Waveform reconstruction and identification method of pulse wave

Abstract

The invention provides a pulse wave waveform reconstruction and recognition method. The method comprises the steps that Fourier expansion is conducted on the basis of a multi-cycle pulse wave in a first pulse wave sequence, a threshold value is set, harmonic waves with the preset number of times are extracted, inverse Fourier transformation is conducted on harmonic wave components according to the cycle of the pulse wave, and a single-cycle pulse wave is reconstructed; the harmonic amplitude value of the preset number of times of the reconstructed single-cycle pulse wave is used as an overall morphological characteristic parameter; extracting a wavelet coefficient of 0-15Hz as a local detail characteristic parameter by using discrete wavelet transform; taking the overall form characteristic parameters and the local detail characteristic parameters as input of a three-classification model, when the three-classification model judges pulse waves, obtaining the maximum value of the pulse waves, storing the maximum value in a preset array, obtaining the position of the maximum value in the preset array, and obtaining the position of the pulse waves in the preset array; the pulse wave of the corresponding column of the preset array is extracted based on the position to serve as the optimal single-cycle pulse wave, the reconstructed single-cycle pulse wave is judged, and effective recognition of the pulse wave sequence is achieved.

Description

Technical Field

[0001] This application relates to the field of pulse wave signal recognition technology, specifically a method for pulse wave waveform reconstruction and recognition. Background Technology

[0002] Pulse signals contain rich physiological and pathological information about the cardiovascular system and are important indicators reflecting heart and blood vessel function, playing a crucial role in blood pressure monitoring. In recent years, researchers in traditional Chinese medicine diagnosis and hemodynamics have successfully utilized the information contained in single-cycle pulse waves to assess patients' cardiovascular health status, including heart rhythm, stroke volume, and vascular elasticity. Modern blood pressure measurement technology is continuously developing in three directions: non-invasive, continuous, and accurate. Because pulse waves are relatively easy to obtain and meet the requirements of being non-invasive and continuous, researchers have conducted extensive research on blood pressure measurement methods based on pulse signals, hoping to improve accuracy. Blood pressure measurement methods based on pulse waves are divided into two categories: those based on conduction time and those based on morphological features, both of which require feature analysis using single-cycle pulse wave signals. Therefore, it is crucial to segment the extracted full-length pulse signal into multiple complete single-cycle signals.

[0003] Typically, pulse wave signals are quasi-periodic. Therefore, before performing data mining on single-cycle pulse waves, it is necessary to segment the entire pulse wave sequence into multiple single cycles. Although pulse signals are relatively easy to acquire compared to other physiological signals, they are weak physiological signals from the body surface and are easily affected by the acquisition environment, human movement, the equipment itself, and inherent defects in the current cycle segmentation method. They often contain a large amount of noise and interference, frequently exhibiting abnormal cycles. These noises mainly originate from two aspects: human physiology and acquisition technology. Physiologically, noise primarily includes baseline drift, electromyographic interference, and motion artifacts; technically, noise mainly includes power frequency interference, interference from the skin surface and electrode pads, and adjacent channel interference. Baseline drift, electromyographic interference, power frequency interference, and interference from the skin surface and electrode pads can be filtered out using low-frequency, high-frequency, and wavelet transform methods. Therefore, this application primarily distinguishes between single-cycle pulse wave signals, adjacent channel interference, and motion artifacts. The pulsation of the artery in the human wrist causes the sensor at the pulsation point to move up and down. This movement forces adjacent sensors to move up and down as well, rapidly altering the contact force between these sensors and the skin, thus generating interference signals. These interference signals are 180 degrees out of phase with the pulse wave, and their rising edge is slightly flatter while their falling edge is steeper compared to the pulse wave. Currently, pulse wave signal recognition methods are mainly divided into traditional detection and machine learning-based detection. Li Qiao et al. proposed using the DTW (Dynamic Time Warping) algorithm to remove interference segments from the pulse wave. The main idea is to calculate the similarity between two time series by extending and shortening them, calculate the DTW distance, and set a threshold to determine whether a segment is interference. The DTW method can effectively remove irregular noise. The greater the similarity between the pulse wave and the target data, the higher the matching degree. The DTW algorithm is affected by the difference in pulse wave amplitude; the smaller the difference, the higher the matching degree, and the greater the difference, the lower the matching degree. When the interference signal and the pulse wave have certain similar contour characteristics, they cannot be distinguished. Wang Haitao et al. applied SAX (Symptomization of Time Series) to detect abnormal data in pulse waves, achieving good results. The main idea is to first reduce the dimensionality of the time series using a piecewise linear representation algorithm based on piecewise aggregation approximation. Then, based on the characteristics of the normal distribution, a symbolization threshold is set to transform the dimensionality-reduced data sequence into a corresponding symbol sequence. Finally, a method of matching similar pulse signals is used to distinguish the interference segment from the normal segment of the signal. The threshold used in this method is an empirical threshold; the accuracy of the recognition is related to the sliding window length. When the sliding window length is short, the pulse wave time period in a single cycle is longer, resulting in higher resolution. For signals with less obvious changes in period and morphology between periods, the recognition rate is higher. When the sliding window length is long, the resolution of the pulse wave in a single cycle decreases, and the matching degree between pulse waves between periods increases. However, pulse waves with significant changes in period between periods are prone to misjudgment.For noisy signals, due to their strong repetition, they may be misjudged as regular pulse wave signals. Guo Wei et al. proposed an algorithm based on multi-feature parameter comprehensive analysis of pulse signal distortion, which achieved good results in pulse wave anomaly identification. However, this method requires the position of the R-wave peak of the ECG signal as an identifier for segmenting the pulse wave during the process of judging pulse wave waveform distortion, which cannot be achieved in the analysis of a single pulse signal. Huang.L. analyzed the generation mechanism of interference signals in MEMS sensor array acquisition systems by establishing a simple mechanical structure model. Based on the morphological differences between pulse wave signals, noise signals, and interference signals, a convolutional neural network was used to classify the three. Although the above methods can identify multiple continuous pulse waves, when there are jitters or other interferences in multiple continuous pulse waves, the entire pulse wave is easily judged as noise or interference signals. Summary of the Invention

