Method for extracting blood flow sound signal features of arteriovenous fistula and electronic device

Abstract

The application discloses a method and electronic equipment for feature extraction of arteriovenous fistula blood flow sound signals, and relates to the technical field of artificial intelligence medical information processing. The method comprises the following steps: positioning the starting and ending positions of each cycle of the blood flow sound signals; calculating the logarithmic mel spectrum of the blood flow sound signals and generating a local energy ratio spectrum graph; obtaining a deep feature matrix of the logarithmic mel spectrum and the local energy ratio spectrum graph, and calculating a cosine similarity matrix of the two deep feature matrices; and performing channel fusion on the two deep feature matrices with the same dimensions. The application can better obtain deep features of the blood flow sound signals representing arteriovenous fistula stenosis, so as to train a deep neural network with better performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information processing based on artificial intelligence, in particular to a method for extracting features of arteriovenous fistula blood flow sound signals and an electronic device. BACKGROUND

[0002] Arteriovenous fistula (AVF) is formed by anastomosing peripheral arteries and superficial veins through surgical operation, and is called the lifeline of hemodialysis patients. AVF stenosis or occlusion will affect the normal progress of hemodialysis, which is the primary problem that troubles doctors and patients.

[0003] The current gold standard for monitoring AVF is Doppler ultrasound examination, digital angiography technology, etc., but these detection methods have obvious shortcomings, such as expensive equipment, the need for professional operation, large size, etc. In routine clinical practice, the function of arteriovenous fistula is usually evaluated by physical examination, including palpation of arteriovenous fistula blood vessels and auscultation, but palpation and auscultation also test the experience and level of relevant personnel. With the continuous progress of artificial intelligence technology, the application of deep learning technology in the medical field has also become a research hotspot at present. By using a stethoscope to listen to the sound, combined with deep learning technology, more robust features can be extracted, which can reduce the error caused by the experience of relevant personnel in normal auscultation and reduce the work burden of medical personnel.

[0004] There are mainly two methods for identifying AVF stenosis through auscultation audio at present:

[0005] One is to extract the statistical features of the audio and establish quantitative indicators to distinguish between stenosis and non-stenosis. The quantitative indicators include frequency correlation coefficient, high-low frequency energy ratio, etc.

[0006] The other is to extract the time-frequency domain features of the audio and input them into a deep neural network for classification. Since different populations have certain differences in blood flow patterns even when they all have stenosis, the deep neural network will affect the classification results when capturing the difference information. Although many deep neural network models consider the local characteristics of audio features and the mutual relationship between features when training, they have certain advantages in extracting robust deep features, but they are also affected by difference information in the case of small data volume.

[0007] In view of the above problems, there is an urgent need for a technology for extracting features of arteriovenous fistula blood flow sound signals, so that the deep neural network model can learn deep features that can distinguish stenosis in the training process, thereby improving the anti-interference ability of the deep neural network model. SUMMARY

[0008] In view of the above defects or deficiencies in the prior art, the present application provides a method and an electronic device for feature extraction of an arteriovenous fistula blood flow sound signal, which utilizes a high-low frequency energy ratio at each frequency of each cardiac cycle in a time-frequency domain of the blood flow sound signal to guide deep feature extraction of a time-frequency feature, so that the model learns a deep feature capable of distinguishing stenosis in a biased manner in a training process, and the anti-interference capability of a deep neural network model is greatly improved.

[0009] In one aspect of the present application, a method for feature extraction of an arteriovenous fistula blood flow sound signal is provided, comprising:

[0010] Periodically segmenting and positioning the blood flow sound signal at an arteriovenous anastomosis port, locating the start and end positions of each cycle;

[0011] Calculating a log mel spectrum of the blood flow sound signal, mapping the cycle segmentation position to the log mel spectrum, calculating a highest and lowest amplitude ratio of each frequency within each cycle in the time-frequency domain of the log mel spectrum, and generating a local energy ratio spectrum graph;

[0012] Respectively performing convolutional transformation on the log mel spectrum and the local energy ratio spectrum graph to obtain two deep features with the same dimensions, and calculating a cosine similarity matrix of the two deep features with the same dimensions;

[0013] Fusing the two deep features with the same dimensions, and then inputting to a deep neural network to be trained, wherein a loss function of the deep neural network to be trained comprises two parts of a feature alignment loss and a classification loss, and the feature alignment loss is determined according to the cosine similarity matrix.

[0014] In another aspect of the present application, an electronic device is also provided, comprising one or more processors, a storage device for storing one or more computer programs, and when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the above-mentioned method for feature extraction of an arteriovenous fistula blood flow sound signal.

