A method and system for patch position correction

Abstract

This invention discloses a patch position correction method and system, relating to the fields of biomedical engineering and digital health technology. The invention employs hardware-level time synchronization and motion-artifact coupling modeling technology, effectively solving the problems of time delay drift and motion interference in signal acquisition under dynamic environments, significantly improving the basic quality of ECG signals. Secondly, by establishing a body surface reference coordinate system and a pose compensation parameter generation algorithm, the micro-displacement state of the patch can be accurately identified and intuitive position adjustment guidance can be generated, ensuring the patch remains in the optimal measurement position and fundamentally reducing signal attenuation due to poor contact. The multi-dimensional signal quality evaluation system, through an intelligent weighted fusion mechanism, adapts to the quality discrimination requirements under different motion states. Combined with adaptive filtering and deep learning enhancement technology, it maintains the morphological integrity of the QRS waveform while effectively suppressing various interference components, enabling the system to maintain excellent ECG signals even under strenuous exercise.

Description

Technical Field

[0001] This invention relates to the fields of biomedical engineering and digital health technology, and in particular to a method and system for patch position correction. Background Technology

[0002] In the field of dynamic health monitoring, wearable ECG devices continuously collect electrocardiogram signals through skin-contact electrode patches. These devices need to maintain signal quality during daily human activities. Their electrode patches are usually made of flexible materials that adhere to human skin and integrate wireless transmission modules to achieve mobile monitoring. Current systems are generally equipped with basic position detection functions, which determine the contact status between the patch and the skin through impedance measurement or simple optical sensing.

[0003] In existing technologies, some solutions improve signal stability through multi-sensor fusion, while others introduce inertial measurement units to capture motion posture and combine machine learning algorithms to distinguish motion artifacts from ECG signal characteristics; or they use adaptive filtering technology to dynamically adjust signal processing parameters according to motion state; and they automatically select the optimal signal channel by combining multi-electrode redundancy with signal quality assessment algorithms.

[0004] However, in practical applications, such methods still face several limitations. The spectral overlap between motion artifacts and physiological signals limits the effectiveness of traditional filtering methods; insufficient time synchronization accuracy of multi-sensor data introduces new signal distortions; individual skin characteristics make it difficult to unify the criteria for judging contact status; baseline drift caused by micro-displacement between electrodes and skin in dynamic environments has not been effectively suppressed; these factors affect the monitoring reliability of wearable ECG devices in real-world scenarios. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a patch position correction method and system to solve the problems of motion artifact interference, multi-sensor data asynchrony, poor individual adaptability, and signal quality degradation caused by patch micro-displacement in existing wearable ECG devices under dynamic environments.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a patch position correction method, which includes,

[0009] Step S1: Acquire multimodal sensing data, which includes at least one electrocardiogram signal and a three-axis acceleration signal;

[0010] Step S2: Time synchronization and alignment of multi-source data are performed by using hardware timestamps or a unified clock and combining resampling to eliminate delay and drift.

[0011] Step S3: Construct a motion-artifact coupling model based on acceleration, and use separation / adaptive noise cancellation of the reference channel constraint to separate the motion artifact component from the ECG;

[0012] Step S4: Generate pose compensation parameters based on the separation results and output prompt information for guiding patch wearing;

[0013] Step S5: Based on the signal quality index and motion state, perform adaptive filtering to adjust the ECG processing parameters in real time;

[0014] Step S6: When the signal quality is below the threshold, the recalibration process is triggered.

[0015] As a preferred embodiment of the patch position correction method of the present invention, in step S3, separation is achieved by constrained independent component analysis or adaptive noise cancellation, wherein acceleration is used as a reference input, and the coupling model includes characteristic parameters of motion intensity, main frequency band and phase relationship.

[0016] As a preferred embodiment of the patch position correction method of the present invention, the method includes: constructing a motion feature dictionary and performing pattern recognition. The dictionary is obtained by feature extraction from typical scene data. The motion patterns include stationary, walking, running, and climbing stairs. A classification algorithm with transition suppression and confidence assessment is adopted.

[0017] As a preferred embodiment of the patch position correction method of the present invention, the signal quality index includes at least one of the following: signal-to-noise ratio, baseline drift degree, power frequency interference intensity, heartbeat morphology stability and R wave detection integrity. The index is weighted and fused to obtain an interval score and a threshold and hysteresis are set.

[0018] The steps for calculating the signal quality index and weighting the fusion include:

[0019] 1) Perform windowing and preprocessing, within a length of Within the sliding window, basic preprocessing of the ECG sequence with bandpass and power frequency notch is performed, and robust statistics of R wave position and QRS amplitude are retained within the same window.

[0020] 2) Define the calculation method for individual indicators, and obtain Sub-ratings with consistent direction The details are as follows:

[0021] Within the window, the signal power of the QRS neighborhood and the noise power of the QRS exclusion region are taken to obtain:

[0022] ,

[0023] ,

[0024] in, Let be the signal-to-noise ratio at time t. The mean square power after bandpass in the QRS neighborhood. The mean square power after removing the QRS neighborhood within the window. It is a very small constant. For the upper and lower limits of the linear mapping, For the reason The self-score obtained by linear normalization This represents the interval clipping operator;

[0025] The baseline is extracted using a low-pass filter, and its root mean square to QRS amplitude ratio is calculated:

[0026] And mapped as:

[0027] ,

[0028] in, It is the ratio of baseline drift amplitude to heart rate amplitude. Baseline components extracted by low-pass extraction The root mean square, The median peak-to-trough amplitude of the QRS complex within the window. The lower and upper limits of the tolerance range, Baseline drift sub-score;

[0029] Calculate power frequency interference intensity: Assume power frequency ,bandwidth ;Pick for The spectral power of its first harmonic neighborhood and, for Total power, get the ratio And mapped as:

[0030] ,

[0031] in, For power frequency interference sub-scores, This is the tolerance range. This represents the proportion of power frequency;

[0032] Defining cardiac morphological stability:

[0033] Equal-length heartbeat segments aligned with R And the amplitude is normalized, using the running template. Median of correlation coefficient:

[0034] ,Pick ,

[0035] in, The Pearson correlation coefficient is used. For the median operator, The median value of the correlation coefficient. For morphological stability sub-scores, this sub-score is not included in the fusion when the effective number of heartbeats is less than 3.

