A method for extracting frequency-domain features of pulse wave signals

Abstract

The present invention relates to a method for extracting frequency-domain features of pulse wave signals, belonging to the field of signal processing. It includes: acquiring pulse signals and performing preprocessing, cycle segmentation, quality screening, cycle normalization, and cycle extension; respectively extracting the first group, second group, third group, and fourth group of frequency-domain features through Fourier transform, power spectrum transform, cepstrum transform, and fractional Fourier transform; constructing an initial feature matrix and performing dimensionality reduction to obtain the dimensionality-reduced feature matrix and its evaluation score, and screening the dimensionality-reduced feature matrix based on the evaluation score to form the first data set; selecting the eigenvector corresponding to the minimum KL divergence between the feature matrix in the first data set and the initial feature matrix as the final frequency-domain feature. The present invention extracts the frequency-domain features of pulse signals through Fourier transform, power spectrum transform, cepstrum transform, and fractional Fourier transform, and selects the final frequency-domain features through the evaluation scores of each dimensionality reduction and KL divergence. The extracted frequency-domain features are more comprehensive and more effective.

Description

Technical Field

[0001] The present invention relates to the technical field of physiological signal detection, and particularly relates to a method for extracting frequency-domain features of pulse wave signals. Background Art

[0002] Pulse signals are important human physiological signals and are external reflections of important information such as the state of the heart and blood vessels. In traditional Chinese medicine theory, by means of "feeling the pulse", the characteristics of pulse signals such as frequency, rhythm, depth, and strength can be obtained, the types of pulse signals can be determined, and thus the changes in the human physiological system can be diagnosed.

[0003] With the development of biomedical engineering, the research on pulse signals has been ongoing. Researchers are committed to converting the subjective judgments of physicians into machine biometric recognition with reference basis, so as to provide unified judgment criteria and norms. With the development of machine learning and deep learning, how to achieve more accurate and detailed classification of pulse signals has become a new research topic. Therefore, how to extract more comprehensive and effective pulse signal features to provide a basis for the classification and recognition of pulse signals is particularly important.

[0004] At present, the existing methods for extracting frequency-domain features of pulse signals mainly extract single spectra, and the extracted frequency-domain features are not comprehensive. Therefore, providing a method that can extract more comprehensive and effective frequency-domain features of pulse signals is an urgent problem to be solved. Summary of the Invention

[0005] In view of the above analysis, the present invention aims to provide a method for extracting frequency-domain features of pulse wave signals to solve the problems that the existing methods for extracting frequency-domain features of pulse signals are not comprehensive and have low effectiveness.

[0006] The present invention provides a method for extracting frequency-domain features of pulse wave signals, and the method includes the following steps:

[0007] Obtain pulse signals of a plurality of human bodies through a sensor, and preprocess the pulse signals to obtain the preprocessed pulse signals of each human body;

[0008] Perform cycle segmentation, quality screening, cycle normalization, and cycle extension on the preprocessed pulse signals of each human body to obtain the cycle-extended pulse signals of each human body;

[0009] Extract a first group of frequency-domain features from the cycle-extended pulse signals of each human body through Fourier transform, extract a second group of frequency-domain features through power spectrum transform, extract a third group of frequency-domain features through cepstrum transform, and extract a fourth group of frequency-domain features through fractional Fourier transform;

[0010] Construct an initial feature matrix based on the frequency-domain features of all human pulse signals, perform dimensionality reduction on the initial feature matrix multiple times to obtain the feature matrices after each dimensionality reduction, obtain the evaluation scores of the frequency-domain eigenvalue matrices corresponding to the feature matrices after each dimensionality reduction based on the SVM model, and screen the feature matrices after dimensionality reduction based on the evaluation scores to form the first data set;

[0011] Calculate the KL divergence between each feature matrix in the first data set and the initial feature matrix, and select the frequency-domain features corresponding to the feature matrix when the KL divergence is the smallest as the final frequency-domain features of each human pulse signal.

[0012] Furthermore, perform eigen decomposition on the initial feature matrix to obtain eigenvectors, subtract the eigenvectors corresponding to the smallest eigenvalue from the eigenvectors in turn to obtain the eigenvectors after each dimensionality reduction, and construct the feature matrices after each dimensionality reduction based on the eigenvectors after each dimensionality reduction.

[0013] Furthermore, the performing eigen decomposition on the initial feature matrix to obtain eigenvectors includes:

[0014] Perform centering on the frequency-domain features in the initial feature matrix to obtain the centered initial feature matrix;

[0015] Obtain its covariance matrix based on the centered initial feature matrix;

[0016] Perform eigen decomposition on the covariance matrix to obtain eigenvectors, and sort the features in the eigenvectors in descending order of eigenvalues.