[0015] The method and the electronic device for feature extraction of an arteriovenous fistula blood flow sound signal provided by the present application extract features through a log mel spectrum and a local energy ratio spectrum graph, dynamically enhance the features, design effective audio features capable of reflecting characteristics of arteriovenous fistula stenosis, and utilize a feature alignment loss to enhance the similarity between the sound signal features and the log mel spectrum in the process of training a deep neural network model, thereby greatly improving the anti-interference capability of the deep neural network model. BRIEF DESCRIPTION OF DRAWINGS

[0016] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments made with reference to the following drawings:

[0017] Figure 1 is a flowchart of a method for extracting features of an arteriovenous fistula blood flow sound signal provided by an embodiment of the present application Figure 1 ;

[0018] Figure 2 is a flowchart of a method for extracting features of an arteriovenous fistula blood flow sound signal provided by an embodiment of the present application Figure 2 ;

[0019] Figure 3 is a waveform diagram of a cycle positioning result of a blood flow sound signal provided by an embodiment of the present application

[0020] Figure 4 is a spectrogram of a log-mel spectrum of a blood flow sound signal provided by an embodiment of the present application

[0021] Figure 5 is a structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of the present application.

[0023] The terms used in the embodiments of the present application are merely for the purpose of describing particular embodiments and are not intended to limit the present application. The singular forms "a", "an", and "the" used in the embodiments of the present application are also intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should be understood that although the terms first, second, third, etc. may be used in the embodiments of the present application to describe obtaining modules, these obtaining modules should not be limited to these terms. These terms are only used to distinguish the obtaining modules from each other.

[0025] Referring to Figure 1 , an embodiment of the present application provides a method for extracting features of an arteriovenous fistula blood flow sound signal, mainly including four stages of blood flow sound signal cycle segmentation, log-mel spectrum and local energy ratio spectrum diagram double-path feature extraction, feature dynamic enhancement, and loss function used for deep neural network training. The method can run in the form of a computer program on a special medical device, a PC, a smart terminal, or a server.

[0026] Referring toFigure 2 The method for extracting blood flow sound signal features of an arteriovenous fistula in this embodiment includes the following steps:

[0027] Step S101, the blood flow sound signal at the anastomotic stoma of the arteriovenous fistula is periodically segmented and positioned to locate the start and end positions of each period.

[0028] The blood flow turbulence at the anastomotic stoma of the arteriovenous fistula is the most significant, and the abnormal sound such as high-frequency noise generated by the stenosis of the arteriovenous fistula is most easily detected at the anastomotic stoma. Therefore, the blood flow sound signal is collected from the anastomotic stoma of the arteriovenous fistula. Due to the fluctuation of the heart, the blood flow sound signal at the anastomotic stoma presents a certain quasi-periodicity. In this step, the start and end positions of each period are located by a blood flow sound signal period segmentation module to facilitate the positioning of each period at the feature level in the subsequent steps.

[0029] Specifically, the period segmentation and positioning of the blood flow sound signal includes the following processes:

[0030] Step S1011, eliminate the direct current offset introduced during signal acquisition.

[0031] Since the collected blood flow sound signal may include natural background noise or bias, which may not be caused by the blood flow itself, it will make the entire blood flow sound signal drift above or below the zero line. In this embodiment, the average value of all sound data of the entire collected blood flow sound signal is taken, and this average value represents the constant background noise or bias. Subtracting the average value from the sound value of each point in the blood flow sound signal can pull the entire sound waveform back to the position centered on the zero line, and at this time the ups and downs of the waveform truly represent the changes of the blood flow sound itself. Specifically, this step calculates the amplitude mean value of all sampling points of the original blood flow sound signal, and subtracts the amplitude mean value from the original blood flow sound signal to eliminate the direct current offset such as device baseline drift introduced during signal acquisition, to obtain a first signal.

[0032] Step S1012, scale the amplitude of the above first signal to the range of [-1, 1] to facilitate subsequent step processing.

[0033] Since the collected blood flow sound signal may have a very small amplitude (low sound), it may also have a very large amplitude (high sound). Different volumes will make the numerical range of subsequent calculations differ greatly, which is not convenient for unified processing and is easy to make some calculations inaccurate or complex. Therefore, this step detects the maximum and minimum values of the amplitude of the entire blood flow sound signal, and then scales the amplitude of each point in the blood flow sound signal to between -1 and 1 according to the maximum and minimum values, to obtain the standard value of the first signal.

[0034] Step S1013, the first signal is divided into three segments, and autocorrelation operation is performed on each segment of signal. After ignoring the global maximum value at zero delay, the first secondary peak position is taken as the estimated period value of each segment. The median of the estimated period values of each segment is taken as the final period length of the first signal.