[0036] Define R-wave detection integrity: Let the number of R-waves detected within the window be... Estimate the expected number based on the median of RR ,get:

[0037] , ,

[0038] in, Based on the median RR interval Estimated expected heart rate The number of R waves detected within the window. The median of the RR interval. The length of the sliding window. To round down, R-wave integrity sub-score;

[0039] 3) After obtaining the sub-scores within the window, perform weighted fusion:

[0040] ,

[0041] in, For a moment Signal quality zonal scoring For the set of indicators participating in the integration, For the first Item rating, , For the first Item weights, satisfying , For availability factor, It is a very small number;

[0042] right Define two thresholds and hysteresis: low threshold With high threshold ,when If judged as poor, when If the condition is deemed "good," and falls between the two, the previous condition remains unchanged, while a minimum dwell time is set. ;in, These are low and high thresholds, respectively. This is the minimum dwell time.

[0043] As a preferred embodiment of the patch position correction method described in this invention, the adaptive filtering includes a segmented filtering strategy based on motion patterns, a parameter-motion intensity mapping relationship, and a smoothing / soft-start mechanism. When the score deteriorates, a stronger suppression is adopted to preserve the integrity of the QRS waveform.

[0044] As a preferred embodiment of the patch position correction method of the present invention, the pose compensation parameters are obtained through a displacement and attitude estimation model, and the estimated quantities in the sensor coordinate system are mapped to the body surface reference system using a coordinate transformation matrix to calculate the compensation amount, and the output is a position adjustment suggestion or signal alignment parameters.

[0045] The pose compensation parameters are obtained through a displacement and attitude estimation model, and the steps include:

[0046] a. Establish four coordinate systems:

[0047] Sensor System That is, the coordinates of the accelerometer itself;

[0048] Patch Panel , and Fixed connection, take ;

[0049] Body surface reference system In the body surface reference system, Along the head-to-toe direction, Along the left and right directions, The outer normal to the body surface;

[0050] Earth system In the Earth system Pointing in the opposite direction of gravity;

[0051] The state is defined within the sliding window as follows: ,in For patch relative The tilt angle, For the center of gravity of the surface mount Small displacements, with the tangential component being dominant;

[0052] b. Perform gravity vector extraction and attitude calculation:

[0053] triaxial acceleration Low-pass filter ,by Pitch / roll and yaw are calculated, but are not observable when acceleration is used alone; use the initial calibration values. ; thus constructing , The rotation; at the initial, correctly fitted, and stationary calibration moment. storage And at that time and relative rotation , For one-time calibration, the settings are provided by patch geometry and body position.

[0054] c. Perform differential coordinate transformation and mapping:

[0055] ,

[0056] in, Indicated in the body surface reference system The target vector below, For sensor system The mapped quantity below, For calibration time from arrive The rotation matrix, For calibration time from arrive The rotation matrix, For the current from arrive The rotation matrix;

[0057] d. Perform attitude deviation extraction and displacement estimation:

[0058] In step c, Received With calibration value Compare and extract tilt deviation Small angle from exist The change in the tangential plane is approximated;

[0059] Displacement is achieved using a gravity-bandpass-integral-stationary zero-velocity correction link:

[0060] ,

[0061] in, In order to be in Linear acceleration in the system, For the measured total acceleration, , Represents the magnitude of a vector, after After bandpass, map to step c. Within the static / easing mode window, a zero-velocity constraint is applied to reset the velocity drift, resulting in a small displacement. tangential component;

[0062] e. Compensation vector and output aperture:

[0063] The attitude error is combined with the displacement to form the compensation amount. ;in, For the compensation triplet, Pick The tangential component, Pick exist The resultant tilt angle on the plane;

[0064] The interface indicates movement towards the head / foot. Millimeters, move left / right Millimeters, clockwise / counterclockwise micro-rotation The degree of prompt, when When the value is below the display lower limit, it remains silent; when the value exceeds the recalibration threshold, it is incorporated into the quality judgment and recalibration is triggered.

[0065] As a preferred embodiment of the patch position correction method of the present invention, the time synchronization management includes hardware timestamp alignment, link delay jitter estimation and sampling clock drift correction, and unified resampling is performed after alignment.

[0066] As a preferred embodiment of the patch position correction method of the present invention, the method further includes inputting the separated ECG segments and reference features into a deep learning network for signal purification. The network performs end-to-end enhancement or residual correction of ECG quality and adopts a combination of offline training and online fine-tuning.

[0067] Secondly, the present invention provides a patch position correction system, comprising,

[0068] The multimodal acquisition module is used to acquire at least one ECG signal and a three-axis acceleration signal at a preset sampling rate, and output a raw data stream with timestamps.

[0069] The time synchronization management module is used to align ECG and acceleration data based on hardware timestamps or a unified clock, estimate and compensate for link delay jitter and sampling clock drift, and generate equally spaced aligned sequences.

[0070] The motion pattern recognition module is used to extract time-domain / frequency-domain / time-frequency features from acceleration data and output motion pattern labels and their confidence scores for subsequent parameter selection.

[0071] The coupling modeling and separation module is used to establish the motion-artifact coupling relationship based on acceleration as a reference, and uses constrained independent component analysis and / or adaptive noise cancellation to separate motion artifacts in the ECG, thus obtaining a preliminary purified ECG.

[0072] The signal quality assessment module is used to calculate at least one of the following: signal-to-noise ratio, baseline drift, power frequency interference, morphological stability, and R-wave integrity, and to generate a signal quality score and threshold determination.