[0004] To overcome the aforementioned problems in the prior art, this application provides a method for waveform reconstruction and recognition of pulse waves, employing the following technical solution:

[0005] In a first aspect, this application provides a method for waveform reconstruction and recognition of pulse waves, including:

[0006] Fourier expansion is performed on a multi-cycle pulse wave from the first pulse wave sequence. By setting a threshold, a preset number of harmonics are extracted. The harmonic components are then subjected to inverse Fourier transform according to the period of the pulse wave to reconstruct the single-cycle pulse wave.

[0007] The harmonic amplitude value of the preset number of reconstructed single-cycle pulse wave is used as the overall morphological characteristic parameter of single-cycle pulse wave, interference signal and noise.

[0008] Using discrete wavelet transform, wavelet coefficients from 0 to 15 Hz in the reconstructed single-cycle pulse wave are extracted as local detail feature parameters of the single-cycle pulse wave, interference signal, and noise.

[0009] The overall morphological feature parameters and local detail feature parameters are used as inputs to a three-classification model to determine whether the reconstructed single-cycle pulse wave is a pulse wave. When the output of the three-classification model is 0, it is determined to be a pulse wave. The maximum value of the pulse wave is obtained and stored in a preset array. The position of the maximum value in the preset array is obtained, and the pulse wave of the corresponding column of the preset array is extracted based on the position as the optimal single-cycle pulse wave.

[0010] Further, it includes: inputting multiple consecutive pulse wave sequences, retaining pulse wave sequences with frequencies above a preset frequency through a high-pass filter, removing power frequency noise using a notch filter, and obtaining a first pulse wave sequence.

[0011] Furthermore, a Fourier expansion is performed based on a multi-cycle pulse wave from the first pulse wave sequence, including:

[0012] (2)

[0013] Formula (2) is the Fourier series expansion of the pulse wave, used to decompose the time-domain pulse wave signal into frequency-domain components. Its specific form is:

[0014] (3)

[0015] in This represents the mean of the filtered pulse wave sequence, reflecting the baseline level of the pulse wave. It is the first The sinusoidal component coefficients of the second harmonic reflect the... The contribution intensity of the subsine harmonic component to the original pulse wave; It is the first The cosine component coefficient of the second harmonic reflects the first harmonic's cosine component coefficient. The contribution intensity of the second cosine harmonic component to the original pulse wave.

[0016] Furthermore, by performing an inverse Fourier transform on the harmonic components based on the period of the pulse wave, a single-period pulse wave is reconstructed, including pulse wave reconstruction using Equation 2, resulting in:

[0017] (6)

[0018] By obtaining By analyzing the nearest peaks and troughs of the two peak points, a single-cycle pulse wave is obtained, where yy represents the time-domain signal amplitude of the reconstructed single-cycle pulse wave. These are the DC component coefficients of the Fourier series. This represents the total number of sampling points in the filtered pulse wave sequence. The harmonic order ranges from , This represents the harmonic sampling point number, with a value range of [value range missing]. , The harmonic error constant is... , The first The sine and cosine component coefficients of the subharmonic.

