A deep learning-based remote pulse information acquisition method

Abstract

The application discloses a kind of remote pulse information acquisition methods based on deep learning, including the following steps: collecting remote user wrist pulse association data and preprocessing, generate standardized pulse acquisition data sequence;Fragment division and collection state synchronous arrangement are carried out, and candidate pulse segment set is generated;Waveform state, pressure state, posture state and position state are acquired, and remote pulse segment state feature set is generated;Input pulse Mamba-KAN quality identification network, and generate collection quality identification result;Remove low-quality pulse segment, and generate effective pulse segment set;Period alignment and collection disturbance correction are carried out, and corrected effective pulse segment set is generated;Structured feature is extracted, and remote pulse feature set is generated;Structured arrangement generates remote pulse information acquisition result.The application can improve the stability and reliability of remote pulse acquisition result.

Description

Technical Field

[0001] This invention relates to the field of remote medical data acquisition technology, and in particular to a method for remote pulse information acquisition based on deep learning. Background Technology

[0002] In recent years, with the development of telemedicine, wearable sensing, mobile health monitoring, and digital technologies for TCM pulse diagnosis, pulse information collection has gradually shifted from on-site manual palpation to remote terminal collection and intelligent analysis. Current technologies typically collect pulse wave signals and collection status data from the user's wrist using wrist pressure sensors, pulse wave acquisition modules, posture sensors, or mobile terminals. The collected data undergoes noise reduction, normalization, period segmentation, and feature extraction. Traditional machine learning or deep learning models are then used to identify pulse rhythm, waveform amplitude, peak and trough variations, and pulse morphology characteristics, generating pulse data results for remote health monitoring or TCM-assisted analysis. Some solutions also integrate cloud platforms to upload, store, and analyze the collected data, enabling remote users to complete pulse data collection in non-on-site diagnostic environments.

[0003] However, existing technologies still suffer from insufficient quality control when remotely collecting pulse data. On one hand, remote collection lacks on-site doctor guidance, making it easy for the collection position to deviate from the target pulse location. Wrist contact pressure can fluctuate, and wrist posture can change with user operation, leading to discontinuous, distorted, periodically unstable, or abnormal peak and trough positions in the pulse waveform. On the other hand, existing solutions often directly extract features from the collected pulse wave data, lacking joint modeling of waveform, pressure, posture, and position states. This makes it difficult to accurately distinguish between genuine pulse changes and pseudo-abnormal changes caused by collection disturbances. Low-quality pulse segments are prone to enter subsequent analysis, affecting the stability and reliability of remote pulse feature extraction and remote pulse information collection results. Summary of the Invention

[0004] One objective of this invention is to propose a remote pulse information acquisition method based on deep learning. This method fully utilizes remote medical acquisition and deep learning quality recognition technology, and has the advantages of accurate identification of low-quality pulse segments, stable and reliable remote acquisition results, and a high degree of structure in pulse information.

[0005] A remote pulse information acquisition method based on deep learning according to an embodiment of the present invention includes the following steps: Collect pulse data associated with the wrists of remote users, preprocess the pulse data associated with the wrists of remote users, and generate a standardized pulse data sequence. The standardized pulse data acquisition sequence is divided into segments according to the continuous sampling time, and the pulse waveform and acquisition status are synchronized and organized to generate a set of candidate pulse segments. Based on the candidate pulse segment set, the waveform state, pressure state, attitude state and position state of each candidate pulse segment are obtained to generate a remote pulse segment state feature set. The set of state features of remote pulse segments is input into the pulse Mamba-KAN quality recognition network to perform selective state-space temporal modeling, local waveform state extraction and spline-gated state fusion to generate acquisition quality recognition results. Based on the acquisition quality identification results, low-quality pulse segments caused by acquisition position offset, pressure fluctuation or wrist posture change are identified, and low-quality pulse segments are removed from the candidate pulse segment set to generate a valid pulse segment set. Periodic alignment and acquisition perturbation correction are performed on the set of effective pulse segments to generate a corrected set of effective pulse segments. Based on the corrected set of effective pulse segments, structured pulse features are extracted to generate a set of remote pulse features; The remote pulse feature set, the corrected effective pulse segment set, and the acquisition quality recognition results are structured and organized to generate remote pulse information acquisition results.

[0006] Optionally, the remote user's wrist pulse correlation data includes pulse wave sampling data, wrist contact pressure data, wrist posture data, acquisition location data, and acquisition time data. Preprocessing includes timestamp unification, sampling frequency matching, abnormal sampling point removal, signal denoising, baseline drift correction, pressure drift calibration, posture coordinate unification, acquisition location calibration, amplitude normalization, and continuous sampling sequence processing.

[0007] Optionally, the generation of the candidate pulse segment set includes: The standardized pulse acquisition data sequence is arranged in chronological order of sampling time, the continuity between adjacent sampling times is checked, and the segment boundaries are marked at the locations where sampling is interrupted. Based on segment boundary marking, the standardized pulse image acquisition data sequence is initially segmented to obtain continuous pulse image data segments; Within a continuous pulse data segment, segments are divided according to the periodic changes of the pulse waveform to generate candidate pulse segments. Synchronize and organize the pulse waveform, pressure state, attitude state and position state at the same sampling time within each candidate pulse segment; The candidate pulse segments are sorted according to the sampling time order to generate a set of candidate pulse segments.

[0008] Optionally, the generation of the remote pulse segment state feature set includes: Read the candidate pulse segments in the candidate pulse segment set one by one, extract the pulse waveform changes in the candidate pulse segments according to the sampling time order, and count the continuity, periodic changes and peak and valley distribution of the pulse waveform to generate the waveform status. Read wrist contact pressure data within the same candidate pulse segment, organize the pressure change amplitude and pressure maintenance according to the sampling time sequence, and generate pressure status; Read wrist posture data within the same candidate pulse segment, organize the direction and amplitude of wrist posture changes, and generate posture state; Read the acquisition location data within the same candidate pulse segment, organize the acquisition location offset direction and acquisition location offset magnitude, and generate the location status; The waveform state, pressure state, attitude state, and position state of each candidate pulse segment are combined according to the sampling time order to generate a remote pulse segment state feature set.