[0017] Furthermore, the KL divergence between the feature matrix in the first data set and the initial feature matrix is obtained by the following method:

[0018]

[0019] where N1 is the initial feature matrix, N2 is the feature matrix in the first data set, k is the dimension of the feature vector corresponding to the feature matrix in the first data set, μ1 is the mean corresponding to N1, μ2 is the mean corresponding to N2, ∑1 is the variance corresponding to N1, and ∑2 is the variance corresponding to N2.

[0020] Furthermore, extract the resonance frequency, resonance amplitude, amplitude difference, frequency difference, and amplitude ratio as the first group of frequency-domain features of each human pulse signal from the pulse signal of each human after period extension through Fourier transform; the resonance frequency is the frequency corresponding to the Fourier transform harmonic, the resonance amplitude is the amplitude corresponding to the Fourier transform harmonic, the amplitude difference is the amplitude difference between adjacent harmonics, the frequency difference is the frequency difference between adjacent harmonics, and the amplitude ratio is the ratio of the amplitude of the current harmonic to the amplitude of the fundamental harmonic.

[0021] Further, the energy ratio, peak area, and peak ratio are extracted from the pulse signals after the human body cycle extension through power spectrum transformation as the second set of frequency domain features of each human body pulse signal; the pulse signals are divided into several segments of equal length, the periodogram of each segment of the pulse signal is calculated, and the arithmetic mean of the periodograms of all segments of the pulse signals is calculated to obtain the power spectrum.

[0022] Further, the cepstrum peak amplitude is extracted from the pulse signals after the human body cycle extension through cepstrum transformation as the third set of frequency domain features of each human body pulse signal; the cepstrum peak amplitude is the amplitude values corresponding to the 10 peaks of the cepstrum of the pulse signal.

[0023] Further, the lowest point value, the highest point value, the difference in abscissa between the highest point and the lowest point, the difference between adjacent highest points, and the difference between adjacent lowest points of each fractional domain spectrum are extracted from the pulse signals after the human body cycle extension through fractional Fourier transform as the fourth set of frequency domain features of each human body pulse signal.

[0024] Further, the preprocessing of the pulse signals to obtain the preprocessed pulse signals of each human body includes:

[0025] Set the noise threshold, determine the wavelet basis function and the decomposition level, perform wavelet decomposition on the pulse signals to obtain multiple wavelet components, denoise the wavelet components according to the noise threshold, and then reconstruct the signal using wavelet coefficients to obtain the denoised pulse signals.

[0026] Extract the pacing points from the denoised pulse signals by searching for local extrema, perform cubic spline interpolation on all pacing points to fit the baseline noise, and remove the baseline noise from the denoised pulse signals to obtain the preprocessed pulse signals.

[0027] Further, the period segmentation, quality screening, period normalization, and period extension of the preprocessed pulse signals of each human body to obtain the pulse signals after the human body cycle extension of each human body include:

[0028] Perform the following steps on the preprocessed pulse signals of each human body:

[0029] Step S21: Perform period segmentation on the preprocessed pulse signals based on the pacing points;

[0030] Step S22: Calculate the peak-valley slope, peak-valley distance, and peak-valley difference of the pulse signals of each period after segmentation. Based on the peak-valley slopes and peak-valley distances of all the segmented pulse signals, obtain the arithmetic mean of the peak-valley slopes and the arithmetic mean of the peak-valley distances respectively. Remove the periodic pulse signals whose peak-valley slope deviates from the peak-valley slope average by more than the threshold or whose peak-valley distance deviates from the peak-valley distance average by more than the threshold from each period signal to obtain the remaining periodic pulse signals;

[0031] Step S23: Obtain the average cycle length of each remaining cycle pulse signal; perform cycle normalization by extending or compressing each remaining cycle pulse signal.

[0032] Step S24: Perform cycle extension on the cycle-normalized pulse signal to obtain the cycle-extended pulse signal.

[0033] Compared with the prior art, the present invention can at least achieve one of the following beneficial effects:

[0034] 1. The present invention extracts the frequency-domain features of the pulse signal through Fourier transform, power spectrum transform, cepstrum transform, and fractional Fourier transform, realizing a more comprehensive extraction of the frequency-domain features of the pulse signal, providing a basis for more accurate and detailed classification and recognition of the pulse signal.

[0035] 2. The present invention selects the corresponding feature matrix through the evaluation scores of each dimensionality reduction, and selects the final frequency-domain features through KL divergence, reducing the dimension of the frequency-domain features as much as possible without losing the main information of the pulse signal, providing a basis for improving the speed of neural network model training and the accuracy of model output based on the frequency-domain features.

