Pulse condition signal time-frequency domain feature extraction method

Through wavelet decomposition, local mean decomposition and empirical modal decomposition, seven sets of time-frequency domain characteristics of pulse signals are extracted, which solves the problem of incomplete features in the existing technology, and achieves more accurate pulse signal classification recognition and feature reflection.

CN118902409APending Publication Date: 2024-11-08BEIJING INST OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410965062.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing pulse signal time frequency domain feature extraction methods are not comprehensive and cannot accurately reflect pulse information.

Method used

The sensors obtain the three-channel pulse signals at the human body's cub, cube and cube positions, perform wavelet decomposition, local mean decomposition and empirical modal decomposition, and extract seven sets of time-frequency domain features, including energy, correlation numbers and features after Hilbert transformation , weighted average and normalization to obtain more comprehensive time-frequency domain features.

Benefits of technology

It realizes more accurate and detailed feature extraction of pulse signals, improves the accuracy of classification recognition, avoids aliasing of decomposition results, and ensures that features can more accurately reflect pulse information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118902409A_ABST
    Figure CN118902409A_ABST
Patent Text Reader

Abstract

The invention relates to a pulse condition signal time-frequency domain feature extraction method, and belongs to the technical field of signal processing. Comprising the steps that pulse signals are obtained and preprocessed; extracting energy corresponding to the detail signal through wavelet decomposition as a first time-frequency domain feature; extracting the average intensity, the spectrum centroid and the average energy contribution rate corresponding to the PF component as a second time-frequency domain feature, a third time-frequency domain feature and a fourth time-frequency domain feature by using local mean decomposition; utilizing empirical mode decomposition to extract energy corresponding to the eigenmode function as a fifth time-frequency domain feature; and extracting Hilbert spectrum energy and marginal spectrum energy corresponding to the transformed eigenmode function through Hilbert transform as a sixth time-frequency domain feature and a seventh time-frequency domain feature. According to the method, seven groups of time-frequency domain features of the pulse condition signals are obtained through wavelet decomposition, local mean decomposition and empirical mode decomposition, more comprehensive time-frequency domain feature extraction of the pulse condition signals is realized, and a basis is provided for more accurate and more detailed classification and recognition of the pulse condition signals.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of signal processing, and in particular to a method for extracting time-frequency domain features of a pulse signal. Background Art

[0002] Pulse diagnosis is one of the main methods of TCM in disease diagnosis. It is used to observe the changes in the human body's pulse. The pulse wave is formed by the heartbeat propagating along the arteries and blood flow. The pulse wave is affected by the interaction of multiple physiological factors such as the elasticity of the blood vessel wall and vascular resistance. It contains rich information and can well reflect the physical condition of the human body.

[0003] Traditional pulse diagnosis relies on the doctor's diagnosis experience and subjective judgment by pressing the pulse points of the patient with fingers to feel the pulse under different pressures. Along with the development of biomedical engineering, it becomes a research hotspot to automatically obtain human pulse signal and carry out biometric identification and feature parameter extraction to pulse signal. The feature extracted by the existing pulse signal time-frequency domain feature extraction method is not comprehensive, thereby can not accurately reflect the pulse information. Therefore, it is particularly important to provide a method that can extract more comprehensive pulse signal time-frequency domain feature. Summary of the invention

[0004] In view of the above analysis, the present invention aims to provide a method for extracting time-frequency domain features of a pulse signal, so as to solve the problem that the features extracted by the existing time-frequency domain feature extraction method of a pulse signal are incomplete and thus cannot accurately reflect the pulse information.

[0005] The invention provides a method for extracting time-frequency domain features of a pulse signal, the method comprising the following steps:

[0006] Acquire three-channel pulse signals at three positions of the human body, namely, Cun, Guan and Chi, through sensors, and pre-process the three-channel pulse signals to obtain pre-processed pulse signals of each channel;

[0007] Decomposing the pulse signals preprocessed by each channel by wavelet to obtain the detail signal and approximate signal of each channel, extracting the energy corresponding to the detail signal of each channel and performing weighted average as the first time-frequency domain feature;

[0008] Decomposing the pulse signals preprocessed by each channel by using local mean decomposition to obtain the PF components of each channel, and obtaining the second time-frequency domain feature, the third time-frequency domain feature, and the fourth time-frequency domain feature based on the PF components of each channel;

[0009] The preprocessed pulse signals of each channel are decomposed by empirical mode decomposition to obtain the initial intrinsic mode function of each channel, and the initial intrinsic mode function with a correlation coefficient greater than a threshold is selected as the intrinsic mode function of each channel, and the fifth time-frequency domain feature is obtained based on the intrinsic mode function of each channel; the intrinsic mode function of each channel is subjected to Hilbert transform to obtain the transformed intrinsic mode function, and the sixth time-frequency domain feature and the seventh time-frequency domain feature are obtained based on the transformed intrinsic mode function.