[0073] The pose compensation parameter generation module is used to estimate the displacement / attitude deviation of the patch in the body surface reference frame based on the motion mode and coupling model, and output pose compensation parameters or position adjustment suggestions.

[0074] The adaptive filtering module is used to select the filter type and parameters based on the motion pattern and signal quality score, and to perform parameter smoothing and hysteresis control to output the target purified ECG.

[0075] The guidance and prompting module is used to prompt the pose compensation parameters and recalibration commands through the user interface, sound and light or vibration, or to send the parameters to the host computer / mobile terminal;

[0076] The modules are executed under the control of the processor in the order of acquisition-alignment-separation-evaluation-compensation-filtering-prompt or a combination thereof.

[0077] As a preferred embodiment of the patch position correction system of the present invention, the system further includes a processor and a memory, the memory storing a computer program, and the processor being configured to execute the computer program to implement the steps of the patch position correction method described above.

[0078] The multimodal acquisition module includes flexible surface electrodes, an analog front-end and an analog-to-digital converter, and a triaxial inertial measurement unit;

[0079] The time synchronization management module uses hardware timestamp alignment combined with resampling;

[0080] The guidance and prompting module includes at least one of sound, light, or vibration devices;

[0081] The system is equipped with a wireless communication unit for uploading data and receiving parameters;

[0082] The coupling modeling and separation module includes a deep learning purification subunit for enhancing the separated ECG; when the signal quality score is below a threshold, the guidance prompt module triggers a recalibration prompt.

[0083] The beneficial effects of this invention are as follows: This invention adopts hardware-level time synchronization and motion-artifact coupling modeling technology, which effectively solves the problems of time delay drift and motion interference in signal acquisition under dynamic environment, and significantly improves the basic quality of ECG signals; secondly, by establishing a body surface reference coordinate system and a pose compensation parameter generation algorithm, it can accurately identify the micro-displacement state of the patch and generate intuitive position adjustment guidance, so that the patch is always kept in the optimal measurement position, fundamentally reducing the phenomenon of signal attenuation caused by poor contact.

[0084] The multi-dimensional signal quality evaluation system proposed in this invention adapts to the quality discrimination requirements under different motion states through an intelligent weighted fusion mechanism, avoiding the limitations of single-index evaluation. Combining adaptive filtering and deep learning enhancement techniques, it maintains the morphological integrity of the QRS waveform while effectively suppressing various interference components, enabling the system to maintain excellent ECG signal quality even under strenuous exercise.

[0085] This invention offers excellent individual adaptability and ease of operation. Its one-time calibration and automatic recalibration mechanisms lower the barrier to entry, while multimodal cues such as sound, light, and vibration provide intuitive guidance for patch application. Without increasing hardware complexity, it achieves synergistic optimization of patch pose compensation and signal quality enhancement, effectively ensuring the reliability and practicality of dynamic ECG monitoring, making it particularly suitable for long-term home health monitoring and clinical diagnostic applications. Attached Figure Description

[0086] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0087] Figure 1 This is a flowchart illustrating a patch position correction method in one embodiment.

[0088] Figure 2 This is a schematic diagram of the frame of a patch position correction system in one embodiment. Detailed Implementation

[0089] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0090] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0091] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0092] This application proposes a patch position correction method, combined with Figure 1 As shown, the method includes:

[0093] Step S1: Acquire multimodal sensing data, which includes at least one ECG signal and a three-axis acceleration signal;

[0094] In this embodiment, at least one ECG signal refers to the raw analog / digital sequence of a single-lead patch ECG, and the three-axis acceleration signal refers to the x / y / z axis acceleration sequence in the sensor's body coordinate system. Both are timestamped with the sampling clock. Numerical specifications: ECG default sampling rate 250 Hz (adjustable from 125-500 Hz to meet QRS morphology fidelity and power consumption trade-offs), acceleration default 100 Hz (adjustable from 50-200 Hz to balance attitude / glide observation and bandwidth), timestamp resolution 1 ms (adjustable from 0.1-5 ms, set according to the main control crystal oscillator / RTC capability). Necessary anomaly handling: When the acceleration channel experiences short-term saturation or missing measurements, that segment is marked as unusable for reference, and subsequent separation / compensation steps are skipped or downgraded using signal quality gating.

[0095] Step S2: Time synchronization and alignment of multi-source data are performed by using hardware timestamps or a unified clock and combining resampling to eliminate delay and drift.

[0096] Specifically, time synchronization and alignment refer to correcting sampling timing differences and link delay jitter under the same time base, outputting parallel sequences with equal intervals and in phase. Numerical specifications: The default synchronization update period is 1 s (adjustable from 0.5-5 s), the maximum allowed relative clock deviation is 50 ppm (adjustable from 20-100 ppm), and the resampling step size for each alignment does not exceed 20% of the original sampling period. Optional implementations: When hardware timestamps are missing, a short-window cross-correlation estimation of the relative delay is used before linear resampling; under unified clock conditions, only equal-interval resampling and phase alignment are performed. Necessary exception handling: When the estimated jitter exceeds the set upper limit, an alignment failure flag is output, and the alignment parameters of the previous version are maintained until the next cycle.

[0097] Step S3: Construct a motion-artifact coupling model based on acceleration, and use separation / adaptive noise cancellation of the reference channel constraint to separate the motion artifact component from the ECG;

[0098] For example, the reference channel constraint refers to constructing a reference input with the same origin as motion artifacts using acceleration, driving the separator / canceller to estimate and eliminate noise components related to the reference, while preventing excessive suppression of the ECG principal components. Numerical specifications: The adaptive noise canceller has a default length of 16 orders (adjustable from 8 to 64 orders), a default step size coefficient of 0.05 (adjustable from 0.01 to 0.20), and an update period consistent with acceleration sampling; the correlation trigger threshold defaults to |r|≥0.10 (adjustable from 0.05 to 0.30) to enable cancellation, avoiding false suppression at low correlation. Optional implementations: When ICA prior is insufficient, it degenerates into NLMS / RLS single-reference cancellation; when the reference is strongly non-stationary, it enables step size upper limit and leakage term restriction updates. Necessary anomaly handling: When a near-neighbor heartbeat (e.g., R ± several milliseconds) is detected, the weight update is frozen for a short window to reduce the risk of QRS morphology being eroded by adaptive learning.