[0036] 3. The present invention suppresses noise through wavelet transform and cubic spline interpolation fitting, improving the fitting accuracy, so that the extracted features can more accurately reflect the pulse information.

[0037] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent specification, and some advantages can be made obvious from the specification, or understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The drawings are only for the purpose of showing specific embodiments, and are not considered as limiting the present invention. Throughout the drawings, the same reference signs denote the same components.

[0039] Figure 1 It is a flowchart of the method for extracting the frequency-domain features of the pulse signal in the embodiment of the present invention;

[0040] Figure 2 It is a cepstrum diagram of the pulse signal in the embodiment of the present invention;

[0041] Figure 3 It is a fractional-domain frequency spectrum diagram of the pulse signal when p = 0.3 in the embodiment of the present invention;

[0042] Figure 4Schematic diagram of the fractional domain spectrum when the pulse signal p = 0.5 in the embodiment of the present invention;

[0043] Figure 5 Schematic diagram of the fractional domain spectrum when the pulse signal p = 0.7 in the embodiment of the present invention. Detailed implementation manners

[0044] The following will specifically describe the preferred embodiments of the present invention with reference to the accompanying drawings. The accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principle of the present invention, rather than to limit the scope of the present invention.

[0045] A specific embodiment of the present invention discloses a method for extracting frequency domain features of a pulse wave signal. As Figure 1 shown, the method includes the following steps:

[0046] Step S1: Obtain the pulse signals of a number of human bodies through a sensor, and preprocess the pulse signals to obtain the preprocessed pulse signals of each human body;

[0047] Step S2: Perform cycle segmentation, quality screening, cycle normalization, and cycle extension on the preprocessed pulse signals of each human body to obtain the cycle-extended pulse signals of each human body;

[0048] Step S3: Extract the first set of frequency domain features from the cycle-extended pulse signals of each human body through Fourier transform, extract the second set of frequency domain features through power spectrum transform, extract the third set of frequency domain features through cepstrum transform, and extract the fourth set of frequency domain features through fractional Fourier transform;

[0049] Step S4: Construct an initial feature matrix based on the frequency domain features of all human body pulse signals, perform multiple dimensionality reductions on the initial feature matrix to obtain the feature matrices after each dimensionality reduction, obtain the evaluation scores of the frequency domain eigenvalue matrices corresponding to the feature matrices after each dimensionality reduction based on the SVM model, and screen the feature matrices after dimensionality reduction based on the evaluation scores to form a first data set;

[0050] Step S5: Calculate the KL divergence between each feature matrix in the first data set and the initial feature matrix, and select the frequency domain features corresponding to the feature matrix when the KL divergence is the smallest as the final frequency domain features of each human body pulse signal.

[0051] Specifically, in step S1, the preprocessing of the pulse signals to obtain the preprocessed pulse signals of each human body includes:

[0052] Set a noise threshold, determine the wavelet basis function and the decomposition level, perform wavelet decomposition on the pulse signal to obtain multiple wavelet components, denoise the wavelet components according to the noise threshold, and then reconstruct the signal using wavelet coefficients to obtain the denoised pulse signal;

[0053] The pacemaker points are extracted from the denoised pulse signal by searching for local extrema, and cubic spline interpolation is performed on all the pacemaker points to fit the baseline noise. The baseline noise is removed from the denoised pulse signal to obtain the preprocessed pulse signal.

[0054] Specifically, a number of human pulse signals are acquired through a sensor, and then the pulse signals of each human body are preprocessed respectively. The pulse signal is the trajectory of vascular pulsation and is a periodic signal. The pulse signal contains a main wave, a tidal wave (pre-diastolic wave), and a dicrotic wave within each period. The amplitude of the main wave reflects left ventricular ejection and large artery compliance, the amplitude of the tidal wave reflects vascular elasticity and peripheral resistance, and the amplitude of the dicrotic wave reflects large artery elasticity and the state of the aortic valve. The time-domain characteristics of the pulse signal can reflect the change of its frequency over time, as well as the instantaneous frequency and amplitude corresponding to each moment. The pulse signal of the human body is sampled through a sensor at a sampling frequency of 125 Hz, and a total of 10,000 sampling data are obtained.

[0055] Furthermore, a low-frequency noise threshold and a high-frequency noise threshold are set. The pulse signal is decomposed by 7-layer wavelet using the sym8 wavelet basis function, and the wavelet components with frequencies lower than the low-frequency noise threshold and higher than the high-frequency noise threshold are removed. The remaining wavelet components are used to reconstruct the signal by wavelet coefficients to obtain the denoised pulse signal.

