Pulse condition feature extraction system and method based on adaptive filtering

Abstract

This invention belongs to the field of signal filtering and processing, and particularly relates to a pulse feature extraction system and method based on adaptive filtering, comprising: a signal acquisition module for periodically acquiring raw pulse signals and performing segmented preprocessing, identifying and recording signal frames containing the target waveform contour; an optimization module for dynamically adjusting the starting point of the next acquisition cycle based on waveform structure characteristics after receiving signal frames, calling an adaptive filter bank to reduce noise on new signal frames, extracting multi-dimensional features, comparing them with the features of the previous qualified signal frame, and determining whether to add them to a high-quality signal set based on a preset similarity threshold, wherein the filter type and parameters are dynamically selected according to the spectral characteristics of the current signal frame; and a feature modeling module for continuously monitoring the output, terminating the iteration when multiple consecutive new signal frames are not added to the set, integrating and analyzing the high-quality signal set, extracting the position and quantization morphology parameters of the target waveform, and constructing a high-quality signal set describing individual-specific waveform changes.

Description

Technical Field

[0001] This invention belongs to the field of signal filtering and processing, and particularly relates to a pulse feature extraction system and method based on adaptive filtering. Background Technology

[0002] In existing technologies, the process of extracting features from traditional Chinese medicine pulse patterns usually relies on traditional filtering algorithms, fixed parameter models, or single-modal data processing methods to complete the noise reduction and feature extraction of pulse signals. However, traditional Chinese medicine pulse signals are easily affected by various environmental and physiological factors such as respiratory interference, muscle tremors, and fluctuations in sensor contact pressure. Furthermore, there are significant differences in the pulse morphology among different individuals, which often exhibit characteristics such as waveform overlap, blurred feature points, and strong dynamic variability. It is difficult to achieve accurate filtering of interference through static filtering rules and fixed parameter configurations.

[0003] All of the above-mentioned related technologies suffer from the problems mentioned in this background technology: they lack an adaptive response mechanism to dynamic interference of pulse signals and do not fully consider the differences in individual pulse characteristics, resulting in insufficient accuracy, residual noise, or loss of key features in the extracted pulse feature parameters. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a pulse feature extraction system and method based on adaptive filtering. The system includes: a signal acquisition module for periodically acquiring raw pulse signals and performing segmented preprocessing, identifying and recording signal frames containing the target waveform contour; an optimization module that, upon receiving a signal frame, dynamically adjusts the starting point of the next acquisition cycle based on waveform structure characteristics, applies an adaptive filter bank to denoise new signal frames, extracts multi-dimensional features, compares them with the features of the previous qualified signal frame, and determines whether to add them to a high-quality signal set based on a preset similarity threshold, wherein the filter type and parameters are dynamically selected based on the spectral characteristics of the current signal frame; and a feature modeling module that continuously monitors the output, terminates iteration when multiple consecutive new signal frames are not added to the set, integrates and analyzes the high-quality signal set, extracts the position and quantization morphology parameters of the target waveform, and constructs a high-quality signal set describing individual-specific waveform changes. This invention achieves adaptive optimization of acquisition and filtering parameters, effectively suppresses noise interference, and significantly improves the accuracy and robustness of pulse feature extraction.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A pulse feature extraction system based on adaptive filtering includes:

[0007] The signal acquisition module is used to periodically acquire raw pulse signals and perform segmented preprocessing on the raw pulse signals until at least a partial outline containing the target waveform features in the signal segment is identified, and the corresponding signal frame is recorded.

[0008] An optimization module is used to receive the signal frame and dynamically adjust the starting point of the next acquisition cycle based on its waveform structure characteristics; for the new signal frame acquired in the next acquisition cycle, an adaptive filter bank is invoked for noise reduction processing, and multi-dimensional features of the noise-reduced signal frame, including time-domain entropy values, are extracted. The multi-dimensional features are then compared with the corresponding features of the previous qualified signal frame recorded by the signal acquisition module, and a preset similarity threshold is used to determine whether to add the new signal frame to the high-quality signal set; wherein the filter type and parameters are dynamically selected based on the spectral characteristics of the current signal frame.

[0009] The feature modeling module is used to continuously monitor the output of the signal acquisition module and the optimization module. When multiple consecutive new signal frames are not added to the high-quality signal set, the iterative optimization of the current acquisition cycle is terminated. All signal frames in the high-quality signal set are integrated and analyzed to extract the position information and quantization morphological parameters of the target waveform features and construct a high-quality signal set describing individual-specific waveform changes.

[0010] Specifically, the raw pulse signal undergoes segmented preprocessing, including:

[0011] The continuously acquired raw pulse signals are orthogonally decomposed to extract in-phase and quadrature components, and the instantaneous phase angle sequence is calculated based on the in-phase and quadrature components.

[0012] Obtain the time deviation of the sampling clock relative to the reference source in the current acquisition period, determine the phase compensation value based on the time deviation, and superimpose the phase compensation value onto each phase angle of the instantaneous phase angle sequence to obtain the aligned phase sequence;

[0013] The time point corresponding to the minimum absolute value of the difference between the instantaneous phase angle in the aligned phase sequence and the pre-stored phase value of the main peak of the target waveform is taken as the segmentation start reference point. A signal segment of a preset dynamic time window length is extracted from the segmentation start reference point as a candidate signal segment, so that the candidate signal segment at least includes the rising branch, the main peak and the falling branch of the target waveform.

[0014] Specifically, segmented preprocessing of the raw pulse signal also includes:

[0015] The candidate signal segment and the pre-stored target waveform template are subjected to normalized sliding cross-correlation in the time domain to extract the maximum cross-correlation number.

[0016] The maximum cross-correlation coefficient is compared with a preset first cross-correlation coefficient threshold. When the maximum cross-correlation coefficient is greater than or equal to the first cross-correlation coefficient threshold, the candidate signal segment is determined to be a qualified signal frame in which the target waveform profile is effectively identified. The alignment phase reference value, start timestamp and original sampling data corresponding to the qualified signal frame are recorded as reference frames for the optimization module.

[0017] When the maximum cross-correlation coefficient is less than the first cross-correlation coefficient threshold, it is determined that the candidate signal segment does not contain a complete and effectively identifiable target waveform profile, and the candidate signal segment is marked as an unqualified signal frame.

[0018] Specifically, a signal segment of a preset dynamic time window length is extracted from the segmentation start reference point as a candidate signal segment, including:

[0019] A queue of historical qualified signal frames is obtained. From each historical qualified signal frame, the rising limb threshold crossing time and the falling limb threshold crossing time of the target waveform are extracted. The time difference between the two is calculated as the waveform width of the signal frame. The waveform width is then accumulated and statistically analyzed using a sliding window to obtain the mean waveform width of the current subject. and waveform width standard deviation ;

[0020] Based on the average waveform width and waveform width standard deviation ,according to Calculate the length component of the first dynamic window, where k is the preset width coverage coefficient;

[0021] Obtain the point-by-point phase difference between the aligned phase sequence of the current acquisition period and the pre-stored main peak phase value of the target waveform, and calculate the root mean square value of the point-by-point phase difference as the phase alignment residual estimate. and will Multiply by the sampling period to convert to time offset compensation. ,in T is the center frequency of the target waveform. s The sampling period is the time interval between adjacent sampling points.

[0022] Specifically, selecting a signal segment of a preset dynamic time window length as a candidate signal segment from the segmentation start reference point further includes:

[0023] Based on the time offset compensation amount and the preset compensation gain coefficient Calculate the second dynamic window length component;

[0024] Perform initial waveform morphology classification on the current candidate signal segment: extract the zero-crossing rate, peak amplitude, and half-peak width ratio of the signal frame, input them into a pre-set lightweight decision tree model, and classify the current waveform as a normal shape or at least one predefined abnormal shape; when the classification result is an abnormal shape, query the waveform width statistics corresponding to this type of abnormal shape from the historical qualified signal frame queue. and and with and Replace the and Recalculate the first dynamic window length component;

[0025] The dynamic time window length is obtained by summing the first dynamic window length component and the second dynamic window length component.

[0026] Specifically, an adaptive filter bank is invoked for noise reduction, and multidimensional features of the denoised signal frame are extracted, including:

[0027] Receive the qualified signal frame and its alignment phase reference value output by the signal acquisition module, and calculate the phase difference between the starting point of the current acquisition cycle and the main peak of the target waveform with the main peak of the qualified signal frame as a reference.

[0028] The starting sampling point offset for the next acquisition cycle is generated based on the phase difference, so that the acquisition window of the next acquisition cycle is aligned with the phase of the target waveform.

[0029] Acquire a new signal frame for the next acquisition cycle. Read the reference spectrum of the previous qualified signal frame after noise reduction from the local buffer as a spectrum template. Calculate the frequency domain cross-correlation coefficient between the power spectral density of the new signal frame and the spectrum template. When the frequency domain cross-correlation coefficient is greater than a preset second cross-correlation coefficient threshold, filter the new signal frame using the filter type and parameters adopted in the previous cycle.

[0030] Specifically, the process of calling an adaptive filter bank for noise reduction and extracting multidimensional features of the denoised signal frame also includes:

[0031] When the frequency domain cross-correlation coefficient is less than or equal to the second cross-correlation coefficient threshold, based on the main peak frequency and energy bandwidth of the spectrum template, a filter type and cutoff parameter that covers the main peak frequency and has stopband attenuation characteristics matching the energy bandwidth are selected from the pre-stored filter library. The group delay response of the selected filter is differentially compensated with the group delay response of the previous cycle filter so that the phase difference between the phase responses of the two cycles filters at the main peak frequency of the target waveform is less than the preset third phase difference threshold. The compensated filter parameters are then used to adaptively filter and reduce noise in the new signal frame.