[0019] Furthermore, after reconstructing the single-cycle pulse wave, it also includes:

[0020] The single-cycle pulse wave sequence segment was reconstructed by resampling using cubic spline interpolation. The length of the reconstructed single-cycle pulse wave sequence segment was unified to a standard length. The amplitude of the reconstructed single-cycle pulse wave signal was standardized using a zero-mean equation.

[0021] Furthermore, the harmonic amplitude values ​​of the reconstructed single-cycle pulse wave at a preset number of times are used as parameters representing the overall morphological characteristics of the single-cycle pulse wave, interference signal, and noise, including:

[0022] Interpolation is performed on the reconstructed single-cycle pulse wave signal. The number of single-cycle sampling points for the pulse wave, interference signal, and noise signal is set to N. Then, when extracting the spectrum of the single-cycle signal, a 10-fold period extension is performed to obtain the spectrum of the single-cycle signal. The macroscopic morphological differences of the three types of signals are transformed into quantifiable harmonic amplitude parameters, that is, the overall morphological characteristic parameters of the reconstructed single-cycle pulse wave are obtained.

[0023] Furthermore, the overall morphological characteristic parameters of the reconstructed single-cycle pulse wave are obtained, including:

[0024]

[0025] In the formula, Represents the Fourier transform coefficients. Indicates the first The amplitude at each point The frequency domain characteristic parameters are as follows:

[0026] Amplitude:

[0027] frequency:

[0028] To remove absolute errors in the data, the amplitude was standardized, as shown in the formula below:

[0029]

[0030] in, Represents the Fourier transform coefficients. This represents the average amplitude of the pulse signal. Represents the relative amplitude, which is a standardized dimensionless amplitude parameter. .

[0031] Furthermore, the overall morphological feature parameters and local detail feature parameters are used as inputs to the three-classification model, which includes classifier A and classifier B. Classifier A distinguishes pulse waves from interference signals + noise, while classifier B distinguishes interference signals from noise.

[0032] Furthermore, based on classifier A, distinguishing pulse waves from interference signals + noise also includes:

[0033] The input features are normalized, and the normalization result is subjected to principal component transformation to obtain the principal component matrix. Then, the dimensionality reduction method is searched. ,satisfy: Based on RBF kernel SVM, its classification rules are as follows: ,in For the front Principal components, For tags, To improve classification accuracy, The decision threshold is set to 0.8. The pulse wave. This is interference plus noise.

[0034] Furthermore, based on classifier B, interference signals are distinguished from noise, including:

[0035] Receive the output of classifier A and search for dimensionality reduction. ,satisfy: ,in For tags, Indicates interference signal. To represent noise, an SVM based on the RBF kernel is used for classification, with the following classification rules: ,in The decision threshold is set to 0.8.

[0036] Final classification output

[0037] in Indicates pulse wave, Indicates interference. Indicates noise.

[0038] This application has the following beneficial effects:

[0039] 1. This invention reconstructs a single-cycle pulse wave by performing a Fourier expansion on a multi-cycle pulse wave from a first pulse wave sequence, extracting harmonics of a preset number by setting a threshold, and performing an inverse Fourier transform on the harmonic components according to the pulse wave's period. The amplitude values ​​of the preset number of harmonics in the reconstructed single-cycle pulse wave are used as overall morphological feature parameters. Discrete wavelet transform is used to extract wavelet coefficients from 0 to 15 Hz as local detail feature parameters. The overall morphological feature parameters and local detail feature parameters are used as inputs to a three-classification model. When the three-classification model identifies the pulse wave as a pulse wave, the maximum value of the pulse wave is obtained and stored in a preset array. The position of the maximum value in the preset array is obtained, and the pulse wave in the corresponding column of the preset array is extracted based on the position as the optimal single-cycle pulse wave. This invention achieves effective identification of pulse wave sequences by judging the reconstructed single-cycle pulse wave, and can provide pulse wave sequences for modeling cardiovascular diseases and traditional Chinese medicine pulse diagnosis.

[0040] 2. In the single-cycle reconstruction stage, this invention, based on Fourier series expansion, accurately calculates the fundamental frequency and harmonic components and sets a reasonable error range, which can adapt to the non-strict periodicity of the pulse wave and achieve accurate reconstruction of the single-cycle pulse wave, providing high-quality data for subsequent analysis.

[0041] 3. In the construction of the three-classification model, this invention first standardizes the signal through cubic spline interpolation and zero-mean normalization, then extracts the harmonic amplitude value (overall shape) and wavelet coefficients (local details), and combines the OVA-SVM classifier and PCA dimensionality reduction optimization to construct the TRI-OVA-SVM three-classification model, which can accurately distinguish between pulse waves, interference signals and noise, avoiding the misjudgment problem of traditional methods. Finally, it can screen out the best single-cycle pulse wave from the multi-pressure pulse wave sequence, providing reliable pulse wave sequence support for TCM pulse modeling. Attached Figure Description

[0042] Figure 1 This is a flowchart of the pulse wave waveform reconstruction and recognition method according to an embodiment of this application;

[0043] Figure 2 The FFT transform diagrams of the pulse wave signal, noise signal, and interference signal are shown.