[0056] Preferably, 0.825 Hz is set as the low-frequency noise threshold, and 31.25 Hz is set as the high-frequency noise threshold, which can remove the low-frequency noise generated by skin friction, 50 Hz power frequency interference, and 30 - 120 Hz high-frequency noise of the electrocardiogram signal, and improve the signal-to-noise ratio.

[0057] Furthermore, a search step size is set, and the local minima of the denoised pulse signal are searched within the range of the search step size as the pacemaker points. All the pacemaker points are used as interpolation points for cubic spline interpolation to fit the baseline noise, and the denoised pulse signal is subtracted by the baseline noise to obtain the preprocessed pulse signal.

[0058] Preferably, since the pulse signal period is usually between 80 and 110, the search step size is set to 50.

[0059] It can be understood that in addition to high and low frequency noises, the pulse signal also contains baseline drift caused by respiration. The cubic spline interpolation method is used to perform curve fitting on the pacemaker points to simulate respiration and remove it, and a smooth pulse signal with a high signal-to-noise ratio can be obtained.

[0060] Specifically, in step S2, the process of performing cycle segmentation, quality screening, cycle normalization, and cycle extension on the preprocessed pulse signal of each human body to obtain the cycle-extended pulse signal of each human body includes:

[0061] Perform the following steps on the preprocessed pulse signal of each human body:

[0062] Step S21: Perform cycle segmentation on the preprocessed pulse signal based on the pacemaker point;

[0063] Step S22: Calculate the peak-valley slope, peak-valley distance, and peak-valley difference for the pulse signals of each segmented cycle. Based on the peak-valley slopes and peak-valley distances of all segmented pulse signals, obtain the arithmetic mean of the peak-valley slopes and the arithmetic mean of the peak-valley distances respectively. Remove the periodic pulse signals from each cycle signal whose peak-valley slope deviates from the peak-valley slope mean by more than the threshold or whose peak-valley distance deviates from the peak-valley distance mean by more than the threshold, to obtain the remaining periodic pulse signals;

[0064] Step S23: Obtain the average cycle length of each remaining periodic pulse signal; perform cycle normalization by extending or compressing each remaining periodic pulse signal;

[0065] Step S24: Perform cycle extension on the cycle-normalized pulse signal to obtain the cycle-extended pulse signal.

[0066] Specifically, in step S21, the signal between adjacent pacemaker points is a cycle signal, and the preprocessed pulse signal is segmented according to the pacemaker point.

[0067] Specifically, in step S22, calculate the peak-valley slope, peak-valley distance, and peak-valley difference for each segmented cycle signal. The peak-valley slope is the slope of the line connecting the peak point of the cycle signal and the pacemaker point, that is, the ratio of the difference in the vertical coordinates of the peak point and the pacemaker point to the difference in the horizontal coordinates of the peak point and the pacemaker point; the peak-valley distance is the difference in the horizontal coordinates between the peak point of the cycle signal and the pacemaker point, and the peak-valley difference is the difference in the vertical coordinates between the peak point of the cycle signal and the pacemaker point.

[0068] Preferably, remove the periodic pulse signals whose peak-valley slope deviates from the peak-valley slope mean by more than 10% or whose peak-valley distance deviates from the peak-valley distance mean by more than 10%.

[0069] Specifically, in step S23, for each removed periodic pulse signal, if the length of the signal is greater than the average cycle length, truncate the signal to the average cycle length; if the length of the signal is less than the average cycle length, pad the signal with zeros to the average cycle length.

[0070] It can be understood that during the process of processing a large number of different pulse signals, amplitude normalization can eliminate the dimensional data differences of the pulse signals and retain the shapes of the main peak, pre-branch wave, and dicrotic wave of the signal; cycle normalization retains the relative positions of characteristic points such as the main peak, pre-branch wave, and dicrotic wave of the signal, facilitating subsequent analysis and processing.

[0071] Specifically, in step S24, a periodic extension with a quantity of 10 is performed, that is, the pulse signal after period normalization is copied 10 times for extension, so as to enhance the intensity of the pulse signal and facilitate better extraction of frequency domain features.

[0072] Specifically, in step S3, for the pulse signal of each human body after periodic extension, the resonant frequency, resonant amplitude, amplitude difference, frequency difference, and amplitude ratio are extracted through Fourier transform as the first group of frequency domain features of each human body's pulse signal; the resonant frequency is the frequency corresponding to the harmonic of the Fourier transform, the resonant amplitude is the amplitude corresponding to the harmonic of the Fourier transform, the amplitude difference is the amplitude difference between adjacent harmonics, the frequency difference is the frequency difference between adjacent harmonics, and the amplitude ratio is the ratio of the amplitude of the current harmonic to the amplitude of the fundamental harmonic.

