Pulse wave time-frequency feature extraction method based on improved Gabor atom dictionary

Abstract

The invention provides a pulse wave time-frequency feature extraction method based on an improved Gabor atom dictionary, and belongs to the technical field of human physiological signal processing analysis and health monitoring. Firstly, an original pulse wave signal is subjected to center moving average and wavelet transform to remove high-frequency noise; then baseline drift removal operation is carried out to eliminate low-frequency noise, and a pulse wave signal without noise is obtained; performing period segmentation and normalization processing on the denoised pulse wave signal to obtain an average period signal; an over-complete Gabor atom dictionary is constructed, sparse representation is conducted on the average pulse wave period signals through an OMP orthogonal matching pursuit algorithm, and time-frequency characteristic components are obtained; the method has the advantages that the pulse wave signals can be effectively denoised, the time-frequency features with certain physiological significance can be extracted, and the method can play an important role in the aspects of human health monitoring, disease prevention and the like.

Description

Technical Field

[0001] The invention relates to the technical field of human physiological signal processing analysis and health monitoring, and in particular to a pulse wave time-frequency feature extraction method based on an improved Gabor atom dictionary. Background Art

[0002] Cardiovascular disease is the world's leading cause of death. According to statistics from the World Health Organization, more than 18 million people die from cardiovascular disease each year, accounting for one-third of the world's total deaths. In my country, the number of patients with cardiovascular disease is still increasing year by year, and nearly 2 out of every 5 deaths are caused by cardiovascular disease, which has become a major health and safety issue. The 2023 "China Cardiovascular Disease Report" pointed out that cardiovascular disease deaths account for the first place in the total cause of death of urban and rural residents in my country, accounting for 44.6% and 42.51% in rural and urban areas respectively. At present, the main means of cardiovascular disease detection in hospitals is electrocardiogram, and the electrocardiogram measurement process is complicated and has high operation requirements. When measuring the electrocardiogram, in order to obtain accurate signals, it is usually necessary to attach multiple electrodes to specific positions on the patient's body, which requires not only operators with professional skills, but also special equipment and environment. In addition, electrocardiogram equipment is usually complex, expensive and less portable, and is not suitable for home use. Therefore, a simple and portable flexible pulse sensor can be used to measure the pulse wave signal instead of the electrocardiogram.

[0003] The pulse wave signal is a physiological signal generated by the periodic beating of the heart and propagated along the artery. It contains a lot of information closely related to the cardiovascular health status. By analyzing the pulse wave signal, important physiological indicators such as heart rate, blood pressure, and blood vessel hardness can be obtained. This information has important application value in clinical diagnosis, chronic disease management, and individual health monitoring.

[0004] Traditional pulse wave signal analysis is usually based on its time domain characteristics, such as the amplitude and conduction time of each time domain feature point of the pulse wave. Although analysis based on time domain characteristics can provide certain physiological information, it cannot reflect the characteristics of the pulse wave signal in the frequency domain. In order to more comprehensively reflect the physiological information contained in the pulse wave signal, time-frequency analysis methods have been widely used in recent years, especially those based on wavelet transform and short-time Fourier transform. However, these methods have some limitations in practical applications. The wavelet transform lacks adaptability when selecting basis functions, and may lose some important details when processing complex physiological signals such as pulse wave signals. Due to its fixed window length, the short-time Fourier transform will have insufficient resolution when processing pulse wave signals, which is not enough to capture accurate time and frequency information at the same time.

[0005] To solve the above problems, pulse wave feature extraction technology based on Gabor atomic dictionary has been gradually applied in recent years. Due to the good time-frequency characteristics of Gabor function, it can effectively extract features of pulse wave signals in the time-frequency domain. However, when using the atomic dictionary constructed by standard Gabor function to extract time-frequency features of complex physiological signals, there are problems such as high dictionary redundancy, poor adaptability, and poor adaptability to nonlinear changing signals. Summary of the invention

[0006] The present invention provides a pulse wave time-frequency feature extraction method based on an improved Gabor atomic dictionary to solve the problems of high redundancy, poor adaptability and poor adaptability to nonlinear change signals of the original Gabor dictionary when extracting time-frequency features of complex physiological signals, so as to more accurately capture the time-frequency features of the pulse wave signal.

