A multi-modal data hierarchical management method and system of an intelligent wearable device

Abstract

This invention relates to the field of smart wearable device technology, and discloses a method and system for hierarchical management of multimodal data in smart wearable devices. The method includes acquiring continuous multimodal data from the wearable device, truncating and filtering it to obtain clean data segments; extracting core feature vectors and classifying them to obtain the current value density; when the value density is less than a density threshold, the backup node goes into hibernation and is powered off, switching to a single-node operating architecture; the master node performs time-series waveform feature analysis on the clean data segments and continuously outputs routine physiological state information, calculating the abnormal fluctuation amplitude based on the routine physiological state information; when the abnormal fluctuation reaches a fluctuation threshold or the value density reaches a density threshold, the backup node is woken up and powered on, switching to a dual-active operating architecture; time synchronization processing is performed through the dual-active operating architecture to obtain aligned data sequences and generate hierarchical early warning information. This invention achieves hierarchical processing of multimodal data and dynamic adaptation of computing resources.

Description

Technical Field

[0001] This invention relates to the field of smart wearable device technology, and in particular to a method and system for hierarchical management of multimodal data in smart wearable devices. Background Technology

[0002] Currently, smart wearable devices are widely used in fitness trackers, smartwatches, wireless earphones, and other terminals. They integrate multiple sensors such as heart rate, accelerometer, gyroscope, and microphone to collect users' physiological signals and motion status data in real time. The ability to fuse and process multimodal information directly determines the device's accuracy in perceiving the user's state and its response speed, and is a core direction for the evolution of wearable product functions.

[0003] In one existing technology, smart wearable devices typically employ a fixed frame rate / fixed feature dimension indiscriminate processing mode. The main processing node and backup processing nodes operate synchronously for extended periods, performing uniform computational analysis on all multimodal sensor data. Internally, the device usually uses an application-specific integrated circuit (ASIC) as the main control chip, working in conjunction with multiple coprocessors to form a multi-node collaborative architecture. All nodes are continuously powered on and running, performing calculations with the same processing intensity regardless of whether the collected data contains critical information. This mode results in the device experiencing high power consumption comparable to vigorous exercise even in stable scenarios (such as sitting or sleeping), significantly shortening battery life. Another approach attempts to switch node operating states based on fixed rules, but these preset rules are difficult to adapt to changes in user scenarios, often leading to insufficient computing power during deep processing and idle resources in simple scenarios. The root cause lies in the lack of dynamic perception of data value density in existing technologies, making it impossible to allocate processing resources in real time based on the actual value of the incoming data.

[0004] In summary, existing technologies suffer from the problem that multimodal data processing resource scheduling cannot be matched with data value density in real time. Summary of the Invention

[0005] This invention provides a method and system for hierarchical management of multimodal data in smart wearable devices, in order to solve the problem that the scheduling of multimodal data processing resources and the data value density cannot be matched in real time.

[0006] In a first aspect, to address the aforementioned technical problems, the present invention provides a method for hierarchical management of multimodal data in smart wearable devices, comprising: Acquire multimodal continuous data from a wearable device containing a master node and a backup node, extract data segments from the multimodal continuous data, and perform low-pass filtering and noise reduction on the data segments to obtain clean data fragments. The core feature vector is extracted from the clean data fragment, the core feature vector is classified to obtain a density label, and the current value density is obtained by numerical conversion based on the density label. When the current value density is less than a preset density threshold, a hibernation control command is generated. According to the hibernation control command, the backup node is controlled to enter hibernation state and the power supply is cut off, switching to a single-node operation architecture. The master node performs time-series waveform feature analysis on the clean data segment, continuously outputs routine physiological state information, and calculates the amplitude of abnormal state fluctuations based on the routine physiological state information of multiple consecutive cycles. When the abnormal fluctuation amplitude of the state is greater than or equal to the preset fluctuation threshold, or when the current value density is greater than or equal to the density threshold, a wake-up coordination instruction is generated, and the backup node is controlled to resume operation and power supply is connected according to the wake-up coordination instruction, switching to a dual-active operation architecture. The clean data fragments are time-synchronized using the dual-active operating architecture to obtain an aligned data sequence. Hidden health features are extracted from the aligned data sequence, and hierarchical early warning information is generated based on the hidden health features.

[0007] In a second aspect, the present invention provides a multimodal data hierarchical management system for a smart wearable device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0008] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains clean data segments by performing sliding window truncation and low-pass filtering on the initial continuous data collected by the sensor. Then, time-domain and frequency-domain features are extracted from these clean data segments. The extracted multi-dimensional features are then dimensionality-reduced to determine the core feature vector. Finally, classification is performed based on the core feature vector to determine the current value density. This scheme can quantify the information richness of each data segment in real time, thus providing accurate judgment criteria for subsequent differentiated processing and avoiding the waste of computing power caused by applying uniform specifications to all data.

[0010] (2) This invention generates a hibernation control command to cut off the power supply to the backup node when the current value density is lower than a preset threshold, and performs time-series waveform feature analysis on the clean data fragment using a single-node operating architecture to obtain normal physiological state information; when the current value density reaches the preset threshold, it generates a wake-up coordination command to turn on the power supply to the backup node, and performs time synchronization processing on the clean data fragment using a dual-active operating architecture to extract hidden health features. This solution realizes on-demand allocation of processing resources, and can adaptively respond to scene changes without relying on pre-set fixed rules. It effectively extends the device's battery life in stable scenarios and ensures sufficient computing power supply in abnormal scenarios.

[0011] (3) In this invention, the primary and backup nodes in a dual-active architecture process clean data fragments of different modalities and align them in the time dimension. Then, the initial feature vectors are extracted from the aligned data sequences and nonlinear mapping is performed to obtain hidden health features. Subsequently, cluster analysis is performed on the hidden health features to obtain outlier feature clusters. When the density of outlier feature clusters exceeds the warning threshold, a graded warning information is generated. This scheme, through the collaborative computing of the primary and backup nodes and deep feature mining, can reveal potential physiological abnormality correlations that cannot be captured by conventional analysis of a single node, thereby improving the sensitivity and reliability of health abnormality warnings. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of a multimodal data hierarchical management method for a smart wearable device provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the standby node hibernation trigger threshold provided in the first embodiment of the present invention; Figure 3 This is a schematic diagram of the dual-active operation architecture provided in the first embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 The first embodiment of the present invention provides a method for hierarchical management of multimodal data in smart wearable devices, comprising the following steps: S11, acquire multimodal continuous data of the wearable device including the master node and the backup node, extract the multimodal continuous data to obtain data segments, perform low-pass filtering and noise reduction processing on the data segments to obtain clean data segments; S12, extract core feature vectors from the clean data fragments, classify the core feature vectors to obtain density labels, and perform numerical conversion based on the density labels to obtain the current value density; S13, when the current value density is less than the preset density threshold, a hibernation control command is generated, and the backup node is controlled to enter hibernation state and the power supply is cut off according to the hibernation control command, switching to a single-node operation architecture; S14, the master node performs time-series waveform feature analysis on the clean data segment, continuously outputs routine physiological state information, and calculates the amplitude of abnormal state fluctuations based on the routine physiological state information of multiple consecutive cycles. S15, when the abnormal fluctuation amplitude of the state is greater than or equal to the preset fluctuation threshold, or when the current value density is greater than or equal to the density threshold, a wake-up coordination instruction is generated, and the backup node is controlled to restore the running state and connect the power supply according to the wake-up coordination instruction, switching to a dual-active running architecture. S16, the clean data fragment is time-synchronized through the dual-active operation architecture to obtain an aligned data sequence, hidden health features are extracted based on the aligned data sequence, and hierarchical early warning information is generated based on the hidden health features.