[0035] The period T of the blood flow sound signal is to be found below. The preferred scheme of the present embodiment is as follows:

[0036] First, the first signal is divided into multiple segments, for example, three segments. This is because the blood flow sound signal is not completely uniform, and some segments may have poor sound quality mixed with noise. Multiple segments can avoid the influence of a bad point on the overall judgment, thereby improving reliability.

[0037] Next, autocorrelation operation is performed on the signal of each segment to obtain the autocorrelation curve of each segment. The principle is that by detecting the similarity of a segment of signal and itself after being shifted by a certain time, it can be known whether the signal has periodicity. This is because if a segment of signal has no periodicity, the shifted signal will not be similar to itself at any time. If a segment of signal has periodicity, the shifted signal will be most similar to itself when the shift time is equal to a period T.

[0038] Next, each autocorrelation curve is observed. The curve at zero delay, i.e., the signal has no delay, is equivalent to comparing the correlation of two identical signals. There will be a global maximum peak, which means a similarity of 100%. However, this peak has no meaning, so the maximum peak at zero delay is ignored.

[0039] Next, after skipping the maximum peak at zero delay, the first significant secondary peak position after the global maximum peak is detected along the time axis. This position indicates the time length required for the sound pattern to repeat once in this segment of signal. In this way, each segment can estimate a global period length T1, T2, T3, … of the blood flow sound signal.

[0040] Finally, the estimated period lengths T1, T2, T3, … may not be exactly the same because the blood flow sound signal always has slight fluctuations or noise. In order to obtain a final period T that is most reliable and least affected by individual abnormal values, the median of all estimated period lengths is taken as the final period length because the median is more resistant to interference than the average.

[0041] Step S1014, the start and end positions of each period are located.

[0042] Firstly, local peak points of the first signal are detected to obtain a sequence of local peak points, and a plurality of candidate peak points with adjacent peak time intervals greater than or equal to 0.7T (in order to avoid detecting adjacent noise peak values) and local peak values greater than a threshold value 0.7 (because in the blood flow sound signal, the maximum amplitude value corresponds to the ejection peak value in the cardiac systolic phase, which is the most reliable physiological marker point for cycle segmentation) are screened out, and the candidate peak points are taken as starting points of the segmented cycle segments.

[0043] Then, it is calculated whether a difference between the starting point of the current cycle segment and the maximum value of the previous time point is stable within a preset amplitude range and maintained for a predetermined length of time.

[0044] Specifically, the difference between the starting point and the signal on the left side of the starting point (the previous time point) is calculated to ensure that the starting point itself is not disturbed by the residual of the previous cycle. If there is still a large fluctuation on the left side of the starting point, it indicates that the signal has not yet stabilized from the fluctuation of the previous cycle into a new cycle starting point, and further verification is needed. Only when the left side of the starting point is relatively stable and continues to be stable to the right, can it be confirmed that the true cycle end point is found. In other words, if there is significant left side interference near the starting point (the difference between the starting point and the previous predetermined range time point continues to be substandard), it indicates that the current starting point may be contaminated by the residual fluctuation of the previous cycle, and the current starting point needs to be discarded and the candidate peak point sequence of the previous step is returned, and the next candidate peak point is selected as a new starting point. Preferably, the above-mentioned preset amplitude range is for a time period, and the predetermined length of time is 0.1s.

[0045] Then, if the difference between the starting point of the current cycle segment and the maximum value of the previous time point is stable within a preset amplitude range and maintained for a predetermined length of time, a search is performed from the starting point of the current cycle segment backward, and when the difference between the starting point and the maximum value of the subsequent time point is stable within the preset amplitude range and maintained for the predetermined length of time, the time point that first reaches the preset amplitude range is taken as the end point of the current cycle segment. Preferably, when searching to the right (subsequent time point), the above-mentioned preset amplitude range is also for a time period, and the predetermined length of time is also 0.1s.

[0046] Referring to Figure 3 , Figure 3 is a waveform diagram of the cycle positioning result of the blood flow sound signal.

[0047] In step S102, the logarithmic mel spectrum of the blood flow sound signal is calculated, the cycle segmentation position is mapped to the logarithmic mel spectrum, the highest and lowest amplitude ratios of each frequency within each cycle in the time-frequency domain of the logarithmic mel spectrum are calculated, and a local energy ratio spectrum diagram is generated.