[0099] Step S4: Generate pose compensation parameters based on the separation results and output prompt information for guiding patch wearing;

[0100] Similarly, the pose compensation parameter is defined as a combination of tangential displacement and small-angle tilt in the body surface reference frame, used to indicate that the patch should move slightly along the head-to-toe / left-right direction and rotate slightly clockwise / counterclockwise. Numerical specifications: The default refresh cycle is 1 second (adjustable 0.5-2 seconds), the minimum display resolution is 1 mm and 1° (adjustable 0.5-2 mm, 0.5-3°), and the hysteresis threshold is 1.5 times the lower limit of the display to suppress jitter. Optional implementation: When the device has both vibration and screen display prompting capabilities, screen display is prioritized; a short vibration is triggered only after three consecutive prompts with no response. Necessary anomaly handling: When the user is in a vigorous motion mode, only the compensation amount is cached without immediate prompting; a concentrated prompt is given after the quality score recovers to the set upper limit.

[0101] Step S5: Based on the signal quality index and motion state, perform adaptive filtering to adjust the ECG processing parameters in real time;

[0102] Optionally, real-time adjustment refers to a limited and smooth update within the parameter space to avoid waveform distortion caused by large transient modifications. Numerical specifications: The parameter smoothing time constant is 2 s by default (adjustable from 0.5-5 s); the minimum dwell time for switching segmented strategies is 2 s by default (adjustable from 1-5 s); and the upper limit of the degradation suppression coefficient when the score deteriorates is 1.5 times the normal value. Necessary anomaly handling: When two consecutive windows of scores are unavailable, the system reverts to a conservative filtering configuration with fixed parameters and records the event for subsequent recalibration.

[0103] Step S6: When the signal quality is below the threshold, a recalibration process is triggered;

[0104] Furthermore, the recalibration process includes restoring the patch-body surface reference relationship and initial quality baseline, and clearing the historical adaptive state. Numerical specifications: The trigger condition is a score below the low threshold for a duration of at least 2 seconds (adjustable from 1 to 5 seconds); the default waiting time for static steady-state detection during recalibration is 3 seconds (adjustable from 2 to 6 seconds); the default retry interval is 30 seconds (adjustable from 10 to 60 seconds). Necessary anomaly handling: If three recalibrations fail, a manual check of patch adhesion and electrode contact is prompted; the algorithm maintains conservative parameters until manual intervention is completed.

[0105] In one embodiment, in step S3, separation is achieved by constrained independent component analysis or adaptive noise cancellation, wherein acceleration is used as a reference input and the coupling model includes characteristic parameters of motion intensity, main frequency band and phase relationship.

[0106] In this embodiment, motion intensity can be robustly statistically represented by the acceleration modulus, and the relationship between the dominant frequency band and phase is given by the coherence of the short window spectrum and the reference-target, which is used to drive the start-up and stop of the separator and the step size scheduling. Numerical caliber: The intensity is divided into three levels by default (stationary, slow, and violent), and the coherence threshold is set in conjunction with the intensity level; the separation start-up and stop criteria are evaluated every 0.5-1 s. Optional implementation: When coherence fluctuations are frequent, a hysteresis loop is formed using a dual threshold and a minimum dwell time to avoid repeated separator resets.

[0107] In one embodiment, a motion feature dictionary is constructed and pattern recognition is performed. The dictionary is obtained by feature extraction from typical scene data. The motion patterns include stationary, walking, running and climbing stairs. A classification algorithm with transition suppression and confidence evaluation is used.

[0108] Specifically, transition suppression refers to introducing hysteresis and short-term majority voting at mode boundaries to reduce jitter, while confidence assessment provides a 0-1 quantized confidence level along with the mode label for subsequent parameter mapping weighting. Numerical specifications: window length defaults to 2 s, step size 0.5 s (both adjustable), minimum dwell time for transition states 0.5 s (adjustable 0.2-1.0 s), low confidence threshold defaults to 0.5 (adjustable 0.4-0.7). Necessary exception handling: when the confidence level remains below the threshold, the mode input is treated as unknown, and the system operates using a conservative filtering and delayed prompting strategy.

[0109] In one embodiment, the signal quality indicators include at least one of the following: signal-to-noise ratio, baseline drift, power frequency interference intensity, heartbeat morphology stability, and R-wave detection integrity. The indicators are weighted and fused to obtain interval scores, and thresholds and hysteresis are set.

[0110] For example, individual indicators are first standardized in a unified direction, and then a total score is obtained by combining availability gating and weights. This score is used to control the filtering intensity and recalibration trigger. Numerical specifications: The indicator update cycle is 0.5 s by default (adjustable from 0.2 to 1.0 s), the sliding window overlap is no less than 50% (adjustable from 50% to 80%), and the weight table maintains independent entries according to motion mode. Necessary anomaly handling: When the R-wave count is insufficient or the template correlation is unavailable, the corresponding sub-score is automatically removed and the weights are recalculated.