[0015] In step S11, multimodal continuous data of the wearable device, including the master node and the backup node, is acquired. The multimodal continuous data is segmented to obtain data segments. The data segments are then subjected to low-pass filtering and noise reduction processing to obtain clean data fragments.

[0016] The data segments are subjected to low-pass filtering for noise reduction to obtain clean data segments, including: Calculate the fluctuation severity index of the data segment. When the fluctuation severity index is greater than a preset variance threshold, filter out the high-frequency signal components of the data segment to obtain smoothed signal data. When the fluctuation intensity index is not greater than the preset variance threshold, the data is segmented as signal smoothing data; Baseline drift correction is performed on the smoothed signal data to obtain clean data segments.

[0017] In this embodiment, the wearable device includes a master node and a backup node. The master node and backup node are two independent processing units within the wearable device. The master node is responsible for basic data acquisition and preliminary processing, while the backup node is responsible for complex feature extraction and deep analysis tasks. This embodiment acquires multimodal continuous data collected by the wearable device's sensors. The wearable device incorporates multiple sensing units, including a triaxial accelerometer and an opto-pulse wave sensor. Each sensing unit continuously outputs sensing signals according to a preset sampling frequency. All sensors use a unified sampling frequency of 50 Hz for synchronous acquisition. The triaxial accelerometer acquires 50 acceleration data points per second, and the opto-pulse wave sensor acquires 50 pulse wave data points per second. Data from each channel is sampled synchronously using the same time base.

[0018] This embodiment truncates the multimodal continuous data to obtain data segments. Wearable devices need to perform real-time analysis on the continuously flowing sensor data. However, limited by finite computing and storage resources, they cannot process the entire continuously growing data stream. Therefore, a sliding window technique is used to truncate the continuous data stream in frames. The sliding window duration is set to 2 seconds, and the sliding step size is set to 1 second. A 2-second window can cover a complete cycle of human motion or physiological signals while including a sufficient number of signal cycles, providing enough sampling points for subsequent time-domain and frequency-domain feature extraction. A 1-second sliding step size results in a 1-second data overlap between adjacent windows. This overlapping truncation method can preserve the contextual information between adjacent data segments and avoid signal feature loss caused by direct truncation. Based on the above parameters, each data segment contains 100 sampling points (50 Hz multiplied by 2 seconds), and there is a 50-sampling-point data overlap between adjacent data segments.

[0019] After obtaining the data segments, this embodiment performs low-pass filtering and denoising on the data segments to obtain clean data segments. For ease of description, this embodiment uses data collected by an accelerometer as an example to detail the specific implementation of low-pass filtering and denoising. Data collected by other types of sensors can be processed using the same procedure, and the cutoff frequency parameter of the filter can be adjusted according to the effective frequency band range of the corresponding signal. During actual wear, the signals collected by the sensors of wearable devices are mixed with various interference components. Taking the acceleration signal as an example, when the user runs or swings their arm, high-frequency mechanical vibration noise is superimposed on the signal; at the same time, low-frequency baseline drift caused by loose wear or human breathing also affects signal quality. Low-pass filtering and denoising is used to eliminate these interferences.

[0020] Low-pass filtering denoising specifically includes two steps. The first step is to filter out high-frequency signal components from the data segments to obtain smoothed signal data. In this embodiment, the variance of the acceleration amplitude within the data segment is first calculated. The acceleration amplitude is the composite amplitude of the three-axis accelerations, i.e., the square root of the sum of the squares of the three-axis components. The calculated variance is compared with a preset variance threshold. When the variance is greater than the preset variance threshold, it indicates that the current data segment is in a high-dynamic motion state, with strong high-frequency vibration noise superimposed, requiring filtering. When the variance is not greater than the preset variance threshold, it indicates that the current data segment is in a stable state with good signal quality, requiring no additional filtering, and the current data segment is directly used as smoothed signal data. For data segments requiring filtering, this embodiment uses a Butterworth low-pass filter. The cutoff frequency of the Butterworth low-pass filter is set to 5 Hz. This filter has the flattest amplitude-frequency response characteristics within the passband, effectively preserving low-frequency components below 5 Hz while suppressing high-frequency mechanical vibration noise above 5 Hz. The data is segmented and input into the filter, and the output is the smoothed signal data after filtering out high-frequency noise.

[0021] It should be noted that the preset variance threshold was set by pre-collecting acceleration data of the wearable device while it was worn on the wrist, with a sample size of no less than 200 sets, each covering four scenarios: sitting, slow walking, fast walking, and running. For each scenario, the variance of acceleration amplitude was calculated segment by segment, and the mean and standard deviation of the variance values ​​for each scenario were statistically obtained. The mean was 0.12 and the standard deviation was 0.08 for the sitting scenario, 0.28 and 0.15 for the slow walking scenario, 0.65 and 0.20 for the fast walking scenario, and 1.20 and 0.35 for the running scenario. The upper limit of variance (mean plus twice the standard deviation) for the sitting and slow walking scenarios was 0.28 and 0.58 respectively, and the lower limit of variance (mean minus twice the standard deviation) for the fast walking and running scenarios was 0.25 and 0.50 respectively. There is a clear distinguishing interval between the two groups; therefore, 0.5 was determined as the critical value for distinguishing between stable and highly dynamic scenarios. When the calculated variance is greater than 0.5, the current data segment is in a high-dynamic motion state, with strong high-frequency vibration noise superimposed, and filtering processing is required; when the variance is not greater than 0.5, it is in a stable state, the signal quality is good, and no additional filtering is required.

[0022] It should be noted that the above thresholds are based on statistical results of wrist wear. If the wearable device is worn on other parts of the body, such as the chest or ankle, the amplitude characteristics of the sensor signal will change. In such cases, the data should be collected again and the corresponding variance threshold should be determined using the same statistical method.