[0032] Extract the time-domain waveform sequence of the new signal frame after noise reduction, calculate its time-domain entropy, zero-crossing rate, peak amplitude and main peak half-peak width, and form an initial multidimensional feature vector;

[0033] An orthogonalization transformation is performed on the initial multidimensional feature vector to obtain independent feature components whose cross-correlation coefficients between each dimension are lower than a preset fourth cross-correlation coefficient threshold, thus forming a feature vector for similarity discrimination.

[0034] Specifically, the process of calling an adaptive filter bank for noise reduction and extracting multidimensional features of the denoised signal frame also includes:

[0035] The first qualified signal frame of the current subject is obtained as the absolute reference frame, and noise reduction, temporal feature extraction and orthogonal transformation are performed on the absolute reference frame in sequence to obtain the absolute reference feature vector.

[0036] Simultaneously, the dynamic reference feature vector obtained after noise reduction and orthogonal transformation of the previous qualified signal frame is read from the local cache; the feature vector used for similarity discrimination is compared with the dynamic reference feature vector and the absolute reference feature vector respectively: the first Euclidean distance between the feature vector used for similarity discrimination and the dynamic reference feature vector, and the second Euclidean distance between the feature vector used for similarity discrimination and the absolute reference feature vector are calculated;

[0037] When the first Euclidean distance is less than the preset fifth threshold and the second Euclidean distance is less than the preset sixth threshold, it is determined that the new signal frame maintains the same features as the historical qualified signal frame and no reference drift has occurred. The new signal frame is stored in the high-quality signal set, and the dynamic reference feature vector is updated with the feature vector used for similarity discrimination.

[0038] When the first Euclidean distance is less than the fifth threshold but the second Euclidean distance is greater than or equal to the sixth threshold, it is determined that the benchmark has undergone asymptotic drift. The absolute benchmark feature vector is then used to cover the dynamic benchmark feature vector, and the drift counter is reset.

[0039] When the first Euclidean distance is greater than or equal to the fifth threshold, the new signal frame is determined to be unqualified, the new signal frame is discarded, and the count of unqualified signal frames is incremented.

[0040] Specifically, differential compensation of the group delay response of the selected filter with the group delay response of the filter in the previous cycle includes:

[0041] Obtain the first phase response value of the filter in the previous cycle at the main peak frequency of the target waveform;

[0042] Obtain the second phase response value of the selected filter at the main peak frequency point;

[0043] The difference between the second phase response value and the first phase response value is calculated as the phase deviation.

[0044] Based on the phase deviation, an all-pass phase compensation filter is generated, and the phase response value of the all-pass phase compensation filter at the main peak frequency is configured to be the opposite of the phase deviation.

[0045] The full-pass phase compensation filter is cascaded with the selected filter to obtain the compensated filter parameters, such that the phase difference between the phase response value of the compensated filter at the main peak frequency point and the first phase response value is less than a preset third phase difference threshold.

[0046] Pulse feature extraction methods based on adaptive filtering include:

[0047] The raw pulse signal is periodically acquired and segmented for preprocessing until at least a partial outline containing the target waveform features is identified in the signal segment, and the corresponding signal frame is recorded.

[0048] The system receives the signal frame and dynamically adjusts the starting point of the next acquisition cycle based on its waveform structure characteristics. For the new signal frame acquired in the next acquisition cycle, an adaptive filter bank is invoked for noise reduction processing. Multidimensional features of the noise-reduced signal frame are extracted, including the time-domain entropy value. The multidimensional features are compared with the corresponding features of the previous qualified signal frame recorded by the signal acquisition module. Based on a preset similarity threshold, it is determined whether to add the new signal frame to the high-quality signal set. The filter type and parameters are dynamically selected based on the spectral characteristics of the current signal frame.

[0049] The output of the signal acquisition module and the optimization module are continuously monitored. When multiple new signal frames are not added to the high-quality signal set, the iterative optimization of the current acquisition cycle is terminated. All signal frames in the high-quality signal set are integrated and analyzed to extract the position information and quantization morphological parameters of the target waveform features and construct a high-quality signal set describing individual-specific waveform changes.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] This invention addresses the shortcomings of existing technologies by actively locking the segment start point and target waveform phase at the acquisition front end through orthogonal decomposition and clock deviation compensation. This fundamentally eliminates waveform fragmentation caused by equal-interval segmentation and acquisition inefficiency caused by passive trial and error. Furthermore, it integrates individual waveform width statistics, phase residual estimation, and morphology recognition to generate a dynamically adaptable window length, simultaneously resolving issues related to individual differences, physiological time variations, residual drift, and mismatch in the truncation of multi-morphological waveforms. In the filtering stage, the clean spectrum of historical qualified frames serves as a template to guide the filter's steady-state decision-making, cutting off noise contamination of the selection path. When switching is necessary, group delay differential compensation enforces consistency in the phase response of the preceding and following periods, eliminating non-physiological jumps in feature vectors. At the feature level, orthogonal decoupling is performed on the initial multi-dimensional features, stripping away noise coupling components and dimensional redundancy, improving the selective response of the similarity criterion to essential distortions. Simultaneously, a dual-threshold joint decision architecture of absolute and dynamic benchmarks is constructed, using the first frame's health features to anchor the physiological variation boundary, blocking alarm-free drift under slow distortion and the benchmark contamination chain caused by pseudo-qualified frames. Ultimately, the feature modeling module accurately extracts individual-specific waveform parameters based on a high-quality signal set that has undergone multiple purification and access checks, and completes the feature model construction. This application enables the system to achieve a comprehensive improvement in phase locking accuracy, window length adaptation capability, filter decision robustness, feature discrimination fidelity, and benchmark traceability under scenarios with individual differences, time-varying states, strong noise interference, and coexistence of multiple forms, significantly enhancing the adaptability of physiological signal acquisition, the purity of feature extraction, and the completeness of clinical information. Attached Figure Description

[0052] Figure 1 This is a block diagram of the pulse feature extraction system based on adaptive filtering in Embodiment 1 of the present invention;

[0053] Figure 2 This is a logic diagram of segmented preprocessing of the original pulse signal in Embodiment 1 of the present invention;

[0054] Figure 3 This is a flowchart of the pulse feature extraction method based on adaptive filtering in Embodiment 2 of the present invention. Detailed Implementation

[0055] Example 1

[0056] Please see Figure 1 The present invention provides an embodiment of a pulse feature extraction system based on adaptive filtering, comprising the following steps:

[0057] The signal acquisition module is used to periodically acquire raw pulse signals and perform segmented preprocessing on the raw pulse signals until at least a partial outline containing the target waveform features in the signal segment is identified, and the corresponding signal frame is recorded.

[0058] An optimization module is used to receive the signal frame and dynamically adjust the starting point of the next acquisition cycle based on its waveform structure characteristics; for the new signal frame acquired in the next acquisition cycle, an adaptive filter bank is invoked for noise reduction processing, and multi-dimensional features of the noise-reduced signal frame, including time-domain entropy values, are extracted. The multi-dimensional features are then compared with the corresponding features of the previous qualified signal frame recorded by the signal acquisition module, and a preset similarity threshold is used to determine whether to add the new signal frame to the high-quality signal set; wherein the filter type and parameters are dynamically selected based on the spectral characteristics of the current signal frame.

[0059] The feature modeling module is used to continuously monitor the output of the signal acquisition module and the optimization module. When multiple consecutive new signal frames are not added to the high-quality signal set, the iterative optimization of the current acquisition cycle is terminated. All signal frames in the high-quality signal set are integrated and analyzed to extract the position information and quantization morphological parameters of the target waveform features and construct a high-quality signal set describing individual-specific waveform changes.

[0060] In existing technologies, the phase decoupling between the equal-interval segmentation and the target waveform is completely decoupled, resulting in the segmentation boundary severing the main peak of the waveform and causing the loss of structural contour information. This is a logical defect that cannot be compensated for by algorithms. The passive trial-and-error acquisition mechanism lacks active phase guidance. Under low signal-to-noise ratio or waveform distortion, it requires 5 to 15 invalid cycles to obtain a qualified frame, resulting in a more than 60% drop in acquisition throughput and an inability to provide timely feedback to the backend module. The admission of qualified frames depends on qualitative descriptions such as the number of zero-crossings of the main body, without quantitative threshold constraints. This causes 18% to 25% of pseudo-qualified frames to be used as reference benchmarks for downstream tasks, leading to filter misselection, phase offset calculation deviations exceeding 30%, and amplified error levels. Furthermore, in determining the segmentation start parameters... The fixed window length truncation strategy adopted after the test point leads to a four-fold mechanistic mismatch: individual morphological differences cause narrow waveform noise to account for more than 45% and wide waveform truncation rate to reach 34%; time-varying physiological state causes feature vector jumps to exceed the similarity threshold; the accumulation of phase compensation residuals causes the window and waveform relative position to drift by 8-12ms per hundred cycles and the cross-correlation peak to decay by more than 0.15; and the complete inclusion rate of abnormal waveforms is less than 20% when multiple waveforms coexist, resulting in selective loss of pathological information. In summary, the existing technology, due to the lack of phase binding for segmentation, the lack of quantitative benchmarks for screening, and the lack of dynamic window width adaptation, cannot achieve robust, high-fidelity, and low-latency target waveform capture and feature modeling in scenarios with individual differences, physiological changes, non-ideal acquisition, and multiple coexistence of waveforms.