[0009] Optionally, the generation of the acquisition quality identification result includes: The set of state features of remote pulse segments is input into the pulse Mamba-KAN quality recognition network. The pulse Mamba-KAN quality recognition network includes a state feature processing layer, a selective state space temporal modeling layer, a local waveform state extraction layer, a gated state fusion layer, and a quality recognition output layer. In the state feature processing layer, the set of state features of remote pulse segments is processed into a sequence of segment state features according to the candidate pulse segments and the sampling time. In the selective state-space temporal modeling layer, the waveform state, pressure state, attitude state and position state are updated temporally based on the fragment state feature sequence to generate pulse temporal quality features. In the local waveform state extraction layer, local changes in the waveform state in the pulse time sequence quality features are extracted to generate local waveform quality features. In the gated state fusion layer, acquisition perturbation gating features are generated based on pressure state, attitude state, and position state, and the acquisition perturbation gating features are fused with local waveform quality features; In the quality recognition output layer, the acquired quality recognition results are generated based on the fused features.

[0010] Optionally, the generation of the effective pulse segment set includes: According to the order of the candidate pulse segment set, the acquisition quality identification results are matched with the candidate pulse segment set; Based on the acquisition quality identification results, determine whether each candidate pulse segment has acquisition position offset, pressure fluctuation or wrist posture change; Candidate pulse segments with sampling position offset, pressure fluctuation or wrist posture change are identified as low-quality pulse segments. Candidate pulse segments that are not identified as low-quality pulse segments and whose pulse waveforms remain continuous are retained in the candidate pulse segment set. Low-quality pulse segments are removed from the candidate pulse segment set, and the remaining candidate pulse segments are rearranged according to the sampling time order to generate a valid pulse segment set.

[0011] Optionally, the generation of the corrected effective pulse segment set includes: Read the effective pulse segments from the set of effective pulse segments, and extract the effective trough and peak positions in each effective pulse segment according to the sampling time order; Using the pulse cycle between adjacent effective trough positions as the alignment unit, determine the cycle start point, cycle end point, and cycle arrangement order of each effective pulse segment; The starting point of each pulse cycle is moved to a unified cycle index, and the differences in cycle length are sorted out on the time axis to generate effective pulse segments after cycle alignment. The range of pressure fluctuation is determined based on the pressure state, the range of wrist posture change is determined based on the posture state, and the range of acquisition position offset is determined based on the position state. Within the range of pressure fluctuation, wrist posture change, and acquisition position offset, the pulse waveform amplitude, pulse cycle boundary, and peak and trough positions are corrected and organized to generate a corrected effective pulse segment. The corrected effective pulse segments are arranged according to the sampling time order to generate a set of corrected effective pulse segments.

[0012] Optionally, the generation of the remote pulse feature set includes: Read the corrected effective pulse segment in the set of corrected effective pulse segments one by one, and organize the pulse waveform amplitude, effective peak position, effective trough position and pulse cycle boundary according to the sampling time order after period alignment. Based on the pulse cycle boundary statistics, the cycle length, cycle arrangement order and changes of adjacent cycles of each pulse cycle are statistically analyzed to generate pulse rhythm characteristics; Based on the amplitude of the pulse waveform, the range of amplitude changes, the changes in the rising segment, and the changes in the falling segment within each pulse cycle are sorted out to generate pulse waveform features; Based on the effective peak and trough positions, the distribution sequence of peaks and troughs, the peak-trough intervals, and the cross-cycle position changes are sorted out to generate pulse morphology features. The pulse rhythm features, pulse waveform features, and pulse morphology features are combined according to the sampling time sequence of the corrected effective pulse segments to generate a remote pulse feature set.

[0013] Optionally, the generation of the remote pulse information acquisition results includes: The corrected effective pulse segment set is arranged in the order of sampling time, and the sampling time, pulse cycle boundary, effective peak position and effective trough position of each corrected effective pulse segment are extracted. The pulse rhythm features, pulse waveform features, and pulse morphology features in the remote pulse feature set are written into the corrected effective pulse segments according to the segment arrangement order. The acquisition quality identification results are matched with the corrected valid pulse segments according to the segment arrangement order, and the corresponding acquisition quality status is recorded. After matching and correction, the effective pulse segments, pulse rhythm features, pulse waveform features, pulse morphology features, and acquisition quality status are merged, sorted by time, and formatted to generate remote pulse information acquisition results.

[0014] The beneficial effects of this invention are: This invention collects wrist pulse data from remote users and processes pulse wave sampling data, wrist contact pressure data, wrist posture data, acquisition location data, and acquisition time data in a unified manner. This enables multi-source pulse data generated in a remote acquisition environment to be organized according to a continuous sampling sequence, reducing data inconsistencies caused by different acquisition devices, different user operating habits, and different acquisition times. This provides a stable data foundation for subsequent candidate pulse segment division and state feature extraction.

[0015] This invention synchronously organizes the waveform state, pressure state, posture state, and position state in candidate pulse segments, and uses the pulse Mamba-KAN quality recognition network to perform selective state-space temporal modeling, local waveform state extraction, and spline-gated state fusion. This allows for the simultaneous consideration of the impact of pressure fluctuations, wrist posture changes, and acquisition position offsets on waveform quality during pulse waveform change analysis, avoiding the misinterpretation of acquisition disturbances as real pulse changes based solely on the pulse waveform itself, and improving the accuracy of low-quality pulse segment recognition.

[0016] This invention eliminates low-quality pulse segments based on acquisition quality identification results and performs periodic alignment and acquisition perturbation correction on valid pulse segments, ensuring that the data entering the pulse structured feature extraction stage has better waveform continuity, periodic consistency, and state stability. By further generating pulse rhythm features, pulse waveform features, and pulse morphology features, and then structurally organizing them with the corrected valid pulse segments and acquisition quality identification results, stable and reliable remote pulse information acquisition results can be obtained, improving the usability of remote pulse diagnosis data in health monitoring, remote diagnostic auxiliary analysis, and the digital application of traditional Chinese medicine pulse diagnosis. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a remote pulse information acquisition method based on deep learning proposed in this invention; Figure 2 This is a schematic diagram illustrating the generation of a remote pulse segment state feature set in a deep learning-based remote pulse information acquisition method proposed in this invention. Figure 3 This is a schematic diagram illustrating the generation of acquisition quality identification results for a remote pulse information acquisition method based on deep learning proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A remote pulse information acquisition method based on deep learning includes the following steps: Collect pulse data associated with the wrists of remote users, preprocess the pulse data associated with the wrists of remote users, and generate a standardized pulse data sequence. The standardized pulse data acquisition sequence is divided into segments according to the continuous sampling time, and the pulse waveform and acquisition status are synchronized and organized to generate a set of candidate pulse segments. Based on the candidate pulse segment set, the waveform state, pressure state, attitude state and position state of each candidate pulse segment are obtained to generate a remote pulse segment state feature set. The set of state features of remote pulse segments is input into the pulse Mamba-KAN quality recognition network to perform selective state-space temporal modeling, local waveform state extraction and spline-gated state fusion to generate acquisition quality recognition results. Based on the acquisition quality identification results, low-quality pulse segments caused by acquisition position offset, pressure fluctuation or wrist posture change are identified, and low-quality pulse segments are removed from the candidate pulse segment set to generate a valid pulse segment set. Periodic alignment and acquisition perturbation correction are performed on the set of effective pulse segments to generate a corrected set of effective pulse segments. Based on the corrected set of effective pulse segments, structured pulse features are extracted to generate a set of remote pulse features; The remote pulse feature set, the corrected effective pulse segment set, and the acquisition quality recognition results are structured and organized to generate remote pulse information acquisition results.