[0023] It is worth noting that baseline drift correction is performed on the smoothed signal data to obtain clean data segments. However, baseline drift may still exist in the smoothed signal data due to loose fitting or human breathing, meaning the signal's reference level shifts slowly over time. Baseline drift is a low-frequency interference, typically below 0.5 Hz, and partially overlaps with the frequency band of the useful signal, making it impossible to completely eliminate using the aforementioned low-pass filter. To address this issue, this embodiment employs a polynomial fitting detrending algorithm for correction. Specifically, using the smoothed signal data as input, a third-order polynomial is used to fit the overall trend of the signal. The resulting polynomial curve characterizes the change in baseline drift over time. The fitted baseline drift curve is subtracted point-by-point from the original smoothed signal data, causing the signal's reference level to return to near the zero mean, resulting in the clean data segment. Thus, this embodiment completes the entire preprocessing flow from the original sensor signal to the clean data segment, providing a data foundation for subsequent feature extraction and value density determination.

[0024] It should be noted that the parameters such as the time length of the sliding window, the sliding step size, and the cutoff frequency of the filter can be adjusted according to the specific sensor type and application scenario of the wearable device.

[0025] In step S12, core feature vectors are extracted from the clean data fragments, and the core feature vectors are classified to obtain density labels. Numerical transformation is then performed based on the density labels to obtain the current value density, including: Calculate the time-domain statistics and frequency-domain spectral features of the clean data segment, and concatenate the time-domain statistics and the frequency-domain spectral features to obtain a multi-dimensional feature vector; A linear transformation is performed on the multidimensional feature vector to obtain the core feature vector; The core feature vectors are classified to obtain density labels; The initial value base is determined based on the density label, and the total number of sampling points of the pure data fragment is counted. The initial value base and the total number of sampling points are weighted and calculated to obtain the current value density.

[0026] This embodiment extracts core feature vectors from the clean data segment and classifies these vectors to obtain density labels. Based on these density labels, a numerical conversion is performed to obtain the current value density. This embodiment first calculates the time-domain statistics and frequency-domain spectral features of the clean data segment. Time-domain statistics reflect the amplitude distribution characteristics of the signal in the time dimension, while frequency-domain spectral features reflect the energy distribution characteristics of the signal in the frequency dimension. In the time domain, this embodiment calculates the mean, variance, and zero-crossing rate of the clean data segment as time-domain statistics. The mean is the arithmetic mean of the amplitudes at each sampling point within the clean data segment, reflecting the DC component level of the signal; the variance is the dispersion of the amplitude at each sampling point relative to the mean, reflecting the intensity of signal fluctuations; and the zero-crossing rate is the ratio of the number of times the signal waveform crosses zero levels to the total number of sampling points, reflecting the oscillation frequency characteristics of the signal. In the frequency domain, this embodiment uses a Fast Fourier Transform algorithm to perform time-frequency conversion on the clean data segment, transforming the time-domain signal to the frequency domain to obtain a spectral distribution. Then, the peak frequency and frequency band energy distribution are extracted from the spectral distribution as frequency-domain spectral features. The frequency domain spectral features include five dimensions: peak frequency and the energy proportion of each of the four sub-bands. The peak frequency is the frequency point corresponding to the maximum amplitude in the spectrum, reflecting the dominant oscillation frequency of the signal. The frequency band energy distribution is the proportion of the sum of squared amplitudes in each sub-band to the total energy of the spectrum after dividing the spectrum into multiple sub-bands according to frequency intervals. It should be noted that time-domain statistics include three dimensions: mean, variance, and zero-crossing rate, while frequency domain spectral features include five dimensions: peak frequency and the energy proportion of each sub-band. Since time-domain statistics and frequency domain spectral features differ in data structure and numerical range, this embodiment performs dimension alignment processing, mapping both types of features to the same feature space. Then, the aligned feature sequences are concatenated to obtain a multi-dimensional feature vector. For example, the multi-dimensional feature vector has eight dimensions.

[0027] To reduce redundant information in the feature dimensions, this embodiment performs a linear transformation on the multidimensional feature vectors to obtain core feature vectors. The linear transformation is implemented using principal component analysis (PCA). Specifically, this embodiment constructs a feature set using multidimensional feature vectors corresponding to multiple historically collected clean data segments. The mean of this feature set in each dimension is calculated, and the mean of each multidimensional feature vector in each dimension is subtracted from the mean of that dimension, making the mean of each dimension zero, resulting in a centered feature matrix. Then, the covariance matrix of the centered feature matrix is ​​calculated, where each element represents the degree of linear correlation between two feature dimensions. Eigenvalue decomposition is performed on the covariance matrix to obtain multiple initial eigenvalues ​​and their corresponding initial feature vectors. The dimensionality reduction projection matrix is ​​obtained through offline training before the wearable device leaves the factory. The training samples are at least 1000 pre-collected multidimensional feature vector samples labeled with density tags, covering various usage scenarios such as sitting, walking slowly, walking briskly, running, and sleeping. During training, the covariance matrix and eigenvalues ​​of the sample set are calculated according to the above process. When the cumulative contribution reaches 85%, the corresponding initial feature vectors are truncated and concatenated to form a dimension-reduced projection matrix. After training, the matrix parameters are stored in the non-volatile memory of the wearable device. In actual use, the current multidimensional feature vector is directly multiplied by the stored dimension-reduced projection matrix to obtain the core feature vector.

[0028] After obtaining the core feature vector, this embodiment performs classification processing on the core feature vector. Specifically, the feature values ​​of each dimension in the core feature vector are weighted and summed to obtain a comprehensive score. The weights are determined by pre-collecting a large number of sample core feature vectors with labeled density. For each feature dimension, the difference between the mean of samples with a first density label and the mean of samples with a second density label in that dimension is calculated. The mean difference of each dimension is normalized and used as the weight of that dimension. After the weights are determined, for the current core feature vector to be classified, the feature values ​​of each dimension are multiplied by the corresponding weights and summed to obtain the comprehensive score. When the comprehensive score is greater than a preset density classification threshold, the density label of the first value is output; when the comprehensive score is not greater than the preset density classification threshold, the density label of the second value is output. The preset density classification threshold is determined by using ten-fold cross-validation during the offline training phase. The pre-collected sample set with labeled density is randomly divided into ten equal parts, and one part is used as the validation set and the remaining nine parts are used as the training set in turn, repeated ten times. In each validation, the mean of the combined scores of the two classes of samples is calculated using the training set. The median value between the two means is taken as a candidate threshold, and the classification accuracy is evaluated on the validation set. The average accuracy of ten validations is calculated, and the candidate threshold corresponding to the highest average accuracy is taken as the density classification threshold. For example, a density label of 1 represents that the clean data segment contains rich exercise physiological information, while a density label of 0 represents that the clean data segment contains relatively simple information. The density label is used to characterize the information richness of the current clean data segment.