[0007] The technical solution adopted by the present invention comprises the following steps:

[0008] Step 1: Remove high-frequency noise by performing center sliding average and wavelet transform on the original pulse wave signal, and then perform Fourier transform to obtain the fundamental frequency f of the signal. b ;

[0009] Step 2: After removing the high-frequency noise, the pulse wave signal is subjected to a baseline drift removal operation according to the valley point of the pulse wave signal and the selected auxiliary point to eliminate the low-frequency noise and obtain a pulse wave signal without noise;

[0010] Step 3: Perform period segmentation on the denoised signal, perform amplitude normalization and length normalization to obtain an average periodic signal;

[0011] Step 4: Construct an overcomplete atomic dictionary D based on the improved Gabor function, using a commonly used discretization method β = (a b ,ca b Δμ,da -b Δω,eΔv), obtain the value range of the four parameters β=(δ,μ,ω,v) in the Gabor function, and construct the atomic dictionary D, where each column vector in D represents an atom;

[0012] Step 5: Use the OMP orthogonal matching pursuit algorithm to sparsely represent the average pulse wave period signal and obtain s sparse representation components (γ1G1,γ2G2,...γ s G s ), where G p represents the pth atom obtained from the atom dictionary, γ pRepresents its coefficient, 1≤p≤s, and then these components are divided into main wave component, tidal wave component, dicrotic wave component and tail component, and a total of 4 time-frequency characteristic components are obtained.

[0013] The specific process of removing high-frequency noise in step 1 of the present invention is as follows:

[0014] (1) Set the window size to w and perform center sliding average denoising on the original pulse wave signal:

[0015]

[0016] Where CMA(t) represents the central moving average of the t-th data point, and x(i) represents the i-th data point in the original pulse wave signal;

[0017] (2) When performing wavelet transform denoising, determine the number of wavelet transform decomposition layers, wavelet basis functions, and threshold functions, and apply the improved threshold function When |w j,k |≤λ, When |w j,k When |≥λ, The function formula is:

[0018]

[0019] In the formula, w j,k is the kth wavelet coefficient of the jth layer obtained after wavelet transform decomposition, λ is the threshold, δ=(median|w j,k |) / 0.6745, N is the number of sampling points of the signal, α is the adjustment factor, α≥0, the larger α is The closer to the hard threshold, the smaller α The closer to the soft threshold, the median|w j,k | is the wavelet coefficient w of the jth layer decomposition j,k The median value of the amplitude;

[0020] After the pulse wave signal is processed by the central sliding average and wavelet transform, it will no longer contain high-frequency noise. The discrete Fourier transform is performed on it to obtain X[n], and the amplitude spectrum of X[n] is calculated:

[0021]

[0022] Where Re(X[n]) and Im(X[n]) are the real and imaginary parts of X[n] respectively;

[0023] The fundamental frequency is the frequency corresponding to the maximum amplitude in the amplitude spectrum, and its calculation formula is:

[0024]

[0025] Where f b is the fundamental frequency, n max is the index corresponding to the maximum amplitude in the amplitude spectrum, f s is the sampling frequency, N is the number of Fourier transform points, which is equal to the number of sampling points of the signal.

[0026] In step 2 of the present invention, when removing the baseline drift, a sliding window is first used to traverse the entire signal to obtain the valley value of the pulse wave signal, and the window size W is set to f s / f b , f s The signal sampling frequency is set, and the step size is set to 2 / 3 of the window size. The signal is detected for valley values. A baseline is initially obtained based on the valley values ​​of the signal using the cubic spline interpolation method. The points in the pulse wave signal that are lower than this baseline are used as auxiliary points. The baseline is refitted using the cubic spline interpolation method based on the valley values ​​and auxiliary points of the signal to obtain a new baseline, which is then removed.