[0020] In this embodiment, the remote user's wrist pulse correlation data includes pulse wave sampling data, wrist contact pressure data, wrist posture data, acquisition location data, and acquisition time data. Preprocessing includes timestamp unification, sampling frequency matching, abnormal sampling point removal, signal denoising, baseline drift correction, pressure drift calibration, posture coordinate unification, acquisition location calibration, amplitude normalization, and continuous sampling sequence processing. Baseline drift correction is used to address waveform reference bias in pulse wave sampling data that slowly shifts upward or downward over time. Specifically, pulse wave sampling data is read and organized into a pulse wave amplitude sequence according to the sampling time. A sliding window analysis is performed on the pulse wave amplitude sequence to extract local reference values ​​reflecting low-frequency variation trends within each window. Each local reference value is continuously smoothed to obtain the pulse wave baseline drift component. The waveform amplitude at each sampling time in the pulse wave amplitude sequence is subtracted and processed according to the pulse wave baseline drift component to generate pulse wave sampling data after removing baseline drift. The amplitude range of the pulse wave sampling data after removing baseline drift is then adjusted so that the pulse waveform is distributed around a stable reference. Pressure drift calibration is used to address pressure reference shifts in wrist contact pressure data caused by sensor zero-point offset, changes in wearing tightness, and prolonged contact. Specifically, wrist contact pressure data is read and organized into a pressure amplitude sequence according to sampling time. Stable pressure segments with gentle pressure changes and weak pulse wave disturbances are selected from the pressure amplitude sequence. A pressure reference is determined based on the stable pressure segments. Slow-shifting portions in the pressure amplitude sequence are identified according to the pressure reference to obtain the pressure drift component. The pressure values ​​at each sampling time in the pressure amplitude sequence are subtracted and organized according to the pressure drift component to generate corrected wrist contact pressure data. Range constraints are applied to the corrected wrist contact pressure data to ensure that the pressure state reflects the actual changes in wrist contact intensity.

[0021] In this embodiment, the generation of the candidate pulse segment set includes: The standardized pulse acquisition data sequence is arranged in chronological order of sampling time, the continuity between adjacent sampling times is checked, and the segment boundaries are marked at the locations where sampling is interrupted. Based on segment boundary marking, the standardized pulse image acquisition data sequence is initially segmented to obtain continuous pulse image data segments; Within a continuous pulse data segment, segments are divided according to the periodic changes of the pulse waveform to generate candidate pulse segments. The generation of candidate pulse segments involves: reading the pulse waveform sequence from a continuous pulse data segment and organizing the pulse waveform amplitude according to the sampling time; smoothing the pulse waveform amplitude to preserve the main pulse wave variation and reduce local jitter; marking the rising and falling segments of the waveform according to the amplitude increase / decrease direction at adjacent sampling times, marking the position from the rising segment to the falling segment as a peak candidate point, and marking the position from the falling segment to the rising segment as a trough candidate point; merging and organizing adjacent peak and trough candidate points that are close in distance and have small amplitude changes to obtain the effective pulse segment. Peak position and effective trough position; determine a single pulse cycle by the waveform interval between adjacent effective trough positions, and check whether the waveform interval contains an effective peak position; retain pulse cycles containing a complete rising segment, an effective peak position, and a complete falling segment as complete pulse cycles; classify consecutive complete pulse cycles into the same candidate pulse segment boundary according to the sampling time sequence; when there is waveform truncation at the candidate pulse segment boundary, adjust the boundary to the adjacent effective trough position; truncate continuous pulse data segments according to the adjusted segment boundary to generate candidate pulse segments; Synchronize and organize the pulse waveform, pressure state, attitude state and position state at the same sampling time within each candidate pulse segment; The candidate pulse segments are sorted according to the sampling time order to generate a set of candidate pulse segments.