[0061] Further explanation is needed; please refer to [link / reference]. Figure 2 This embodiment performs segmented preprocessing on the original pulse signal, including:

[0062] A1. Perform orthogonal decomposition on the continuously acquired raw pulse signals to extract in-phase and quadrature components, and calculate the instantaneous phase angle sequence based on the in-phase and quadrature components, specifically:

[0063] Iterate through each corresponding sampling point of the in-phase component sequence and the orthogonal component sequence, and calculate the ratio of the two using the current amplitude of the orthogonal component sequence as the numerator and the current amplitude of the in-phase component sequence as the denominator.

[0064] Perform an arctangent operation on the ratio to obtain the original principal value of the phase angle corresponding to the sampling point. The range of the original principal value of the phase angle is constrained to the interval from negative π to π.

[0065] The original phase angle principal value sequence is scanned sequentially along the time axis, and the phase jump variable between adjacent sampling points is detected. When the absolute value of the phase jump variable is greater than a preset phase jump variable threshold, phase wrapping is determined to have occurred. An integer multiple of 2π is then added to the original phase angle principal value at the current sampling point, causing the phase angle sequence to continuously and monotonically unfold along the time axis, resulting in a wrap-free instantaneous phase angle sequence. Furthermore, the phase jump variable threshold in this embodiment is preset based on the mathematical characteristics of the instantaneous phase angle and the sampling theorem: since the arctangent operation limits the true phase... Within the principal value range of (-π, π), when the absolute value of the true phase difference between adjacent sampling points exceeds π, a jump from -π to π or from π to -π will inevitably occur. Based on the Nyquist sampling theorem, to avoid phase ambiguity, the sampling frequency must ensure that the true phase change between adjacent sampling points is much smaller than π. Therefore, the phase jump threshold is preset to π or a constant slightly smaller than π (such as π-ε, where ε is a preset minimum constant value, such as 0.01) to distinguish between physiological phase changes and step jumps caused by phase winding, thereby accurately performing phase unwinding operations.

[0066] Linear trend removal is performed on the unwound instantaneous phase angle sequence to reduce the phase value of the first sampling point of the sequence to zero, thereby obtaining a normalized instantaneous phase angle sequence with the waveform starting point as the phase reference.

[0067] A2. Obtain the time deviation of the sampling clock relative to the reference source in the current acquisition period, determine the phase compensation value based on the time deviation, and superimpose the phase compensation value onto each phase angle of the instantaneous phase angle sequence to obtain the aligned phase sequence.

[0068] It should be further noted that the phase compensation value determined in this embodiment includes:

[0069] Obtain the time deviation between the sampling clock and the reference source within the current acquisition period. The time deviation is the cumulative timing error of the sampling clock relative to the reference source, and the unit is seconds.

[0070] Read the pre-stored target waveform fundamental frequency, which is the characteristic frequency of the main peak of the physiological signal to be acquired, and the unit is Hertz;

[0071] Multiply the time deviation by the product of the fundamental frequency of the target waveform and 2π to obtain the original phase compensation value, in radians;

[0072] The original phase compensation value is normalized to modulo 2π and mapped to the principal value range of 0 to 2π to obtain the principal phase compensation value.

[0073] The difference between the main phase compensation value and the historical phase compensation value used in the previous acquisition cycle is obtained and recorded as the phase compensation fluctuation amount;

[0074] When the absolute value of the phase compensation fluctuation is less than the preset phase change threshold, it is determined that the current phase compensation value has not changed substantially, and the historical phase compensation value is used as the phase compensation value for the current period to avoid frequent switching of compensation values ​​due to random jitter. Furthermore, the phase change threshold in this embodiment is preset based on the system's short-term stability and jitter tolerance capability of the phase compensation value. Its value is usually small (e.g., 0.01rad~0.05rad), specifically related to the short-term stability of the sampling clock, the fundamental frequency of the target waveform, and the system's requirements for the continuity of compensation values. It aims to filter out invalid fluctuations caused by quantization errors or random noise, ensure smooth updates of the phase compensation value, and avoid frequent switching of compensation values ​​due to minor jitter.

[0075] When the absolute value of the phase compensation fluctuation is greater than or equal to the phase change threshold, it is determined that the phase compensation value has changed significantly, and the phase compensation value of the current period is updated with the main value phase compensation value.

[0076] The phase compensation value is output and superimposed on each phase angle of the instantaneous phase angle sequence to align the segment start reference point with the main peak phase of the target waveform.

[0077] A3. The time point corresponding to the minimum absolute value of the difference between the instantaneous phase angle in the aligned phase sequence and the pre-stored target waveform main peak phase value is taken as the segmentation start reference point. A signal segment with a preset dynamic time window length is extracted from the segmentation start reference point as a candidate signal segment, so that the candidate signal segment at least includes the rising branch, main peak and falling branch of the target waveform.

[0078] It should be further explained that in this embodiment, a signal segment of a preset dynamic time window length is extracted from the segmentation start reference point as a candidate signal segment, including:

[0079] A31. Obtain the historical qualified signal frame queue. Extract the rising limb threshold crossing time and falling limb threshold crossing time of the target waveform from each historical qualified signal frame. Calculate the time difference between the two as the waveform width of the signal frame. Perform sliding window cumulative statistics on the waveform width to obtain the mean waveform width of the current subject. and waveform width standard deviation Furthermore, in this embodiment, the rising limb threshold crossing time is defined as the time when the waveform amplitude first crosses a preset threshold during the rising phase of the target waveform, wherein the preset threshold is set according to 10% to 20% of the main peak amplitude of the signal frame; the falling limb threshold crossing time is defined as the time when the waveform amplitude last crosses the same preset threshold during the falling phase of the target waveform; by calculating the time difference between the falling limb threshold crossing time and the rising limb threshold crossing time, the waveform width of the signal frame is obtained, and the waveform width is input into the historical qualified signal frame queue for sliding window accumulation statistics to generate the mean waveform width and standard deviation of the waveform width of the current test individual, which is used to dynamically adapt the truncation window length of subsequent signal frames. The preset threshold in this embodiment refers to the amplitude threshold used to define the start of the rising branch and the end of the falling branch of the target waveform. Its value is dynamically determined according to a preset ratio based on the peak amplitude of the current signal frame. This ratio is typically set to 10% to 20% and can be adaptively adjusted based on system configuration or the statistical characteristics of historical signal frames to ensure stable capture of the main body of the waveform under different individual and physiological conditions, thereby accurately calculating the threshold moments of the rising and falling branches. For example, assuming the system has locked a qualified pulse signal frame through cross-correlation calculation, and the peak amplitude is detected to be 1.0 (normalized amplitude) within this frame, the preset threshold is 15% of the peak amplitude, i.e., 0.15. Scanning backward from the waveform's starting point, when the amplitude rises from 0.05 to 0.15, the timestamp corresponding to this sampling point is recorded as the "rising branch threshold moment" (e.g., 10ms). Continuing to scan to the peak, the waveform falls from the peak value of 1.0. When the amplitude falls back to 0.15, the timestamp corresponding to this sampling point is recorded as the "falling branch threshold moment" (e.g., 85ms). The time difference between the two (85ms - 10ms = 75ms) is the waveform width of the signal frame. This width value will be included in the sliding window statistics to dynamically adapt the truncated window length of the current subject to cope with individual differences and changes in physiological state.

[0080] A32. Based on the average waveform width and waveform width standard deviation ,according to The first dynamic window length component is calculated, where k is a preset width coverage coefficient, ranging from 1.5 to 3.0. In this embodiment, the width coverage coefficient k is preset based on the statistical principle of normal distribution and engineering tolerance requirements: assuming the waveform width follows a normal distribution, k is used to expand the waveform width mean by several times the standard deviation to cover all possible width values ​​within a specific confidence interval. A value range of k from 1.5 to 3.0 corresponds to a theoretical coverage probability of approximately 86.6% to 99.7%, where k=2.0 serves as a baseline value that can cover approximately 95.4% of physiological fluctuations. The upper limit of 3.0 is used for extreme individual differences or high-noise scenarios to ensure the window fully encompasses rare wide waveforms; the lower limit of 1.5 is used for low-noise and stable waveform scenarios to minimize the inclusion of invalid noise into the window. This coefficient can be adaptively adjusted based on the waveform width variation coefficient (standard deviation / mean) of historical signal frames. When the variation coefficient increases, the k value is automatically increased to enhance inclusiveness; conversely, the k value is decreased to improve the signal-to-noise ratio, achieving a balance between robustness and accuracy in window length adaptation. The "86.6% to 99.7%" in this embodiment is a standard statistical conclusion based on the cumulative probability characteristics of a normal distribution. It corresponds to the theoretical probability that the waveform width data falls within the range near the mean when the width coverage coefficient k is between 1.5 and 3.0. This is a well-known principle in statistics, widely used to describe the variation range of physiological signals and as a basis for setting dynamic parameters.

[0081] A33. Obtain the point-by-point phase difference between the aligned phase sequence of the current acquisition period and the pre-stored main peak phase value of the target waveform, and calculate the root mean square value of the point-by-point phase difference as the phase alignment residual estimate. and will Multiply by the sampling period to convert to time offset compensation. ,in T is the center frequency of the target waveform. s The sampling period is the time interval between adjacent sampling points.