[0048] Specifically, this step performs double-path feature extraction, including the following processes:

[0049] Step S1021, the blood flow sound signal at the anastomotic stoma is pre-emphasized by a first-order FIR high-pass filter to improve the weight of high-frequency components. The system function of the first-order FIR high-pass filter is preferably:

[0050]

[0051] wherein, represents the complex variable of Z transform, that is, the time shift operator in a discrete time system.

[0052] Step S1022, the pre-emphasized blood flow sound signal is framed by a Hanning window.

[0053] The purpose of framing by the Hanning window in this step is to cut the continuous signal into short time segments to reduce spectral leakage. The short-time Fourier transform (STFT) in the subsequent steps also requires the input to be a finite length signal, and this step of framing is also needed to achieve it. Considering the requirement of time-frequency resolution, the window length of the Hanning window is set to 0.5s and the window shift is set to 0.25s in this step.

[0054] Step S1023, the short-time Fourier transform is performed on each framed signal to obtain a time-frequency graph, and the modulus of each frame of the obtained time-frequency graph is taken to obtain a power spectrum. Considering that the Mel scale is more in line with the human auditory characteristics, the power spectrum is converted from the linear domain to the Mel domain, and then the features in the Mel domain are logarithmically processed to obtain a log Mel spectrum (which is a two-dimensional matrix containing a frequency dimension and a time dimension). The spectrum graph of the log Mel spectrum is shown in Figure 4 .

[0055] Step S1024, the positioning position of each cycle obtained in step S101 is mapped to the cycle segmentation position of the log Mel spectrum according to the design of a window length of 0.5s and a window shift of 0.25s to obtain a feature level positioning.

[0056] Step S1025, the highest and lowest amplitude ratio of each frequency in each cycle in the time-frequency domain of the log Mel spectrum is calculated to generate a local energy ratio spectrum graph.

[0057] It should be pointed out that the local energy ratio spectrum graph is actually an effective representation of the log Mel spectrum in a local cycle that can reflect the characteristics of narrow sounds (narrow lesions can cause turbulent flow, resulting in a dramatic fluctuation of energy in a specific frequency band, and the local energy ratio spectrum graph can amplify such abnormalities). Figure 4 The frequency dimension of the log Mel spectrum in has 64 dimensions and the time dimension has 50 dimensions. Figure 4 After the segmentation position of the cycle segment is positioned, the spectrum graph can be roughly divided into 10 cycle segments, and the ratio of the highest amplitude to the lowest amplitude in each cycle segment under each frequency dimension can be calculated to obtain a a two-dimensional spectrum, i.e., a local energy ratio spectrum.

[0058] In step S103, the log-Mel spectrum and the local energy ratio spectrum are respectively subjected to convolution transformation to obtain two deep features with the same dimension, and a cosine similarity matrix of the two deep features with the same dimension is calculated.

[0059] In order to eliminate the amplitude difference caused by the blood flow rate change in the heart cycle, highlight the pathological related relative distribution mode of the spectrum, and provide a standardized input for subsequent cosine similarity matrix calculation to ensure the effectiveness of the dynamic enhancement mechanism, normalization needs to be performed first. Specifically, the log-Mel spectrum and the local energy ratio spectrum are subjected to feature transformation through multiple convolution layers, and the number of channels of the last convolution layer is 1, so as to finally obtain two deep feature matrices A* and B* with the same dimension. The two deep feature matrices are subjected to minimum-maximum normalization processing by column. Exemplarily, the two deep feature matrices are two-dimensional matrices with the same dimension, the vertical axis still represents the frequency dimension, and the horizontal axis still represents the time dimension. For each vector at a time step, normalization is performed according to the following formula.

[0060]

[0061] wherein, denotes the normalized feature vector at the time step, denotes a vector composed of all the feature values of all the frequencies at the time step in the deep feature matrix, denotes the maximum feature value in the vector at the time step, denotes the minimum feature value in the vector at the time step.

[0062] The two normalized deep feature matrices obtained are subjected to cosine similarity calculation to obtain a similarity matrix M. The similarity matrix M is used to measure the similarity of the feature vectors at different time steps, and the interval is [0, 1]. The cosine similarity calculation method is as follows

[0063]

[0064] wherein is the normalized vector of the i-th column of the deep feature matrix A*; i is the normalized vector of the i-th column of the deep feature matrix B*; denotes the similarity value of the two normalized vectors, which is the value corresponding to the i-th row and the i-th column of the similarity matrix M. j i j

[0065] ​​​​In step S104, the two depth features with the same dimension are fused in a channel, and then input to the depth neural network to be trained, the loss function of the depth neural network to be trained includes two parts of feature alignment loss and classification loss, and the feature alignment loss is determined according to the cosine similarity matrix.