[0029] After obtaining the density label, this embodiment determines the initial value base based on the density label. Specifically, a mapping rule is set: when the density label is a first value, a first initial value base is assigned; when the density label is a second value, a second initial value base is assigned. For example, a density label of 1 represents that the clean data segment contains rich exercise physiological information, and the initial value base is assigned to 90; a density label of 0 represents that the clean data segment contains relatively simple information, and the initial value base is assigned to 40. Simultaneously, this embodiment counts the total number of sampling points in the clean data segment. The clean data segment originates from the truncation of continuous data by a sliding window, and the number of sampling points contained within the window is the total number of sampling points. The total number of sampling points is multiplied by the sampling point weight coefficient and then added to the initial value base to obtain the current value density. Based on the data segment truncation by the sliding window in S11 containing 100 sampling points, the initial value base is either 40 or 90. To ensure that the contribution of the total number of sampling points is on the same order of magnitude as the contribution of the initial value base, the sampling point weight coefficient is set to 5.

[0030] In step S13, when the current value density is less than a preset density threshold, a hibernation control command is generated. According to the hibernation control command, the standby node is controlled to enter a hibernation state and the power supply is cut off, switching to a single-node operating architecture.

[0031] In this embodiment, the current value density is the grading result corresponding to the current clean data segment. When the current value density is less than a preset density threshold, the current clean data segment is determined to belong to the low value level; when the current value density is greater than or equal to the preset density threshold, the current clean data segment is determined to belong to the high value level. The preset density threshold is the boundary value used to divide the low value level and the high value level. In this embodiment, the current value density calculated in step S12 is compared with the preset density threshold. The preset density threshold is used to determine whether the current data segment belongs to a low information content scenario or a high information content scenario. The threshold is determined based on the following: during the calculation of value density, the density label is used to distinguish the degree of information richness. When the density label is 1 (information rich), the initial value base is 90, and when the density label is 0 (information simple), the initial value base is 40. After weighting the initial value base with the total number of sampling points, the value density values ​​corresponding to the low information content scenario and the high information content scenario are distributed in two intervals, forming a transition zone between the two intervals. The middle value of the transition zone is set as the preset density threshold. In actual comparison, when the current value density is less than 565, it indicates that the information content of the current clean data fragment is low, triggering the sleep control process; when the current value density is greater than or equal to 565, it indicates that the information content of the current clean data fragment is high, triggering the wake-up control process. In other implementations, the preset density threshold can also be adjusted according to the specific application scenario and battery life requirements of the wearable device.

[0032] When the current value density is less than the preset density threshold, it indicates that the physiological information contained in the current clean data fragment is relatively simple, and there is no need to call the backup node for processing. At this time, the hibernation control process is triggered. This embodiment generates a hibernation control instruction. The hibernation control instruction is a digital instruction containing an opcode and a parameter field. The opcode is used to identify that the instruction type is a hibernation operation, and the parameter field is used to specify the identification information of the target node.

[0033] After generating the hibernation control command, this embodiment controls the backup node to enter hibernation mode according to the hibernation control command. The wearable device adopts a multi-node collaborative architecture, including a master node and a backup node. The master node is responsible for basic heart rate acquisition and preliminary data processing, while the backup node is responsible for complex gait feature extraction and deep analysis tasks. The wearable device's firmware stores a hardware configuration mapping table, which records the correspondence between each node and the power supply circuit. This embodiment reads the hardware configuration mapping table to find the power supply circuit number corresponding to the backup node.

[0034] After identifying the backup node and its corresponding power supply circuit, this embodiment sends a state switching signal to the backup node via the internal communication bus. The state switching signal instructs the backup node to save its current operating context and enter a sleep state. Upon receiving the state switching signal, the backup node saves its current register state, cached data, and task progress to non-volatile memory, completing the context saving operation. Then, it stops the clock signal and switches its operating state to sleep mode. After confirming that the backup node has entered a sleep state, this embodiment cuts off the power supply to the backup node. Specifically, the control interface of the power management module is invoked to send a disconnect command to the electronic switch on the power supply circuit corresponding to the backup node. This electronic switch is a transistor switch controlled by the power management module. By disconnecting this electronic switch, the power output from the power supply circuit to the backup node is physically cut off. It should be noted that the power supply cut off is to the backup node's dedicated power path; the power supply to the primary node is unaffected, and the primary node continues to operate. After the backup node's power supply is cut off, the wearable device switches to a single-node operating architecture. In this single-node operating architecture, only the primary node is in operation, responsible for performing subsequent data processing and analysis tasks. Thus, this embodiment completes the standby node hibernation and power cut-off process based on value density judgment.

[0035] like Figure 2 As shown in the figure, this diagram illustrates the threshold determination for the change of current value density over time and the triggering of standby node hibernation. The solid blue line represents the curve of current value density over time, the dashed red line represents the preset density threshold (value 565), and the light orange area marks the low-value range where the current value density is below the density threshold. When the current value density is consistently below the density threshold, the system determines that the current data segment belongs to a low-information-content scenario, triggers the standby node hibernation strategy, and switches to a single-node operating architecture to reduce device power consumption.

[0036] In step S14, the master node performs time-series waveform feature analysis on the clean data segment, continuously outputs normal physiological state information, and calculates the amplitude of abnormal state fluctuations based on the normal physiological state information of multiple consecutive cycles.

[0037] Specifically, the master node performs time-series waveform feature analysis on the purified data segment and continuously outputs routine physiological state information, including: The master node extracts the temporal features of the clean data segment to obtain the physiological waveform periodic sequence. Based on the physiological waveform periodic sequence, waveform features are extracted to obtain a multidimensional physiological feature vector; The multidimensional physiological feature vectors are mapped to a preset physiological state classification space to obtain physiological state classification results and classification confidence. When the classification confidence level is greater than the preset confidence level threshold, routine physiological state information is generated and continuously output based on the physiological state classification result.

[0038] This embodiment first extracts the temporal features of the purified data segment through the master node to obtain the physiological waveform periodic sequence. The purified data segment is continuous sampling data of photoplethysmography (PPG). After reading the purified data segment from its internal buffer, the master node uses a sliding window mechanism combined with an autocorrelation algorithm to perform periodic analysis on the signal. The autocorrelation algorithm identifies the periodic structure of the signal by calculating the similarity of the signal at different time delays: for a purified data segment of length N, the original signal is multiplied point-by-point with the signal delayed by τ sampling points and accumulated to obtain the autocorrelation value corresponding to the delay time τ; by traversing multiple delay times, when a local peak appears in the autocorrelation value corresponding to a certain delay time, it indicates that the signal has significant periodic repetition at that delay time, and this delay time is the duration of one cardiac cycle. Based on the identified multiple periodic peak points, the time interval sequence between adjacent peaks is output, which is the physiological waveform periodic sequence. For example, a clean data segment with a length of 5 seconds is input, the sampling frequency is 50 Hz, and a total of 250 sampling points are included. The autocorrelation algorithm detects a significant autocorrelation peak at a delay time of about 1 second, identifies 6 complete cardiac cycles, and outputs a physiological waveform cycle sequence containing 6 time intervals between adjacent peaks.