[0082] A34. Based on the time offset compensation amount and the preset compensation gain coefficient. The second dynamic window length component is calculated as follows: The calculation of the second dynamic window length component is based on the time offset compensation amount and a preset compensation gain coefficient. The specific process is as follows: First, obtain the time offset compensation amount calculated in the current acquisition cycle. This compensation amount is expressed in time units and reflects the potential window offset requirement caused by the phase alignment residual. Then, multiply the time offset compensation amount by the preset compensation gain coefficient to adjust the compensation intensity. The compensation gain coefficient is preset according to system calibration or individual differences and is used to control the aggressiveness of the compensation. Next, divide the product result by the sampling cycle of the current acquisition cycle (i.e., the time interval between adjacent sampling points) to convert it into an offset amount in units of sampling points. Finally, round the offset amount to obtain the integer form of the second dynamic window length component, which is used to sum with the first dynamic window length component to form the dynamic window length component. The time window length is determined to ensure that the window truncation can effectively compensate for the phase alignment error caused by sampling clock drift and physiological jitter. It should be further explained that the preset compensation gain coefficient β in this embodiment is determined jointly through system calibration and adaptive optimization: In the offline calibration stage, typical signal samples covering different signal-to-noise ratios, different physiological states, and different sampling clock drift degrees are collected. The optimization objective is to maximize the cross-correlation coefficient between the candidate signal segment and the target waveform template. A grid search or Bayesian optimization method is used to iteratively solve for the global empirical benchmark value of β. In the online operation stage, β is used as a finely adjustable hyperparameter. Based on the actual statistical correlation between the phase alignment residual and the truncation window offset in the historical qualified signal frames of the current test individual, incremental updates are performed using the minimum mean square error criterion. This allows β to dynamically adapt to individual-specific jitter characteristics, ensuring the time offset compensation amount... It can effectively correct the cumulative cutoff deviation caused by sampling clock drift and physiological jitter.

[0083] A35. Perform initial waveform morphology classification on the current candidate signal segment: Extract the zero-crossing rate, peak amplitude, and half-peak width ratio of the signal frame, input them into a pre-set lightweight decision tree model, and classify the current waveform as a normal shape or at least one predefined abnormal shape; when the classification result is an abnormal shape, query the waveform width statistics corresponding to this type of abnormal shape from the historical qualified signal frame queue. and and with and Replace the and The first dynamic window length component is recalculated. In this embodiment, the lightweight decision tree model is constructed by combining offline training and online fine-tuning: First, historical qualified signal frames covering different individuals, different physiological states, and different signal-to-noise ratio scenarios are collected. For each frame, the zero-crossing rate, peak amplitude, and half-peak width ratio are extracted as input feature vectors. Domain experts then label the waveforms based on waveform morphological features, classifying them into normal forms and at least one predefined abnormal form. For example, a form with a significant secondary peak after the main peak, where the amplitude of the secondary peak is 30% to 70% lower than that of the main peak and there is a clear valley between the two peaks is labeled as a dimer wave. A form with an additional small wave in the rising or falling branch, where the amplitude of the small peak is 15% lower than that of the main peak, is labeled as diphtheria wave enhancement. A form with a flattened main peak and a half-peak width ratio of 15% is labeled as diphtheria wave enhancement. Morphological features exceeding the normal mean plus two standard deviations are labeled as pulse-like changes. The labeled dataset is divided into training and validation sets according to a preset ratio. A classification regression tree algorithm or the C4.5 algorithm is used to train a decision tree model, with the Gini coefficient or information gain ratio as the splitting criterion. The tree depth is controlled by a pruning strategy to prevent overfitting, resulting in an initial pre-trained model. During actual system operation, whenever a high-quality signal frame confirmed by the subsequent optimization module is acquired, its feature vector and the qualified label automatically determined by the optimization module are used as incremental samples and input into the local cache. When the number of cached samples reaches a preset threshold, online fine-tuning is triggered. Random forest or gradient boosting decision tree are used to incrementally update the pre-trained model, enabling the model to dynamically adapt to the specific waveform morphology of the current subject and significantly improve the accuracy of the initial classification.

[0084] A36. Summing the first dynamic window length component and the second dynamic window length component yields the dynamic time window length.

[0085] A4. Perform normalized sliding cross-correlation operation on the candidate signal segment and the pre-stored target waveform template in the time domain, and extract the maximum cross-correlation number;

[0086] It should be further noted that the extraction of the maximum cross-correlation coefficient in this embodiment includes:

[0087] Obtain the candidate signal segment, wherein the candidate signal segment has a length of The discrete-time series with a sampling frequency of The sampling frequency The value range is from 500Hz to 2000Hz; sequence length The dynamic time window length and sampling frequency The product is determined.

[0088] Obtain a pre-stored target waveform template, wherein the target waveform template has a length of The reference waveform sequence, the Corresponding to the expected value of the main peak width of the target waveform, and Less than , and The length ratio is configured to be 1:3 to 1:5; the target waveform template in this embodiment is a reference sequence of the main peak morphology extracted from the standard pulse signal of healthy subjects, and its length is set to the expected value of the width of the main peak of the target waveform; in the normalized sliding cross-correlation operation, the ratio of the candidate signal segment length to the template length is configured to be 1:3 to 1:5. This ratio is determined based on the following criteria: the lower limit of 1:3 ensures that the candidate signal segment, while containing the complete main peak, reserves sufficient sliding matching margin for possible physiological stretching of the waveform (such as heart rate variability, individual differences), and avoids the window being too close to the edge of the main peak. The edge effect can cause the loss of the optimal matching position; the upper limit of 1:5 is based on the trade-off between the peak signal-to-noise ratio of cross-correlation and computational complexity. That is, when the candidate segment is too long, the contribution of random noise in the non-target region will significantly increase the denominator of the cross-correlation coefficient (i.e., the product of signal energy), causing the peak to be submerged by background noise. At the same time, an excessively long sliding window will introduce invalid computational overhead. Therefore, a ratio of 1:3 to 1:5 can control the interference of non-target region noise on peak determination within the preset signal-to-noise ratio tolerance, while ensuring that there is a subsequence that is highly similar to the template, thus achieving an engineering balance between matching accuracy and real-time performance.

[0089] The candidate signal segments are subjected to zero-mean processing, and their arithmetic mean is subtracted to obtain the zero-mean candidate sequence;

[0090] The target waveform template is subjected to zero-mean processing, and its arithmetic mean is subtracted to obtain a zero-mean template sequence;

[0091] Calculate the energy of the zero-mean template sequence. The energy This is the sum of the squares of the amplitudes at each sampling point;

[0092] According to the above and The length difference, with an initial length of . - A cross-correlation array R with a step size of 1, wherein the index i of array R starts from 0 and the step size is 1;

[0093] Starting from i = 0 to i = ... - Perform the following operations in sequence:

[0094] Extract a sequence from the zero-mean candidate sequence starting at index i and of length i. A continuous subsequence, used as a sliding window sequence;

[0095] Calculate the energy E_window of the sliding window sequence;

[0096] Calculate the dot product between the sliding window sequence and the zero-mean template sequence to obtain the cross-correlation value. ;

[0097] Calculate the normalized cross-correlation coefficient between the sliding window sequence and the zero-mean template sequence. The normalized cross-correlation coefficient is equal to the dot product of the sliding window sequence and the zero-mean template sequence, divided by the arithmetic square root of the product of the energy value of the zero-mean template sequence and the energy value of the sliding window sequence. Store it in the i-th position of the array R.

[0098] Scan all elements of the array R and extract the maximum value as the maximum cross-correlation coefficient R. max And record R max Corresponding index i max ;

[0099] Output the maximum cross-correlation coefficient R. max and the index i max The index i max It is used to identify the start time of the waveform that matches the target waveform template the most in the candidate signal segment, and to input the alignment phase reference value into the optimization module as the qualified signal frame.

[0100] A5. Compare the maximum cross-correlation coefficient with a preset first cross-correlation coefficient threshold. When the maximum cross-correlation coefficient is greater than or equal to the first cross-correlation coefficient threshold, determine that the candidate signal segment is a qualified signal frame in which the target waveform contour is effectively identified, and record the alignment phase reference value, start timestamp and original sampling data corresponding to the qualified signal frame as a reference frame for the optimization module.

[0101] A6. When the maximum cross-correlation number is less than the first cross-correlation number threshold, it is determined that the candidate signal segment does not contain a complete and effectively identifiable target waveform profile, and the candidate signal segment is marked as an unqualified signal frame.

[0102] A7. Obtain the count value of the non-qualified signal frames in the current acquisition period, and perform an increment operation on the count value of the non-qualified signal frames. Store the updated count value of the non-qualified signal frames in the local cache.

[0103] A8. Clear the original sampling data, alignment phase reference value and start timestamp corresponding to the candidate signal segment from the temporary memory, and do not send any reference frame data to the adaptive processing and optimization module;

[0104] A9. Read the length of the historical qualified signal frame queue. When the historical qualified signal frame queue is empty and the current unqualified signal frame count exceeds the preset maximum consecutive unqualified signal frame threshold, trigger the acquisition parameter reset command. The setting of the maximum consecutive unqualified signal frame threshold in this embodiment needs to comprehensively consider individual physiological stability, noise environment intensity, and acquisition efficiency requirements. Specifically, it is determined in the following way: First, statistically analyze the historical qualified signal frame interval distribution of the current subject in the resting state, calculate the average probability of consecutive unqualified signal frames and the 95% confidence upper limit; then, combined with the real-time data acquisition requirements of clinical application scenarios, the threshold is preset to 5 to 10 frames. The basis for this is that if 5 consecutive frames are judged as unqualified, it usually indicates that the signal quality has been continuously interfered with (e.g., If sensor displacement, limb tremors, or significant changes in an individual's physiological state occur (such as sudden changes in heart rate or increased respiratory disturbances), continued acquisition may introduce excessive noise and cannot be recovered in a short time. In this case, terminating the iteration can avoid invalid calculations and prompt the operator to intervene. If the threshold is set below 5 frames (such as 3 frames), it is easy to misjudge due to occasional noise (such as coughing or brief muscle tension), prematurely terminating a high-quality acquisition process that could be recovered. If the threshold is above 10 frames, it will remain in an inefficient acquisition state for a long time in a noisy environment, wasting system resources and delaying feature output. Therefore, a threshold range of 5 to 10 frames can achieve a balance between suppressing occasional noise interference and responding promptly to continuous signal deterioration, ensuring that the construction of a high-quality signal set has both statistical reliability and meets the timeliness requirements of clinical acquisition.