[0022] In this embodiment, the generation of the remote pulse segment state feature set includes: Read the candidate pulse segments in the candidate pulse segment set one by one, extract the pulse waveform changes in the candidate pulse segments according to the sampling time order, and count the continuity, periodic changes and peak and valley distribution of the pulse waveform to generate the waveform status. The waveform state generation process is as follows: First, the time interval between adjacent sampling moments is checked, and the interruption points of the time interval and the missing waveform amplitudes are marked to obtain the continuity of the pulse waveform. Second, the pulse waveform amplitude is smoothed, preserving the main wave fluctuations. Based on the direction of amplitude increase / decrease between adjacent sampling moments, the waveform is divided into rising and falling segments. The position where the rising segment transitions into a falling segment is determined as the peak position, and the position where the falling segment transitions into a rising segment is determined as the trough position. Third, the pulse cycle is divided according to the sampling interval between adjacent trough positions. The start and end times, cycle length, waveform amplitude variation range, and cycle arrangement order of each pulse cycle are statistically analyzed to obtain the cycle variation. Fourth, the distribution order of peak and trough positions within each pulse cycle is checked. Repeated peak and trough positions formed by local jitter are deleted, and valid peak and trough positions are organized to obtain the peak and trough distribution. Fifth, the continuity, cycle variation, and peak and trough distribution of the pulse waveform are combined according to the sampling time order of the candidate pulse segments to generate the waveform state of the candidate pulse segments. Read wrist contact pressure data within the same candidate pulse segment, organize the pressure change amplitude and pressure maintenance according to the sampling time sequence, and generate pressure status; The generation of pressure states specifically involves: extracting pressure values ​​at each sampling time and marking locations where pressure values ​​are missing or where pressure abruptly changes; smoothing and organizing wrist contact pressure data to retain the pressure change trend caused by variations in wrist contact intensity; calculating the pressure difference between adjacent sampling times and organizing the pressure rise interval, pressure fall interval, and pressure hold interval according to the direction of pressure difference change; statistically analyzing the concentrated distribution of pressure values, pressure fluctuation range, and continuous sampling length within each pressure hold interval to obtain the pressure hold status; statistically analyzing the maximum range of pressure value changes, the cumulative change of pressure difference between adjacent sampling times, and the pressure change across the pulse cycle within each candidate pulse segment to obtain the pressure change amplitude; and combining the pressure change amplitude, pressure hold status, pressure abruptly changing location, and pressure missing location according to the sampling time sequence to generate the pressure state of the candidate pulse segment. Read wrist posture data within the same candidate pulse segment, organize the direction and amplitude of wrist posture changes, and generate posture state; The generation of posture states is specifically as follows: extract wrist posture values ​​at each sampling time and mark the missing posture value locations and posture abrupt change locations; smooth and organize the wrist posture data, retaining the posture change trend formed by changes in wrist placement; organize the changes in wrist upward deflection, downward deflection, inward deflection, outward deflection, and posture maintenance based on the direction of posture value changes between adjacent sampling times; statistically analyze the concentrated distribution of posture values, the range of posture fluctuations, and the continuous sampling length within the posture maintenance interval to obtain the posture maintenance status; statistically analyze the maximum range of posture value changes, the cumulative change of posture differences between adjacent sampling times, and the posture changes across pulse cycles within each candidate pulse segment to obtain the posture change amplitude; combine the wrist posture change direction, posture change amplitude, posture maintenance status, posture abrupt change location, and posture missing location according to the sampling time order to generate the posture state of the candidate pulse segment; Read the acquisition location data within the same candidate pulse segment, organize the acquisition location offset direction and acquisition location offset magnitude, and generate the location status; The generation of position status is specifically as follows: Extract the acquisition position markers at each sampling time, and mark missing and transitional acquisition positions; smooth the acquisition position data, preserving the position change trend caused by the movement of the acquisition area; using the stable acquisition position at the beginning of the candidate pulse segment as a position reference, calculate the offset direction and distance of the acquisition position relative to the position reference at each sampling time; based on the changes in the acquisition position in the transverse and longitudinal directions of the wrist, organize the changes in ulnar, radial, proximal, and distal offsets of the acquisition position, as well as the changes in position maintenance; statistically analyze the concentrated distribution, position fluctuation range, and continuous sampling length of the acquisition positions within the position maintenance interval to obtain the position maintenance status; statistically analyze the maximum offset range of the acquisition position, the cumulative change in position offset between adjacent sampling times, and the position offset across the pulse cycle within each candidate pulse segment to obtain the acquisition position offset amplitude; combine the acquisition position offset direction, acquisition position offset amplitude, position maintenance status, transitional acquisition positions, and missing acquisition positions according to the sampling time order to generate the position status of the candidate pulse segment. The waveform state, pressure state, attitude state, and position state of each candidate pulse segment are combined according to the sampling time order to generate a remote pulse segment state feature set.

[0023] In this embodiment, the generation of the quality identification results includes: The set of state features of remote pulse segments is input into the pulse Mamba-KAN quality recognition network. The pulse Mamba-KAN quality recognition network includes a state feature processing layer, a selective state space temporal modeling layer, a local waveform state extraction layer, a gated state fusion layer, and a quality recognition output layer. The core of the pulse Mamba-KAN quality recognition network is to add a state control mechanism oriented towards pulse acquisition perturbations to the Mamba selective state-space model, and to add a spline gating fusion mechanism oriented towards pressure state, attitude state, and position state to the KAN network. The state feature processing layer organizes the waveform state, pressure state, attitude state, and position state into a segment state feature sequence according to the candidate pulse segments and sampling time order, so that the pulse waveform changes and acquisition state changes maintain the same temporal structure. The selective state-space temporal modeling layer uses Mamba's selective state update capability to perform long-term temporal modeling of the segment state feature sequence, and adjusts the state update according to pressure... The system adjusts the degree of pulse waveform state preservation based on state, posture, and position, generating pulse temporal quality features. The local waveform state extraction layer extracts local states based on waveform continuity, periodic stability, and peak-valley distribution in the pulse temporal quality features, identifying local low-quality performance caused by short-term jitter, waveform truncation, and periodic anomalies. The gated state fusion layer uses KAN spline functions to perform nonlinear mapping on pressure state, posture state, and position state, generating acquisition perturbation gating features. These features are then used to modulate local waveform quality features, enabling the network to distinguish between real pulse waveform changes and pseudo-abnormal changes caused by acquisition position offset, pressure fluctuations, and wrist posture changes. The training data for the pulse quality recognition network Mamba-KAN comes from remote user wrist pulse association data that has been collected and labeled with quality. The labeling includes valid pulse segment markers, low-quality pulse segment markers, acquisition position offset markers, pressure fluctuation markers, and wrist posture change markers. The remote user wrist pulse association data for training is preprocessed to generate a standardized pulse acquisition data sequence for training. The data is then segmented according to continuous sampling time to generate a set of candidate pulse segments for training. Based on the set of candidate pulse segments for training, waveform state, pressure state, posture state, and position state are obtained to generate a set of state features for remote pulse segments for training. Training parameters include segment length, segment sliding step size, batch size, learning rate, training epochs, weight decay coefficient, hidden state dimension in state space, number of spline basis functions, perturbation gating dimension, and gradient clipping threshold. During training, the set of state features of remote pulse segments used for training is input into the pulse Mamba-KAN quality recognition network. The pulse temporal quality features are updated in the selective state space temporal modeling layer, local waveform quality features are extracted in the local waveform state extraction layer, and acquisition perturbation gating features are generated based on pressure state, posture state, and position state in the gated state fusion layer. The acquisition quality recognition result is output in the quality recognition output layer. The loss function is set as a combination of segment quality classification loss, low quality cause identification loss, perturbation gating constraint loss, and temporal consistency loss. The parameters of the pulse Mamba-KAN quality recognition network are updated in reverse according to the combined loss. After the acquisition quality recognition result on the validation data meets the training stopping condition, the trained pulse Mamba-KAN quality recognition network is saved. In the state feature processing layer, the set of state features of remote pulse segments is processed into a sequence of segment state features according to the candidate pulse segments and the sampling time. In the selective state-space temporal modeling layer, the waveform state, pressure state, attitude state and position state are updated temporally based on the fragment state feature sequence to generate pulse temporal quality features. The generation of pulse temporal quality features is as follows: Based on the pressure state, it is determined whether there is a fluctuation in pressure intensity at the sampling time; based on the posture state, it is determined whether there is a change in wrist posture at the sampling time; based on the position state, it is determined whether there is a shift in the acquisition position at the sampling time. When no fluctuation in pressure intensity, wrist posture change, or acquisition position shift occurs, the current sampling time is determined as a stable sampling time. When fluctuations in pressure intensity, wrist posture change, or acquisition position shift occur, the current sampling time is determined as a perturbation sampling time. For stable sampling times, the waveform state of the current sampling time is written into the pulse temporal state features and updated with the pulse temporal state features retained from the previous sampling time. For perturbation sampling times, the amount of waveform state written at the current sampling time is reduced, the pulse temporal state features formed at the previous stable sampling time are retained, and the pressure state, posture state, and position state of the current sampling time are recorded as acquisition perturbation content. All sampling times within the candidate pulse segment are processed continuously according to the sampling time sequence to generate a pulse temporal state feature sequence. Based on the pulse temporal state feature sequence, the stable waveform continuation content, the perturbation waveform suppression content, and the acquisition perturbation content are organized to generate pulse temporal quality features. In the local waveform state extraction layer, local changes in the waveform state in the pulse time sequence quality features are extracted to generate local waveform quality features. The generation of local waveform quality features is specifically as follows: Extracting pulse waveform changes from the continuous content of a stable waveform according to the sampling time sequence; dividing each pulse cycle into rising, peak, falling, and trough segments based on the effective trough and peak positions in the waveform state; checking the continuity of the rising segment, the integrity of the peak segment, the continuity of the falling segment, and whether the trough segment is truncated within each pulse cycle; sequentially comparing the effective peak and trough positions and cycle lengths in adjacent pulse cycles, and sorting out peak position offsets, trough position offsets, and cycle length changes; performing a smoothing check on the local waveform amplitude changes within each pulse cycle, marking local spikes caused by short-term jitter, local gaps caused by signal discontinuity, and local distortions caused by acquisition disturbances; and organizing the continuous rising segment, intact peak segment, continuous falling segment, truncated trough segment, peak position offset, trough position offset, cycle length changes, local spikes, local gaps, and local distortions according to the sampling time sequence to generate local waveform quality features. In the gated state fusion layer, acquisition perturbation gating features are generated based on pressure state, attitude state, and position state, and the acquisition perturbation gating features are fused with local waveform quality features; The gated state fusion layer processes the following steps: It reads the pressure state, posture state, and position state from the fragment state feature sequence; extracts the pressure change amplitude, pressure maintenance status, wrist posture change direction, posture change amplitude, acquisition position offset direction, and acquisition position offset amplitude according to the sampling time order; concatenates the pressure state, posture state, and position state at the same sampling moment to generate an acquisition perturbation state vector; uses a KAN spline mapping structure to perform nonlinear mapping on the pressure change content, posture change content, and position offset content in the acquisition perturbation state vector to generate pressure perturbation weights, posture perturbation weights, and position perturbation weights; and then applies these weights... The features are combined according to the sampling time sequence to generate acquisition disturbance gating features; the continuity of the rising segment, the integrity of the peak segment, the continuity of the falling segment, the truncation of the valley segment, the change in period length, local glitch, local gap, and local distortion are read from the local waveform quality features at the same sampling time; the acquisition disturbance gating features are used to perform weighted modulation on the local waveform quality features at the same sampling time, so that local waveform anomalies caused by pressure fluctuations, attitude changes, and position offsets are marked as acquisition disturbance-affected content, and local waveform changes not affected by acquisition disturbances are retained as waveform quality judgment content; the weighted modulated local waveform quality features are sorted according to the sampling time sequence to generate fused features; In the quality recognition output layer, the acquired quality recognition results are generated based on the fused features.