[0010] Further, the method of utilizing empirical mode decomposition to decompose the pretreated pulse signal of each channel to obtain the initial intrinsic mode function of each channel, and selecting the initial intrinsic mode function whose correlation coefficient is greater than a threshold value as the eigenmode function of each channel comprises the following steps:

[0011] Step S41, generating a lower envelope based on the pacemaker point of the preprocessed pulse signal of each channel, generating an upper envelope based on the peak point of the preprocessed pulse signal of each channel, and generating a mean envelope of the pulse signal of each channel based on the upper and lower envelopes;

[0012] Step S42, subtracting the corresponding mean envelope from the pre-processed pulse signal of each channel to obtain an initial eigenmode function;

[0013] Step S43, subtracting the initial intrinsic mode function from the pre-processed pulse signal of each channel to obtain a difference signal;

[0014] Step S44, taking the difference signal as the pulse signal after preprocessing of each channel, repeating steps S41 to S43 until the difference signal is a monotonic function, and taking the difference signal obtained at this time as the residual signal of empirical mode decomposition;

[0015] Step S45, normalize all the initial eigenmode functions and the preprocessed pulse signals of each channel using the min-max method, obtain the correlation coefficient based on the normalized eigenmode function and the preprocessed pulse signals of each channel, and select the eigenmode function corresponding to the correlation coefficient greater than the threshold as the eigenmode function of each channel.

[0016] Further, the method of utilizing local mean decomposition to decompose the pretreated pulse signal of each channel to obtain each channel PF component comprises the following steps:

[0017] Step S31, calculating the local mean of each cycle pulse signal after preprocessing of each channel, that is, half of the sum of the peak point and the pacemaker point of each cycle pulse signal x(t), and smoothing the local mean by a sliding average method within the cycle to obtain a local mean signal m1;

[0018] Step S32, calculating the local amplitude of each cycle pulse signal, that is, half of the difference between the peak point and the pacing point of each cycle pulse signal, and smoothing the local amplitude within the cycle by a sliding average method to obtain a local envelope signal b1;

[0019] Step S33, from each periodic pulse signal, subtract its corresponding local mean signal and then divide by the local envelope signal to obtain the demodulated periodic pulse signal

[0020] Step S34, if the demodulated periodic pulse signal is not a pure frequency modulation signal, the demodulated periodic pulse signal is used as the updated periodic pulse signal, and the process returns to step S31 until a pure frequency modulation signal is obtained;

[0021] Step S35, calculating the product of the local envelope signal and the pure FM signal to obtain the PF component; the local envelope signal is the local envelope signal corresponding to the pure FM signal;

[0022] Step S36, subtract the PF component from the periodic pulse signal to obtain a residual signal. If the residual signal is not monotonic, use the residual signal as the updated periodic pulse signal and return to step S31 until a monotonic residual signal is obtained. The monotonic residual signal is a local mean decomposition residual signal.

[0023] Furthermore, the wavelet basis function and the number of decomposition layers are determined to perform wavelet decomposition on the preprocessed pulse signals of each channel to obtain detail signals and approximate signals, the detail signals and approximate signals corresponding to each layer of wavelet decomposition are extracted, the energy corresponding to the detail signals of each channel is extracted, the energy corresponding to the detail signals of each channel is normalized and then weighted averaged to obtain the first time-frequency domain feature.

[0024] Furthermore, the correlation coefficient is obtained by the following method:

[0025]

[0026] Among them, λ i is the correlation coefficient of the ith eigenmode function, x n is the value of the pulse signal at the nth sampling point after single-channel preprocessing, is the mean value of the pulse signal after single channel preprocessing, IMF in is the value of the i-th eigenmode function of a single channel at the n-th sampling point, is the mean of the single-channel intrinsic mode function, and N is the number of sampling points.

[0027] Furthermore, the energy corresponding to the intrinsic mode function of each channel is extracted as the fifth time-frequency domain feature; the energy of the Hilbert spectrum and the energy of the Hilbert marginal spectrum corresponding to the transformed intrinsic mode function are extracted as the sixth time-frequency domain feature and the seventh time-frequency domain feature.

[0028] Furthermore, the average intensity corresponding to the PF component of each channel is extracted and weighted averaged as the second time-frequency domain feature, the spectral centroid corresponding to the PF component of each channel is extracted and weighted averaged as the third time-frequency domain feature, and the average energy contribution rate corresponding to the PF component of each channel is extracted and weighted averaged as the fourth time-frequency domain feature.

[0029] Further, the three-channel pulse signal is pre-processed to obtain the pre-processed pulse signal of each channel, comprising:

[0030] Setting a low-frequency noise threshold and a high-frequency noise threshold, determining a wavelet basis function and a decomposition layer number, performing wavelet decomposition on the pulse signal of each channel to obtain a plurality of wavelet components, removing the wavelet components whose frequencies are lower than the low-frequency noise threshold and whose frequencies are higher than the high-frequency noise threshold, and then reconstructing the signal using the wavelet coefficients to obtain a denoised pulse signal;

[0031] Extracting the pacing point and the peak point of the denoised pulse signal by searching for local extreme values, fitting all pacing points with cubic spline interpolation to obtain baseline noise, and removing the baseline noise from the pulse signal to obtain the pulse signal after the baseline noise is removed;

[0032] The pulse signal after removing the baseline noise is subjected to period normalization to obtain the pulse signal after preprocessing of each channel.