[0044] Figure 3 This is a three-layer decomposition diagram of pulse wave, noise, and interference signal in an embodiment of this application. Detailed Implementation

[0045] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used in the description of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application; the terms "comprising" and "having," and any variations thereof, in the description, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the description, claims, or accompanying drawings of this application are used to distinguish different objects, not to describe a particular order.

[0046] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0047] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0048] It should be noted that this invention can provide pulse wave sequences for modeling cardiovascular diseases and traditional Chinese medicine pulse patterns.

[0049] Please refer to Figure 1 The following is a flowchart of a pulse wave waveform reconstruction and recognition method provided in an embodiment of this application. The specific steps are as follows:

[0050] Step S1: Perform Fourier expansion on a multi-cycle pulse wave from the first pulse wave sequence, extract harmonics of a preset number of times by setting a threshold, and perform inverse Fourier transform on the harmonic components according to the period of the pulse wave to reconstruct a single-cycle pulse wave.

[0051] In this embodiment of the application, obtaining the first pulse wave sequence includes:

[0052] Input multiple consecutive pulse wave sequences, use a high-pass filter to retain pulse wave sequences with frequencies above a preset frequency, and use a notch filter to remove power frequency noise to obtain the first pulse wave sequence.

[0053] It should be noted that this application uses a combination of a high-pass filter and a notch filter for noise reduction, filtering out common noises such as baseline drift and power frequency interference. Compared with traditional single filtering methods, the purification effect is better. The first pulse wave sequence is the pulse wave sequence after filtering.

[0054] It should be noted that this application uses a high-pass filter to retain pulse wave sequences with frequencies above a preset frequency, where the preset frequency is 0.5Hz.

[0055] It should be noted that the preset number of harmonics extracted in this embodiment is 12.

[0056] In this embodiment of the application, it is assumed that the filtered first pulse wave sequence can be represented as:

[0057] (1)

[0058] Where y(n) represents the amplitude of the pulse wave signal corresponding to the nth sampling point after noise reduction. This value reflects the intensity of the pulse wave at that moment and is a discretized time-domain signal data; n represents the sampling point number, which takes the value of a positive integer 1, 2, 3, ... N, and its value corresponds to the time sequence of signal acquisition; N represents the total number of sampling points in the pulse wave sequence, and its value is determined by the sampling frequency (f s The signal acquisition duration and the signal acquisition time together determine (N=f) s ×Collection duration).

[0059] Subsequently, the series expansion was obtained using Fourier series transform:

[0060] (2)

[0061] Formula (2) is the Fourier series expansion of the pulse wave, used to decompose the time-domain pulse wave signal into frequency-domain components. Its specific form is:

[0062] (3)

[0063] in This represents the mean of the filtered pulse wave sequence, reflecting the baseline level of the pulse wave. It is the first The sinusoidal component coefficients of the second harmonic reflect the... The contribution intensity of the subsine harmonic component to the original pulse wave. It is the first The cosine component coefficient of the second harmonic reflects the first harmonic's cosine component coefficient. The contribution intensity of the second cosine harmonic component to the original pulse wave It is a gradient-based integral function.

[0064] Assuming the pulse rate range is Let the range of the first harmonic frequency of the pulse wave be... ,but:

[0065] (4)

[0066] Seeking in The peak value of the Fourier transform of the pulse wave within the range is P1, and its corresponding frequency is F1. Find the value within... The peak value of the Fourier transform of the pulse wave within the range is P2, and its corresponding frequency is F2. If P2 > If P1 is true, then the first harmonic of the pulse wave is P=P2, F=F2; otherwise, P=P1, F=F1. Therefore, the frequency of the Mth harmonic of the pulse wave is M×F, where M×F < P1. f s .

[0067] The number of sampling points S for the Mth harmonic frequency of a single-cycle pulse wave M :

[0068] (5)

[0069] Since the pulse wave is a non-periodic signal, the multiple relationships of its Fourier series transform harmonic components have a certain error. Assume the error range of the Mth harmonic component is [S]. M -K,S M +K], where K<(S) M -SM-1 ) / 2. Based on the above, reconstructing the pulse wave from Formula 2 yields:

[0070] (6)

[0071] The result can be obtained using formula (6). A single-cycle pulse wave is obtained by taking the nearest peak and trough points between the two peak points. This represents the time-domain signal amplitude of the reconstructed single-cycle pulse wave. The DC component coefficient of the Fourier series (reflecting the baseline level of the pulse wave, calculated by formula 3). This represents the total number of sampling points in the filtered pulse wave sequence (determined by the sampling frequency and acquisition duration). The harmonic order ranges from , This represents the harmonic sampling point number, with a value range of [value range missing]. , Harmonic error constant ( ), , The first The sine and cosine component coefficients of the subharmonics are calculated by formula (3).