[0105] The acquisition parameter reset command is used to revert the phase reference of the segmented starting reference point to the initial default phase value, restore the dynamic time window length to the system-preset initial window length and width, and reset the cumulative statistical values ​​of the waveform width mean and waveform width standard deviation. Further, in this embodiment, the system-preset initial window length and width are set jointly based on population statistical characteristics and sampling parameters: firstly, based on large-scale clinical pulse sample statistics, the distribution range of the target waveform main peak width in normal adult populations is obtained (e.g., a 95% confidence interval of 60ms to 120ms), and extreme individual differences and possible widening under pathological conditions are considered. For waveforms (e.g., up to 150ms), the initial window length is set to cover the upper limit of the distribution with an additional engineering margin, such as using 200ms as the baseline value. At the same time, the time length is converted into the number of sampling points (i.e., 200 points) by combining the system sampling frequency (e.g., 1000Hz). This ensures that when individual historical statistical information is lacking, the initial window length can completely encompass the target waveform of any random subject with a probability of more than 99%, avoiding the failure of the first qualified frame capture due to the window being too short. At the same time, by limiting the upper limit of the window length (e.g., not exceeding 250ms), the introduction of invalid noise segments is suppressed, providing a reasonable starting point for subsequent dynamic window length adaptation.

[0106] A10. When the historical qualified signal frame queue is not empty and the current unqualified signal frame count value does not exceed the maximum consecutive unqualified signal frame threshold, the phase compensation value, dynamic time window length and filter parameters of the current acquisition cycle remain unchanged, and the acquisition and judgment process of the next candidate signal segment continues.

[0107] A11. Output the updated count of non-compliant signal frames to the feature modeling module to monitor the number of consecutive non-compliant signal frames and trigger the iteration termination condition of the acquisition cycle.

[0108] This embodiment addresses the systemic deficiencies in existing technologies caused by the lack of phase binding in segmentation, the absence of quantitative benchmarks for screening, and the lack of dynamic window width adaptation by constructing a segmented preprocessing architecture that integrates active phase alignment, dynamic window width adaptation, and quantization threshold admission. First, it generates an aligned phase sequence using orthogonal decomposition and sampling clock deviation compensation. The point where the instantaneous phase angle differs least from the target waveform's main peak is used as the starting reference point for segmentation. This ensures that candidate signal segments are actively phase-bound to the target waveform from the very beginning, fundamentally eliminating the waveform's main peak fragmentation and structural loss of contour information caused by equal-interval segmentation. Simultaneously, it completely abandons the traditional passive trial-and-error acquisition method. The inefficient mode, which requires numerous invalid cycles to obtain qualified frames, achieves a substantial improvement in acquisition throughput and establishes a precise phase reference feedback link for backend modules. Secondly, normalized sliding cross-correlation calculation is used, and a first cross-correlation coefficient threshold is set as the quantization admission criterion for qualified signal frames. This effectively blocks pseudo-qualified frames that are mistakenly judged as qualified due to accidental local coupling of waveforms in noisy scenarios during the coarse screening stage. This avoids the risk of filter type misselection, significant deviation in phase offset calculation, and reference contamination caused by using such distorted frames as reference inputs to downstream tasks, ensuring the correctness of feature space convergence and the fidelity of the reference frame. Furthermore, addressing the four-fold mechanistic mismatch exposed by the fixed window length truncation strategy in clinical applications, this embodiment constructs a dynamic window length calculation framework integrating individual morphological statistics, phase residual compensation, and waveform morphology recognition. Specifically, it generates the first dynamic window length component by real-time statistical analysis of the mean and standard deviation of the waveform width of the current subject using a queue of historical qualified signal frames. This significantly reduces the proportion of invalid noise within the window for narrow waveform individuals, effectively suppresses signal-to-noise ratio degradation, essentially eliminates the truncation phenomenon of the main peak descent branch for wide waveform individuals, and significantly improves the template matching pass rate. The second dynamic window is generated by estimating the phase alignment residual using the root mean square value of the point-by-point phase difference and jointly compensating for the gain coefficient. The long component effectively suppresses the metastable cumulative drift caused by sampling clock drift quantization error and physiological baseline displacement, reducing the drift between the window and waveform relative position to a negligible range and significantly controlling the attenuation of cross-correlation peaks, thereby enhancing the temporal coherence and stability of template matching; a lightweight decision tree model is used to perform initial waveform morphology classification of candidate signal segments, and the width statistics of the corresponding morphology are automatically called to reconstruct the long component of the first dynamic window for pathological waveforms with time-limited abnormalities, improving the complete coverage rate of such abnormal waveforms from severely insufficient to a clinically acceptable level, completely eliminating the risk of selective loss of pathological information in the coarse screening stage, and ensuring the integrity of diagnostic information;In summary, this embodiment achieves forced binding of segmentation and waveform phase through active phase alignment, establishes an objective and accurate qualified frame selection benchmark through quantization threshold admission, and comprehensively covers the complex scenario requirements of individual differences, physiological time variations, residual drift, and the coexistence of multiple morphologies through a dynamic window length adaptation mechanism. This enables the physiological signal adaptive acquisition and feature modeling system to achieve robust, high-fidelity, and low-latency target waveform capture and feature modeling even under complex and non-ideal acquisition conditions, significantly improving the system's clinical adaptability, acquisition efficiency, and signal fidelity.

[0109] In the scenario of adaptive acquisition and feature modeling of physiological signals, there are four core technical defects in the adaptive filtering and feature extraction stages: First, the causal relationship between filter selection and processing target is reversed. Filter decisions are made based on the noisy spectrum without denoising, resulting in a selection accuracy of less than 60% when the signal-to-noise ratio is below 6dB, frequently leading to errors such as misselecting bandpass filters as notch filters and low-pass filters as high-pass filters. Second, filter switching lacks phase and timing consistency constraints. Different periods use filters with group delays and drastically different amplitude-frequency characteristics, causing non-physiological jumps in the feature vectors of the same waveform after filtering. These jumps are 3-5 times the magnitude of physiological fluctuations, triggering subsequent similarity threshold mechanisms and resulting in the misjudgment and discarding of a large number of qualified frames. Third, multiple... Feature extraction suffers from noise coupling and dimensional redundancy. Temporal entropy is significantly affected by residual non-Gaussian noise, contributing 30%–40%. Feature discriminative power is decoupled from waveform morphology. Furthermore, the entropy value is strongly negatively correlated with the zero-crossing rate, yet it is used as an independent dimension for weighted Euclidean distance calculation. This leads to excessive sensitivity of similarity determination to waveform steepness while ignoring fundamental quality issues such as truncation distortion and baseline drift. Fourthly, the feature comparison benchmark is non-statically degraded. Using the previous qualified frame updated via sliding as the sole benchmark results in gradual benchmark drift due to slow distortions such as electrode polarization, leading to system convergence to the distorted feature space without alarms. Additionally, pseudo-qualified frames contaminate the benchmark, creating a chain of errors and lacking an absolute benchmark for independent assessment of feature health. Therefore, it is necessary to further explain that this embodiment uses an adaptive filter bank for noise reduction, extracting multi-dimensional features of the denoised signal frame, including:

[0110] B1. Receive the qualified signal frame and its alignment phase reference value output by the signal acquisition module, and calculate the phase difference between the starting point of the current acquisition cycle and the main peak of the target waveform with the main peak of the qualified signal frame as a reference.

[0111] B2. Generate the starting sampling point offset of the next acquisition cycle based on the phase difference, so that the acquisition window of the next acquisition cycle is aligned with the phase of the target waveform;

[0112] It should be further explained that, in this embodiment, the generation of the starting sampling point offset for the next acquisition cycle based on the phase difference includes:

[0113] The sampling frequency used in the current acquisition cycle is obtained. The sampling frequency is a fixed sampling rate preset by the system or a sampling rate adaptively configured according to the signal frequency band, and its value ranges from 500 Hz to 2000 Hz.

[0114] The fundamental frequency of the target waveform is obtained. The fundamental frequency is the characteristic frequency of the main peak of the physiological signal to be collected. It is determined based on the statistical mean of the main peak interval of the historical qualified signal frames of the current test individual, or by using the pre-stored standard physiological signal fundamental frequency value.

[0115] Divide the phase difference by the product of 2π and the fundamental frequency to convert it into a time offset in seconds. The phase difference is the difference in radians between the starting point of the current acquisition cycle and the main peak phase position of the target waveform.

[0116] Multiply the time offset by the sampling frequency to obtain the original sampling point offset in units of sampling points;

[0117] The original sampling point offset is rounded to the nearest integer to obtain the integer sampling point offset.

[0118] Determine whether the absolute value of the integer sampling point offset exceeds a preset maximum sampling point adjustment threshold, wherein the maximum sampling point adjustment threshold is 10% to 20% of the total number of sampling points in a single acquisition cycle;

[0119] When the absolute value of the integer sampling point offset is less than or equal to the maximum sampling point adjustment threshold, the integer sampling point offset is used as the starting sampling point offset of the next acquisition cycle.

[0120] When the absolute value of the integer sampling point offset is greater than the maximum sampling point adjustment threshold, the integer sampling point offset is limited by the maximum sampling point adjustment threshold, and the limited integer sampling point offset is used as the starting sampling point offset of the next acquisition cycle.