[0111] The steps for calculating signal quality metrics and weighting fusion include:

[0112] 1) Perform windowing and preprocessing, within a length of Within the sliding window (default) Adjustable Bandpassing of ECG sequences (default) The basic preprocessing of power frequency notch filtering preserves robust statistics (such as median peak-to-trough amplitude, denoted as ) of R-wave position and QRS amplitude within the same window. );

[0113] 2) Define the calculation method for individual indicators, and obtain Sub-ratings with consistent direction The details are as follows:

[0114] Within the window, the signal power of the QRS neighborhood and the noise power of the QRS exclusion region are taken to obtain:

[0115] ,

[0116] ,

[0117] in, Let be the signal-to-noise ratio at time t. The mean square power after pass-through in the QRS neighborhood (e.g., R±60ms). The mean square power after removing the QRS neighborhood within the window. Minimal constant (default) ), Upper and lower bounds for linear mapping (default) Adjustable ), For the reason The self-score obtained by linear normalization This represents the interval clipping operator;

[0118] With low-pass (cutoff) Extract the baseline and calculate its root mean square to QRS amplitude ratio:

[0119] And mapped as:

[0120] ,

[0121] in, It is the ratio of baseline drift amplitude to heart rate amplitude. Baseline components extracted by low-pass extraction The root mean square, The median peak-to-trough amplitude of the QRS complex within the window. The lower and upper limits of the tolerance range (default) Adjustable ), Baseline drift sub-score;

[0122] Calculate power frequency interference intensity: Assume power frequency (default In the 60Hz region, take 60 (bandwidth). (default );Pick for The spectral power of its first harmonic neighborhood and, for Total power, get the ratio And mapped as:

[0123] ,

[0124] in, For power frequency interference sub-scores, Tolerance range (default) Adjustable ), This represents the proportion of power frequency;

[0125] Defining cardiac morphological stability:

[0126] Equal-length heartbeat segments aligned with R And the amplitude is normalized, using the running template. Median of correlation coefficient:

[0127] ,Pick ,

[0128] in, The Pearson correlation coefficient is used. For the median operator, The median value of the correlation coefficient. For morphological stability sub-scores, this sub-score is not included in the fusion when the effective number of heartbeats is less than 3.

[0129] Define R-wave detection integrity: Let the number of R-waves detected within the window be... Estimate the expected number based on the median of RR ,get:

[0130] , ,

[0131] in, Based on the median RR interval Estimated expected heart rate The number of R waves detected within the window. The median of the RR interval. The length of the sliding window. To round down, R-wave integrity sub-score;

[0132] 3) After obtaining the sub-scores within the window, perform weighted fusion:

[0133] ,

[0134] in, For a moment Signal quality zonal scoring For the set of indicators participating in the integration, For the first Item rating, , For the first Item weights, satisfying (Still mode default) Vigorous exercise can improve Weights to each ), Availability factor (e.g., effective heart rate < 3) If the R test fails... ), Minimal number (default) );

[0135] right Define two thresholds and hysteresis: low threshold With high threshold (default Adjustable ),when If judged as poor, when If the condition is deemed "good," and falls between the two, the previous condition remains unchanged, while a minimum dwell time is set. (default Adjustable To reduce frequent switching; among which, These are low and high thresholds, respectively. Minimum dwell time; the threshold source is recommended to use the quantile method of offline validation sets: manually labeled "available / unavailable" samples. The distribution is determined by taking the upper quantile of the unusable distribution and the lower quantile of the usable distribution as... The initial values ​​are determined, and a set of threshold tables is maintained according to different motion modes;

[0136] Specifically, signal quality is broken down into five observable dimensions, and consistent sub-scores are given within a unified time window. In noise-dominated scenarios, signal-to-noise ratio and power frequency interference are more sensitive to short-term availability within the window. In slow-drift scenarios, the amplitude ratio of baseline drift better reflects the recoverability of subsequent filtering, while morphological stability and R-wave integrity are related to clinical interpretability. The weights are not fixed as a set of constants, but are linked to motion patterns to adapt to different degradation mechanisms of patches in resting and dynamic states. The fusion process adopts a normalized weighting method with availability gating to avoid lowering the overall score due to the failure of individual indicators. The threshold and hysteresis settings emphasize the quantile method derived from the data distribution and introduce minimum dwell time to suppress state jitter. The resulting interval-based score can directly drive the intensity switching of adaptive filtering and can also serve as a criterion for recalibration triggering.

[0137] In one embodiment, adaptive filtering includes a segmented filtering strategy based on motion patterns, a parameter-motion intensity mapping relationship, and a smoothing / soft-start mechanism, which downgrades to stronger suppression while preserving the integrity of the QRS waveform when the score deteriorates.

[0138] Similarly, the slow start mechanism refers to gradually reaching the target parameter in a linear or exponential manner over a set time when switching to the strong suppression level, reducing abrupt changes. Numerical definition: The slow start time is 1 second by default (adjustable from 0.5 to 3 seconds), and the maximum step change in parameter cannot exceed 20% of the target value. Optional implementation: When the score recovers to above the high threshold for three consecutive windows, automatically revert to the normal suppression level and execute the same slow exit strategy.

[0139] In one embodiment, the pose compensation parameters are obtained through a displacement and attitude estimation model, and the estimated values ​​in the sensor coordinate system are mapped to the body surface reference system using a coordinate transformation matrix to calculate the compensation amount. The output is a position adjustment suggestion or signal alignment parameters.

[0140] Optionally, the output aperture is discretized in millimeters / degrees with accompanying directional text, and compensation values ​​below the display lower limit are silenced to reduce user interference. Numerical aperture: The display lower limit is 2 mm and 2° by default (adjustable to 1-3 mm and 1-3°), and the minimum interval for continuous prompts in the same direction is 10 s by default (adjustable to 5-30 s). When the recalibration threshold is exceeded, the prompts are merged into a single high-priority prompt. Necessary anomaly handling: When the compensation value alternates signs in three consecutive evaluations and the amplitude is close to the display lower limit, the prompts are suppressed and only the log is logged.