[0027] In step 3 of the present invention, the average pulse wave period signal is obtained. First, the interval signal between two adjacent valley values ​​of the pulse wave signal is divided into a period, and each period signal is normalized. The amplitude normalization calculation formula is:

[0028]

[0029] Where X(i) represents the value of each point of each pulse wave cycle signal, X min is the minimum value of each cycle, X max is the maximum value of each cycle;

[0030] Length normalization is to first find the maximum length T of each pulse wave cycle signal max , through cubic spline interpolation, each signal period is T max , define a new time sampling point x in each cycle new , Where N i is the number of sampling points in each original cycle;

[0031] After normalizing the amplitude and length, multiple pulse wave periodic signals are obtained. Assume that there are M pulse wave periodic signals, and each periodic signal has a length of T. max The discrete signal sequence is represented by X m , 1≤m≤M, sum and average all pulse wave period signals to obtain the average period signal, the calculation formula is:

[0032]

[0033] The improved Gabor function in step 4 of the present invention has the following function formula:

[0034]

[0035] Where G is the improved Gabor function, which is controlled by a quaternary group β = (δ, μ, ω, v) for morphology control, x is the data point position of the average period signal of the pulse wave, and the value range of the quaternary group parameter is determined by a commonly used discretization method β = (a b ,ca b Δμ,da -b Δω, eΔv) is determined, where a=2, Δμ=1 / 2, Δω=π, Δv=π / 6,0 <b≤log2L,0≤c<L2 -b+1 ,0≤d<2 b+1 ,0≤e≤12,b,c,d,e∈Z, L is the length of the pulse wave signal for sparse representation;

[0036] For each set of parameters β = (δ, μ, ω, v), a corresponding Gabor atom G is generated β (x), by traversing all combinations of parameters, the generated multiple Gabor atoms are stored in the dictionary. The final atom dictionary D can be expressed as: D = {G β (x)|β=(δ,μ,ω,v)}, each column in D represents an atom.

[0037] The specific process of using the OMP orthogonal matching pursuit algorithm to sparsely represent the average pulse wave period signal in step 5 of the present invention is:

[0038] (1) Initialization: residual r = y, atomic set Coefficient Set y is the signal to be reconstructed after standard deviation normalization;

[0039] (2) Atom selection: Calculate the inner product of the current residual r and all atoms in the dictionary D, and find the atom G with the largest absolute value of the inner product with the current residual j , the calculation formula is:

[0040] G p =argmax|r T D q |

[0041] Where D q is the basis vector of the qth column in the atomic dictionary D;

[0042] (3) Update the atom set: Update the selected atom G p Add to the atomic set G;

[0043] (4) Use the least squares method to update the coefficient set γ: that is, find a set of coefficients that makes y in the currently selected atom G p The coefficient γ with the smallest reconstruction error p , the calculation formula is:

[0044] γ p =argmin||y-γ p G p || 2

[0045] The calculated coefficient γ p Add to the coefficient set γ;

[0046] (5) Update residual: The updated residual is the difference between the current signal and the reconstructed signal. The calculation formula is:

[0047] r=y-γ p G p

[0048] (6) Repeat steps (2) to (5) to obtain s decomposition components (γ1G1,γ2G2,...γ s G s ).

[0049] The four time-frequency characteristic components in step 5 of the present invention are based on each component γ p G p The position of the maximum amplitude point is classified in combination with the time domain characteristic points of the pulse wave signal. If the time domain horizontal coordinate corresponding to the maximum amplitude point of the component is between the starting point of the pulse wave and the starting point of the tidal wave, it is the main wave component; between the starting point of the tidal wave and the descending gorge, it is the tidal wave component; between the descending gorge and the end point of the dibolic wave, it is the dibolic wave component; between the end point of the dibolic wave and the end point of the pulse wave, it is the tail component. In this way, four categories of main wave component, tidal wave component, dibolic wave component and tail component are obtained, and the components of each category are added and summed to obtain 4 time-frequency characteristic components.