[0024] In this embodiment, the generation of the effective pulse segment set includes: According to the order of the candidate pulse segment set, the acquisition quality identification results are matched with the candidate pulse segment set; Based on the acquisition quality identification results, determine whether each candidate pulse segment has acquisition position offset, pressure fluctuation or wrist posture change; Candidate pulse segments with sampling position offset, pressure fluctuation or wrist posture change are identified as low-quality pulse segments. Candidate pulse segments that are not identified as low-quality pulse segments and whose pulse waveforms remain continuous are retained in the candidate pulse segment set. Low-quality pulse segments are removed from the candidate pulse segment set, and the remaining candidate pulse segments are rearranged according to the sampling time order to generate a valid pulse segment set.

[0025] In this embodiment, the generation of the corrected effective pulse segment set includes: Read the effective pulse segments from the set of effective pulse segments, and extract the effective trough and peak positions in each effective pulse segment according to the sampling time order; Using the pulse cycle between adjacent effective trough positions as the alignment unit, determine the cycle start point, cycle end point, and cycle arrangement order of each effective pulse segment; The starting points of each pulse cycle are moved to a unified cycle index, and the differences in cycle length are rectified along the time axis to generate valid pulse segments with aligned cycles. Specifically, the starting point of each pulse cycle is used as the starting index of the current cycle, and the starting points of the cycles are uniformly adjusted to the starting index within the cycle. The sampling points within the cycle are then rearranged to their relative index positions within the cycle according to the sampling time order. The cycle lengths of each pulse cycle within the valid pulse segment are calculated, and the most concentrated cycle lengths are selected as the unified cycle length. For pulse cycles with a cycle length greater than the unified cycle length, the cycle length is maintained... With the starting point, cycle end point, and effective peak position unchanged, adjacent sampling points are compressed and arranged at equal intervals. For pulse cycles with a cycle length shorter than the uniform cycle length, sampling points are added between the cycle starting point and the effective peak position, and between the effective peak position and the cycle end point, according to the waveform change trend. The pulse cycles after compression and arrangement or addition of sampling points are arranged according to the uniform cycle length, so that the cycle starting point, effective peak position, and cycle end point of each pulse cycle are within the uniform cycle index range. The pulse cycles are spliced ​​together according to their original time order to generate a cycle-aligned effective pulse segment. The range of pressure fluctuation is determined based on the pressure state, the range of wrist posture change is determined based on the posture state, and the range of acquisition position offset is determined based on the position state. Within the pressure fluctuation range, wrist posture change range, and acquisition position offset range, the pulse waveform amplitude, pulse cycle boundary, and peak and trough positions are corrected and refined to generate corrected effective pulse segments. Specifically, sampling intervals with continuous pressure are selected before and after the pressure fluctuation range, and the pulse waveform amplitude range within these intervals is refined. Pulse waveform amplitudes deviating from this range within the pressure fluctuation range are then shifted and compressed. Similarly, sampling intervals with continuous wrist posture are selected before and after the wrist posture change range, and the amplitudes within these intervals are refined... The pulse cycle boundary position is determined, and the pulse cycle boundary that shifts within the wrist posture change range is adjusted to the nearest effective trough position. A sampling interval maintaining continuous positional status is selected before and after the offset sampling interval. The distribution order of effective peak and trough positions within this sampling interval is organized, and peak and trough positions that are repeated, have abrupt changes in position, or are out of order within the offset sampling interval are deleted, retained, or relocated. After completing the correction of pulse waveform amplitude, pulse cycle boundary, and peak and trough positions, the effective pulse segments are reorganized according to the sampling time order to generate corrected effective pulse segments. The corrected effective pulse segments are arranged according to the sampling time order to generate a set of corrected effective pulse segments.