[0121] The starting sampling point offset is output to the acquisition control unit, which is used to shift the starting point of the acquisition window forward or backward relative to the starting point of the current cycle by the number of sampling points corresponding to the starting sampling point offset, so as to achieve dynamic alignment of the acquisition window with the phase of the target waveform.

[0122] B3. Acquire a new signal frame for the next acquisition cycle. Read the reference spectrum of the previous qualified signal frame after noise reduction from the local buffer as a spectrum template. Calculate the frequency domain cross-correlation coefficient between the power spectral density of the new signal frame and the spectrum template. When the frequency domain cross-correlation coefficient is greater than a preset second cross-correlation coefficient threshold, filter the new signal frame using the same filter type and parameters as in the previous cycle. Further, in this embodiment, the second cross-correlation coefficient threshold is used to determine the similarity between the power spectral density of the new signal frame and the reference spectrum of the previous cycle, and its setting is based on spectral stability. The requirement is to determine the cross-correlation coefficients between historical qualified signal frames of the current subject under resting conditions in conjunction with the signal-to-noise ratio tolerance: First, statistically analyze the frequency domain cross-correlation coefficient distribution between the current subject and historical qualified signal frames, calculate its mean and standard deviation, and preset the threshold as the mean minus one to two times the standard deviation (e.g., 0.85 to 0.95). When the cross-correlation coefficient between the new signal frame and the reference spectrum is greater than this threshold, it indicates that the spectral structure has not changed substantially, and using the previous cycle filter can ensure the continuity of the phase response. When it is lower than this threshold, it is determined that the spectrum has changed significantly (e.g., noise spectrum intrusion, main peak drift), and the filter needs to be switched to adapt to the new spectral characteristics. For example, if the mean of the cross-correlation coefficient between historical frames of a subject is 0.92 and the standard deviation is 0.03, then the threshold is set to 0.89. When the correlation coefficient of the new frame drops to 0.88, the filter is switched to avoid characteristic instability caused by frequent switching due to small fluctuations, while responding promptly to substantial spectral variations, achieving a balance between robustness and adaptability in filter decision-making.

[0123] B4. When the frequency domain cross-correlation coefficient is less than or equal to the second cross-correlation coefficient threshold, based on the peak frequency and energy bandwidth of the spectrum template, a filter type and cutoff parameters that cover the peak frequency in the passband and match the energy bandwidth in the pre-stored filter library are selected. The group delay response of the selected filter is differentially compensated with the group delay response of the previous cycle filter, so that the phase difference between the phase responses of the two cycles filter at the peak frequency of the target waveform is less than the preset third phase difference threshold. The compensated filter parameters are then used to adaptively filter and denoise the new signal frame. It should be further noted that the third phase difference threshold in this embodiment is used to constrain the phase response deviation of the two cycles filter at the peak frequency. Its setting is determined jointly based on the waveform morphology fidelity requirement and the sampling time resolution: firstly, the peak period of the target waveform is used as a reference. To ensure that the time offset corresponding to the phase deviation is much smaller than the sampling interval, avoid the main peak position jitter or waveform distortion caused by phase mismatch. For example, if the sampling frequency is 1000Hz and the sampling interval is 1ms, and the main peak frequency is 100Hz (period 10ms), then the phase difference threshold is set to 5 degrees (about 0.14ms time offset). This ensures that the offset is less than 15% of the sampling interval, and the main peak alignment error is still within an acceptable range after the waveform is processed by different filters. At the same time, combined with the human eye's sensitivity to waveform morphology, the threshold range is preset to 3 to 8 degrees. When the actual deviation exceeds this range (such as a phase difference of 12 degrees calculated after a switch, corresponding to a time offset of 0.33ms), differential compensation is triggered to force correction to within the threshold. This allows for dynamic switching of filters while ensuring that the output waveforms of the previous and subsequent cycles have visual consistency and comparability at the clinical feature extraction level.

[0124] B5. Extract the time-domain waveform sequence of the new signal frame after noise reduction, calculate its time-domain entropy, zero-crossing rate, peak amplitude and main peak half-width, and form an initial multidimensional feature vector.

[0125] B6. Perform orthogonalization transformation on the initial multidimensional feature vector to obtain independent feature components whose cross-correlation coefficients between dimensions are lower than a preset fourth cross-correlation coefficient threshold, thus forming a feature vector for similarity discrimination. It should be further explained that the fourth cross-correlation coefficient threshold in this embodiment is used to determine the degree of independence between the feature components of each dimension after orthogonalization transformation. Its setting is determined jointly based on feature decoupling requirements and information retention rate: First, using the commonly used threshold for significant correlation in statistics as a reference, a cross-correlation coefficient absolute value greater than 0.5 is considered a strong correlation, and less than 0.3 is considered a weak correlation or no correlation. To ensure that the information carried by each dimension of the transformed feature vector is not repeated, this threshold is preset to between 0.2 and 0.3, so that the absolute value of the cross-correlation coefficients of any two independent feature components is lower than this threshold, thereby eliminating dimensional redundancy in the original features caused by the negative correlation between entropy and zero-crossing rate, and the positive correlation between peak amplitude and half-peak width. Simultaneously, by monitoring the cumulative contribution rate of variance during the orthogonalization transformation process, it is ensured that the first few independent feature components can retain more than 90% of the energy information of the original features, avoiding the loss of effective discrimination information due to excessive pursuit of low correlation. For example, after performing principal component analysis or independent component analysis on the initial four-dimensional feature vector extracted from a signal frame, four new feature components are obtained. The cross-correlation coefficients of each pair are calculated, and the maximum value is found to be 0.85. After retaining the first two components with a variance contribution rate greater than 95%, the maximum cross-correlation coefficient is recalculated, and the maximum cross-correlation coefficient drops to 0.12, which is lower than the preset threshold of 0.25. That is, the feature vector used for similarity discrimination is formed by the first two components, which not only ensures the independence between dimensions, but also fully retains the main discriminative ability of the original features.

[0126] B7. Obtain the first qualified signal frame of the current subject as the absolute reference frame, and sequentially perform noise reduction processing, temporal feature extraction, and orthogonalization transformation on the absolute reference frame to obtain the absolute reference feature vector, specifically:

[0127] The first qualified signal frame of the current subject is obtained as the absolute baseline frame. First, the raw sampled data of this frame is denoised using an adaptive filter bank to obtain a clean time-domain waveform sequence. Then, the time-domain entropy, zero-crossing rate, peak amplitude, and half-peak width of the main peak are extracted from this waveform sequence to form an initial multidimensional feature vector. An orthogonal transformation (such as principal component analysis or independent component analysis) is performed on this initial multidimensional feature vector to obtain independent feature components whose cross-correlation coefficients between dimensions are lower than a preset threshold. These independent feature components are then combined to obtain the absolute baseline feature vector. For example, after denoising, the first qualified signal frame of a subject has a time-domain entropy of 2.5, a zero-crossing rate of 5, a peak amplitude of 0.95, and a half-peak width of 35 milliseconds, forming a four-dimensional initial feature vector. After principal component analysis, the cumulative variance contribution rate of the first two principal components reaches 96%, and their cross-correlation coefficient is 0.18, lower than the preset threshold of 0.25. Therefore, the first two principal components are used as the absolute baseline feature vector for this individual, for feature comparison and drift monitoring of subsequent signal frames.

[0128] B8. Simultaneously, the dynamic reference feature vector, obtained after denoising and orthogonalization transformation of the previous qualified signal frame, is read from the local cache. Specifically, the dynamic reference feature vector is an independent feature component obtained after denoising, feature extraction, and orthogonalization transformation of the previous qualified signal frame, used as a sliding reference for real-time tracking of recent waveform features. Its calculation process is completely consistent with that of the absolute reference feature vector: First, the previous qualified signal frame is denoised using an adaptive filter bank to obtain a clean time-domain waveform sequence; then, the time-domain entropy, zero-crossing rate, peak amplitude, and main peak half-peak width are extracted from this sequence to form an initial multi-dimensional feature vector; finally, the same orthogonalization transformation as the absolute reference is performed on this initial vector (using the same transformation matrix or projection coefficients to ensure the comparability of the feature space), obtaining independent feature components whose cross-correlation coefficients in each dimension are lower than a preset fourth threshold, and combining them yields the dynamic reference feature vector. For example, the initial features extracted from the previous qualified signal frame after noise reduction are a time-domain entropy of 2.4, a zero-crossing rate of 5, a peak amplitude of 0.94, and a main peak half-peak width of 36 milliseconds. After principal component analysis transformation (using the transformation matrix of the absolute reference), the coordinate values ​​of the first two principal components are 1.52 and 0.31, respectively, and their cross-correlation coefficient is 0.15, which is lower than the threshold of 0.25. Therefore, [1.52, 0.31] is used as the dynamic reference feature vector of this frame, which is used to compare the first Euclidean distance with the feature vector of the next new signal frame.

[0129] The feature vector used for similarity determination is compared with the dynamic benchmark feature vector and the absolute benchmark feature vector respectively: the first Euclidean distance between the feature vector used for similarity determination and the dynamic benchmark feature vector, and the second Euclidean distance between the feature vector used for similarity determination and the absolute benchmark feature vector are calculated.