[0050] The beneficial effects of the present invention are:

[0051] 1. In the pulse wave time-frequency feature extraction method based on the improved Gabor atom dictionary of the present invention, the original pulse wave signal is first preprocessed by the central sliding average method; by the sliding average preprocessing, part of the high-frequency noise in the signal is eliminated, and the signal waveform is made smoother, so that more accurate wavelet coefficients can be obtained during the wavelet transform, and waveform distortion is prevented; in addition, the central sliding average denoising has little effect on the low-frequency components of the signal, and therefore will not significantly change the main features of the pulse wave signal.

[0052] 2. In the pulse wave time-frequency feature extraction method based on the improved Gabor atomic dictionary of the present invention, an improved wavelet transform threshold function is adopted. In the process of wavelet transform denoising, the use of this threshold function can effectively balance signal retention and noise suppression, reduce artifacts, improve the signal-to-noise ratio, provide more flexible adaptive processing capabilities, and ultimately achieve a smoother and more natural signal denoising effect.

[0053] 3. In the pulse wave time-frequency feature extraction method based on the improved Gabor atomic dictionary of the present invention, when constructing an over-complete atomic dictionary, the frequency and amplitude of the Gabor function are improved to solve the problems of high redundancy, poor adaptability, and poor adaptability to nonlinear changing signals of the atomic dictionary; at the same time, during feature extraction, the pulse wave signal is decomposed into components and classified to obtain four characteristic components with certain physiological significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic diagram of the process of the present invention;

[0055] Figure 2 is a pulse wave signal after high-frequency noise is removed in an embodiment of the present invention;

[0056] Figure 3 is a pulse wave signal after removing the baseline drift in an embodiment of the present invention;

[0057] Figure 4 is a sparse component of the present invention's example of sparsely representing the average pulse wave period signal;

[0058] Figure 5 is the time domain characteristic point of the pulse wave signal of the example of the present invention;

[0059] Figure 6 It is the time-frequency characteristic component obtained by the example of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, features and principles of the present invention more obvious and easy to understand, the present invention is further described below with reference to specific examples.

[0061] like Figure 1 As shown, the present invention discloses a pulse wave time-frequency feature extraction method based on an improved Gabor atom dictionary, which includes the following five steps:

[0062] Step 1: Remove high-frequency noise by performing center sliding average and wavelet transform on the original pulse wave signal. Figure 2 The pulse wave signal after high-frequency noise is removed is shown; then Fourier transform is performed to obtain the fundamental frequency f of the signal b ;

[0063] Use the pulse sensor to collect pulse wave signals, the sampling frequency is f s =250Hz, get the original pulse wave signal;

[0064] (1) Set the window size to w, w = 11, and perform center sliding average denoising on the original pulse wave signal:

[0065]

[0066] Where CMA(t) represents the central moving average of the t-th data point, and x(i) represents the i-th data point in the original pulse wave signal;

[0067] (2) Perform wavelet transform denoising, select the wavelet transform decomposition level as 4, the wavelet basis function as Sym8, and adopt an improved threshold function When |w j,k |≤λ, When |w j,k When |≥λ, The function formula is:

[0068]

[0069] In the formula, w j,k is the kth wavelet coefficient of the jth layer obtained after wavelet transform decomposition, λ is the threshold, δ=(median|w j,k |) / 0.6745, N is the signal length, α is the adjustment factor, α≥0, the larger α is The closer to the hard threshold, the smaller α The closer it is to the soft threshold; in this example, the adjustment factor α is selected to be 0.8, median|w j,k | is the wavelet coefficient w of the jth layer decomposition j,k The median value of the amplitude;

[0070] The use of improved threshold function can overcome the problems of discontinuous reconstructed signal and mutation caused by hard threshold function and low reconstruction accuracy and large mean square error caused by soft threshold function.