[0033] Furthermore, a search step is set, and a local minimum value of the denoised pulse signal is searched within the search step range as a pacing point, and a local maximum value of the denoised pulse signal is searched within the search step range as a peak point. All pacing points are used as interpolation points and cubic spline interpolation fitting is performed on them to obtain baseline noise, and the baseline noise is subtracted from the denoised pulse signal to obtain a pulse signal after removing the baseline noise.

[0034] Further, the pulse signal after the baseline noise removal is carried out to cycle normalization through the following steps:

[0035] Step S11, performing period segmentation on the pulse signal after removing the baseline noise based on the pacemaker point;

[0036] Step S12, calculating the peak-to-valley slope, the peak-to-valley distance, and the peak-to-valley difference for each segmented periodic pulse signal, obtaining the peak-to-valley slope average value and the peak-to-valley distance average value based on the peak-to-valley slope and the peak-to-valley distance of all segmented pulse signals, respectively, removing the periodic pulse signals whose peak-to-valley slope deviates from the peak-to-valley slope average value by more than a threshold value or whose peak-to-valley distance deviates from the peak-to-valley distance average value by more than a threshold value from each periodic signal, and obtaining the remaining periodic pulse signals;

[0037] Step S13, obtaining the average cycle length of the remaining periodic pulse signals; extending or compressing each remaining periodic pulse signal to normalize the cycle.

[0038] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0039] 1. The present invention obtains seven groups of time-frequency domain features of pulse signals through wavelet decomposition, local mean decomposition, empirical mode decomposition and Hilport transform, realizes more comprehensive time-frequency domain feature extraction of pulse signals, and provides a basis for more accurate and detailed classification and identification of pulse signals.

[0040] 2. The present invention obtains the intrinsic mode function through empirical mode decomposition, obtains the correlation relationship based on the preprocessed pulse signal and the intrinsic mode function, and selects the effective intrinsic mode function based on the correlation coefficient, thereby improving the accuracy of the fifth time-frequency domain feature.

[0041] 3. The present invention selects an effective intrinsic mode function based on the correlation coefficient, and performs Hilbert transform based on the effective intrinsic mode function, thereby avoiding the aliasing of the decomposition result, so that the extracted features can more accurately reflect the pulse information.

[0042] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0044] Figure 1 Be the flow chart of the pulse signal time-frequency domain feature extraction method of the embodiment of the present invention;

[0045] Figure 2 It is the detail signal after eight-layer wavelet decomposition in the embodiment of the present invention;

[0046] Figure 3 These are 5 groups of PF signals decomposed by local mean according to an embodiment of the present invention;

[0047] Figure 4 These are six groups of IMF components derived from the empirical mode decomposition according to the embodiment of the present invention. DETAILED DESCRIPTION

[0048] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0049] A specific embodiment of the present invention discloses a method for extracting time-frequency domain features of a pulse signal. Figure 1 As shown, the method comprises the following steps:

[0050] Step S1, obtaining three-channel pulse signals at three positions of Cun, Guan and Chi of the human body through a sensor, and pre-processing the three-channel pulse signals to obtain pre-processed pulse signals of each channel;

[0051] Step S2, decomposing the pre-processed pulse signals of each channel by wavelet to obtain detail signals and approximate signals of each channel, extracting the energy corresponding to the detail signals of each channel and performing weighted averaging as the first time-frequency domain feature;

[0052] Step S3, using local mean decomposition to decompose the pulse signal after pretreatment of each channel to obtain the PF component of each channel, and obtaining the second time-frequency domain feature, the third time-frequency domain feature, and the fourth time-frequency domain feature based on the PF component of each channel;

[0053] Step S4, using empirical mode decomposition to decompose the preprocessed pulse signal of each channel to obtain the initial intrinsic mode function of each channel, select the initial intrinsic mode function with a correlation coefficient greater than a threshold as the intrinsic mode function of each channel, and obtain the fifth time-frequency domain feature based on the intrinsic mode function of each channel; perform Hilbert transform on the intrinsic mode function of each channel to obtain the transformed intrinsic mode function, and obtain the sixth time-frequency domain feature and the seventh time-frequency domain feature based on the transformed intrinsic mode function.

[0054] Specifically, in step S1, the three-channel pulse signal is pre-processed to obtain the pre-processed pulse signal of each channel, comprising:

[0055] Setting a low-frequency noise threshold and a high-frequency noise threshold, determining a wavelet basis function and a decomposition layer number, performing wavelet decomposition on the pulse signal of each channel to obtain a plurality of wavelet components, removing the wavelet components whose frequencies are lower than the low-frequency noise threshold and whose frequencies are higher than the high-frequency noise threshold, and then reconstructing the signal using the wavelet coefficients to obtain a denoised pulse signal;

[0056] Extracting the pacing point and the peak point of the denoised pulse signal by searching for local extreme values, fitting all pacing points with cubic spline interpolation to obtain baseline noise, and removing the baseline noise from the pulse signal to obtain the pulse signal after the baseline noise is removed;

[0057] The pulse signal after removing the baseline noise is subjected to period normalization to obtain the pulse signal after preprocessing of each channel.