[0026] In this embodiment, the generation of the remote pulse feature set includes: Read the corrected effective pulse segment in the set of corrected effective pulse segments one by one, and organize the pulse waveform amplitude, effective peak position, effective trough position and pulse cycle boundary according to the sampling time order after period alignment. Based on the pulse cycle boundary statistics, the cycle length, cycle arrangement order and changes of adjacent cycles of each pulse cycle are statistically analyzed to generate pulse rhythm characteristics; The generation of pulse rhythm features is specifically as follows: The start and end points of each pulse cycle are determined according to the sampling time sequence; the cycle length of each pulse cycle is obtained based on the sampling length between the start and end points; the cycle arrangement order is organized according to the sequential position of each pulse cycle in the corrected effective pulse segment; the changes in cycle length, cycle start interval, and cycle end interval of adjacent pulse cycles are compared one by one to obtain the changes in adjacent cycles; pulse cycles with continuous cycle length distribution and stable changes in adjacent cycles are marked as rhythmically stable cycles, while pulse cycles with abrupt changes in cycle length, abnormal cycle start interval, or abnormal cycle end interval are marked as rhythmically abnormal cycles; the distribution positions of rhythmically stable cycles and rhythmically abnormal cycles in the corrected effective pulse segment are statistically analyzed according to the sampling time sequence; the cycle length, cycle arrangement order, changes in adjacent cycles, distribution positions of rhythmically stable cycles, and distribution positions of rhythmically abnormal cycles are combined to generate pulse rhythm features. Based on the amplitude of the pulse waveform, the range of amplitude changes, the changes in the rising segment, and the changes in the falling segment within each pulse cycle are sorted out to generate pulse waveform features; The generation of pulse waveform features is specifically as follows: The pulse waveform amplitude is extracted one by one according to the pulse cycle boundary within each pulse cycle; within each pulse cycle, the pulse waveform amplitude between the effective trough position and the effective peak position is extracted, and the continuous rise in amplitude, the start position of the rise, the end position of the rise, and the magnitude of the rise are organized to generate the rise segment variation; the pulse waveform amplitude between the effective peak position and the next effective trough position is extracted, and the continuous fall in amplitude, the start position of the fall, the end position of the fall, and the magnitude of the fall are organized to generate the fall segment variation; the maximum and minimum values ​​of the pulse waveform amplitude within each pulse cycle, as well as the amplitude difference between the maximum and minimum values, are calculated to generate the amplitude variation range; the rise and fall segments are checked for abrupt amplitude jumps, local discontinuities, and abnormal flatness, and the results are incorporated into the rise and fall segment variation; the amplitude variation range, rise segment variation, and fall segment variation of each pulse cycle are combined according to the sampling time sequence to generate the pulse waveform features; Based on the effective peak and trough positions, the distribution sequence of peaks and troughs, the peak-trough intervals, and the cross-cycle position changes are sorted out to generate pulse morphology features. The generation of pulse morphology features specifically involves: reading the effective peak positions, effective trough positions, and pulse cycle boundaries from the corrected effective pulse segment; arranging the effective trough positions, effective peak positions, and the next effective trough position within each pulse cycle according to the sampling time sequence; checking the order of the effective trough positions, effective peak positions, and the next effective trough position within each pulse cycle; and determining the arrangement that satisfies the condition that the effective trough position precedes the effective peak position and the effective peak position precedes the next effective trough position as the peak-trough distribution order; and statistically analyzing the sampling interval and effective wave length between the effective trough position and the effective peak position within the same pulse cycle. The sampling interval between the peak position and the next effective trough position, as well as the sampling interval between adjacent effective peak positions, are used to generate peak-trough intervals. The relative positional changes of effective peak positions and effective trough positions within each pulse cycle are compared according to the order of adjacent pulse cycles to generate cross-cycle positional changes. Pulse cycles with advanced effective peak positions, delayed effective peak positions, offset effective trough positions, or abnormal peak-trough sequences are marked. The peak-trough distribution order, peak-trough intervals, cross-cycle positional changes, and marking results are combined according to the sampling time order to generate pulse morphology features. The pulse rhythm features, pulse waveform features, and pulse morphology features are combined according to the sampling time sequence of the corrected effective pulse segments to generate a remote pulse feature set.

[0027] In this embodiment, the generation of remote pulse information acquisition results includes: The corrected effective pulse segment set is arranged in the order of sampling time, and the sampling time, pulse cycle boundary, effective peak position and effective trough position of each corrected effective pulse segment are extracted. The pulse rhythm features, pulse waveform features, and pulse morphology features in the remote pulse feature set are written into the corrected effective pulse segments according to the segment arrangement order. The acquisition quality identification results are matched with the corrected valid pulse segments according to the segment arrangement order, and the corresponding acquisition quality status is recorded. After matching and correction, the effective pulse segments, pulse rhythm features, pulse waveform features, pulse morphology features, and acquisition quality status are merged, sorted by time, and formatted to generate remote pulse information acquisition results.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a remote pulse information collection scenario conducted at a traditional Chinese medicine hospital. In this scenario, some users live far from the offline clinic and need to submit pulse information through a remote consultation platform during follow-up visits. Since users independently complete wrist collection in their home environment or at community health service stations, issues such as unstable wrist placement angles, deviation of the collection end from the cun, guan, and chi areas, and changes in pressure with breathing and hand movements can easily occur during the collection process. This results in the uploaded pulse wave signal containing discontinuous waveforms, peak shifts, unstable periods, and localized distortions. When remote doctors review this type of pulse data, they often need to repeatedly determine which segments are usable for analysis and which are merely pseudo-anomalies caused by the collection action, increasing the difficulty of processing remote pulse diagnosis data and affecting the reliability of the pulse information collection results.