[0130] B9. When the first Euclidean distance is less than a preset fifth threshold and the second Euclidean distance is less than a preset sixth threshold, it is determined that the new signal frame maintains feature consistency with the historical qualified signal frame and no benchmark drift has occurred. The new signal frame is stored in the high-quality signal set, and the dynamic benchmark feature vector is updated with the feature vector used for similarity discrimination. It should be further explained that the fifth threshold in this embodiment is used to determine the feature similarity between the new signal frame and the previous qualified signal frame. Its setting is based on the sliding window statistics of the Euclidean distance between the historical qualified signal frames of the current test individual: the Euclidean distance between each pair of feature vectors of the most recent N consecutive qualified frames is calculated in real time, and their mean μ_dyn and standard deviation σ_dyn are obtained. The fifth threshold is preset to μ_dyn + ... α×σ_dyn, where α is a relaxation coefficient of 1.5 to 2.5, allows the threshold to dynamically follow short-term physiological fluctuations; the sixth threshold is used to determine the degree of feature deviation between the new signal frame and the first qualified signal frame (absolute reference). Its setting is based on the maximum allowable range of physiological state variation for the individual: starting from the absolute reference, the Euclidean distance between the feature vector of the historical qualified frames of the individual in a stable state and the absolute reference is statistically calculated, and the upper limit of confidence of 95% is taken as the base value of the sixth threshold, and verified in combination with the clinical tolerance for feature drift (such as the feature space distance corresponding to the main peak amplitude change of no more than 20% and the half-peak width change of no more than 15%). The correction ensures that the sixth threshold can both encompass normal physiological fluctuations and effectively capture reference drift caused by progressive distortion or changes in sensor state. It should be further explained that the dynamic reference feature vector is updated in this embodiment as follows: when a new signal frame satisfies the condition that the first Euclidean distance is less than the fifth threshold and the second Euclidean distance is less than the sixth threshold, the feature vector obtained after noise reduction, feature extraction and orthogonal transformation of the new signal frame is directly overwritten with the dynamic reference feature vector corresponding to the previous qualified signal frame stored in the local cache, so that the dynamic reference always follows the latest confirmed qualified signal frame and maintains the ability to track recent waveform feature changes in real time.

[0131] B10. When the first Euclidean distance is less than the fifth threshold but the second Euclidean distance is greater than or equal to the sixth threshold, it is determined that the reference has undergone asymptotic drift. The absolute reference feature vector is used to cover the dynamic reference feature vector and the drift counter is reset. When the first Euclidean distance is greater than or equal to the fifth threshold, the new signal frame is determined to be unqualified. The new signal frame is discarded and the count of unqualified signal frames is incremented.

[0132] It should be further explained that the differential compensation of the group delay response of the selected filter with the group delay response of the filter in the previous cycle in this embodiment includes:

[0133] B41. Obtain the first phase response value of the filter in the previous cycle at the main peak frequency of the target waveform;

[0134] B42. Obtain the second phase response value of the selected filter at the main peak frequency point;

[0135] B43. Calculate the difference between the second phase response value and the first phase response value, as the phase deviation amount;

[0136] B44. Generate an all-pass phase compensation filter based on the phase deviation amount, wherein the phase response value of the all-pass phase compensation filter at the main peak frequency point is configured as the opposite of the phase deviation amount;

[0137] B45. Cascade the all-pass phase compensation filter with the selected filter to obtain the compensated filter parameters, such that the phase difference between the phase response value of the compensated filter at the main peak frequency point and the first phase response value is less than a preset third phase difference threshold.

[0138] This embodiment addresses four core defects in adaptive filtering and feature extraction: inverted causal basis for selection, disruption of timing consistency during filter switching, noise coupling and redundancy in feature dimensions, and non-static degradation of the comparison benchmark. It constructs a filter stabilization decision mechanism anchored by a reference spectrum template, a phase response consistency constraint using group delay differential compensation, a feature decoupling and dimensionality reduction architecture with orthogonalization transformation as its core, and a feature health assessment system guaranteed by dual threshold decisions using both absolute and dynamic benchmarks. At the filter selection level, it abandons the traditional paradigm of directly relying on the spectrum of noisy new signal frames. Instead, it uses the reference spectrum of a historical qualified signal frame after noise reduction as a spectral template. It determines spectral similarity by calculating the frequency domain cross-correlation coefficient between the power spectral density of the new signal frame and the spectral template. When the cross-correlation coefficient exceeds the second cross-correlation coefficient threshold, it forces the use of the previous cycle's filter parameters, fundamentally cutting off the path dependence of the noise spectrum on filter decisions. This significantly improves the selection accuracy in low signal-to-noise ratio scenarios from insufficient to a high level, completely eliminating causal inversions such as bandpass filter misselection as notch filter and low-pass filter misselection as high-pass filter. When a substantial change in the spectrum necessitates filter switching, it uses frequency... The main peak frequency and energy bandwidth of the spectral template are selected using a passband-matched filter. Differential compensation is applied to the group delay response of the selected filter and the previous cycle filter, ensuring that the phase difference between the two cycles at the main peak frequency of the target waveform is constrained within the third phase difference threshold. This suppresses non-physiological jumps in the eigenvectors caused by filter type transitions, significantly compressing the jump amplitude to a lower level within the physiological fluctuation range. This ensures that waveforms from the same source maintain temporal morphology comparability and feature stability even after processing by different filters. At the feature extraction level, the contribution of residual non-Gaussian noise to the temporal entropy value is addressed. The discriminative power degradation and dimensional redundancy problems caused by the high proportion of entropy and the strong negative correlation between entropy and zero-crossing rate are addressed by performing an orthogonalization transformation on the initial multidimensional feature vector to obtain independent feature components whose cross-correlation coefficients of each dimension are lower than the fourth cross-correlation coefficient threshold. This compresses the noise energy to the low-contribution dimension after the orthogonal transformation, significantly reducing the proportion of noise components decoupled from waveform morphology in the feature vector. At the same time, it eliminates the information overlap between entropy and zero-crossing rate, effectively reducing the sensitivity of similarity judgment to the steepness of waveform changes, while greatly improving the response sensitivity to essential quality problems such as truncation distortion and baseline drift.At the feature comparison benchmark level, the first qualified signal frame of the current test individual is introduced as the absolute benchmark frame, and its absolute benchmark feature vector is extracted. This absolute benchmark feature vector, together with the dynamically updated feature vector, forms a dual-benchmark joint decision architecture. A new signal frame is considered qualified only when the first Euclidean distance between its feature vector and the dynamic benchmark feature vector is less than the fifth threshold, and the second Euclidean distance between its feature vector and the absolute benchmark feature vector is less than the sixth threshold. Thus, even in slow distortion scenarios such as electrode polarization, if the dynamic benchmark gradually drifts across multiple consecutive frames, the absolute benchmark can still limit the cumulative offset within the physiological variation range through the second Euclidean distance threshold, completely preventing the risk of the system converging to the distorted feature space without alarm. Simultaneously, when a false qualified frame contaminates the dynamic benchmark... When the second Euclidean distance deviates from the absolute reference, it triggers a threshold overflow. The system immediately overwrites the contaminated reference with the absolute reference feature vector and resets the drift counter, breaking the chain of reference contamination. In summary, this embodiment, through filter stabilization decision guided by spectrum templates, phase consistency constraints of group delay differential compensation, feature decoupling and dimensionality reduction of orthogonal transformation, and drift suppression and contamination repair through joint decision of dual references, enables the adaptive processing module to achieve accurate filter selection, good waveform fidelity, strong feature discrimination, and stable reference source tracing in non-stationary noise, individual differences, and long-term acquisition scenarios. The spectral consistency, morphological comparability, feature orthogonality, and reference absoluteness of the output signal frame are all substantially improved.

[0139] Example 2

[0140] Please see Figure 3 Another embodiment of the present invention provides a pulse feature extraction method based on adaptive filtering, comprising:

[0141] S1. Periodically acquire raw pulse signals and perform segmented preprocessing on the raw pulse signals until at least part of the contour containing the target waveform features in the signal segment is identified, and the corresponding signal frame is recorded.

[0142] S2. Receive the signal frame and dynamically adjust the starting point of the next acquisition cycle based on its waveform structure characteristics; for the new signal frame acquired in the next acquisition cycle, call the adaptive filter bank for noise reduction processing, extract the multidimensional features of the noise-reduced signal frame, including the time-domain entropy value, and compare the multidimensional features with the corresponding features of the previous qualified signal frame recorded by the signal acquisition module, and determine whether to add the new signal frame to the high-quality signal set based on a preset similarity threshold; wherein the filter type and parameters are dynamically selected according to the spectral characteristics of the current signal frame;

[0143] S3. Continuously monitor the output of the signal acquisition module and the optimization module. When multiple new signal frames are not added to the high-quality signal set, terminate the iterative optimization of the current acquisition cycle. Integrate and analyze all signal frames in the high-quality signal set, extract the position information and quantization morphological parameters of the target waveform features, and construct a high-quality signal set describing individual-specific waveform changes.

[0144] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

[0145] If the technical solution disclosed herein involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of explicit consent. For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. A pulse feature extraction system based on adaptive filtering, characterized in that, include: The signal acquisition module is used to periodically acquire raw pulse signals and perform segmented preprocessing on the raw pulse signals until at least a partial outline containing the target waveform features in the signal segment is identified, and the corresponding signal frame is recorded. An optimization module is used to receive the signal frame and dynamically adjust the starting point of the next acquisition cycle according to its waveform structure characteristics. For the new signal frame acquired in the next acquisition cycle, an adaptive filter bank is called to perform noise reduction processing, and multidimensional features of the noise-reduced signal frame are extracted, including the temporal entropy value. The multidimensional features are then compared with the corresponding features of the previous qualified signal frame recorded by the signal acquisition module. Based on a preset similarity threshold, it is determined whether the new signal frame should be added to the high-quality signal set. The filter type and parameters are dynamically selected based on the spectral characteristics of the current signal frame; The feature modeling module is used to continuously monitor the output of the signal acquisition module and the optimization module. When multiple consecutive new signal frames are not added to the high-quality signal set, the iterative optimization of the current acquisition cycle is terminated. All signal frames in the high-quality signal set are integrated and analyzed to extract the position information and quantization morphological parameters of the target waveform features and construct a high-quality signal set describing individual-specific waveform changes.