[0039] After obtaining the physiological waveform cycle sequence, this embodiment extracts waveform features based on the physiological waveform cycle sequence to obtain a multidimensional physiological feature vector. Specifically, for each cardiac cycle in the physiological waveform cycle sequence, waveform morphological parameters within that cycle are extracted, such as the peak value of the main wave, the trough value of the dicrotic wave, and the slope of the rising branch of the waveform within each cycle. The peak value of the main wave is the maximum signal amplitude within that cycle; the trough value of the dicrotic wave is the minimum signal amplitude after the main wave; and the slope of the rising branch of the waveform is the amplitude difference between the peak value and the previous trough value divided by the corresponding time difference. All extracted waveform morphological parameters are combined in cycle order to construct a multidimensional physiological feature vector. For example, three feature dimensions are extracted for each cycle, resulting in a total of 18 feature dimensions for 6 cardiac cycles. To further simplify the processing, the features from each cycle can be averaged to obtain a fixed-dimensional multidimensional physiological feature vector.

[0040] After obtaining the multidimensional physiological feature vector, this embodiment maps the multidimensional physiological feature vector to a preset physiological state classification space to obtain the physiological state classification result and classification confidence. The preset physiological state classification space is a pre-constructed multidimensional feature space in which different regions correspond to different physiological state categories, including resting state, movement state, sleep state, etc. The preset physiological state classification space is implemented through a pre-constructed physiological feature parameter interval mapping table. Specifically, multidimensional physiological feature vector samples with known physiological state labels are pre-collected. For each physiological state category, the feature values ​​of all samples under that category are collected. The mean and standard deviation of each feature under each state category are calculated. The numerical interval corresponding to the mean plus or minus twice the standard deviation is taken as the normal interval of that state category in that feature dimension. The correspondence between each physiological state category and the numerical interval of each feature dimension is stored as a physiological feature parameter interval mapping table. The preset physiological state classification space is defined by the above physiological feature parameter interval mapping table. In this space, each dimension corresponds to a feature term of the multidimensional physiological feature vector. Each state category has a corresponding normal numerical interval in each dimension. The normal intervals of all state categories in all dimensions together define the category partition boundary in this space.

[0041] For example, the heart rate range corresponding to the resting state is 60 to 80 beats per minute, the heart rate range corresponding to the exercise state is 100 to 150 beats per minute, and the heart rate range corresponding to the sleep state is 50 to 65 beats per minute. In actual classification, each feature value in the multidimensional physiological feature vector is compared one by one with the numerical range corresponding to each state category in the physiological feature parameter range mapping table. The number of hits for each feature value falling into each state category range is counted, and the state category with the most hits is output as the physiological state classification result. Simultaneously, the proportion of hits to the total feature dimension is calculated, and this proportion is mapped to a numerical range of 0 to 100 as the classification confidence level.

[0042] After obtaining the classification confidence score, this embodiment compares the classification confidence score with a preset confidence threshold. This threshold is determined by pre-collecting a large number of multi-dimensional physiological feature vector samples with known physiological state labels when constructing the physiological feature parameter interval mapping table, and using leave-one-out cross-validation to verify the classification accuracy of the interval mapping table. Specifically, the sample set is divided into multiple subsets, with one subset used as the validation set in turn, and the remaining subsets used to determine the feature intervals for each state category. The samples in the validation set are input into the interval mapping table for classification, and the classification confidence score of each sample and whether the classification result is correct are recorded. After collecting all the records of the validation samples, the validation samples are sorted in descending order of classification confidence score, and the classification accuracy corresponding to different confidence scores is calculated sequentially. When the classification accuracy reaches a preset standard, the confidence score value is determined as the confidence threshold of this embodiment.

[0043] When the classification confidence level is greater than a preset confidence threshold, it indicates that the current classification result is reliable. At this point, routine physiological state information is generated based on the physiological state classification result. The routine physiological state information includes the currently identified physiological state category label and key physiological values ​​extracted from the clean data fragment. For example, when the classification result indicates a resting state, the current heart rate value and state label are extracted and packaged into a standard format physiological state information package as routine physiological state information. When the classification confidence level is not greater than the preset confidence threshold, it indicates that the current classification result is not reliable. In this embodiment, the routine physiological state information output from the previous cycle is used. The above process is continuously executed by the master node in a single-node operating architecture, continuously generating and outputting routine physiological state information at fixed time intervals.

[0044] This embodiment obtains the normal physiological state information for multiple consecutive cycles, and calculates the amplitude of abnormal state fluctuations based on the normal physiological state information for multiple consecutive cycles.

[0045] The calculation of abnormal fluctuation amplitude based on the conventional physiological state information over multiple consecutive cycles includes: Physiological values ​​are extracted from the conventional physiological state information of multiple consecutive cycles to obtain a physiological value sequence; The average of the absolute values ​​of the differences between adjacent values ​​in the physiological numerical sequence is calculated as the amplitude of abnormal fluctuations in the state.

[0046] Specifically, physiological values ​​from multiple consecutive cycles are extracted from continuously output routine physiological state information. For example, heart rate values ​​are extracted from five consecutively output routine physiological state information packets, resulting in a physiological value sequence containing five heart rate values. Then, the absolute value of the difference between adjacent values ​​in this physiological value sequence is calculated. The sum of these absolute values ​​is divided by the number of differences to obtain the average value, which is used as the amplitude of abnormal state fluctuations. This amplitude is used to subsequently determine whether to trigger the wake-up operation of the backup node.

[0047] In step S15, when the abnormal fluctuation amplitude of the state is greater than or equal to the preset fluctuation threshold, or when the current value density is greater than or equal to the density threshold, a wake-up coordination instruction is generated. According to the wake-up coordination instruction, the backup node is controlled to restore the running state and connect the power supply, switching to a dual-active running architecture.

[0048] The process of controlling the backup node to resume operation and connecting its power supply according to the wake-up coordination command, and switching to a dual-active operating architecture, includes: According to the wake-up coordination command, the power supply to the backup node is turned on and the backup node is controlled to complete the power-on initialization; Establish a data synchronization channel between the backup node and the primary node, and switch to a dual-active operating architecture.

[0049] This embodiment compares the abnormal state fluctuation amplitude calculated in step S14 with the preset fluctuation threshold. Simultaneously, this embodiment also compares the current value density calculated in step S12 with the preset density threshold mentioned in step S13. When the abnormal state fluctuation amplitude is greater than or equal to the preset fluctuation threshold, it indicates a significant change in the current physiological state, and the conventional basic analysis under the single-node operating architecture is insufficient to capture potential health risks, requiring the activation of a backup node for in-depth collaborative analysis. Alternatively, when the current value density is greater than or equal to the preset density threshold, it indicates that the current clean data fragment itself contains rich physiological information, requiring the invocation of a backup node for in-depth processing. This embodiment triggers the wake-up control process when either of the above two conditions is met.