[0029] In this scenario, users complete data collection via a wrist pulse acquisition terminal. The terminal simultaneously acquires pulse wave sampling data, wrist contact pressure data, wrist posture data, acquisition location data, and acquisition time data. The acquisition time covers the user's self-collection phase before follow-up visits, the data confirmation phase before online consultations, and the data processing phase before remote review by doctors. After receiving the remote user's wrist pulse data, the system preprocesses the data from different acquisition channels to generate a standardized pulse acquisition data sequence. Since the acquisition action is not entirely stable in a remote environment, it is difficult to determine the source of abnormalities based solely on the pulse waveform. Therefore, the system divides the standardized pulse acquisition data sequence into segments according to continuous sampling times, synchronously organizing the pulse waveform within each candidate pulse segment with pressure, posture, and position states, ensuring that waveform changes at the same sampling time are consistent with changes in the acquisition state. Based on the candidate pulse segment set, the system extracts waveform, pressure, posture, and position states to generate a remote pulse segment state feature set, which is then fed into the pulse Mamba-KAN quality recognition network. The system removes low-quality pulse segments based on the acquisition quality identification results, performs periodic alignment and acquisition perturbation correction on the remaining valid pulse segments to obtain a set of corrected valid pulse segments, and then extracts pulse rhythm features, pulse waveform features and pulse morphology features to form a remote pulse feature set. Finally, the remote pulse feature set, the set of corrected valid pulse segments and the acquisition quality identification results are structured and organized to generate remote pulse information acquisition results.

[0030] In practical applications, common problems encountered by users when collecting data independently at home mainly involve changes in wrist posture and fluctuations in pressure. When data is collected with staff assistance at community health service stations, the risk of sampling position deviation is relatively reduced, but local waveform distortion can still occur due to slight wrist movements. While sampling conditions are relatively stable at remote sampling service points in internet hospitals, some users still experience brief pressure changes during the sampling process. For data collected in these different locations, this invention does not rely solely on pulse waveforms to determine segment quality. Instead, it analyzes the pulse waveform and sampling status simultaneously, enabling the system to identify low-quality pulse segments caused by sampling actions and retain valid pulse segments with continuous waveforms, stable periods, and relatively stable sampling states. When remote doctors view the remote pulse information collection results generated by the system, they can directly obtain the corrected set of valid pulse segments after quality identification and perturbation correction, and simultaneously view pulse rhythm characteristics, pulse waveform characteristics, pulse morphological characteristics, and sampling quality status, without needing to manually screen segments significantly affected by sampling position deviation, pressure fluctuations, or wrist posture changes.

[0031] Table 1. Performance Comparison of the Invention Method and Traditional Remote Pulse Acquisition Method

[0032] As shown in Table 1, the method of this invention offers a more stable performance improvement compared to traditional remote pulse acquisition methods. Traditional methods primarily rely on the pulse waveform itself for quality judgment. When faced with acquisition position shifts, pressure fluctuations, and wrist posture changes, they are prone to misinterpreting waveform anomalies caused by acquisition disturbances as genuine pulse changes. The method of this invention, by simultaneously utilizing waveform state, pressure state, posture state, and position state, increases the accuracy of low-quality pulse segment identification from 82.4% to 87.1%, and reduces the false negative rate of low-quality pulse segments from 13.6% to 10.2%, demonstrating its ability to more effectively distinguish between genuine pulse waveform changes and pseudo-abnormal changes caused by remote acquisition disturbances.

[0033] In terms of effective pulse segment selection, the false rejection rate of effective pulse segments using traditional methods was 10.8%, while the method of this invention reduced it to 8.4%; the retention rate of effective pulse segments increased from 76.5% to 80.7%. This result demonstrates that this invention does not simply increase the rejection intensity, but rather, while identifying low-quality pulse segments, it better preserves effective pulse segments with analytical value. The performance improvement mainly stems from the selective state-space temporal modeling and spline-gated state fusion in the pulse Mamba-KAN quality recognition network. Selective state-space temporal modeling maintains a continuous and stable pulse waveform state, while spline-gated state fusion combines pressure, attitude, and position states to determine the source of waveform anomalies, thereby reducing the false deletion of effective pulse segments.

[0034] Regarding peak and trough location and waveform continuity, the average deviation of peak and trough location in this invention is reduced from 32.8 ms to 27.6 ms, and the waveform continuity pass rate after correction is increased from 80.9% to 85.2%. While this improvement is not exaggerated, it reflects the actual effect of period alignment and acquisition disturbance correction on the processing of remote pulse segments. Traditional methods typically have limited ability to handle waveform discontinuities, local distortions, and slight period misalignments when performing peak and trough location. This invention, after removing low-quality pulse segments, performs period alignment on the effective pulse segments and corrects acquisition disturbances based on pressure, attitude, and position states, making the effective peak position, effective trough position, and pulse cycle boundary more stable.

[0035] Regarding the usability of the final results, the consistency rate between the acquisition quality status and manual review of the method of this invention increased from 83.7% to 88.0%, and the usability rate of remote pulse information acquisition results increased from 78.2% to 83.6%. This indicates that the remote pulse information acquisition results generated by this invention not only include corrected effective pulse segments, but also further integrate pulse rhythm features, pulse waveform features, pulse morphology features, and acquisition quality status, making the output results easier for remote doctors or subsequent analysis systems to use. Overall, the performance improvement of this invention mainly comes from the synchronous modeling of multi-source acquisition status and pulse waveform, the identification of low-quality segments by the pulse Mamba-KAN quality recognition network, the period alignment and acquisition perturbation correction of effective pulse segments, and the structured organization of remote pulse features.

[0036] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for remote pulse information acquisition based on deep learning, characterized in that, Includes the following steps: Collect pulse data associated with the wrists of remote users, preprocess the pulse data associated with the wrists of remote users, and generate a standardized pulse data sequence. The standardized pulse data acquisition sequence is divided into segments according to the continuous sampling time, and the pulse waveform and acquisition status are synchronized and organized to generate a set of candidate pulse segments. Based on the candidate pulse segment set, the waveform state, pressure state, attitude state and position state of each candidate pulse segment are obtained to generate a remote pulse segment state feature set. The set of state features of remote pulse segments is input into the pulse Mamba-KAN quality recognition network to perform selective state-space temporal modeling, local waveform state extraction and spline-gated state fusion to generate acquisition quality recognition results. Based on the acquisition quality identification results, low-quality pulse segments caused by acquisition position offset, pressure fluctuation or wrist posture change are identified, and low-quality pulse segments are removed from the candidate pulse segment set to generate a valid pulse segment set. Periodic alignment and acquisition perturbation correction are performed on the set of effective pulse segments to generate a corrected set of effective pulse segments. Based on the corrected set of effective pulse segments, structured pulse features are extracted to generate a set of remote pulse features; The remote pulse feature set, the corrected effective pulse segment set, and the acquisition quality recognition results are structured and organized to generate remote pulse information acquisition results.