[0073] Further, for the pulse signal of each human body after periodic extension, the energy ratio, peak area, and peak ratio are extracted through power spectrum transform as the second group of frequency domain features of each human body's pulse signal.

[0074] Further, the pulse signal is divided into several segments of equal length, the periodogram of each segment of the pulse signal is calculated, and the arithmetic mean of the periodograms of all segments of the pulse signal is calculated to obtain the power spectrum.

[0075] Specifically, the following method is used to obtain several segments of pulse signals of equal length:

[0076] x i (n) = x[n+(i - 1)M] 0 ≤ n ≤ M - 1, 1 ≤ i ≤ L,

[0077] where L is the number of divided segments, M is the length of each segment of the pulse signal, x[] is the pulse signal after periodic extension, and x i (n) is the pulse signal of each segment.

[0078] Further, the following method is used to obtain the periodogram of each segment of the pulse signal:

[0079]

[0080] Further, the arithmetic mean of the periodograms of all segments of the pulse signal, that is, the power spectrum is:

[0081]

[0082] Specifically, the energy ratio is obtained through the following formula:

[0083] P Ei = E i / E a ,

[0084] where E iis the energy corresponding to the frequencies in the ranges of 5, 10, 15, 20, and 40 Hz respectively, E a is the total energy of the power spectrum.

[0085] Specifically, the peak area is obtained through the following formula:

[0086]

[0087] where X i is the sampling point corresponding to the i-th peak.

[0088] Specifically, the peak ratio is obtained through the following formula:

[0089]

[0090] where, A i is the amplitude value corresponding to the i-th peak, and A1 is the amplitude value corresponding to the first peak.

[0091] It can be understood that the power spectrum obtained by the above method has higher resolution, thus extracting more and more detailed frequency-domain features of the pulse signal to provide a basis for subsequent classification and recognition.

[0092] Furthermore, as Figure 2 shown, the cepstrum peak amplitude is extracted from the pulse signal after extension of each human body cycle through cepstrum transformation as the third group of frequency-domain features of each human body pulse signal; the cepstrum peak amplitude is the amplitude values corresponding to the 10 peaks of the cepstrum of the pulse signal.

[0093] Specifically, the cepstrum transformation is performed through the following formula:

[0094]

[0095] where, F(ω) is the spectrum obtained by Fourier transformation of the original pulse signal.

[0096] Furthermore, the lowest point value, the highest point value, the difference in abscissa between the highest point and the lowest point, the difference between adjacent highest points, and the difference between adjacent lowest points of the fractional-domain spectra are extracted from the pulse signal after extension of each human body cycle through fractional Fourier transformation as the fourth group of frequency-domain features of each human body pulse signal.

[0097] Specifically, the fractional Fourier transformation is performed using the following formula:

[0098]

[0099] K α (u,t) = K α+2π (u,t)

[0100]

[0101] Among them, K α (u, t) is the kernel function of the fractional Fourier transform, f(t) is the pulse signal after the periodic extension of each human body, α is the rotation angle, p is the order of the fractional Fourier transform, u is the variable in the fractional transform domain, δ(t) is the impulse function.

[0102] Specifically, as Figure 3 、 Figure 4 、 Figure 5 shown, the fractional Fourier transform of the pulse wave signal is performed with p = 0.3, p = 0.5, and p = 0.7 using the above formula. When p = 0.3, p = 0.5, and p = 0.7, the corresponding α values are 0.15π, 0.25π, and 0.35π respectively.

[0103] It can be understood that the fractional Fourier transform maps the signal to the fractional domain between the time domain and the frequency domain with the order p as a parameter. The fractional domain Fourier transform can rotate the coordinate axes of the time-frequency plane from the perspective of viewing the time-frequency plane, so as to analyze information from the perspective of observing the frequency domain. By introducing the parameter α to control the transition of the signal between the time domain and the frequency domain. The value of α can be a real number or a complex number, and different α values will result in different transformation results. This enables it to more finely control the trade-off relationship between the time domain and the frequency domain of the signal, and thus has higher flexibility when processing non-stationary signals. Since the pulse wave signal belongs to a non-stationary signal, the fractional domain Fourier transform and the global basis function can be used to well analyze the properties and characteristics of the pulse wave signal, so as to extract more detailed features and provide a basis for subsequent classification and recognition.

