECG data, the second IMU data, and the first PPG data, and the corresponding type is the data in the initial data set within a preset sliding window.

[0114] In one implementation, both the target interval and the target conduction time are set with preset intervals, the preset interval of the target interval is 300-1100 ms, and the preset interval of the target conduction time is 80-300 ms.

[0115] In one implementation, ECG (electrocardiography), IMU (inertia), and PPG (photoplethysmography) signals are collected and processed simultaneously, allowing the system to conduct comprehensive analysis from multiple dimensions of electrophysiology, movement, and physiological circulation; this helps to more accurately model the individual's physiological state.

[0116] In one implementation, Pan-Tompkins is a classic and efficient ECG R-wave detection algorithm, which can be used for heart rate variability analysis when combined with the RR interval. PTT (pulse transit time) calculated jointly by PPG and ECG is an important physiological indicator of blood pressure and arterial compliance. Using it as a filtering condition for data validity, screening criteria for vascular health dimensions are further introduced to increase the depth of physiological signal analysis.

[0117] In one implementation, time domain features (such as mean, variance, maximum value, and volatility) capture the trend and fluctuation of the signal; frequency domain features (such as power spectral density, dominant frequency, and energy distribution) reflect periodicity and rhythmicity. The time-frequency features formed by the fusion of the two can more comprehensively reflect the intrinsic characteristics of the signal.

[0118] In one embodiment, the preset lightweight classification model construction includes:

[0119] Obtain the time-frequency feature set corresponding to different motion states; the time-frequency feature set includes ECG time-frequency features, IMU time-frequency features, and PPG time-frequency features;

[0120] For each motion state, the corresponding ECG time-frequency features are substituted into the ECG input branch, the IMU time-frequency features are substituted into the IMU input branch, and the PPG time-frequency features are substituted into the PPG input branch;

[0121] All input features are convolved across channels and then weighted fused to obtain the target fusion features;

[0122] Substitute the target fusion features into the gated recurrent unit to obtain the temporal dependency features;

[0123] The temporal dependent features are globally averaged and pooled before being fed into the fully connected layer, and the probability distribution of each motion state is obtained through the Softmax activation function.

[0124] In one implementation, traditional multimodal fusion methods (such as simple concatenation) only achieve feature superposition and are unable to model complex inter-modal relationships (such as the correlation between ECG heart rate variability and IMU exercise intensity). Cross-channel convolution uses convolution kernels with shared weights to traverse different modal feature channels and explore interaction patterns between channels. Weighted fusion uses an attention mechanism to assign higher weights to key channels (for example, increasing the weight of IMU features in motion and ECG features in resting state).

[0125] In one implementation, the generated fusion features include "modal collaborative information" (such as the joint features of "increased heart rate + acceleration fluctuations when running"), which provides a more comprehensive basis for classification decisions; reduces redundant feature interference and improves the compactness of feature expression (such as suppressing non-motion-related PPG noise channels).

[0126] In one implementation, the weight of weighted fusion is to substitute the ECG time-frequency features, IMU time-frequency features and PPG time-frequency features into the attention mechanism to obtain the weight corresponding to each feature.

[0127] In one implementation, the weights in weighted fusion are obtained through the attention mechanism; global average pooling takes the average of the spatial dimensions of the time-dependent features, compressing the multi-dimensional features into a one-dimensional vector, avoiding the position information redundancy caused by the flattening operation of the traditional fully connected layer; the Softmax function maps the output to a normalized probability distribution, intuitively reflecting the confidence of each motion state.

[0128] Based on the same inventive concept, the present invention also provides an intelligent monitoring system for smart jewelry. Figure 2 , Figure 2 A framework diagram of an intelligent monitoring system for smart jewelry provided in an embodiment of the present invention includes:

[0129] The data time alignment module is used to collect target type flow data in real time through multi-source sensors and perform time alignment on the target flow data using a dynamic time warping algorithm to obtain the original data set; the target type flow data includes ECG data, IMU data, and PPG data;

[0130] The data preprocessing module is used to preprocess all types of traffic data in the original data set to obtain the initial data set;

[0131] The first feature extraction module is used to extract the time-frequency features corresponding to each type of traffic data in the initial data set based on a preset sliding window;

[0132] The user status recognition module is used to substitute the time-frequency features corresponding to all types of traffic data into the preset lightweight classification model to obtain the user status;

[0133] The sampling strategy determination module is used to dynamically configure the sensor sampling strategy according to the user status.