[0050] It should be noted that the preset fluctuation threshold is used to determine whether the fluctuation of the current physiological state has reached an abnormal level. This threshold is determined by collecting physiological value sequence samples from the wearable device in a normal, stable state. Stable states include scenarios such as sitting, walking slowly, and sleeping, which do not require triggering in-depth analysis. For each sample, the average of the absolute values ​​of the differences between adjacent heart rate values ​​is calculated in the same way as in step S14 to obtain the fluctuation amplitude corresponding to that sample. Statistical analysis is performed on the fluctuation amplitudes of all collected stable state samples to calculate the mean and standard deviation of these fluctuation amplitudes. Based on the statistical distribution characteristics, the value corresponding to the mean plus twice the standard deviation is taken as the upper bound of the normal range of fluctuation amplitude in a stable state, i.e., the preset fluctuation threshold. In other implementations, the preset fluctuation threshold can also be adjusted according to the specific application scenario and user type of the wearable device.

[0051] This embodiment generates a wake-up coordination instruction. The wake-up coordination instruction is a digital instruction containing an opcode and a parameter field. The opcode identifies the instruction type as a wake-up operation, and the parameter field specifies the identification information of the target node. Exemplarily, the wake-up coordination instruction is generated by the master node and sent to the standby node in a dormant state via the internal control bus. After generating the wake-up coordination instruction, this embodiment controls the standby node to resume operation and connect to power supply according to the wake-up coordination instruction. Specifically, after receiving the wake-up coordination instruction, the standby node parses the opcode and parameter field in the instruction to confirm that a wake-up operation needs to be performed. The power management module inside the standby node responds to the instruction by sending a connection signal to the electronic switch on its power supply circuit. The electronic switch is a transistor switch controlled by the power management module. After receiving the connection signal, the electronic switch switches to a closed state, restoring power output to the standby node, thus providing power to the standby node. After power is connected, this embodiment controls the standby node to complete power-on initialization. Specifically, the standby node's internal clock circuit starts, generating a stable operating clock signal; the standby node reads the boot program from non-volatile memory and loads the firmware image; the standby node resets its internal registers to their default state and clears the context information saved before hibernation; the standby node performs a self-test to verify the functional integrity of critical hardware modules. After completing the above initialization operations, the standby node enters a ready state, waiting to establish data communication with the master node.

[0052] After the standby node completes power-on initialization, this embodiment establishes a data synchronization channel between the standby node and the master node. Specifically, the master node sends a synchronization request message to the standby node via its internal control bus. This synchronization request message includes the current system timestamp, data frame format definition, and buffer allocation information. Upon receiving the synchronization request message, the standby node parses the configuration parameters and calibrates its own system time to match that of the master node; it configures its communication interface parameters according to the data frame format definition; and it reserves a shared data area between the master node and the standby node based on the buffer allocation information. After confirming that the communication link between the two nodes is normal, the master node and the standby node enter a synchronized working state.

[0053] After the data synchronization channel between the backup node and the primary node is established, the wearable device switches to a dual-active operating architecture. In this architecture, both the primary and backup nodes are operational simultaneously. The primary node continues to collect and process basic data, while the backup node participates in subsequent deep collaborative computing tasks. Thus, this embodiment completes the entire process of backup node wake-up based on fluctuation anomalies or high-value density, power-on, power-on initialization, data synchronization channel establishment, and dual-active operating architecture switching.

[0054] like Figure 3 As shown in the figure, this diagram illustrates the collaborative operation of value density and abnormal fluctuation amplitude over time after waking up the standby node and switching to a dual-active operating architecture. The blue curve represents the trend of value density change, the red curve represents the trend of abnormal fluctuation amplitude change, the blue dashed line is the density threshold reference line, and the red dashed line is the fluctuation threshold reference line. When the abnormal fluctuation amplitude reaches the fluctuation threshold or the value density reaches the density threshold, the standby node is triggered to wake up, and the dual-active operating architecture is entered. The primary node and the standby node work together to achieve in-depth analysis and hierarchical early warning under abnormal conditions.

[0055] In step S16, the clean data fragment is time-synchronized using the dual-active operating architecture to obtain an aligned data sequence. Hidden health features are extracted from the aligned data sequence, and hierarchical early warning information is generated based on the hidden health features.

[0056] Specifically, the clean data fragment is time-synchronized using the dual-active operating architecture to obtain an aligned data sequence, and hidden health features are extracted from the aligned data sequence, including: The clean data fragments are processed by the master node in the dual-active operating architecture to obtain the first data sequence; The clean data fragments are processed by the backup nodes in the dual-active operating architecture to obtain the second data sequence; The first data sequence and the second data sequence are time-aligned to obtain an aligned data sequence; Feature parameters are extracted from the aligned data sequence to form an initial feature vector; The initial feature vector is nonlinearly mapped to obtain hidden health features.

[0057] In this embodiment, the master node in the dual-active architecture processes the clean data segment to obtain a first data sequence. After receiving the clean data segment, the master node extracts the photoplethysmography (PPG) signal component from it and arranges them in chronological order to form the first data sequence. In this embodiment, the backup node in the dual-active architecture synchronously processes the clean data segment to obtain a second data sequence. After waking up and completing power-on initialization, the backup node synchronously acquires the clean data segment with the master node, extracts the acceleration signal component from it, and arranges it in chronological order to form the second data sequence.

[0058] After obtaining the first data sequence and the second data sequence, this embodiment performs time alignment on the first data sequence and the second data sequence to obtain an aligned data sequence. Since the master node and the backup node process signals from different channels, there may be a slight time offset between the two data sequences. To eliminate this offset, specifically, the first data sequence is used as a reference sequence, and the second data sequence is used as a sliding sequence. The sliding sequence is gradually slid along the time axis relative to the reference sequence by a preset number of offset steps. At each offset step, the cross-correlation coefficient between the reference sequence and the slid sequence is calculated. The cross-correlation coefficient is the sum of the products of the data values ​​of the two sequences at the same time point, representing the similarity between the two sequences at that offset. After traversing all offset steps, the offset step corresponding to the maximum value of the cross-correlation coefficient is determined as the time offset between the first data sequence and the second data sequence. Based on the time offset, a translation operation is performed on the first data sequence and the second data sequence: the lagging one is translated forward by the time offset on the time axis, so that the feature points corresponding to the same physiological event in the two sequences are aligned in time, and an aligned data sequence with consistent time dimension is output. Based on the time offset, the first data sequence and the second data sequence are shifted and resampled to make the two sequences completely aligned in the time dimension, and the aligned data sequence is output.

[0059] After obtaining the aligned data sequence, this embodiment extracts feature parameters from the aligned data sequence to form an initial feature vector. The feature parameters include peak spacing and energy distribution parameters. The peak spacing is extracted by performing peak detection on the photoplethysmography (PPG) signal channel in the aligned data sequence, identifying R-wave peak points, and calculating the time interval between adjacent R-wave peak points. The energy distribution parameters are extracted by performing frequency domain transformation on the acceleration signal channel in the aligned data sequence and calculating the energy proportion within multiple frequency bands. The extracted peak spacing and energy distribution parameters are arranged in a fixed order and combined to form the initial feature vector.