[0104] Specifically, in step S4, the frequency domain features of all human pulse signals are normalized to construct an initial feature matrix. The frequency domain features of all human pulse signals are normalized through the following formula to map the frequency domain feature values to the interval (0, 1):

[0105]

[0106] Among them, x is the frequency domain feature value.

[0107] Specifically, one column of the initial feature matrix is the frequency domain feature after the above normalization of the pulse signal of one human body. Exemplarily, the pulse signals of m human bodies (i.e., m samples) are obtained, and the frequency domain features of each human pulse signal are n-dimensional, then the size of the initial feature matrix is n * m.

[0108] Further, perform eigen - decomposition on the initial feature matrix to obtain eigen - vectors, successively subtract the eigen - vector corresponding to the minimum eigenvalue from the eigen - vectors to obtain the eigen - vectors after each dimensionality reduction, and construct the feature matrices after each dimensionality reduction based on the eigen - vectors after each dimensionality reduction.

[0109] Further, the performing eigen - decomposition on the initial feature matrix to obtain eigen - vectors includes:

[0110] Decentralize the frequency - domain features in the initial feature matrix to obtain a decentralized initial feature matrix;

[0111] Obtain its covariance matrix based on the decentralized initial feature matrix;

[0112] Perform eigen - decomposition on the covariance matrix to obtain eigen - vectors, and sort the features in the eigen - vectors in descending order of eigenvalues.

[0113] Further, the centralized initial feature matrix is obtained by the following method:

[0114]

[0115] where x ij is the frequency - domain feature (i.e., frequency - domain eigenvalue) at the i - th row and j - th column in the initial feature matrix, m is the number of columns in the initial feature matrix, and x′ ij is the element at the i - th row and j - th column in the decentralized initial feature matrix.

[0116] Specifically, the covariance matrix is obtained by the following method:

[0117] R = XX T ,

[0118] where X is the decentralized initial feature matrix, and X T is the transpose matrix of the decentralized initial feature matrix.

[0119] Specifically, the eigen - decomposition of the covariance matrix is performed by the following method:

[0120] R = V∧V -1

[0121] where V is the matrix composed of eigen - vectors, V=(v1, v2…v n ), V -1 is the inverse matrix of V, and ∧ represents the diagonal matrix composed of eigenvalues.

[0122] It can be understood that the eigen - vector is expressed as v i =(v i1 , v i2 …v in ), where vin is the n-th eigenvalue of the i-th eigenvector, arranged in descending order of eigenvalues. The smaller the eigenvalue, the less eigeninformation it contains. Therefore, the smallest eigenvalue is taken each time for dimensionality reduction, and the eigenvector after the first dimensionality reduction is denoted as v i ′ = (v i1 , v i2 … v in-1 ).

[0123] Furthermore, the frequency-domain features corresponding to the dimensionality-reduced feature matrix are obtained by the following method:

[0124] z ij = W T x ij ,

[0125] where z ij is the frequency-domain feature at the i-th row and j-th column in the dimensionality-reduced feature matrix, W is the dimensionality-reduced feature matrix, i is the number of rows of the dimensionality-reduced feature matrix, that is, the dimension of the eigenvector corresponding to this dimensionality-reduced feature matrix.

[0126] It can be understood that after the first dimensionality reduction, W is a matrix with n - 1 rows and m columns, that is, a matrix composed of n - 1-dimensional eigenvectors of m samples. Based on the above formula, the frequency-domain eigenvalue matrix corresponding to the dimensionality-reduced feature matrix can be obtained, that is, a matrix composed of the frequency-domain eigenvalues after dimensionality reduction of m samples.

[0127] Furthermore, the frequency-domain eigenvalue matrices after each dimensionality reduction are respectively input into the trained SVM model to obtain their corresponding accuracy, precision, recall, F1 score, and compression rate; based on their corresponding accuracy, precision, recall, F1 score, and compression rate, the evaluation scores of the frequency-domain eigenvalue matrices corresponding to the feature matrices after each dimensionality reduction are obtained.

[0128] It can be understood that by respectively inputting the frequency-domain eigenvalue matrices after each dimensionality reduction into the trained SVM model for pulse condition prediction of m samples, the prediction result of each sample can be obtained, and based on the prediction result and the true value, their accuracy, precision, recall, F1 score, and compression rate can be obtained.

[0129] Specifically, the evaluation score is obtained by the following formula:

[0130] S = P Acc + P Pre + P Rec + F1 + t,

[0131] where P Acc is the accuracy corresponding to this dimensionality reduction, P Pre is the precision corresponding to this dimensionality reduction, P RecRecall is the recall rate corresponding to this dimensionality reduction, F1 is the F1 score corresponding to this dimensionality reduction, and t is the compression rate corresponding to this dimensionality reduction.