[0072] It should be noted that, in the single-cycle reconstruction stage of this application, based on Fourier series expansion, by accurately calculating the fundamental frequency and harmonic components and setting a reasonable error range, it can adapt to the non-strict periodicity characteristics of the pulse wave, realize the accurate reconstruction of the single-cycle pulse wave, and provide high-quality data for subsequent analysis.

[0073] Step S2: The harmonic amplitude value of the preset number of reconstructed single-cycle pulse wave is used as the overall morphological characteristic parameter of single-cycle pulse wave, interference signal and noise.

[0074] It should be noted that the amplitude and length of the single-cycle pulse wave, interference signal, and noise are different from each other. In order to reduce the impact on the classification process and make the classifier consider the waveform morphology, the data was standardized. For the signal length, cubic spline interpolation was used to resample the segments and unify their lengths to the standard length, as shown in formula (7).

[0075] (7)

[0076] In formula (7), It is the amplitude of the continuous signal after interpolation. To standardize the sampling point locations, The interpolation segment number. It is the original interpolation node position. , , , These are the coefficients of the constant term in the interpolation polynomial.

[0077] The instrument sampling frequency is 225Hz, and the normal cardiac cycle is 0.6-1s. Therefore, in order to ensure that a complete pulse wave cycle is included, the upsampling uniform signal length is 256 sampling points, approximately 1.14s. For the signal amplitude, a zero-mean equation is used, as shown in formula (8):

[0078] (8)

[0079] In formula (8), These are standardized data points. The original signal amplitude, It is the mean of the dataset. It is the standard deviation. It is the mean square error. It is the total number of data points. Indicates the first Data points.

[0080] It should be noted that the pulse wave is a weak physiological signal from the body surface. During data acquisition, it is susceptible to interference from various factors, resulting in abnormal cycles. The main sources of interference are adjacent channel interference and noise. The pulsation of the arteries in the human wrist causes the sensor at the point of pulsation to move up and down. This movement forces nearby sensors to move up and down as well, rapidly altering the contact force between the nearby sensors and the skin, thus generating interference signals. These interference signals are 180 degrees out of phase with the pulse wave, and compared to the pulse wave, their rising edge is slightly flatter, while their falling edge is steeper. Noise interference is mainly caused by physical movements and mental stress. It has the following two characteristics:

[0081] ① High amplitude: Body movements, especially finger movements, can easily interfere with the waveform of the radial artery pulse wave at the wrist. The amplitude (peak-to-trough ratio) of this interference signal is generally more than twice that of the pulse wave. For pulse waves in the same pressure range, the amplitude of the pulse wave vibration fluctuates within a specified range, which is ±0.5 times.

[0082] ② Irregularity: Motion interference can disrupt the regularity of the pulse wave, making the data segment unable to meet the basic pulse characteristics.

[0083] In general, interference signals and noise clutter differ from pulse waves in their time-domain morphology. The above analysis shows that the difference between pulse waves and interference / noise signals lies in the overall morphology of the pulse wave and the differences in its detailed features.

[0084] Therefore, this application uses Fourier transform to extract the amplitude value of the 12th harmonic as a parameter for the overall morphological characteristics of single-cycle pulse wave, interference signal and noise.

[0085] The difference between pulse waves and interference / noise signals lies in the differences in the overall shape and detailed features of the pulse wave. Therefore, this application uses Fourier transform to extract the amplitude values ​​of the 12th harmonic as parameters for the overall shape characteristics of single-cycle pulse waves, interference signals, and noise. First, the single-cycle signal is interpolated, and the number of single-cycle sampling points for the pulse wave, interference signal, and noise signal is set to 256. Then, when extracting the spectrum of the single-cycle signal, a 10-fold periodic extension is performed to obtain the spectrum of the single-cycle signal. The resulting frequency domain characteristic parameter diagram is shown below. Figure 2 As shown in the figure, the horizontal axis represents frequency and the vertical axis represents the relative amplitude parameter.