[0134] An intelligent monitoring system for smart jewelry provided by an embodiment of the present invention uses a dynamic time warping (DTW) algorithm to synchronize the time of multiple sensor data (ECG, IMU, PPG) before preprocessing and extracting time-frequency features, thereby improving data accuracy. The extracted time-frequency features are then substituted into a preset lightweight classification model to determine the user status, thereby determining the sensor sampling strategy, achieving low-power detection while maintaining accurate measurement, and improving the monitoring efficiency of smart jewelry.

[0135] In one embodiment, the data preprocessing module includes:

[0136] A first feature set extraction module is used to perform median filtering on the IMU data to obtain first IMU data, and perform feature extraction on the first IMU data to obtain a first feature set; the first feature set includes the mean, variance, root mean square, and frequency domain features of the acceleration;

[0137] A first feature splicing module is used to extract waveform features of ECG data to obtain a second feature, splice the first feature set and the second feature to obtain a fusion feature, and substitute the fusion feature into a preset regression model to obtain interference data;

[0138] An ECG data correction module, configured to perform bandpass filtering on the ECG data to obtain first ECG data, and correct the first ECG data according to the interference data to obtain second ECG data;

[0139] A low-pass filtering processing module, configured to perform low-pass filtering on the first IMU data to obtain second IMU data;

[0140] The initial data set determination module is used to process the PPG data using a least mean square algorithm to obtain first PPG data, and to obtain an initial data set based on the second ECG data, the second IMU data, and the first PPG data.

[0141] In one embodiment, the preset regression model construction includes:

[0142] Data synchronization acquisition module, used to synchronously collect IMU data and ECG data under different motion states through preset protocols;

[0143] The second feature extraction module is used to extract features from each set of IMU data and ECG data to obtain IMU features and ECG features;

[0144] A noise ECG data extraction module is used to subtract ECG data from preset ECG data to obtain noisy ECG data; the preset ECG data is ECG data collected in a non-motion scene;

[0145] The feature fusion module is used to fuse the IMU features and ECG features to obtain a fused feature, and use the noisy ECG data as the output label of the fused feature;

[0146] The preset regression model construction module is used to initialize the support vector regression model, substitute the fusion features of all groups and the corresponding noisy ECG data into the support vector regression model for training to obtain training parameters, and update the support vector regression model according to the training parameters to obtain the preset regression model.

[0147] In one embodiment, the first feature extraction module includes:

[0148] An identification data set determination module is used to obtain data from the initial data set according to a preset sliding window to obtain an identification data set; the identification data set includes identifying ECG data, identifying IMU data, and identifying PPG data;

[0149] A target interval determination module is used to detect and identify the R wave of ECG data using the Pan-Tompkins algorithm and calculate the adjacent intervals between adjacent R waves to obtain the target interval;

[0150] A target conduction time determination module is used to calculate the corresponding pulse conduction time according to the identified PPG data to obtain the target conduction time;

[0151] a valid data set determining module, configured to record the identified data set as a valid data set if the target interval and the target conduction time are both within corresponding preset ranges;

[0152] The time-frequency feature extraction module is used to calculate the time domain features of all types of traffic data in the valid data set, convert the time domain signals of the traffic data into frequency domain signals, and perform frequency domain feature extraction;

[0153] The time-frequency feature fusion module is used to fuse the extracted time domain features and frequency domain features to form time-frequency features.

[0154] In one embodiment, the preset lightweight classification model construction includes:

[0155] A time-frequency feature set determination module is used to obtain the corresponding time-frequency feature sets under different motion states; the time-frequency feature sets include ECG time-frequency features, IMU time-frequency features, and PPG time-frequency features;

[0156] The feature input module is used to substitute the corresponding ECG time-frequency features into the ECG input branch, the IMU time-frequency features into the IMU input branch, and the PPG time-frequency features into the PPG input branch for each motion state;

[0157] The target fusion feature determination module is used to perform cross-channel convolution on all input features and then perform weighted fusion to obtain the target fusion feature;

[0158] The temporal dependency feature extraction module is used to substitute the target fusion feature into the gated recurrent unit to obtain the temporal dependency feature;

[0159] The probability distribution determination module is used to perform global average pooling on the time-dependent features and then feed them into the fully connected layer, and obtain the probability distribution of each motion state through the Softmax activation function.