[0132] Specifically, a first data set is formed by selecting a feature matrix corresponding to when the evaluation score of the frequency domain eigenvalue matrix after dimensionality reduction is greater than a threshold value.

[0133] It can be understood that in this application, without losing the main information of the pulse signal, the dimension of the frequency domain features is reduced as much as possible, redundant information is removed, which provides a basis for improving the speed of neural network model training based on frequency domain features and the accuracy of model output in the subsequent process.

[0134] Specifically, in step S5, the KL divergence between the feature matrix in the first data set and the initial feature matrix is obtained by the following method:

[0135]

[0136] Among them, N1 is the initial feature matrix, N2 is the feature matrix in the first data set, k is the dimension of the eigenvector corresponding to the feature matrix in the first data set, μ1 is the mean value corresponding to N1, μ2 is the mean value corresponding to N2, Σ1 is the variance corresponding to N1, and Σ2 is the variance corresponding to N2.

[0137] Specifically, the frequency domain eigenvalue matrix corresponding to the feature matrix when the KL divergence is the smallest is selected, and each column in this frequency domain eigenvalue matrix is used as the final frequency domain feature of each human pulse signal.

[0138] It can be understood that the KL divergence is also known as relative entropy, which is a measure of the asymmetry of the difference between the probability distributions of two data. In this application, the KL divergence is used to select the feature with the highest similarity to the original pulse signal as the final frequency domain feature, and more feature information of the pulse signal is retained while reducing the dimension.

[0139] Compared with the prior art, the beneficial effects of the pulse signal frequency domain feature extraction method provided by the present invention are as follows:

[0140] 1. The present invention extracts the frequency domain features of the pulse signal through Fourier transform, power spectrum transform, cepstrum transform, and fractional Fourier transform, realizing a more comprehensive extraction of the frequency domain features of the pulse signal, and providing a basis for more accurate and detailed classification and recognition of the pulse signal.

[0141] 2. The present invention selects the corresponding feature matrix through the evaluation scores of each dimensionality reduction, and selects the final frequency domain feature through the KL divergence. Without losing the main information of the pulse signal, the dimension of the frequency domain features is reduced as much as possible, which provides a basis for improving the speed of neural network model training based on frequency domain features and the accuracy of model output in the subsequent process.

[0142] 3. The present invention suppresses noise through wavelet transform and cubic spline interpolation fitting, improving the fitting accuracy, so that the extracted features can more accurately reflect the pulse information.

[0143] Those skilled in the art can understand that all or part of the processes of implementing the methods of the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a magnetic disk, an optical disk, a read-only memory or a random access memory, etc.

[0144] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for extracting frequency-domain features of pulse wave signals, characterized in that, The method includes the following steps: Obtain the pulse signals of a number of human bodies through sensors, and preprocess the pulse signals to obtain the preprocessed pulse signals of each human body; Perform cycle segmentation, quality screening, cycle normalization, and cycle extension on the preprocessed pulse signals of each human body to obtain the pulse signals of each human body after cycle extension; Extract the first set of frequency-domain features from the pulse signals of each human body after cycle extension through Fourier transform, extract the second set of frequency-domain features through power spectrum transform, extract the third set of frequency-domain features through cepstrum transform, and extract the fourth set of frequency-domain features through fractional Fourier transform; Construct an initial feature matrix based on the frequency-domain features of all human body pulse signals, perform dimensionality reduction on the initial feature matrix multiple times to obtain the feature matrices after each dimensionality reduction, obtain the evaluation scores of the frequency-domain eigenvalue matrices corresponding to the feature matrices after each dimensionality reduction based on the SVM model, and screen the feature matrices after dimensionality reduction based on the evaluation scores to form the first data set; Calculate the KL divergence between each feature matrix in the first data set and the initial feature matrix, and select the frequency-domain features corresponding to the feature matrix when the KL divergence is the smallest as the final frequency-domain features of each human body pulse signal.

2. The method for extracting the frequency-domain characteristics of the pulse wave signal according to claim 1, wherein Perform eigen-decomposition on the initial feature matrix to obtain eigenvectors, subtract the eigenvectors corresponding to the smallest eigenvalue from the eigenvectors in turn to obtain the eigenvectors after each dimensionality reduction, and construct the feature matrices after each dimensionality reduction based on the eigenvectors after each dimensionality reduction.