[0086] The basic principle of Fourier transform is to match the pulse wave with sine and cosine functions to obtain the coefficients of the sine and cosine functions at different periods, thereby achieving frequency domain decomposition of the pulse wave. Since sine and cosine functions are periodic signals, the overall characteristics of periodic signals can be obtained based on Fourier transform. This application first interpolates the single-cycle signal, setting the number of single-cycle sampling points for the pulse wave, interference signal, and noise signal to N. Then, when extracting the spectrum of the single-cycle signal, a 10-fold periodic extension is performed to obtain the spectrum of the single-cycle signal.

[0087] The formula for the discrete FFT transform is shown in equation (10):

[0088] (10)

[0089] In the formula, Represents the Fourier transform coefficients. Indicates the first The amplitude at each point The frequency domain characteristic parameters are as follows:

[0090] Amplitude:

[0091] frequency:

[0092] To remove absolute errors in the data, this application standardizes the amplitude, as shown in formula (11):

[0093] (11)

[0094] in, Represents the Fourier transform coefficients. This represents the average amplitude of the pulse signal. Represents the relative amplitude, which is a standardized dimensionless amplitude parameter. .

[0095] It should be noted that the embodiments of this application extract the overall morphological feature parameters through Fourier transform. The overall morphological feature parameters can reflect the global frequency distribution law of the signal, and transform the macroscopic morphological differences of the three types of signals into quantifiable harmonic amplitude parameters, so that the three-classification model can distinguish between "periodic pulse waves" and "irregular noise", providing macroscopic basic data for the three-classification model.

[0096] Step S3: Using discrete wavelet transform, extract the wavelet coefficients from 0 to 15 Hz in the reconstructed single-cycle pulse wave as local detail feature parameters of the single-cycle pulse wave, interference signal and noise.

[0097] It should be noted that the total number of feature parameters in this embodiment is W. This embodiment extracts features through discrete wavelet transform, focusing on the microscopic fluctuations of key spectra, capturing the local specificity of signals at peaks, valleys, and rises and falls, supplementing the microscopic details, further distinguishing interference signals from noise signals, and ultimately ensuring that the three types of signals can be accurately identified in the three-class classification model for subsequent optimal pulse wave sequence selection.

[0098] The db6 wavelet is selected as the wavelet basis. Since the sampling frequency of the pulse wave is 225Hz, and the main energy of the signal is concentrated within 15Hz, according to the Nyquist sampling theorem in the time domain, the wavelet decomposition has 3 levels. Figure 3 The figure shows the three-level decomposition effect of pulse wave, noise, and interference signal. As can be seen from the figure, the characteristics of pulse wave, noise, and interference signal are completely preserved in the range of 0~14.05Hz, and the total number of characteristic parameters is W=54.

[0099] Based on steps S2 and S3, it should be noted that this application standardizes the reconstructed single-cycle pulse wave by using cubic spline interpolation and zero-mean normalization, then extracts the harmonic amplitude value (overall shape) and wavelet coefficients (local details), and combines the OVA-SVM classifier and PCA dimensionality reduction optimization to construct the TRI-OVA-SVM three-classification model, which can accurately distinguish between pulse waves, interference signals and noise, avoid the misjudgment problem of traditional methods, and finally select the best single-cycle pulse wave from the multi-pressure pulse wave sequence, providing reliable pulse wave sequence support for TCM pulse modeling.

[0100] Step S4: Use the overall morphological feature parameters and local detail feature parameters as input to the three-classification model to determine whether the reconstructed single-cycle pulse wave is a pulse wave. When the output of the three-classification model is 0, it is determined to be a pulse wave. Obtain the maximum value of the pulse wave and store the maximum value in a preset array. Obtain the position of the maximum value in the preset array and extract the pulse wave of the corresponding column of the preset array based on the position as the best single-cycle pulse wave.

[0101] In this embodiment, the three-classification model consists of two classifiers, A and B. Classifier A distinguishes between pulse waves and interference signals + noise; classifier B distinguishes between interference signals and noise. During training, classifier A uses PCA to remove parameter redundancy, and the binary classification model classifies pulse wave signals and interference + noise signals. Classifier B uses PCA to remove parameter redundancy and classifies interference signals and noise signals using a binary classification model. PCA is a principal component classification method.

[0102] It should be noted that for classifier A, the dimensionality reduction parameter of PCA is increased from 1 to W, and the maximum accuracy is calculated as the dimensionality reduction dimension of PCA, using accuracy as the evaluation parameter. For classifier B, the dimensionality reduction parameter of PCA is increased from 1 to W, and the maximum accuracy is calculated as the dimensionality reduction dimension of PCA, using accuracy as the evaluation parameter. This constructs an optimal three-class classification model for pulse wave, interference signal, and noise, assuming the model is named TRI-OVA-SVM. The labels for pulse wave, interference signal, and noise are 0, 1, and 2, respectively.