[0058] Specifically, the pulse signal is the trajectory of vascular pulsation and is a periodic signal. The pulse signal comprises a main wave, a tidal wave (pre-dicrotic wave) and a dicrotic wave in each cycle. The main wave amplitude reflects left ventricular ejection and aortic compliance, the tidal wave amplitude reflects vascular elasticity and peripheral resistance, and the dicrotic wave amplitude reflects aortic elasticity and aortic valve state. The time-frequency domain characteristics of the pulse signal can reflect the situation of its frequency changing over time and the instantaneous frequency and amplitude corresponding to each moment. The pulse signals at the three positions of Cun, Guan and Chi of the human body are sampled by a sensor, and the sampling frequency is 125Hz, and a three-channel pulse signal of 10,000 sampled data is obtained, and the three-channel pulse signals are processed separately in the follow-up.

[0059] Furthermore, the sym8 wavelet basis function is used to perform 7-layer wavelet decomposition on the three-channel pulse signal to remove wavelet components with frequencies lower than the low-frequency noise threshold and higher than the high-frequency noise threshold, and the remaining wavelet components are reconstructed using wavelet coefficients to obtain the denoised pulse signal.

[0060] Preferably, setting 0.825 Hz as the low-frequency noise threshold and setting 31.25 Hz as the high-frequency noise threshold can remove low-frequency noise generated by skin friction, 50 Hz power frequency interference, and 30-120 Hz high-frequency noise of ECG signals, thereby improving the signal-to-noise ratio.

[0061] Furthermore, a search step is set, and a local minimum value of the denoised pulse signal is searched within the search step range as a pacing point, and a local maximum value of the denoised pulse signal is searched within the search step range as a peak point. All pacing points are used as interpolation points and cubic spline interpolation fitting is performed on them to obtain baseline noise, and the baseline noise is subtracted from the denoised pulse signal to obtain a pulse signal after removing the baseline noise.

[0062] Preferably, since the pulse signal period is usually between 80-110, the search step length is set to 50.

[0063] It can be understood that the pulse signal contains not only high and low frequency noise, but also baseline drift caused by breathing. By using cubic spline interpolation to curve fit all pacemaker points to simulate breathing and remove it, a smooth pulse signal with a high signal-to-noise ratio can be obtained.

[0064] Further, the pulse signal after the baseline noise removal is carried out to cycle normalization through the following steps:

[0065] Step S11, performing period segmentation on the pulse signal after removing the baseline noise based on the pacemaker point;

[0066] Step S12, calculating the peak-to-valley slope, the peak-to-valley distance, and the peak-to-valley difference for each segmented periodic pulse signal, obtaining the peak-to-valley slope average value and the peak-to-valley distance average value based on the peak-to-valley slope and the peak-to-valley distance of all segmented pulse signals, respectively, removing the periodic pulse signals whose peak-to-valley slope deviates from the peak-to-valley slope average value by more than a threshold value or whose peak-to-valley distance deviates from the peak-to-valley distance average value by more than a threshold value from each periodic signal, and obtaining the remaining periodic pulse signals;

[0067] Step S13, obtaining the average cycle length of the remaining periodic pulse signals; extending or compressing each remaining periodic pulse signal to normalize the cycle.

[0068] Specifically, in step S11, the signal between adjacent pacemaker points is a periodic signal, and the pulse signal after amplitude normalization is periodically segmented according to the pacemaker points.

[0069] Specifically, in step S12, the peak-to-valley slope, peak-to-valley distance, and peak-to-valley difference are calculated for each segmented periodic signal, where the peak-to-valley slope is the slope of the line connecting the peak point and the pacemaker point of the periodic signal, that is, the ratio of the difference between the vertical coordinates of the peak point and the pacemaker point to the difference between the horizontal coordinates of the peak point and the pacemaker point; the peak-to-valley distance is the difference between the horizontal coordinates of the peak point and the pacemaker point of the periodic signal, and the peak-to-valley difference is the difference between the vertical coordinates of the peak point and the pacemaker point of the periodic signal.

[0070] Preferably, the periodic pulse signals whose peak-to-valley slope deviates from the average value of the peak-to-valley slope by more than 10% or whose peak-to-valley distance deviates from the average value of the peak-to-valley distance by more than 10% are removed.

[0071] Specifically, in step S13, for each periodic pulse signal after removal, if the length of the signal is greater than the average cycle length, the signal is truncated according to the average cycle length; if the length of the signal is less than the average cycle length, the signal is padded with zeros according to the average cycle length.

[0072] It can be understood that in the process of processing a large number of different pulse signals, amplitude normalization can eliminate the dimensional data differences of the pulse signal and retain the shape of the signal's main peak, pre-dicrotic wave, and dicrotic wave; period normalization retains the relative positions of characteristic points such as the signal's main peak, pre-dicrotic wave, and dicrotic wave, which facilitates subsequent analysis and processing.

[0073] Specifically, in step S2, the wavelet basis function and the number of decomposition layers are determined to perform wavelet decomposition on the preprocessed pulse signals of each channel to obtain detail signals and approximate signals, the detail signals and approximate signals corresponding to each layer of wavelet decomposition are extracted, the energy corresponding to the detail signals of each channel is extracted, the energy corresponding to the detail signals of each channel is normalized and then weighted averaged to obtain the first time-frequency domain feature.