[0071] The fundamental frequency f of the signal b The normal range is between 1 and 2 Hz. After the center sliding average and wavelet transform processing, the pulse wave signal will no longer contain high-frequency noise. The discrete Fourier transform is performed to obtain X[n], and the amplitude spectrum of X[n] is calculated:

[0072]

[0073] Where Re(X[n]) and Im(X[n]) are the real and imaginary parts of X[n] respectively;

[0074] The fundamental frequency is the frequency corresponding to the maximum amplitude in the amplitude spectrum, and its calculation formula is:

[0075]

[0076] Where f b is the fundamental frequency, n max is the index corresponding to the maximum amplitude in the amplitude spectrum, f s is the sampling frequency, N is the number of Fourier transform points (equal to the number of sampling points of the signal);

[0077] Step 2: After removing the high-frequency noise, the pulse wave signal is subjected to a baseline drift removal operation according to the valley point of the pulse wave signal and the selected auxiliary point to eliminate the low-frequency noise and obtain a pulse wave signal without noise;

[0078] When removing the baseline drift, first use a sliding window to traverse the entire signal to obtain the valley value of the pulse wave signal, and set the window size W = f s / f b , f s The signal sampling frequency is set, and the step size is set to 2 / 3 of the window size. The valley value of the signal is detected. A baseline is initially obtained based on the valley value of the signal using the cubic spline interpolation method. The points in the pulse wave signal that are lower than this baseline are used as auxiliary points. The baseline is refitted using the cubic spline interpolation method based on the valley value points and auxiliary points of the signal to obtain a new baseline. This baseline is removed, as shown in Figure 3 As shown;

[0079] Cubic spline interpolation is a commonly used interpolation method, which is used to construct a series of smooth cubic polynomial functions through given data points, so that these polynomials pass through each data point and have continuous derivatives; the key to cubic spline interpolation is to ensure that the interpolation function not only passes through each node, but also guarantees the continuity of the first-order and second-order derivatives, so as to obtain a smooth interpolation curve;

[0080] Step 3: Perform period segmentation on the denoised signal, perform amplitude normalization and length normalization to obtain an average periodic signal;

[0081] To obtain the average periodic signal of the pulse wave, firstly, the interval signal between two adjacent valley values ​​of the pulse wave signal is divided into a period, and each periodic signal is normalized. The amplitude normalization calculation formula is:

[0082]

[0083] Where X(i) represents the value of each point of each pulse wave cycle signal, X min is the minimum value of each cycle, X maxis the maximum value of each cycle;

[0084] Length normalization is to first find the maximum length T of each pulse wave cycle signal max , through cubic spline interpolation, each signal period is T max , define a new time sampling point x in each cycle new , Where N i is the number of sampling points in each original cycle;

[0085] After normalizing the amplitude and length, the length of the M periodic signals is T. max The pulse wave period signal is expressed as X m , 1≤m≤M, sum and average all pulse wave period signals to obtain the average period signal, the calculation formula is:

[0086]

[0087] Step 4: Construct an overcomplete atomic dictionary D based on the improved Gabor function, using a commonly used discretization method β = (a b ,ca b Δμ,da -b Δω,eΔv), obtain the value range of the four parameters β=(δ,μ,ω,v) in the Gabor function, and construct the atomic dictionary D, where each column vector in D represents an atom;

[0088] The improved Gabor function has the following function formula:

[0089]

[0090] Where G is the improved Gabor function, which is controlled by a quaternary group β = (δ, μ, ω, v) for morphology control, x is the data point position of the average period signal of the pulse wave, and the value range of the quaternary group parameter is determined by a commonly used discretization method β = (a b ,ca b Δμ,da -b Δω, eΔv) is determined, where a=2, Δμ=1 / 2, Δω=π, Δv=π / 6,0 <b≤log2L,0≤c<L2 -b+1 ,0≤d<2 b+1 ,0≤e≤12,b,c,d,e∈Z, L is the length of the pulse wave signal for sparse representation;