[0103] In this embodiment of the application, classifier A is used to distinguish pulse waves from interference signals + noise, including:

[0104] The input features are normalized, and the normalization result is subjected to principal component transformation to obtain the principal component matrix. Then, the dimensionality reduction method is searched. ,satisfy: Based on RBF kernel SVM, its classification rules are as follows: ,in For the front Principal components, For tags, To improve classification accuracy, The decision threshold is set to 0.8. The pulse wave. This is interference plus noise.

[0105] For example, suppose the input features are , , among them For the sample size, , The mean, For the standard deviation, see formula (8).

[0106] right Perform PCA transformation to obtain the principal component matrix. Thus, the search is reduced in dimensionality. ,satisfy: ,in: For the front Principal components, For tags ( Indicates pulse wave, (Indicates interference + noise) For classification accuracy.

[0107] Based on RBF kernel SVM, its classification rules are as follows: ,in The decision threshold is set to 0.8. The pulse wave. This is interference plus noise.

[0108] In this embodiment of the application, classifier B is used to distinguish between interference signals and noise, including:

[0109] Receive the output of classifier A and search for dimensionality reduction. ,satisfy: ,in For tags, Indicates interference signal. To represent noise, an SVM based on the RBF kernel is used for classification, with the following classification rules: ,in The decision threshold is set to 0.8.

[0110] Final classification output

[0111] in Indicates pulse wave, Indicates interference. Indicates noise.

[0112] It should be noted that multiple compressions are typically used during wrist pulse wave extraction. Assuming there are J compression segments, each compression segment contains a pulse wave sequence. By reconstructing the pulse waves, a pulse wave sequence can be obtained. A three-class classification model is used to determine whether a pulse wave sequence is a pulse wave. If the output of the three-class classification model is 0, the pulse wave sequence is determined to be a pulse wave, and the maximum value of the pulse wave is obtained and stored in MAX[j]. If the output of the three-class classification model is 1 or 2, the pulse wave sequence is determined to be a non-pulse wave, and MAX[j] = 0. When it is a pulse wave, the position of the maximum value in MAX is calculated, and the pulse wave of the corresponding column of the sequence MAX is extracted based on the position to obtain the optimal single-cycle pulse wave.

[0113] It should be noted that in the formula It is the single-cycle pulse wave candidate sequence obtained after noise reduction and reconstruction of the j-th intensified segment, where n is the sampling point number. yes The classification results after inputting the TRI-OVA-SVM three-class classification model are output as three-class labels: 0 (pulse wave), 1 (interference signal), and 2 (noise). MAX[j] is the element that stores the maximum effective pulse wave value of the j-th compression segment. If it is not a pulse wave, it is set to 0.

[0114] This application utilizes a high-pass filter to retain pulse wave sequences with frequencies above 0.5Hz, employs a notch filter to remove power frequency noise, performs a Fourier series expansion on the filtered pulse wave sequence, and reconstructs a single-cycle pulse wave by obtaining the corresponding harmonic components through setting the fundamental frequency component of the pulse wave. A machine learning method is then used to establish a recognition model for the single-cycle pulse wave, and the reconstructed pulse wave is evaluated to achieve pulse wave sequence recognition. This application can provide pulse wave sequences for modeling cardiovascular diseases and traditional Chinese medicine pulse diagnosis.

[0115] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A method for waveform reconstruction and recognition of pulse waves, characterized in that, include: Fourier expansion is performed on a multi-cycle pulse wave from the first pulse wave sequence. By setting a threshold, a preset number of harmonics are extracted. The harmonic components are then subjected to inverse Fourier transform according to the period of the pulse wave to reconstruct the single-cycle pulse wave. The harmonic amplitude value of the preset number of reconstructed single-cycle pulse wave is used as the overall morphological characteristic parameter of single-cycle pulse wave, interference signal and noise. Using discrete wavelet transform, wavelet coefficients of 0-15Hz in the reconstructed single-cycle pulse wave are extracted as local detail feature parameters of the single-cycle pulse wave, interference signal and noise; The overall morphological feature parameters and local detail feature parameters are used as inputs to a three-classification model to determine whether the reconstructed single-cycle pulse wave is a pulse wave. When the output of the three-classification model is 0, it is determined to be a pulse wave. The maximum value of the pulse wave is obtained and stored in a preset array. The position of the maximum value in the preset array is obtained, and the pulse wave of the corresponding column of the preset array is extracted based on the position as the optimal single-cycle pulse wave.

2. The pulse wave waveform reconstruction and recognition method according to claim 1, characterized in that, Obtaining the first pulse wave sequence includes: inputting multiple consecutive pulse wave sequences, retaining pulse wave sequences with frequencies above a preset frequency using a high-pass filter, removing power frequency noise using a notch filter, and obtaining the first pulse wave sequence.