2. The pulse feature extraction system based on adaptive filtering as described in claim 1, characterized in that, The segmented preprocessing of the original pulse signal includes: The continuously acquired raw pulse signals are orthogonally decomposed to extract in-phase and quadrature components, and the instantaneous phase angle sequence is calculated based on the in-phase and quadrature components. Obtain the time deviation of the sampling clock relative to the reference source in the current acquisition period, determine the phase compensation value based on the time deviation, and superimpose the phase compensation value onto each phase angle of the instantaneous phase angle sequence to obtain the aligned phase sequence; The time point corresponding to the minimum absolute value of the difference between the instantaneous phase angle in the aligned phase sequence and the pre-stored phase value of the main peak of the target waveform is taken as the segmentation start reference point. A signal segment of a preset dynamic time window length is extracted from the segmentation start reference point as a candidate signal segment, so that the candidate signal segment at least includes the rising branch, the main peak and the falling branch of the target waveform.

3. The pulse feature extraction system based on adaptive filtering as described in claim 2, characterized in that, The segmented preprocessing of the original pulse signal also includes: The candidate signal segment and the pre-stored target waveform template are subjected to normalized sliding cross-correlation in the time domain to extract the maximum cross-correlation number. The maximum cross-correlation coefficient is compared with a preset first cross-correlation coefficient threshold. When the maximum cross-correlation coefficient is greater than or equal to the first cross-correlation coefficient threshold, the candidate signal segment is determined to be a qualified signal frame in which the target waveform profile is effectively identified. The alignment phase reference value, start timestamp and original sampling data corresponding to the qualified signal frame are recorded as reference frames for the optimization module. When the maximum cross-correlation coefficient is less than the first cross-correlation coefficient threshold, it is determined that the candidate signal segment does not contain a complete and effectively identifiable target waveform profile, and the candidate signal segment is marked as an unqualified signal frame.

4. The pulse feature extraction system based on adaptive filtering as described in claim 3, characterized in that, A signal segment of a preset dynamic time window length is extracted from the starting reference point of the segmentation as a candidate signal segment, including: A queue of historical qualified signal frames is obtained. From each historical qualified signal frame, the rising limb threshold crossing time and the falling limb threshold crossing time of the target waveform are extracted. The time difference between the two is calculated as the waveform width of the signal frame. The waveform width is then accumulated and statistically analyzed using a sliding window to obtain the mean waveform width of the current subject. and waveform width standard deviation ; Based on the average waveform width and waveform width standard deviation ,according to Calculate the length component of the first dynamic window, where k is the preset width coverage coefficient; Obtain the point-by-point phase difference between the aligned phase sequence of the current acquisition period and the pre-stored main peak phase value of the target waveform, and calculate the root mean square value of the point-by-point phase difference as the phase alignment residual estimate. and will Multiply by the sampling period to convert to time offset compensation. ,in T is the center frequency of the target waveform. s The sampling period is the time interval between adjacent sampling points.

5. The pulse feature extraction system based on adaptive filtering as described in claim 4, characterized in that, The method further includes extracting a signal segment of a preset dynamic time window length from the segmentation start reference point as a candidate signal segment, and also includes: Based on the time offset compensation amount and the preset compensation gain coefficient Calculate the second dynamic window length component; Perform initial waveform morphology classification on the current candidate signal segment: extract the zero-crossing rate, peak amplitude, and half-peak width ratio of the signal frame, input them into a pre-set lightweight decision tree model, and classify the current waveform as a normal shape or at least one predefined abnormal shape; when the classification result is an abnormal shape, query the waveform width statistics corresponding to this type of abnormal shape from the historical qualified signal frame queue. and and with and Replace the and Recalculate the first dynamic window length component; The dynamic time window length is obtained by summing the first dynamic window length component and the second dynamic window length component.

6. The pulse feature extraction system based on adaptive filtering as described in claim 5, characterized in that, The step of calling an adaptive filter bank for noise reduction and extracting multi-dimensional features of the denoised signal frame includes: Receive the qualified signal frame and its alignment phase reference value output by the signal acquisition module, and calculate the phase difference between the starting point of the current acquisition cycle and the main peak of the target waveform with the main peak of the qualified signal frame as a reference. The starting sampling point offset for the next acquisition cycle is generated based on the phase difference, so that the acquisition window of the next acquisition cycle is aligned with the phase of the target waveform. Acquire a new signal frame for the next acquisition cycle. Read the reference spectrum of the previous qualified signal frame after noise reduction from the local buffer as a spectrum template. Calculate the frequency domain cross-correlation coefficient between the power spectral density of the new signal frame and the spectrum template. When the frequency domain cross-correlation coefficient is greater than a preset second cross-correlation coefficient threshold, filter the new signal frame using the filter type and parameters adopted in the previous cycle.

7. The pulse feature extraction system based on adaptive filtering as described in claim 6, characterized in that, The step of calling an adaptive filter bank for noise reduction and extracting multidimensional features of the denoised signal frame also includes: When the frequency domain cross-correlation coefficient is less than or equal to the second cross-correlation coefficient threshold, based on the main peak frequency and energy bandwidth of the spectrum template, a filter type and cutoff parameter that covers the main peak frequency and has stopband attenuation characteristics matching the energy bandwidth are selected from the pre-stored filter library. The group delay response of the selected filter is differentially compensated with the group delay response of the previous cycle filter so that the phase difference between the phase responses of the two cycles filters at the main peak frequency of the target waveform is less than the preset third phase difference threshold. The compensated filter parameters are then used to adaptively filter and reduce noise in the new signal frame. Extract the time-domain waveform sequence of the new signal frame after noise reduction, calculate its time-domain entropy, zero-crossing rate, peak amplitude and main peak half-peak width, and form an initial multidimensional feature vector; An orthogonalization transformation is performed on the initial multidimensional feature vector to obtain independent feature components whose cross-correlation coefficients between each dimension are lower than a preset fourth cross-correlation coefficient threshold, thus forming a feature vector for similarity discrimination.

8. The pulse feature extraction system based on adaptive filtering as described in claim 7, characterized in that, The step of calling an adaptive filter bank for noise reduction and extracting multidimensional features of the denoised signal frame also includes: The first qualified signal frame of the current subject is obtained as the absolute reference frame, and noise reduction, temporal feature extraction and orthogonal transformation are performed on the absolute reference frame in sequence to obtain the absolute reference feature vector. Simultaneously, the dynamic reference feature vector obtained after noise reduction and orthogonal transformation of the previous qualified signal frame is read from the local cache; the feature vector used for similarity discrimination is compared with the dynamic reference feature vector and the absolute reference feature vector respectively: the first Euclidean distance between the feature vector used for similarity discrimination and the dynamic reference feature vector, and the second Euclidean distance between the feature vector used for similarity discrimination and the absolute reference feature vector are calculated; When the first Euclidean distance is less than the preset fifth threshold and the second Euclidean distance is less than the preset sixth threshold, it is determined that the new signal frame maintains the same features as the historical qualified signal frame and no reference drift has occurred. The new signal frame is stored in the high-quality signal set, and the dynamic reference feature vector is updated with the feature vector used for similarity discrimination. When the first Euclidean distance is less than the fifth threshold but the second Euclidean distance is greater than or equal to the sixth threshold, it is determined that the benchmark has undergone asymptotic drift. The absolute benchmark feature vector is then used to cover the dynamic benchmark feature vector, and the drift counter is reset. When the first Euclidean distance is greater than or equal to the fifth threshold, the new signal frame is determined to be unqualified, the new signal frame is discarded, and the count of unqualified signal frames is incremented.

9. The pulse feature extraction system based on adaptive filtering as described in claim 8, characterized in that, The step of differentially compensating the group delay response of the selected filter with the group delay response of the filter in the previous cycle includes: Obtain the first phase response value of the filter in the previous cycle at the main peak frequency of the target waveform; Obtain the second phase response value of the selected filter at the main peak frequency point; The difference between the second phase response value and the first phase response value is calculated as the phase deviation. Based on the phase deviation, an all-pass phase compensation filter is generated, and the phase response value of the all-pass phase compensation filter at the main peak frequency is configured to be the opposite of the phase deviation. The full-pass phase compensation filter is cascaded with the selected filter to obtain the compensated filter parameters, such that the phase difference between the phase response value of the compensated filter at the main peak frequency point and the first phase response value is less than a preset third phase difference threshold.

10. A pulse feature extraction method based on adaptive filtering, implemented based on the pulse feature extraction system based on adaptive filtering as described in any one of claims 1-9, characterized in that, include: The raw pulse signal is periodically acquired and segmented for preprocessing until at least a partial outline containing the target waveform features is identified in the signal segment, and the corresponding signal frame is recorded. Receive the signal frame and dynamically adjust the starting point of the next acquisition cycle according to its waveform structure characteristics; For the new signal frame acquired in the next acquisition cycle, an adaptive filter bank is called to perform noise reduction processing, and multidimensional features of the noise-reduced signal frame are extracted, including the temporal entropy value. The multidimensional features are then compared with the corresponding features of the previous qualified signal frame recorded by the signal acquisition module. Based on a preset similarity threshold, it is determined whether the new signal frame should be added to the high-quality signal set. The filter type and parameters are dynamically selected based on the spectral characteristics of the current signal frame; The output of the signal acquisition module and the optimization module are continuously monitored. When multiple new signal frames are not added to the high-quality signal set, the iterative optimization of the current acquisition cycle is terminated. All signal frames in the high-quality signal set are integrated and analyzed to extract the position information and quantization morphological parameters of the target waveform features and construct a high-quality signal set describing individual-specific waveform changes.

Images