[0091] The average pulse cycle signal is downsampled. According to the sampling frequency and human heart rate, the length of a single-cycle pulse wave signal is between 125 and 250 sampling points. Since the main frequency distribution of the pulse wave signal is below 20 Hz, according to the Nyquist theorem, to ensure that the signal information is not excessively lost during the sampling process, the sampling frequency should be at least twice the maximum frequency of the signal. In this case, the signal can be shortened to 6.25 times the original length at most. The pulse wave signal length L for sparse representation is set to 64, and the calculation formula is:

[0092]

[0093] Where Z(i) represents the i-th point after normalization, t0 is the largest integer less than t, t1 is the smallest integer greater than t, and N is the length of the original single-cycle signal;

[0094] According to the above formula, the value range of each parameter in the quaternion β=(δ,μ,ω,v) is calculated, 2≤σ≤64,0≤μ≤64,0≤ω≤2π,0≤v≤2π, and an overcomplete atomic dictionary D is constructed; for each set of parameters β=(δ,μ,ω,v), a corresponding Gabor atom G is generated β (x), by traversing all combinations of parameters, the generated multiple Gabor atoms are stored in the dictionary. The final atom dictionary D can be expressed as: D = {G β (x)|β=(δ,μ,ω,v)}, each column in D represents an atom;

[0095] Step 5: Use the OMP orthogonal matching pursuit algorithm to sparsely represent the average pulse wave period signal and obtain s sparse representation components (γ1G1,γ2G2,...γ s G s ), where G p represents the pth atom obtained from the atom dictionary, γ p represents its coefficient, 1≤p≤s, and then these components are divided into main wave component, tidal wave component, dicrotic wave component and tail component, and a total of 4 time-frequency characteristic components are obtained;

[0096] The specific process of using the OMP orthogonal matching pursuit algorithm to sparsely represent the average pulse wave period signal is as follows:

[0097] (1) Initialization: residual r = y, atomic set Coefficient Set y is the signal to be reconstructed after standard deviation normalization;

[0098] (2) Atom selection: Calculate the inner product of the current residual r and all atoms in the dictionary D, and find the atom G with the largest absolute value of the inner product with the current residual j, the calculation formula is:

[0099] G p =argmax|r T D q |

[0100] Where D q is the basis vector of the qth column in the atomic dictionary D;

[0101] (3) Update the atom set: Update the selected atom G p Add to the atomic set G;

[0102] (4) Use the least squares method to update the coefficient set γ: that is, find a set of coefficients that makes y in the currently selected atom G p The coefficient γ with the smallest reconstruction error p , the calculation formula is:

[0103] γ p =argmin||y-γ p G p || 2

[0104] The calculated coefficient γ p Add to the coefficient set γ;

[0105] (5) Update residual: The updated residual is the difference between the current signal and the reconstructed signal. The calculation formula is:

[0106] r=y-γ p G p

[0107] (6) Repeat steps (2) to (5) to obtain s decomposition components (γ1G1,γ2G2,...γ s G s ).

[0108] The algorithm stops in two cases:

[0109] (1) The number of sparse components has reached the predetermined number;

[0110] (2) The size of the residual is smaller than the preset tolerance, indicating that the signal reconstruction is accurate enough;

[0111] like Figure 4 As shown, 12 sparse components of sparse decomposition of the pulse wave signal are displayed;