[0141] Pose compensation parameters are obtained through a displacement and attitude estimation model, and the steps include:

[0142] a. Establish four coordinate systems:

[0143] Sensor System That is, the coordinates of the accelerometer itself;

[0144] Patch Panel , and Fixed connection, acceptable ;

[0145] Body surface reference system In the body surface reference system, Along the head-to-toe direction, Along the left and right directions, The outer normal to the body surface;

[0146] Earth system In the Earth system Pointing in the opposite direction of gravity;

[0147] The state is defined within the sliding window as follows: ,in For patch relative The tilt angle, For the center of gravity of the surface mount Small displacements, with the tangential component being dominant;

[0148] b. Perform gravity vector extraction and attitude calculation:

[0149] triaxial acceleration Low-pass filter (Default window duration 0.5-1.0s, cutoff 2-3Hz), with Pitch / roll and yaw are calculated, but are not observable when acceleration is used alone; use the initial calibration values. (Static calibration time record); thus constructing , The rotation; at the initial, correctly fitted, and stationary calibration moment. storage And at that time and relative rotation , For one-time calibration, the settings are provided by patch geometry and body position.

[0150] c. Perform differential coordinate transformation and mapping:

[0151] ,

[0152] in, Indicated in the body surface reference system The target vector below, For sensor system The mapped quantity below, For calibration time from arrive The rotation matrix, For calibration time from arrive The rotation matrix, For the current from arrive The rotation matrix; this formula eliminates the influence of overall pose changes on the first-order results in a relatively calibrated difference form;

[0153] d. Perform attitude deviation extraction and displacement estimation:

[0154] In step c, Received With calibration value Compare and extract tilt deviation Small angle from exist The change in the tangential plane is approximated;

[0155] Displacement is achieved using a gravity-bandpass-integral-stationary zero-velocity correction link:

[0156] ,

[0157] in, In order to be in Linear acceleration in the system, For the measured total acceleration, , Represents the magnitude of a vector, after After bandpass, map to step c. Within the static / easing mode window, a zero-velocity constraint is applied to reset the velocity drift, resulting in a small displacement. tangential component;

[0158] e. Compensation vector and output aperture:

[0159] The attitude error is combined with the displacement to form the compensation amount. ;in, For the compensation triplet, Pick The tangential component, Pick exist The resultant tilt angle on the plane;

[0160] The interface indicates movement towards the head / foot. Millimeters, move left / right Millimeters, clockwise / counterclockwise micro-rotation The degree of prompt, when When the value is below the display lower limit, remain silent; when the value exceeds the recalibration threshold, incorporate the quality judgment and trigger recalibration.

[0161] Low-pass window 0.5-1.0s (adjustable 0.3-2.0s), bandpass (Adjustable) Zero velocity criterion: The magnitude of acceleration deviates from gravity by less than 100%. And the variance is lower than Lasting 0.4 seconds, From calibration (remains constant when there is no magnetic sensing) The display lower limit and recalibration threshold are maintained separately in the three modes of stationary / walking / running;

[0162] Specifically, the above process uses a one-time calibration to lock the relative attitude between the patch and the body surface, and uses the low-frequency component of acceleration to recover pitch and roll during operation. Yaw is handled by the calibration item, thus bypassing the difficulty of observing yaw using only acceleration. The core approach is to map the current attitude to the body surface reference frame through a calibration-current differential transformation. This way, even if the overall body position of the subject changes, it will not directly affect the estimation of the relative attitude. The displacement is obtained by integrating the low-frequency acceleration after gravity removal, and zero-velocity constraints are applied to suppress drift during static or slow-moving segments. Therefore, a small displacement at the millimeter level is obtained, which corresponds exactly to the slippage of the patch on the tangential plane of the skin. The compensation amount is output as a combination of tangential displacement and small angular tilt, which can be easily converted into a visual guide to move a few millimeters in a certain direction and rotate a few degrees. At the same time, it is linked with the signal quality score-threshold-hysteresis to avoid frequent prompts during slight jitter.

[0163] In one embodiment, time synchronization management includes hardware timestamp alignment, link latency jitter estimation and sampling clock drift correction, and unified resampling after alignment;

[0164] Furthermore, drift correction estimates the sampling period deviation using a linear model and updates it periodically. After alignment, unified resampling employs band-limited interpolation or piecewise linear interpolation to control spectral leakage. Numerical scope: The drift estimation window is 30 s by default (adjustable from 10-60 s), and the period change in a single correction does not exceed 10 ppm. Necessary anomaly handling: When link jitter exceeds the correctable range, it is marked as unreliable for alignment, and the reference weight is reduced in subsequent fusion.

[0165] In one embodiment, the method further includes inputting the separated ECG segments and reference features into a deep learning network for signal purification, the network performing end-to-end enhancement or residual correction of ECG quality, and employing a combination of offline training and online fine-tuning.

[0166] In this embodiment, the network input is an ECG segment and a spliced ​​vector of motion-related reference features, and the output is a purified ECG or residual term. Numerical caliber: The learning rate and batch size for online fine-tuning are set to conservatively small values, and the fine-tuning data is limited to segments with good quality scores. Necessary anomaly handling: When computational resources are insufficient or temperature rise exceeds limits, network inference is automatically disabled and the network degenerates into a traditional filtering path, maintaining the same input / output interface.

[0167] This embodiment also provides a patch position correction system, combined with Figure 2 As shown, the system includes:

[0168] The multimodal acquisition module is used to acquire at least one ECG signal and a three-axis acceleration signal at a preset sampling rate, and output a raw data stream with timestamps.

[0169] The time synchronization management module is used to align ECG and acceleration data based on hardware timestamps or a unified clock, estimate and compensate for link delay jitter and sampling clock drift, and generate equally spaced aligned sequences.

[0170] The motion pattern recognition module is used to extract time-domain / frequency-domain / time-frequency features from acceleration data and output motion pattern labels and their confidence scores for subsequent parameter selection.

[0171] The coupling modeling and separation module is used to establish the motion-artifact coupling relationship based on acceleration as a reference, and uses constrained independent component analysis and / or adaptive noise cancellation to separate motion artifacts in the ECG, thus obtaining a preliminary purified ECG.