[0160] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. An intelligent monitoring method for smart jewelry, characterized in that: The method comprises: The target type flow data is collected in real time by multi-source sensors, and the target type flow data is time-aligned by a dynamic time warping algorithm to obtain an original data set; the target type flow data includes ECG data, IMU data and PPG data; Preprocessing all types of traffic data in the original data set to obtain an initial data set; For each type of traffic data in the initial data set, the time-frequency features corresponding to the type of traffic data are extracted based on the preset sliding window; Substitute the time-frequency features corresponding to all types of traffic data into the preset lightweight classification model to obtain the user status; The sensor sampling strategy is dynamically configured according to the user status.

2. The intelligent monitoring method for smart jewelry according to claim 1, characterized in that: The initial data set obtained by preprocessing all types of traffic data in the original data set includes: Performing median filtering on the IMU data to obtain first IMU data, and performing feature extraction on the first IMU data to obtain a first feature set; the first feature set includes a mean, a variance, a root mean square, and a frequency domain feature of acceleration; Extracting waveform features of the ECG data to obtain a second feature, concatenating the first feature set and the second feature to obtain a fusion feature, and substituting the fusion feature into a preset regression model to obtain interference data; performing bandpass filtering on the ECG data to obtain first ECG data, and correcting the first ECG data according to the interference data to obtain second ECG data; Performing low-pass filtering on the first IMU data to obtain second IMU data; The PPG data is processed by a least mean square algorithm to obtain first PPG data, and an initial data set is obtained according to the second ECG data, the second IMU data and the first PPG data.

3. The intelligent monitoring method for smart jewelry according to claim 2, characterized in that: The preset regression model construction includes: Synchronously collect IMU data and ECG data under different motion states through preset protocols; For each set of IMU data and ECG data, feature extraction is performed on the IMU data and ECG data to obtain IMU features and ECG features; Subtracting the ECG data from the preset ECG data to obtain noisy ECG data; the preset ECG data is ECG data collected in a non-motion scene; The IMU features and ECG features are fused to obtain fused features, and the noisy ECG data is used as the output label of the fused features; Initialize the support vector regression model, substitute the fusion features of all groups and the corresponding noisy ECG data into the support vector regression model for training to obtain training parameters, and update the support vector regression model according to the training parameters to obtain a preset regression model.

4. The intelligent monitoring method for smart jewelry according to claim 2, characterized in that: For each type of traffic data in the initial data set, the time-frequency features corresponding to the type of traffic data are extracted based on the preset sliding window, including: Acquire data from the initial data set according to a preset sliding window to obtain an identification data set; the identification data set includes identification ECG data, identification IMU data, and identification PPG data; detecting R waves of the identified ECG data using a Pan-Tompkins algorithm, and calculating adjacent intervals between adjacent R waves to obtain a target interval; Calculating the corresponding pulse transit time according to the identified PPG data to obtain a target transit time; If the target interval and the target conduction time are both within the corresponding preset ranges, the identified data set is recorded as a valid data set; For all types of traffic data in the valid data set, the time domain features of the traffic data of this type are calculated, and the time domain signals of the traffic data of this type are converted into frequency domain signals for frequency domain feature extraction; The extracted time domain features and frequency domain features are fused to form time-frequency features.

5. The intelligent monitoring method for smart jewelry according to claim 1, characterized in that: The preset lightweight classification model construction includes: Obtaining a time-frequency feature set corresponding to different motion states; the time-frequency feature set includes ECG time-frequency features, IMU time-frequency features, and PPG time-frequency features; For each motion state, the corresponding ECG time-frequency features are substituted into the ECG input branch, the IMU time-frequency features are substituted into the IMU input branch, and the PPG time-frequency features are substituted into the PPG input branch; All input features are convolved across channels and then weighted fused to obtain the target fusion features; Substituting the target fusion feature into the gated recurrent unit to obtain a temporal dependency feature; The temporal dependency features are globally averaged pooled and fed into the fully connected layer, and the probability distribution of each motion state is obtained through the Softmax activation function.