[0060] After obtaining the initial feature vector, this embodiment performs a nonlinear mapping on the initial feature vector to obtain hidden health features. The nonlinear mapping is implemented using a multilayer perceptron model. The multilayer perceptron model is a pre-trained neural network model, whose training process is completed offline before the wearable device leaves the factory. After training, the model parameters are stored in the wearable device's non-volatile memory for direct loading and calling during subsequent actual use. The training of the multilayer perceptron model involves collecting a large amount of sample data, such as no less than 10,000 samples. Each sample contains a set of aligned data sequences and corresponding real health status labels. Real health status labels are obtained by simultaneously collecting the wearable device's sensor data and using a synchronously connected 12-lead electrocardiograph to collect the user's electrocardiogram (ECG). Professional medical personnel interpret the waveforms recorded by the ECG and, combined with the user's actual activity records, label each time window as either 'normal' or 'abnormal'. After labeling, a second medical professional reviews the labeling results; samples with consistent labeling results are retained, while samples with inconsistent labeling results are discarded.

[0061] For each retained training sample, an initial feature vector is extracted from the aligned data sequence as described above. Using the extracted initial feature vector as input and its corresponding real health status label as a supervision signal, the model parameters are iteratively updated using backpropagation and gradient descent optimizer. The training objective is to minimize the difference between the model output and the real label. After training, the multilayer perceptron model has the following structure: an input layer with the same number of neurons as the initial feature vector; at least one hidden layer, each containing multiple neurons, with neurons in adjacent layers fully connected by learnable weight and bias parameters. Each neuron linearly weights and sums its inputs, then processes the sum using a non-linear activation function to obtain the output; and a bottleneck layer, whose output vector has a lower dimension than the input vector. This low-dimensional vector output by the bottleneck layer is the hidden health feature. In this embodiment, the dimension of the hidden health feature is... It is set to 16 dimensions.

[0062] It is worth noting that the multilayer perceptron model in this embodiment includes an input layer, two hidden layers, and a bottleneck layer. The number of neurons in the input layer is consistent with the dimension of the initial feature vector. The first hidden layer contains 256 neurons, the second hidden layer contains 128 neurons, and the bottleneck layer contains 16 neurons. Each hidden layer uses a modified linear unit as the activation function, and the bottleneck layer uses a linear activation function.

[0063] The generation of tiered early warning information based on the hidden health characteristics includes: Cluster analysis was performed on the hidden health features to obtain outlier feature clusters; The cluster density is obtained by calculating the ratio of the total number of feature samples within the outlier cluster to the distribution volume. When the cluster density is greater than a preset warning threshold, the timestamp and feature extreme value corresponding to the outlier feature cluster are extracted, and the warning level identifier is determined according to the density level range in which the current value density is located, and the hierarchical warning information is encapsulated and generated.

[0064] After obtaining the hidden health features, this embodiment performs cluster analysis on the hidden health features to obtain outlier feature clusters. Specifically, this embodiment collects hidden health features from multiple consecutive time windows as a sample set, and uses a density-based spatial clustering algorithm to divide the sample set. For each sample point in the sample set, a neighborhood region is formed with that point as the center and a preset neighborhood radius as the radius, and the number of sample points contained in this region is counted; when the number of sample points in the neighborhood reaches a preset minimum number of samples, the point is marked as a core point; the core points are connected into clusters through the neighborhood overlap relationship between the core points; sample points that do not meet the core point condition and do not belong to any cluster are marked as outliers. The outliers are aggregated and output as outlier feature clusters.

[0065] The preset neighborhood radius is determined using the elbow rule. Specifically, in the offline phase, the distance between each sample point and all other sample points is calculated and sorted in ascending order. A curve showing the change in distance with the nearest neighbor index is plotted, and the distance value corresponding to the inflection point with the largest curvature change in the curve is determined as the neighborhood radius. If the inflection point cannot be clearly determined, 30% of the average distance between all pairs of sample points is taken as the default value of the neighborhood radius. The preset minimum number of included samples is set to the feature dimension plus one. The basis for this is that in density-based clustering algorithms, the minimum number of included samples should be at least the feature dimension plus one to ensure that the core points have statistical significance in the feature space. In this embodiment, the hidden health feature dimension is 16-dimensional, so the minimum number of included samples is set to 17. In other implementations, if the feature dimension changes, the minimum number of included samples should be adjusted synchronously to the new feature dimension plus one.

[0066] After obtaining the outlier feature clusters, this embodiment calculates the ratio of the total number of feature samples within the outlier feature cluster to the distribution volume to obtain the cluster density. The total number of feature samples within the outlier feature cluster is the number of sample points contained in that cluster. The distribution volume is determined by calculating the covariance matrix of all sample points within the outlier feature cluster, performing eigenvalue decomposition on the covariance matrix to obtain each eigenvalue, multiplying the square roots of each eigenvalue together, and then multiplying by a constant factor to obtain the distribution volume of the cluster in the feature space. In this embodiment, the constant factor is [missing information]. The volume factor of a 3D hypersphere is calculated using the following formula: ; in To conceal dimensions of health characteristics, The gamma function is used. The cluster density is obtained by dividing the total number of feature samples by the distribution volume.

[0067] When the cluster density exceeds a preset warning threshold, this embodiment extracts the timestamp and feature extreme value corresponding to the outlier cluster, determines the warning level identifier based on the density level range of the current value density, and encapsulates it to generate graded warning information. The preset warning threshold is determined by collecting historical data from wearable devices before and after known abnormal events. The abnormal events are labeled using the interpretation results of a synchronously collected 12-lead electrocardiograph as the gold standard, and include at least 50 abnormal event records. At the same time, at least 200 sets of historical data under normal conditions are collected. The cluster density value corresponding to each set of data is calculated according to the aforementioned steps, and the cluster density distribution under normal conditions and the cluster density distribution under abnormal conditions are obtained respectively. The median value between the 95th quantile of the cluster density distribution under normal conditions and the 5th quantile of the cluster density distribution under abnormal conditions is determined as the preset warning threshold.