3. The method for extracting the frequency domain characteristics of the pulse wave signal according to claim 2, wherein, The performing eigen-decomposition on the initial feature matrix to obtain eigenvectors includes: Perform de-centralization on the frequency-domain features in the initial feature matrix to obtain a de-centralized initial feature matrix; Obtain its covariance matrix based on the de-centralized initial feature matrix; Perform eigen-decomposition on the covariance matrix to obtain eigenvectors, and sort the features in the eigenvectors in descending order of eigenvalues.

4. The method for extracting the frequency domain characteristics of the pulse wave signal according to claim 1, wherein, Obtain the KL divergence between the feature matrix in the first data set and the initial feature matrix through the following method: Where, N1 is the initial feature matrix, N2 is the feature matrix in the first data set, k is the dimension of the eigenvector corresponding to the feature matrix in the first data set, μ1 is the mean corresponding to N1, μ2 is the mean corresponding to N2, ∑1 is the variance corresponding to N1, and ∑2 is the variance corresponding to N2.

5. The method for extracting the frequency-domain characteristics of the pulse wave signal according to claim 1, characterized in that, Extract the resonance frequency, resonance amplitude, amplitude difference, frequency difference, and amplitude ratio as the first set of frequency-domain features of each human body pulse signal through Fourier transform on the pulse signals of each human body after cycle extension; The resonance frequency is the frequency corresponding to the Fourier transform harmonic, the resonance amplitude is the amplitude corresponding to the Fourier transform harmonic, the amplitude difference is the amplitude difference between adjacent harmonics, the frequency difference is the frequency difference between adjacent harmonics, and the amplitude ratio is the ratio of the current harmonic amplitude to the fundamental harmonic amplitude.

6. The method for extracting the frequency domain characteristics of the pulse wave signal according to claim 1, wherein, Extract the energy ratio, peak area, and peak ratio as the second set of frequency-domain features of each human body pulse signal through power spectrum transform on the pulse signals of each human body after cycle extension; Divide the pulse signal into several segments of equal length, calculate the periodogram of each segment of the pulse signal, and calculate the arithmetic mean of the periodograms of all segments of the pulse signal to obtain the power spectrum.

7. The method for extracting the frequency domain characteristics of the pulse wave signal according to claim 1, wherein For the pulse signals after the human body cycle extension, the cepstrum peak amplitude is extracted through cepstrum transformation as the third group of frequency domain features of each human body pulse signal; The cepstrum peak amplitude is the amplitude values corresponding to the 10 peaks of the cepstrum of the pulse signal.

8. The method for extracting the frequency-domain characteristics of the pulse wave signal according to claim 1, wherein For the pulse signals after the human body cycle extension, the lowest point value, the highest point value, the difference in abscissa between the highest point and the lowest point, the difference between adjacent highest points, and the difference between adjacent lowest points of each fractional domain spectrum are extracted through fractional Fourier transform as the fourth group of frequency domain features of each human body pulse signal.

9. The method for extracting the frequency-domain characteristics of the pulse wave signal according to claim 1, wherein The preprocessing of the pulse signal to obtain the preprocessed pulse signal for each human body includes: Setting a noise threshold, determining the wavelet basis function and the decomposition level, performing wavelet decomposition on the pulse signal to obtain multiple wavelet components, denoising the wavelet components according to the noise threshold, and then reconstructing the signal using wavelet coefficients to obtain the denoised pulse signal; Extracting the pacing points from the denoised pulse signal by searching for local extrema, performing cubic spline interpolation on all pacing points to fit the baseline noise, and removing the baseline noise from the denoised pulse signal to obtain the preprocessed pulse signal.

10. The method for extracting the frequency domain characteristics of the pulse wave signal according to claim 1, characterized in that, The cycle segmentation, quality screening, cycle normalization, and cycle extension of the preprocessed pulse signal for each human body to obtain the pulse signal after the human body cycle extension include: Performing the following steps on the preprocessed pulse signal for each human body: Step S21: Based on the pacing points, perform cycle segmentation on the preprocessed pulse signal; Step S22: Calculate the peak-valley slope, peak-valley distance, and peak-valley difference of the pulse signals in each segmented cycle. Based on the peak-valley slopes and peak-valley distances of all segmented pulse signals, obtain the arithmetic mean of the peak-valley slopes and the arithmetic mean of the peak-valley distances respectively. Remove the periodic pulse signals in each cycle whose peak-valley slope deviates from the peak-valley slope average by more than the threshold or whose peak-valley distance deviates from the peak-valley distance average by more than the threshold to obtain the remaining periodic pulse signals; Step S23: Obtain the average cycle length of each remaining periodic pulse signal; perform extension or compression on each remaining periodic pulse signal for cycle normalization; Step S24: Perform cycle extension on the cycle-normalized pulse signal to obtain the pulse signal after cycle extension.

Images