[0074] It should be noted that, similar to the first time-frequency domain feature, the second, third, and fourth time-frequency domain features are all obtained by weighted averaging the corresponding channel features of each channel. The approximate signal is composed of the wavelet coefficients of the low-frequency part of each layer of wavelet decomposition, and the detail signal is composed of the wavelet coefficients of the high-frequency part of each layer of wavelet decomposition. The approximate signal of the current layer is decomposed into the approximate signal and detail signal of the next layer.

[0075] Furthermore, the energy corresponding to the single-channel detail signal is extracted by the following method:

[0076]

[0077] Among them, E a is the energy corresponding to the single-channel detail signal, a i,j is the i-th wavelet coefficient of the detail signal corresponding to the j-th wavelet decomposition, E aj is the energy of the detail signal corresponding to the j-th layer of wavelet decomposition, and n is the number of wavelet coefficients in each layer of wavelet decomposition.

[0078] Specifically, the sym8 wavelet basis function is used to perform 8-layer wavelet decomposition on the pulse signal after preprocessing of each channel, and the detail signal corresponding to each layer of wavelet decomposition is extracted. Figure 2 As shown. The square sum of different scale coefficients is extracted from the single channel detail signal, that is, the square sum of each wavelet coefficient of each layer of wavelet decomposition is obtained to obtain the energy corresponding to the wavelet detail signal. The energy of the 8 wavelet detail signals obtained is normalized by the min-max method, and the energy value is limited to the range of [0,1]. The energy of this group of normalized detail signals is the first time-frequency domain feature of the single channel pulse signal. Due to the different importance of the pulse signals at the three positions of Cun, Guan, and Chi of the human body, weight values ​​are set for the pulse signals at the three positions of Cun, Guan, and Chi (i.e., three-channel pulse signals), and the first time-frequency domain features of the three-channel pulse signals are weighted averaged as the first time-frequency domain features, reflecting the differences in different scale frequency signals corresponding to the pulse signals in different frequency ranges.

[0079] Specifically, in step S3, the process of decomposing the pulse signal after pretreatment of each channel by local mean decomposition to obtain the PF component of each channel comprises the following steps:

[0080] Step S31, calculating the local mean of each cycle pulse signal after preprocessing of each channel, that is, half of the sum of the peak point and the pacemaker point of each cycle pulse signal x(t), and smoothing the local mean by a sliding average method within the cycle to obtain a local mean signal m1;

[0081] Step S32, calculating the local amplitude of each cycle pulse signal, that is, half of the difference between the peak point and the pacing point of each cycle pulse signal, and smoothing the local amplitude within the cycle by a sliding average method to obtain a local envelope signal b1;

[0082] Step S33, from each periodic pulse signal, subtract its corresponding local mean signal and then divide by the local envelope signal to obtain the demodulated periodic pulse signal

[0083] Step S34, if the demodulated periodic pulse signal is not a pure frequency modulation signal, the demodulated periodic pulse signal is used as the updated periodic pulse signal, and the process returns to step S31 until a pure frequency modulation signal is obtained;

[0084] Step S35, calculating the product of the local envelope signal and the pure FM signal to obtain the PF component; the local envelope signal is the local envelope signal corresponding to the pure FM signal;

[0085] Step S36, subtract the PF component from the periodic pulse signal to obtain a residual signal. If the residual signal is not monotonic, use the residual signal as the updated periodic pulse signal and return to step S31 until a monotonic residual signal is obtained. The monotonic residual signal is a local mean decomposition residual signal.

[0086] It can be understood that the single-channel periodic pulse signal is decomposed into a PF signal by local mean value as follows Figure 3 As shown, it can be seen that the periodic pulse signal can be decomposed into 5 groups of PF signals.

[0087] Furthermore, the average intensity corresponding to the PF component of each channel is extracted and weighted averaged as the second time-frequency domain feature, the spectral centroid corresponding to the PF component of each channel is extracted and weighted averaged as the third time-frequency domain feature, and the average energy contribution rate corresponding to the PF component of each channel is extracted and weighted averaged as the fourth time-frequency domain feature.

[0088] Further, obtain the mean intensity of the single channel periodic pulse signal by the following method:

[0089]

[0090] Among them, A b is the average intensity of the single-channel periodic pulse signal, K is the number of PF components of the single-channel periodic pulse signal, N is the data length of the PF signal, and abi (j) is the amplitude of the i-th PF component at the j-th sampling point.

[0091] Further, obtain the spectrum centroid of single channel periodic pulse signal by the following method:

[0092]

[0093] Among them, F b is the spectral centroid of the single-channel periodic pulse signal, K is the number of PF components of the single-channel periodic pulse signal, and f bi (j) is the instantaneous frequency of the i-th PF component at the j-th sampling point, E bi (j) is the instantaneous energy of the i-th PF component at the j-th sampling point.

[0094] Further, obtain the average energy contribution rate of the single channel periodic pulse signal by the following method:

[0095]

[0096] Among them, E b is the average energy contribution rate of the single-channel periodic pulse signal, and K is the number of PF components of the single-channel periodic pulse signal.

[0097] It can be understood that the average intensity reflects the influence of factors such as heartbeat and reflected waves on the amplitude of each PF component; the spectral centroid reflects the frequency center of gravity of the PF component, that is, it reflects the frequency characteristics of the pulse signal; the energy contribution rate reflects the inherent physiological information and characteristics of the periodic pulse signal.