[0112] According to each component γ p G p The location of the maximum value point, combined with Figure 5 The marked pulse wave signal time domain feature points are classified, γp G p The position of the maximum amplitude point corresponding to the time domain horizontal coordinate is between A and C, which is the main wave component, between C and E is the tidal wave component, between E and (F+G) / 2 is the dicrotic wave component, and between (F+G) / 2 and G is the tail component; and each type of component is added and summed to obtain 4 time-frequency characteristic components, such as Figure 6 shown.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit it. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A pulse wave time-frequency feature extraction method based on an improved Gabor atom dictionary, characterized in that: The following steps are involved: Step 1: Remove high-frequency noise by performing center sliding average and wavelet transform on the original pulse wave signal, and then perform Fourier transform to obtain the fundamental frequency f of the signal. b ; Step 2: After removing the high-frequency noise, the pulse wave signal is subjected to a baseline drift removal operation according to the valley point of the pulse wave signal and the selected auxiliary point to eliminate the low-frequency noise and obtain a pulse wave signal without noise; Step 3: Perform period segmentation on the denoised signal, perform amplitude normalization and length normalization to obtain an average periodic signal; Step 4: Construct an overcomplete atomic dictionary D based on the improved Gabor function, using a commonly used discretization method β = (a b ,ca b Δμ,da -b Δω,eΔv), obtain the value range of the four parameters β=(δ,μ,ω,v) in the Gabor function, and construct the atomic dictionary D, where each column vector in D represents an atom; Step 5: Use the OMP orthogonal matching pursuit algorithm to sparsely represent the average pulse wave period signal and obtain s sparse representation components (γ1G1,γ2G2,...γ s G s ), where G p represents the pth atom obtained from the atom dictionary, γ p Represents its coefficient, 1≤p≤s, and then these components are divided into main wave component, tidal wave component, dicrotic wave component and tail component, and a total of 4 time-frequency characteristic components are obtained.

2. The method for extracting pulse wave time-frequency features based on an improved Gabor atom dictionary according to claim 1, characterized in that: The specific process of removing high-frequency noise described in step 1 is as follows: (1) Set the window size to w and perform center sliding average denoising on the original pulse wave signal: Where CMA(t) represents the central moving average of the t-th data point, and x(i) represents the i-th data point in the original pulse wave signal; (2) When performing wavelet transform denoising, determine the number of wavelet transform decomposition layers, wavelet basis functions, and threshold functions, and apply the improved threshold function When |w j，k |≤λ, When |w j,k When |≥λ, The function formula is: In the formula, w j,k is the kth wavelet coefficient of the jth layer obtained after wavelet transform decomposition, λ is the threshold, δ=(median|w j,k |) / 0.6745, N is the number of sampling points of the signal, α is the adjustment factor, α≥0, the larger α is The closer to the hard threshold, the smaller α The closer to the soft threshold, the median|w j,k | is the wavelet coefficient w of the jth layer decomposition j,k The median value of the amplitude; After the pulse wave signal is processed by the central sliding average and wavelet transform, it will no longer contain high-frequency noise. The discrete Fourier transform is performed on it to obtain X[n], and the amplitude spectrum of X[n] is calculated: Where Re(X[n]) and Im(X[n]) are the real and imaginary parts of X[n] respectively; The fundamental frequency is the frequency corresponding to the maximum amplitude in the amplitude spectrum, and its calculation formula is: Where f b is the fundamental frequency, n max is the index corresponding to the maximum amplitude in the amplitude spectrum, f s is the sampling frequency, N is the number of Fourier transform points, which is equal to the number of sampling points of the signal.

3. The pulse wave time-frequency feature extraction method based on an improved Gabor atom dictionary according to claim 1, characterized in that: When removing the baseline drift described in step 2, first use a sliding window to traverse the entire signal to obtain the valley value of the pulse wave signal, and set the window size W = f s / f b , f s The signal sampling frequency is set, and the step size is set to 2 / 3 of the window size. The signal is detected for valley values. A baseline is initially obtained based on the valley values ​​of the signal using the cubic spline interpolation method. The points in the pulse wave signal that are lower than this baseline are used as auxiliary points. The baseline is refitted using the cubic spline interpolation method based on the valley values ​​and auxiliary points of the signal to obtain a new baseline, which is then removed.