[0172] The signal quality assessment module is used to calculate at least one of the following: signal-to-noise ratio, baseline drift, power frequency interference, morphological stability, and R-wave integrity, and to generate a signal quality score and threshold determination.

[0173] The pose compensation parameter generation module is used to estimate the displacement / attitude deviation of the patch in the body surface reference frame based on the motion mode and coupling model, and output pose compensation parameters or position adjustment suggestions.

[0174] The adaptive filtering module is used to select the filter type and parameters based on the motion pattern and signal quality score, and to perform parameter smoothing and hysteresis control to output the target purified ECG.

[0175] The guidance and prompting module is used to prompt the pose compensation parameters and recalibration commands through the user interface, sound and light or vibration, or to send the parameters to the host computer / mobile terminal;

[0176] Each module, under the control of the processor, executes in the order of acquisition-alignment-separation-evaluation-compensation-filtering-prompt or a combination thereof, in order to achieve patch pose compensation and motion artifact suppression;

[0177] Specifically, to ensure real-time performance, each module employs fixed time slots and priority scheduling, with the priority order being: Acquisition / Alignment > Separation / Evaluation > Compensation / Filtering > Prompt. Numerical specifications: The end-to-end processing latency target is less than 200 ms (adjustable from 100-300 ms), and the CPU utilization target is less than 60% at the maximum sampling rate configuration. Necessary anomaly handling: When the latency exceeds the target threshold, non-critical steps (such as deep purification) are shut down in a preset order to meet the real-time requirements of the core link.

[0178] In one embodiment, the multimodal acquisition module includes flexible surface electrodes, an analog front-end and an analog-to-digital converter, and a triaxial inertial measurement unit.

[0179] The time synchronization management module uses hardware timestamp alignment combined with resampling;

[0180] The guidance and prompting module includes at least one of sound, light, or vibration devices;

[0181] The system is equipped with a wireless communication unit for uploading data and receiving parameters;

[0182] The coupling modeling and separation module includes a deep learning purification subunit for enhancing the separated ECG; when the signal quality score is below the threshold, a recalibration prompt is triggered by the guidance prompt module.

[0183] Furthermore, the deep learning purification subunit and the adaptive filtering module share the same quality scoring gate to avoid over-filtering caused by superimposed processing under poor conditions; the guidance prompting module suppresses high-frequency repetitive prompts to reduce interference. Numerical specifications: the minimum interval for high-frequency prompt suppression is 30 s (adjustable from 10-60 s), and a maximum of three consecutive prompts are given in the same motion mode before entering a cool-down period.

[0184] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0185] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A patch position correction method characterized by, Comprising, Step S1, collecting multi-modal sensor data, the multi-modal sensor data including at least one electrocardiogram signal and three-axis acceleration signal; Step S2, time synchronization and alignment of multi-source data, using hardware timestamp or unified clock and combining resampling to eliminate time delay and drift; Step S3, constructing a motion-artifact coupling model based on acceleration, using reference channel constraint separation / adaptive noise cancellation to separate motion artifact components from electrocardiogram; Step S4, generating pose compensation parameters according to the separation results, and outputting prompt information for patch wearing guidance; Step S5, performing adaptive filtering to adjust electrocardiogram processing parameters in real time according to signal quality indicators and motion states; Step S6, triggering a re-calibration process when the signal quality is below a threshold value; The pose compensation parameters are obtained through a displacement and attitude estimation model, and a coordinate transformation matrix is used to map the estimated quantities in the sensor coordinate system to the body surface reference system to calculate the compensation amount, and the output is a position adjustment suggestion or signal alignment parameter; The pose compensation parameters are obtained through a displacement and attitude estimation model, and the steps include: a. Establishing four coordinate systems: Sensor system i.e. the accelerometer's own coordinates; Patch system , With Fixed connection, take ; body surface reference frame in the body surface reference frame, in the cranial-caudal direction, in the left-right direction, is the outward normal to the body surface; earth system in the earth system against the direction of gravity The state vector is defined in a sliding window as where is the tilt angle of the patch relative to is the small displacement of the patch centroid in with the tangential component prevailing;​ b. Extracting gravity vector and solving attitude: on three axes low pass , to get pitch / roll, yaw are not observable from accelerations alone, take initial calibration values ; from this construct , rotation of the patch; at calibration time when correctly worn and stationary store and at that time relative rotation of , , is a one-time calibration, given by patch geometry and body position setup c. Differential coordinate transformation and mapping: ， wherein denotes a target vector in a body frame of reference , is a mapped quantity in a sensor frame , is a rotation matrix from to at the calibration time, is a rotation matrix from to at the calibration time, is a current rotation matrix from to ; d. Attitude deviation extraction and displacement estimation: The tilt deviation is extracted by comparing the obtained values with calibration values the change in the tangential plane is approximated​​​​​ Displacement adopts a gravity-elimination-band-pass-integration-stationary zero-speed correction chain: ， wherein, is the linear acceleration in the direction under the linear acceleration, is the measured total acceleration, , denotes the vector length, the band-passed, mapped to and the zero-velocity constraint is applied in the rest / pan mode window, resulting in a small displacement of the tangential component; e. Compensation vector and output aperture: Combining attitude errors with displacements into compensation quantities ; wherein is a compensation quantity triad, takes the tangential component of , takes the normal component of , the resultant inclination on the plane; Interface gives a prompt to move head / foot side millimeters, move left / right millimeters, micro-rotate clockwise / counterclockwise degrees, when Silent below display lower limit, over-re-calibration threshold incorporates quality determination to trigger re-calibration.

2. A patch position correction method as claimed in claim 1, characterized in that, In step S3, constraint independent component analysis or adaptive noise cancellation is used to realize separation, wherein acceleration is used as a reference input, and the coupling model includes characteristic parameters of motion intensity, main frequency band and phase relationship.