[0098] Specifically, in step S4, the pulse signal after the pretreatment of each channel is decomposed by utilizing empirical mode decomposition to obtain the initial intrinsic mode function of each channel, and the initial intrinsic mode function whose correlation coefficient is greater than the threshold value is selected as the eigenmode function of each channel comprises the following steps:

[0099] Step S41, generating a lower envelope based on the pacemaker point of the preprocessed pulse signal of each channel, generating an upper envelope based on the peak point of the preprocessed pulse signal of each channel, and generating a mean envelope of the pulse signal of each channel based on the upper and lower envelopes;

[0100] Step S42, subtracting the corresponding mean envelope from the pre-processed pulse signal of each channel to obtain an initial eigenmode function;

[0101] Step S43, subtracting the initial intrinsic mode function from the pre-processed pulse signal of each channel to obtain a difference signal;

[0102] Step S44, taking the difference signal as the pulse signal after preprocessing of each channel, repeating steps S41 to S43 until the difference signal is a monotonic function, and taking the difference signal obtained at this time as the residual signal of empirical mode decomposition;

[0103] Step S45, normalize all the initial eigenmode functions and the preprocessed pulse signals of each channel using the min-max method, obtain the correlation coefficient based on the normalized eigenmode function and the preprocessed pulse signals of each channel, and select the eigenmode function corresponding to the correlation coefficient greater than the threshold as the eigenmode function of each channel.

[0104] Specifically, the correlation coefficient is obtained by the following method:

[0105]

[0106] Among them, λ i is the correlation coefficient of the ith eigenmode function, x n is the value of the pulse signal at the nth sampling point after single-channel preprocessing, is the mean value of the pulse signal after single channel preprocessing, IMF in is the value of the i-th eigenmode function of a single channel at the n-th sampling point, is the mean of the single-channel intrinsic mode function, and N is the number of sampling points.

[0107] Specifically, the standard deviation of the correlation coefficient is set as a threshold, and the eigenmode function corresponding to the correlation coefficient being greater than the threshold is selected as the three-channel eigenmode function, and the other eigenmode functions are discarded.

[0108] It can be understood that the pre-processed pulse signal is decomposed into the intrinsic mode function (i.e., IMF component) by empirical mode decomposition as follows: Figure 4 As shown, it can be seen that the pre-treated pulse signal can be decomposed into 6 groups of IMF components, that is, the components with different frequencies in the pulse signal can be decomposed in order from high to low frequencies. The present invention obtains the intrinsic mode function through empirical mode decomposition, obtains the correlation relationship based on the pre-treated pulse signal and the intrinsic mode function, and selects the effective intrinsic mode function based on the correlation coefficient, thereby improving the accuracy of the fifth time-frequency domain feature.

[0109] Furthermore, the energy corresponding to the intrinsic mode function of each channel is extracted as the fifth time-frequency domain feature; the energy of the Hilbert spectrum and the energy of the Hilbert marginal spectrum corresponding to the transformed intrinsic mode function are extracted as the sixth time-frequency domain feature and the seventh time-frequency domain feature.

[0110] Further, obtain the energy corresponding to the intrinsic mode function of the single channel periodic pulse signal by the following method:

[0111]

[0112] Among them, E c is the energy corresponding to the intrinsic mode function of the single-channel periodic pulse signal, L is the number of intrinsic mode functions of the single-channel periodic pulse signal, IMF i (j) is the value of the i-th eigenmode function at the j-th sampling point.

[0113] Furthermore, the Hilbert transform of the eigenmode function is performed by the following method:

[0114]

[0115] Among them, d i (t) is the Hilbert transform of the i-th eigenmode function.

[0116] It can be understood that the present invention selects effective intrinsic mode functions based on the correlation coefficient, and performs Hilbert transform based on the effective intrinsic mode functions, thereby avoiding aliasing of the decomposition results, so that the extracted features can more accurately reflect the pulse information.

[0117] Further, obtain the energy of the Hilbert spectrum corresponding to the eigenmode function after the single channel periodic pulse signal conversion by the following method:

[0118]

[0119] Among them, E d is the energy of the Hilbert spectrum corresponding to the eigenmode function after the transformation of the single-channel periodic pulse signal, and L is the number of eigenmode functions of the single-channel periodic pulse signal.

[0120] Specifically, the energy of the Hilport spectrum corresponding to the transformed eigenmode function is normalized by a min-max method, and the energy is limited to the range of [0, 1].

[0121] Further, obtain the marginal spectrum corresponding to the intrinsic mode function after the single channel periodic pulse signal transformation by the following method, i.e. the real part integration of the Hilbert time-frequency spectrum in time:

[0122]

[0123] Among them, H mi (ω, t) is the amplitude spectrum of the Hilport transform of the i-th intrinsic mode function in a three-dimensional graph with time, amplitude and frequency as coordinates, and T is the length of the data sequence.

[0124] Further, obtain the energy of the marginal spectrum corresponding to the intrinsic mode function after the single channel periodic pulse signal conversion by the following method:

[0125]

[0126] Among them, E m is the energy of the marginal spectrum corresponding to the intrinsic mode function after the transformation of the single-channel periodic pulse signal, and L is the number of eigenmode functions of the single-channel periodic pulse signal.