4. The method for extracting pulse wave time-frequency features based on an improved Gabor atom dictionary according to claim 1, characterized in that: The step 3 of obtaining the average pulse wave period signal is to first divide the interval signal between two adjacent valley values ​​of the pulse wave signal into a period, and normalize each period signal. The amplitude normalization calculation formula is: Where X(i) represents the value of each point of each pulse wave cycle signal, X min is the minimum value of each cycle, X max is the maximum value of each cycle; Length normalization is to first find the maximum length T of each pulse wave cycle signal max , through cubic spline interpolation, each signal period is T max , define a new time sampling point x in each cycle new , Where N i is the number of sampling points in each original cycle; After normalizing the amplitude and length, multiple pulse wave periodic signals are obtained. Assume that there are M pulse wave periodic signals, and each periodic signal has a length of T. max The discrete signal sequence is represented by X m , 1≤m≤M, sum and average all pulse wave period signals to obtain the average period signal, the calculation formula is:

5. The method for extracting pulse wave time-frequency features based on an improved Gabor atom dictionary according to claim 1, characterized in that: The improved Gabor function described in step 4 has the following function formula: Where G is the improved Gabor function, which is controlled by a quaternary group β = (δ, μ, ω, v) for morphology control, x is the data point position of the average period signal of the pulse wave, and the value range of the quaternary group parameter is determined by a commonly used discretization method β = (a b ,ca b Δμ,da -b Δω, eΔv) is determined, where a=2, Δμ=1 / 2, Δω=π, Δv=π / 6,0 <b≤log2L,0≤c<L2 -b+1 ,0≤d<2 b+1 ,0≤e≤12,b,c,d,e∈Z, L is the length of the pulse wave signal for sparse representation; For each set of parameters β = (δ, μ, ω, v), a corresponding Gabor atom G is generated β (x), by traversing all combinations of parameters, the generated multiple Gabor atoms are stored in the dictionary. The final atom dictionary D can be expressed as: D = {G β (x)|β=(δ,μ,ω,v)}, each column in D represents an atom.

6. The method for extracting pulse wave time-frequency features based on an improved Gabor atom dictionary according to claim 1, characterized in that: The specific process of using the OMP orthogonal matching pursuit algorithm to sparsely represent the average pulse wave period signal described in step 5 is: (1) Initialization: residual r = y, atomic set Coefficient Set y is the signal to be reconstructed after standard deviation normalization; (2) Atom selection: Calculate the inner product of the current residual r and all atoms in the dictionary D, and find the atom G with the largest absolute value of the inner product with the current residual j , the calculation formula is: G p =arg max|r T D q | Where D q is the basis vector of the qth column in the atomic dictionary D; (3) Update the atom set: Update the selected atom G p Add to the atomic set G; (4) Use the least squares method to update the coefficient set γ: that is, find a set of coefficients that makes y in the currently selected atom G p The coefficient γ with the smallest reconstruction error p , the calculation formula is: c p =arg min||y-γ p G p || 2 The calculated coefficient γ p Add to the coefficient set γ; (5) Update residual: The updated residual is the difference between the current signal and the reconstructed signal, and the calculation formula is: r=y-γ p G p (6) Repeat steps (2) to (5) to obtain s decomposition components (γ1G1,γ2G2,...γ s G s ).

7. The method for extracting pulse wave time-frequency features based on an improved Gabor atom dictionary according to claim 1, characterized in that: The four time-frequency characteristic components described in step 5 are based on each component γ p G p The position of the maximum amplitude point is classified in combination with the time domain characteristic points of the pulse wave signal. If the time domain horizontal coordinate corresponding to the maximum amplitude point of the component is between the starting point of the pulse wave and the starting point of the tidal wave, it is the main wave component; between the starting point of the tidal wave and the descending gorge, it is the tidal wave component; between the descending gorge and the end point of the dibolic wave, it is the dibolic wave component; between the end point of the dibolic wave and the end point of the pulse wave, it is the tail component. In this way, four categories of main wave component, tidal wave component, dibolic wave component and tail component are obtained, and the components of each category are added and summed to obtain 4 time-frequency characteristic components.

Images