3. The pulse wave waveform reconstruction and recognition method according to claim 1, characterized in that, Fourier expansion based on a multi-cycle pulse wave from the first pulse wave sequence includes: （2） Formula (2) is the Fourier series expansion of the pulse wave, used to decompose the time-domain pulse wave signal into frequency-domain components. Its specific form is: （3） in This represents the mean of the filtered pulse wave sequence, reflecting the baseline level of the pulse wave. It is the first The sinusoidal component coefficients of the second harmonic reflect the... The contribution intensity of the subsine harmonic component to the original pulse wave; It is the first The cosine component coefficient of the second harmonic reflects the first harmonic's cosine component coefficient. The contribution intensity of the second cosine harmonic component to the original pulse wave.

4. The pulse wave waveform reconstruction and recognition method according to claim 2, characterized in that, Performing an inverse Fourier transform on the harmonic components based on the period of the pulse wave to reconstruct the single-period pulse wave includes reconstructing the pulse wave using Equation 2, resulting in: （6） By obtaining By analyzing the nearest peak and trough points of the two peak points, a single-cycle pulse wave is obtained, where yy represents the time-domain signal amplitude of the reconstructed single-cycle pulse wave. These are the DC component coefficients of the Fourier series. This represents the total number of sampling points in the filtered pulse wave sequence. The harmonic order ranges from , This represents the harmonic sampling point number, with a value range of [value range missing]. , The harmonic error constant is... , The first The sine and cosine component coefficients of the subharmonic.

5. The pulse wave waveform reconstruction and recognition method according to claim 1, characterized in that, After reconstructing the single-cycle pulse wave, it also includes: The single-cycle pulse wave sequence segment was reconstructed by resampling using cubic spline interpolation. The length of the reconstructed single-cycle pulse wave sequence segment was unified to a standard length. The amplitude of the reconstructed single-cycle pulse wave signal was standardized using a zero-mean equation.

6. The pulse wave waveform reconstruction and recognition method according to claim 3, characterized in that, The harmonic amplitude values ​​of the reconstructed single-cycle pulse wave at a preset number of times are used as parameters representing the overall morphological characteristics of the single-cycle pulse wave, interference signal, and noise, including: Interpolation is performed on the reconstructed single-cycle pulse wave signal. The number of single-cycle sampling points for the pulse wave, interference signal, and noise signal is set to N. Then, when extracting the spectrum of the single-cycle signal, a 10-fold period extension is performed to obtain the spectrum of the single-cycle signal. The macroscopic morphological differences of the three types of signals are transformed into quantifiable harmonic amplitude parameters, that is, the overall morphological characteristic parameters of the reconstructed single-cycle pulse wave are obtained.

7. The pulse wave waveform reconstruction and recognition method according to claim 4, characterized in that, Obtain the overall morphological feature parameters of the reconstructed single-cycle pulse wave, including: In the formula, Represents the Fourier transform coefficients. Indicates the first The amplitude at each point The frequency domain characteristic parameters are as follows: Amplitude: frequency: To remove absolute errors in the data, the amplitude was standardized, as shown in the formula below: in, Represents the Fourier transform coefficients. This represents the average amplitude of the pulse signal. Represents relative amplitude, which is a standardized dimensionless amplitude parameter. .

8. The pulse wave waveform reconstruction and recognition method according to claim 1, characterized in that, The overall morphological feature parameters and local detail feature parameters are used as inputs to the three-classification model, which includes classifier A and classifier B. Classifier A distinguishes pulse waves from interference signals + noise, while classifier B distinguishes interference signals from noise.

9. The pulse wave waveform reconstruction and recognition method according to claim 8, characterized in that, Based on classifier A, distinguishing pulse waves from interference signals + noise also includes: The input features are normalized, and the normalization result is subjected to principal component transformation to obtain the principal component matrix. Then, the dimensionality reduction method is searched. ,satisfy: Based on RBF kernel SVM, its classification rules are as follows: ,in For the front Principal components, For tags, To improve classification accuracy, The decision threshold is set to 0.

8. The pulse wave. This is interference plus noise.

10. The pulse wave waveform reconstruction and recognition method according to claim 8, characterized in that, Based on classifier B, interference signals are distinguished from noise, including: Receive the output of classifier A and search for dimensionality reduction. ,satisfy: ,in For tags, Indicates interference signal. To represent noise, an SVM based on the RBF kernel is used for classification, with the following classification rules: ,in The decision threshold is set to 0.

8. Final classification output in Indicates pulse wave, Indicates interference. Indicates noise.

Images