[0068] When the cluster density exceeds the preset warning threshold, a clustered physiological abnormality is identified. Based on this determination, the timestamp corresponding to each sample point in the outlier cluster is extracted, along with the extreme values ​​of the cluster that deviate furthest from the center in each feature dimension. The timestamps and feature extreme values ​​are then encapsulated in a preset format to generate tiered warning information. The preset format is a pre-agreed data transmission format between the wearable device and the monitoring terminal. This format defines the data structure of the warning information, including a header identifier field to identify the information type as a health warning, a timestamp field to record the time of the anomaly, a feature extreme value field to carry the abnormal deviation values ​​of each feature dimension, and a check field to verify data integrity. The construction of this preset format determines the required information fields based on the processing capabilities and display requirements of the monitoring terminal, allocating a fixed byte length to each field and arranging them in the order of header identifier, timestamp, feature extreme value, and check field. The header identifier field is fixed as a two-digit hexadecimal value, where the first digit identifies the information type and the second digit identifies the protocol version number. The timestamp field uses a standard time format. The feature extreme value fields are arranged sequentially according to each feature dimension. The check field is generated using a cyclic redundancy check algorithm and is used by the receiving end to verify data integrity. After the format is determined, the field definitions, byte allocation, and check rules are embedded in the communication protocol stack of both the wearable device and the monitoring terminal. After arranging each field in order according to this format and filling in the corresponding values, the encapsulation of the graded warning information is completed. Specifically, when the current value density is less than the density threshold, the warning level is identified as L1; when the current value density is greater than or equal to the density threshold, the warning level is identified as L2; ​​and the graded warning information is sent to the monitoring terminal.

[0069] In summary, this invention obtains value density by analyzing multimodal sensor data in real time, and dynamically switches the power supply status and operating architecture of backup nodes according to the value density, thereby achieving adaptive matching of processing resources and data value, and taking into account both equipment endurance and anomaly warning sensitivity.

[0070] The second embodiment of the present invention provides a multimodal data hierarchical management system for a smart wearable device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0071] It should be noted that the multimodal data hierarchical management system for smart wearable devices provided in this embodiment of the invention is used to execute all the process steps of the multimodal data hierarchical management method for smart wearable devices described in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0072] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0073] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for hierarchical management of multimodal data in a smart wearable device, characterized in that, include: Acquire multimodal continuous data from a wearable device containing a master node and a backup node, extract data segments from the multimodal continuous data, and perform low-pass filtering and noise reduction on the data segments to obtain clean data fragments. The core feature vector is extracted from the clean data fragment, the core feature vector is classified to obtain a density label, and the current value density is obtained by numerical conversion based on the density label. When the current value density is less than a preset density threshold, a hibernation control command is generated. According to the hibernation control command, the backup node is controlled to enter hibernation state and the power supply is cut off, switching to a single-node operation architecture. The master node performs time-series waveform feature analysis on the clean data segment, continuously outputs routine physiological state information, and calculates the amplitude of abnormal state fluctuations based on the routine physiological state information of multiple consecutive cycles. When the abnormal fluctuation amplitude of the state is greater than or equal to the preset fluctuation threshold, or when the current value density is greater than or equal to the density threshold, a wake-up coordination instruction is generated, and the backup node is controlled to resume operation and power supply is connected according to the wake-up coordination instruction, switching to a dual-active operation architecture. The clean data fragments are time-synchronized using the dual-active operating architecture to obtain an aligned data sequence. Hidden health features are extracted from the aligned data sequence, and hierarchical early warning information is generated based on the hidden health features.

2. The multimodal data hierarchical management method for smart wearable devices according to claim 1, characterized in that, The step of performing low-pass filtering and noise reduction on the data segments to obtain clean data segments includes: Calculate the fluctuation severity index of the data segment. When the fluctuation severity index is greater than a preset variance threshold, filter out the high-frequency signal components of the data segment to obtain smoothed signal data. When the fluctuation intensity index is not greater than the preset variance threshold, the data is segmented as signal smoothing data; Baseline drift correction is performed on the smoothed signal data to obtain clean data segments.

3. The multimodal data hierarchical management method for smart wearable devices according to claim 1, characterized in that, The process of extracting core feature vectors from the clean data fragments, classifying the core feature vectors to obtain density labels, and performing numerical conversion based on the density labels to obtain the current value density includes: Calculate the time-domain statistics and frequency-domain spectral features of the clean data segment, and concatenate the time-domain statistics and the frequency-domain spectral features to obtain a multi-dimensional feature vector; A linear transformation is performed on the multidimensional feature vector to obtain the core feature vector; The core feature vectors are classified to obtain density labels; The initial value base is determined based on the density label, and the total number of sampling points of the pure data fragment is counted. The initial value base and the total number of sampling points are weighted and calculated to obtain the current value density.

4. The multimodal data hierarchical management method for smart wearable devices according to claim 1, characterized in that, The step involves performing time-series waveform feature analysis on the clean data segment through the master node, and continuously outputting routine physiological state information, including: The master node extracts the temporal features of the clean data segment to obtain the physiological waveform periodic sequence. Based on the physiological waveform periodic sequence, waveform features are extracted to obtain a multidimensional physiological feature vector; The multidimensional physiological feature vectors are mapped to a preset physiological state classification space to obtain physiological state classification results and classification confidence. When the classification confidence level is greater than the preset confidence level threshold, routine physiological state information is generated and continuously output based on the physiological state classification result.

5. The multimodal data hierarchical management method for smart wearable devices according to claim 1, characterized in that, The calculation of the abnormal fluctuation amplitude of the state based on the normal physiological state information of multiple consecutive cycles includes: Physiological values ​​are extracted from the conventional physiological state information of multiple consecutive cycles to obtain a physiological value sequence; The average of the absolute values ​​of the differences between adjacent values ​​in the physiological numerical sequence is calculated as the amplitude of abnormal fluctuations in the state.

6. The multimodal data hierarchical management method for smart wearable devices according to claim 1, characterized in that, The step of controlling the backup node to resume operation and connect power supply according to the wake-up coordination command, and switching to a dual-active operation architecture, includes: According to the wake-up coordination command, the power supply to the backup node is turned on and the backup node is controlled to complete the power-on initialization; Establish a data synchronization channel between the backup node and the primary node, and switch to a dual-active operating architecture.

7. The multimodal data hierarchical management method for smart wearable devices according to claim 1, characterized in that, The step of performing time synchronization processing on the clean data fragment using the dual-active operating architecture to obtain an aligned data sequence, and extracting hidden health features based on the aligned data sequence, includes: The clean data fragments are processed by the master node in the dual-active operating architecture to obtain the first data sequence; The clean data fragments are processed by the backup nodes in the dual-active operating architecture to obtain the second data sequence; The first data sequence and the second data sequence are time-aligned to obtain an aligned data sequence; Feature parameters are extracted from the aligned data sequence to form an initial feature vector; The initial feature vector is nonlinearly mapped to obtain hidden health features.

8. The multimodal data hierarchical management method for smart wearable devices according to claim 1, characterized in that, The step of generating tiered early warning information based on the hidden health characteristics includes: Cluster analysis was performed on the hidden health features to obtain outlier feature clusters; The cluster density is obtained by calculating the ratio of the total number of feature samples within the outlier cluster to the distribution volume. When the cluster density is greater than a preset warning threshold, the timestamp and feature extreme value corresponding to the outlier feature cluster are extracted, and the warning level identifier is determined according to the density level range in which the current value density is located, and the hierarchical warning information is encapsulated and generated.

9. A multimodal data hierarchical management system for a smart wearable device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method described in any one of claims 1 to 8.

Images