[0127] Specifically, the energy of the marginal spectrum corresponding to the transformed intrinsic mode function is normalized by the min-max method, and the energy is limited to the range of [0, 1].

[0128] It is understandable that the energy corresponding to the intrinsic mode function reflects the energy corresponding to different frequencies of the pulse signal; the energy of the Hilbert spectrum and the energy of the marginal spectrum reflect the changes in instantaneous frequency and amplitude. The present invention obtains seven groups of time-frequency domain features of the pulse signal through wavelet decomposition, local mean decomposition, and empirical mode decomposition, realizes a more comprehensive feature extraction of the pulse signal, and provides a basis for more accurate and detailed classification and identification of the pulse signal.

[0129] Compared with prior art, the beneficial effect of pulse signal time-frequency domain feature extraction method provided by the invention is as follows:

[0130] 1. The present invention obtains seven groups of time-frequency domain features of pulse signals through wavelet decomposition, local mean decomposition and empirical mode decomposition, thereby realizing more comprehensive time-frequency domain feature extraction of pulse signals and providing a basis for more accurate and detailed classification and identification of pulse signals.

[0131] 2. The present invention obtains the intrinsic mode function by local mean decomposition, obtains the correlation relationship based on the preprocessed pulse signal and the intrinsic mode function, and selects the effective intrinsic mode function based on the correlation coefficient, thereby improving the accuracy of the fifth time-frequency domain feature.

[0132] 3. The present invention selects an effective intrinsic mode function based on the correlation coefficient, and performs Hilbert transform based on the effective intrinsic mode function, thereby avoiding the aliasing of the decomposition result, so that the extracted features can more accurately reflect the pulse information.

[0133] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0134] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A pulse signal time-frequency domain feature extraction method, characterized in that, The method comprises the following steps: Acquire three-channel pulse signals at three positions of the human body, namely, Cun, Guan and Chi, through sensors, and pre-process the three-channel pulse signals to obtain pre-processed pulse signals of each channel; Decomposing the pulse signals preprocessed by each channel by wavelet to obtain the detail signal and approximate signal of each channel, extracting the energy corresponding to the detail signal of each channel and performing weighted average as the first time-frequency domain feature; Decomposing the pulse signals preprocessed by each channel by using local mean decomposition to obtain the PF components of each channel, and obtaining the second time-frequency domain feature, the third time-frequency domain feature, and the fourth time-frequency domain feature based on the PF components of each channel; The preprocessed pulse signals of each channel are decomposed by empirical mode decomposition to obtain the initial intrinsic mode function of each channel, and the initial intrinsic mode function with a correlation coefficient greater than a threshold is selected as the intrinsic mode function of each channel, and the fifth time-frequency domain feature is obtained based on the intrinsic mode function of each channel; the intrinsic mode function of each channel is subjected to Hilbert transform to obtain the transformed intrinsic mode function, and the sixth time-frequency domain feature and the seventh time-frequency domain feature are obtained based on the transformed intrinsic mode function.

2. pulse signal time-frequency domain feature extraction method according to claim 1, is characterized in that, Described utilizing empirical mode decomposition decomposes the pretreated pulse signal of each channel to obtain the initial intrinsic mode function of each channel, and selecting the initial intrinsic mode function whose correlation coefficient is greater than threshold value as each channel eigenmode function comprises the following steps: Step S41, generating a lower envelope based on the pacemaker point of the preprocessed pulse signal of each channel, generating an upper envelope based on the peak point of the preprocessed pulse signal of each channel, and generating a mean envelope of the pulse signal of each channel based on the upper and lower envelopes; Step S42, subtracting the corresponding mean envelope from the pre-processed pulse signal of each channel to obtain an initial eigenmode function; Step S43, subtracting the initial intrinsic mode function from the pre-processed pulse signal of each channel to obtain a difference signal; Step S44, taking the difference signal as the pulse signal after preprocessing of each channel, repeating steps S41 to S43 until the difference signal is a monotonic function, and taking the difference signal obtained at this time as the residual signal of empirical mode decomposition; Step S45, normalize all the initial eigenmode functions and the preprocessed pulse signals of each channel using the min-max method, obtain the correlation coefficient based on the normalized eigenmode function and the preprocessed pulse signals of each channel, and select the eigenmode function corresponding to the correlation coefficient greater than the threshold as the eigenmode function of each channel.