6. An intelligent monitoring system for smart jewelry, characterized in that: The system comprises: A data time alignment module is used to collect target type flow data in real time through multi-source sensors and perform time alignment on the target flow data using a dynamic time warping algorithm to obtain an original data set; the target type flow data includes ECG data, IMU data, and PPG data; A data preprocessing module, configured to preprocess all types of traffic data in the original data set to obtain an initial data set; The first feature extraction module is used to extract the time-frequency features corresponding to each type of traffic data in the initial data set based on a preset sliding window; The user status recognition module is used to substitute the time-frequency features corresponding to all types of traffic data into the preset lightweight classification model to obtain the user status; The sampling strategy determination module is used to dynamically configure the sensor sampling strategy according to the user status.

7. The intelligent monitoring system for smart jewelry according to claim 6, characterized in that: The data preprocessing module includes: a first feature set extraction module, configured to perform median filtering on the IMU data to obtain first IMU data, and perform feature extraction on the first IMU data to obtain a first feature set; the first feature set includes the mean, variance, root mean square, and frequency domain features of acceleration; A first feature splicing module is used to extract waveform features of the ECG data to obtain a second feature, splice the first feature set and the second feature to obtain a fusion feature, and substitute the fusion feature into a preset regression model to obtain interference data; An ECG data correction module, configured to perform bandpass filtering on the ECG data to obtain first ECG data, and correct the first ECG data according to the interference data to obtain second ECG data; A low-pass filtering processing module, configured to perform low-pass filtering on the first IMU data to obtain second IMU data; An initial data set determination module is configured to process the PPG data using a least mean square algorithm to obtain first PPG data, and to obtain an initial data set based on the second ECG data, the second IMU data, and the first PPG data.

8. The intelligent monitoring system for smart jewelry according to claim 7, characterized in that: The preset regression model construction includes: Data synchronization acquisition module, used to synchronously collect IMU data and ECG data under different motion states through preset protocols; The second feature extraction module is used to extract features from each set of IMU data and ECG data to obtain IMU features and ECG features; A noise ECG data extraction module is used to subtract ECG data from preset ECG data to obtain noisy ECG data; the preset ECG data is ECG data collected in a non-motion scene; The feature fusion module is used to fuse the IMU features and ECG features to obtain a fused feature, and use the noisy ECG data as the output label of the fused feature; The preset regression model construction module is used to initialize the support vector regression model, substitute the fusion features of all groups and the corresponding noisy ECG data into the support vector regression model for training to obtain training parameters, and update the support vector regression model according to the training parameters to obtain the preset regression model.

9. The intelligent monitoring system for smart jewelry according to claim 7, characterized in that: The first feature extraction module includes: an identification data set determination module, configured to obtain data from the initial data set according to a preset sliding window to obtain an identification data set; the identification data set includes identification ECG data, identification IMU data, and identification PPG data; a target interval determination module, configured to detect the R wave of the identified ECG data using a Pan-Tompkins algorithm and calculate adjacent intervals between adjacent R waves to obtain a target interval; a target conduction time determination module, configured to calculate the corresponding pulse conduction time according to the identified PPG data to obtain a target conduction time; a valid data set determining module, configured to record the identified data set as a valid data set if the target interval and the target conduction time are both within corresponding preset ranges; The time-frequency feature extraction module is used to calculate the time domain features of all types of traffic data in the valid data set, convert the time domain signals of the traffic data into frequency domain signals, and perform frequency domain feature extraction; The time-frequency feature fusion module is used to fuse the extracted time domain features and frequency domain features to form time-frequency features.

10. The intelligent monitoring system for smart jewelry according to claim 6, characterized in that: The preset lightweight classification model construction includes: A time-frequency feature set determination module is used to obtain the time-frequency feature set corresponding to different motion states; the time-frequency feature set includes ECG time-frequency features, IMU time-frequency features and PPG time-frequency features; The feature input module is used to substitute the corresponding ECG time-frequency features into the ECG input branch, the IMU time-frequency features into the IMU input branch, and the PPG time-frequency features into the PPG input branch for each motion state; The target fusion feature determination module is used to perform cross-channel convolution on all input features and then perform weighted fusion to obtain the target fusion feature; A time-dependent feature extraction module, configured to substitute the target fusion feature into the gated loop unit to obtain a time-dependent feature; The probability distribution determination module is used to perform global average pooling on the temporal dependency features and then input them into the fully connected layer, and obtain the probability distribution of each motion state through the Softmax activation function.

Images