2. The method for remote pulse information acquisition based on deep learning according to claim 1, characterized in that, The remote user wrist pulse correlation data includes pulse wave sampling data, wrist contact pressure data, wrist posture data, acquisition location data, and acquisition time data. Preprocessing includes timestamp unification, sampling frequency matching, abnormal sampling point removal, signal denoising, baseline drift correction, pressure drift calibration, posture coordinate unification, acquisition location calibration, amplitude normalization, and continuous sampling sequence processing.

3. The remote pulse information acquisition method based on deep learning according to claim 1, characterized in that, The generation of the candidate pulse segment set includes: The standardized pulse acquisition data sequence is arranged in chronological order of sampling time, the continuity between adjacent sampling times is checked, and the segment boundaries are marked at the locations where sampling is interrupted. Based on segment boundary marking, the standardized pulse image acquisition data sequence is initially segmented to obtain continuous pulse image data segments; Within a continuous pulse data segment, segments are divided according to the periodic changes of the pulse waveform to generate candidate pulse segments. Synchronize and organize the pulse waveform, pressure state, attitude state and position state at the same sampling time within each candidate pulse segment; The candidate pulse segments are sorted according to the sampling time order to generate a set of candidate pulse segments.

4. The remote pulse information acquisition method based on deep learning according to claim 1, characterized in that, The generation of the remote pulse segment state feature set includes: Read the candidate pulse segments in the candidate pulse segment set one by one, extract the pulse waveform changes in the candidate pulse segments according to the sampling time order, and count the continuity, periodic changes and peak and valley distribution of the pulse waveform to generate the waveform status. Read wrist contact pressure data within the same candidate pulse segment, organize the pressure change amplitude and pressure maintenance according to the sampling time sequence, and generate pressure status; Read wrist posture data within the same candidate pulse segment, organize the direction and amplitude of wrist posture changes, and generate posture state; Read the acquisition location data within the same candidate pulse segment, organize the acquisition location offset direction and acquisition location offset magnitude, and generate the location status; The waveform state, pressure state, attitude state, and position state of each candidate pulse segment are combined according to the sampling time order to generate a remote pulse segment state feature set.

5. The method for remote pulse information acquisition based on deep learning according to claim 1, characterized in that, The generation of the acquisition quality identification results includes: The set of state features of remote pulse segments is input into the pulse Mamba-KAN quality recognition network. The pulse Mamba-KAN quality recognition network includes a state feature processing layer, a selective state space temporal modeling layer, a local waveform state extraction layer, a gated state fusion layer, and a quality recognition output layer. In the state feature processing layer, the set of state features of remote pulse segments is processed into a sequence of segment state features according to the candidate pulse segments and the sampling time. In the selective state-space temporal modeling layer, the waveform state, pressure state, attitude state and position state are updated temporally based on the fragment state feature sequence to generate pulse temporal quality features. In the local waveform state extraction layer, local changes in the waveform state in the pulse time sequence quality features are extracted to generate local waveform quality features. In the gated state fusion layer, acquisition perturbation gating features are generated based on pressure state, attitude state, and position state, and the acquisition perturbation gating features are fused with local waveform quality features; In the quality recognition output layer, the acquired quality recognition results are generated based on the fused features.

6. The remote pulse information acquisition method based on deep learning according to claim 1, characterized in that, The generation of the effective pulse segment set includes: According to the order of the candidate pulse segment set, the acquisition quality identification results are matched with the candidate pulse segment set; Based on the acquisition quality identification results, determine whether each candidate pulse segment has acquisition position offset, pressure fluctuation or wrist posture change; Candidate pulse segments with sampling position offset, pressure fluctuation or wrist posture change are identified as low-quality pulse segments. Candidate pulse segments that are not identified as low-quality pulse segments and whose pulse waveforms remain continuous are retained in the candidate pulse segment set. Low-quality pulse segments are removed from the candidate pulse segment set, and the remaining candidate pulse segments are rearranged according to the sampling time order to generate a valid pulse segment set.

7. The method for remote pulse information acquisition based on deep learning according to claim 1, characterized in that, The generation of the corrected effective pulse segment set includes: Read the effective pulse segments from the set of effective pulse segments, and extract the effective trough and peak positions in each effective pulse segment according to the sampling time order; Using the pulse cycle between adjacent effective trough positions as the alignment unit, determine the cycle start point, cycle end point, and cycle arrangement order of each effective pulse segment; The starting point of each pulse cycle is moved to a unified cycle index, and the differences in cycle length are sorted out on the time axis to generate effective pulse segments after cycle alignment. The range of pressure fluctuation is determined based on the pressure state, the range of wrist posture change is determined based on the posture state, and the range of acquisition position offset is determined based on the position state. Within the range of pressure fluctuation, wrist posture change, and acquisition position offset, the pulse waveform amplitude, pulse cycle boundary, and peak and trough positions are corrected and organized to generate a corrected effective pulse segment. The corrected effective pulse segments are arranged according to the sampling time order to generate a set of corrected effective pulse segments.

8. The remote pulse information acquisition method based on deep learning according to claim 1, characterized in that, The generation of the remote pulse feature set includes: Read the corrected effective pulse segment in the set of corrected effective pulse segments one by one, and organize the pulse waveform amplitude, effective peak position, effective trough position and pulse cycle boundary according to the sampling time order after period alignment. Based on the pulse cycle boundary statistics, the cycle length, cycle arrangement order and changes of adjacent cycles of each pulse cycle are statistically analyzed to generate pulse rhythm characteristics; Based on the amplitude of the pulse waveform, the range of amplitude changes, the changes in the rising segment, and the changes in the falling segment within each pulse cycle are sorted out to generate pulse waveform features; Based on the effective peak and trough positions, the distribution sequence of peaks and troughs, the peak-trough intervals, and the cross-cycle position changes are sorted out to generate pulse morphology features. The pulse rhythm features, pulse waveform features, and pulse morphology features are combined according to the sampling time sequence of the corrected effective pulse segments to generate a remote pulse feature set.

9. The remote pulse information acquisition method based on deep learning according to claim 1, characterized in that, The generation of the remote pulse information acquisition results includes: The corrected effective pulse segment set is arranged in the order of sampling time, and the sampling time, pulse cycle boundary, effective peak position and effective trough position of each corrected effective pulse segment are extracted. The pulse rhythm features, pulse waveform features, and pulse morphology features in the remote pulse feature set are written into the corrected effective pulse segments according to the segment arrangement order. The acquisition quality identification results are matched with the corrected valid pulse segments according to the segment arrangement order, and the corresponding acquisition quality status is recorded. After matching and correction, the effective pulse segments, pulse rhythm features, pulse waveform features, pulse morphology features, and acquisition quality status are merged, sorted by time, and formatted to generate remote pulse information acquisition results.

Images