3. pulse signal time-frequency domain feature extraction method according to claim 1, is characterized in that, Described utilizing local mean decomposition to decompose the pretreated pulse signal of each channel to obtain each channel PF component comprises the steps: Step S31, calculating the local mean of each cycle pulse signal after preprocessing of each channel, that is, half of the sum of the peak point and the pacemaker point of each cycle pulse signal x(t), and smoothing the local mean by a sliding average method within the cycle to obtain a local mean signal m1; Step S32, calculating the local amplitude of each cycle pulse signal, that is, half of the difference between the peak point and the pacing point of each cycle pulse signal, and smoothing the local amplitude within the cycle by a sliding average method to obtain a local envelope signal b1; Step S33, from each periodic pulse signal, subtract its corresponding local mean signal and then divide by the local envelope signal to obtain the demodulated periodic pulse signal Step S34, if the demodulated periodic pulse signal is not a pure frequency modulation signal, the demodulated periodic pulse signal is used as the updated periodic pulse signal, and the process returns to step S31 until a pure frequency modulation signal is obtained; Step S35, calculating the product of the local envelope signal and the pure FM signal to obtain the PF component; the local envelope signal is the local envelope signal corresponding to the pure FM signal; Step S36, subtract the PF component from the periodic pulse signal to obtain a residual signal. If the residual signal is not monotonic, use the residual signal as the updated periodic pulse signal and return to step S31 until a monotonic residual signal is obtained. The monotonic residual signal is a local mean decomposition residual signal.

4. pulse signal time-frequency domain feature extraction method according to claim 1, is characterized in that, Determine the wavelet basis function and the number of decomposition layers, perform wavelet decomposition on the preprocessed pulse signals of each channel to obtain detail signals and approximate signals, extract the detail signals and approximate signals corresponding to each layer of wavelet decomposition, extract the energy corresponding to the detail signals of each channel, normalize the energy corresponding to the detail signals of each channel, and then perform weighted averaging to obtain the first time-frequency domain feature.

5. pulse signal time-frequency domain feature extraction method according to claim 2, is characterized in that, The correlation coefficient is obtained by the following method: Among them, λ i is the correlation coefficient of the ith eigenmode function, x n is the value of the pulse signal at the nth sampling point after single-channel preprocessing, is the mean value of the pulse signal after single channel preprocessing, IMF in is the value of the i-th eigenmode function of a single channel at the n-th sampling point, is the mean of the single-channel intrinsic mode function, and N is the number of sampling points.

6. pulse signal time-frequency domain feature extraction method according to claim 5, is characterized in that, The energy corresponding to the intrinsic mode function of each channel is extracted as the fifth time-frequency domain feature; the energy of the Hilbert spectrum and the energy of the Hilbert marginal spectrum corresponding to the transformed intrinsic mode function are extracted as the sixth time-frequency domain feature and the seventh time-frequency domain feature.

7. The pulse signal time-frequency domain feature extraction method according to claim 3, wherein: The average intensity corresponding to the PF component of each channel is extracted and weighted averaged as the second time-frequency domain feature, the spectral centroid corresponding to the PF component of each channel is extracted and weighted averaged as the third time-frequency domain feature, and the average energy contribution rate corresponding to the PF component of each channel is extracted and weighted averaged as the fourth time-frequency domain feature.

8. pulse signal time-frequency domain feature extraction method according to claim 1, is characterized in that, The three-channel pulse signal is pre-processed to obtain the pre-processed pulse signal of each channel, comprising: Setting a low-frequency noise threshold and a high-frequency noise threshold, determining a wavelet basis function and a decomposition layer number, performing wavelet decomposition on the pulse signal of each channel to obtain a plurality of wavelet components, removing the wavelet components whose frequencies are lower than the low-frequency noise threshold and whose frequencies are higher than the high-frequency noise threshold, and then reconstructing the signal using the wavelet coefficients to obtain a denoised pulse signal; Extracting the pacing point and the peak point of the denoised pulse signal by searching for local extreme values, fitting all pacing points with cubic spline interpolation to obtain baseline noise, and removing the baseline noise from the pulse signal to obtain the pulse signal after the baseline noise is removed; The pulse signal after removing the baseline noise is subjected to period normalization to obtain the pulse signal after preprocessing of each channel.

9. The pulse signal time-frequency domain feature extraction method according to claim 8, wherein: A search step is set, and a local minimum value of the denoised pulse signal is searched within the search step range as a pacing point, and a local maximum value of the denoised pulse signal is searched within the search step range as a peak point. All pacing points are used as interpolation points to perform cubic spline interpolation fitting on them to obtain baseline noise, and the baseline noise is subtracted from the denoised pulse signal to obtain a pulse signal after removing the baseline noise.

10. The pulse signal time-frequency domain feature extraction method according to claim 9, wherein: After removing the baseline noise, the pulse signal is periodically normalized by the following steps: Step S11, performing period segmentation on the pulse signal after removing the baseline noise based on the pacemaker point; Step S12, calculating the peak-to-valley slope, the peak-to-valley distance, and the peak-to-valley difference for each segmented periodic pulse signal, obtaining the peak-to-valley slope average value and the peak-to-valley distance average value based on the peak-to-valley slope and the peak-to-valley distance of all segmented pulse signals, respectively, removing the periodic pulse signals whose peak-to-valley slope deviates from the peak-to-valley slope average value by more than a threshold value or whose peak-to-valley distance deviates from the peak-to-valley distance average value by more than a threshold value from each periodic signal, and obtaining the remaining periodic pulse signals; Step S13, obtaining the average cycle length of the remaining periodic pulse signals; extending or compressing each remaining periodic pulse signal to normalize the cycle.

Citation Information

Patent Citations

  • Pulse wave analysis method and device

    CN109171678A

  • Empirical mode decomposition apparatus, method and article of manufacture for analyzing biological signals and performing curve fitting

    US6381559B1