3. A patch position correction method as claimed in claim 2, characterized in that A motion feature dictionary library is constructed and pattern recognition is performed, the dictionary library is obtained by feature extraction from typical scene data, and the motion patterns include static, walking, running and stair climbing, and a classification algorithm with transition suppression and confidence evaluation is used.

4. The patch position correction method of claim 1, wherein, The signal quality indicators include at least one of: signal-to-noise ratio, baseline drift degree, power frequency interference intensity, heart beat shape stability and R wave detection integrity, the indicators are intervalized by weight fusion to obtain intervalized scores and set thresholds and hysteresis; The steps of calculating and weight fusing the signal quality indicators include: 1) Perform windowing and pre-processing, band-pass and power-line notch on the ECG sequence within a sliding window of length and retain the R-wave location and robust statistics of QRS amplitude within the same window. 2) Define the calculation range of single indicator, get Sub-scores of the same direction Specific as follows: Taking the signal power of QRS neighborhood and the noise power of QRS exclusion zone in the window, we get: ， ， wherein, is the signal-to-noise ratio at time t, is the mean square power of the QRS neighborhood band-passed, is the mean square power of the QRS neighborhood band-passed, is a very small constant, is the linear mapping upper and lower limits, is the self-score obtained by linear normalization, is the self-score obtained by linear normalization, denotes the interval clipping operator; Extracting the baseline with a low-pass filter, calculating the root mean square and the QRS amplitude ratio: and is mapped as: ， wherein, is the ratio of baseline drift amplitude to heart beat amplitude, is the low pass extracted baseline component is the root mean square, is the median peak-trough amplitude in the QRS within the window, are the lower and upper limits of the tolerance interval, is the baseline drift subscore; Compute the power-line interference strength: let the power-line , bandwidth ; take the sum of the spectral power in the power-line and its first harmonic neighborhood, , and divide by the total power, to get the ratio , and map it to: ， wherein, is a power frequency interference sub-score, is a tolerance interval, is a power frequency power ratio; Defining the heart beat shape stability: R-aligned truncated equal-length beat segments and amplitude normalized, using running templates median of correlation coefficients: , take , wherein, is the Pearson correlation coefficient, is the median operator, is the correlation coefficient median value, is the morphological stability subscore, which does not participate in the fusion when the number of valid heartbeats is less than 3; Define R-wave detection integrity: Let the number of R's detected in a window be Estimate the expected number of R's in RR median We get: ， ， wherein, is the median of the RR intervals, is the estimated expected heart rate, is the number of R-waves detected within the window, is the median of the RR intervals, is the length of the sliding window, is the floor function, is the R-wave completeness subscore; 3) After obtaining each sub-score in the window, weighted fusion is performed: ， wherein, is a signal quality binning score at time , is a set of metrics participating in the fusion, is a th sub-score, , is a th weight satisfying , is an availability factor, is a very small constant; To Set two thresholds and hysteresis: low threshold And high threshold , when Bad, when Good, between the two remain unchanged last determination, while setting the minimum residence time ; wherein, Low, high threshold, Minimum residence time.

5. The patch position correction method of claim 1, wherein, Adaptive filtering includes segmented filtering strategy based on motion pattern, parameter-motion intensity mapping relationship and smoothing / cushioning mechanism, and when the score deteriorates, a stronger suppression is used to retain the integrity of the QRS waveform.

6. The patch position correction method of claim 1, wherein, Time synchronization management includes hardware timestamp alignment, link time delay jitter estimation and sampling clock drift correction, and unified resampling after alignment.

7. The patch position correction method of claim 1, wherein, The method further includes inputting the separated electrocardiogram segment and reference features into a deep learning network for signal purification, the network performs end-to-end enhancement or residual correction on electrocardiogram quality, and an offline training and online fine-tuning combination is used.

8. A patch position correction system based on any one of the patch position correction methods according to claims 1 to 7, characterized in that, Comprising: A multi-modal acquisition module is configured to acquire at least one electrocardiogram signal and three-axis acceleration signal at a preset sampling rate and output a raw data stream with a time stamp; A time synchronization management module is configured to align the electrocardiogram and acceleration data based on a hardware time stamp or a unified clock, estimate and compensate link delay jitter and sampling clock drift, and generate an equally-spaced alignment sequence; A motion pattern recognition module is configured to extract time domain / frequency domain / time-frequency features from the acceleration data and output a motion pattern label and its confidence, which are used for subsequent parameter selection; A coupling modeling and separation module is configured to establish a motion-artifact coupling relationship based on acceleration as a reference, separate motion artifacts in electrocardiogram by using constrained independent component analysis and / or adaptive noise cancellation, and obtain preliminary purified electrocardiogram; A signal quality evaluation module is configured to calculate at least one of signal-to-noise ratio, baseline drift, power frequency interference, morphology stability and R wave integrity, and generate a signal quality score and threshold determination; A pose compensation parameter generation module is configured to estimate the displacement / attitude deviation of the patch in the body surface reference system based on the motion pattern and the coupling model, and output pose compensation parameters or position adjustment suggestions; An adaptive filtering module is configured to select a filter type and parameters based on the motion pattern and the signal quality score, perform parameter smoothing and hysteresis control, and output target purified electrocardiogram; A guidance prompt module is configured to prompt the pose compensation parameters and recalibration instructions in a user interface, sound, light or vibration, or to send the parameters to an upper computer / mobile terminal.

9. The patch position correction system of claim 8, wherein: the multi-modal acquisition module comprises a flexible body surface electrode, an analog front end and an analog-to-digital converter, and a three-axis inertial measurement unit; the time synchronization management module uses hardware time stamp alignment combined with resampling; the guidance prompt module comprises at least one of a sound, light or vibration device; the system is provided with a wireless communication unit for uploading data and receiving parameters; the coupling modeling and separation module comprises a deep learning purification subunit for enhancing the separated electrocardiogram; when the signal quality score is lower than the threshold, the recalibration prompt is triggered by the guidance prompt module.

Images