[0066] Specifically, the two depth feature matrices A* and B* obtained in step S103 are fused in a channel, for example, the dimensions of A* and B* are both M x N, and after the two matrices are fused in the channel, the dimension becomes 2 x M x N. Then the fused feature passes through the convolution layer, the pooling layer and the full connection layer of the depth neural network to be trained, and the output value logits in the network model training process is obtained.

[0067] It should be pointed out that in the training process of the depth neural network, in order to highlight the weak and narrow signal in the complex blood flow sound, the traditional static spectrum analysis is upgraded to a lesion-oriented dynamic feature enhancement, and the feature alignment loss The depth neural network model forces the original spectrum feature and the energy ratio feature to be aligned at the same time step, so that the network weight is adaptively focused on the narrow sensitive area revealed by the depth feature B*, such as the high-frequency mutation band.

[0068] In order to achieve the above technical purposes, the loss function is carefully set by the application . .

[0069] The feature alignment loss is used to constrain the cosine similarity matrix M to approximate the unit matrix , and to force the learning of the features at the same time step and the features at different time steps, that is, the features at the same time step indicate that the basic spectrum feature A (logarithmic mel spectrum) is highly correlated with the energy ratio feature B (local energy ratio), and the features at different time steps indicate that there is no correlation between the features.

[0070] The classification loss is used to minimize the difference between the network prediction result and the true label y, and optimize the performance of the narrow / unobstructed binary classification.

[0071] The mathematical expression of the loss function adopted by the application is:

[0072] ;

[0073] ;

[0074] ​ ;

[0075] wherein, is a hyper-parameter to measure the importance of in the whole loss calculation; denotes a unit diagonal matrix; denotes the cross-entropy loss of calculating and , denotes the cross-entropy loss function of and ; denotes the cross-entropy loss of calculating and , denotes the cross-entropy loss function of and ; denotes the cosine similarity matrix of the two dimensions of the deep features, denotes the label value of the corresponding blood flow sound signal sample, denotes the predicted value output by the model.

[0076] The loss function is linked with the feature alignment loss through the hyper-parameter alpha and the classification loss , so as to realize the cooperation of feature optimization and classification decision, and improve the prediction ability after the deep neural network model is trained.

[0077] In summary, the method for extracting blood flow sound signal features of arteriovenous fistula provided by the application extracts and dynamically enhances the features through the double-path feature extraction of the log mel spectrum and the local energy ratio spectrum diagram, designs effective audio features that can reflect the characteristics of arteriovenous fistula stenosis, and uses the feature alignment loss to enhance the similarity between the sound signal features and the log mel spectrum during the training of the deep neural network model, so as to greatly improve the performance of the deep neural network model.

[0078] Another embodiment of the application provides an electronic device, comprising: one or more processors; a storage device for storing one or more computer programs; and when the one or more programs are executed by the one or more processors, the one or more processors implement the method for extracting blood flow sound signal features of arteriovenous fistula in the above-mentioned method embodiments.

[0079] Referring to Figure 5 , the electronic device 100 can include but is not limited to a special medical device, a PC, an industrial computer, a smart terminal or a server. Figure 5The electronic device 100 shown is merely one example and should not be taken as limiting the scope of the embodiments of the present application. For instance, the electronic device 100 can include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 101 that can perform various appropriate actions and processes to implement the methods of the embodiments as described herein according to programs stored in a read-only memory (ROM) 102 or loaded into a random access memory (RAM) 103 from a storage device 108. Various programs and data required for operation of the electronic device 100 are also stored in the RAM 103. The processing device 101, the ROM 102, and the RAM 103 are connected to each other through a bus 105. An input / output (I / O) interface 104 is also connected to the bus 105.

[0080] Generally, the following devices can be connected to the I / O interface 104: input devices 106 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; output devices 107 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 108 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 109. The communication devices 109 can allow the electronic device 100 to communicate wirelessly or wired with other devices to exchange data. Although Figure 5 The electronic device 100 is shown with various devices, but it should be understood that not all of the shown devices are required to be implemented or present. More or fewer devices can alternatively be implemented or present.

[0081] The above description is merely exemplary of the present application. One skilled in the art will readily recognize from the disclosure herein the broadest scope of the present application, which is not limited to the specific embodiments disclosed. For example, one skilled in the art will recognize that the features of the various embodiments disclosed can be combined, rearranged, and / or eliminated, and that the scope of the application extends to all alternative combinations.

Claims

1. A method for feature extraction of an arteriovenous fistula blood flow sound signal, characterized in that, The method comprises the following steps: Periodic segmentation positioning of the blood flow sound signal at the arteriovenous anastomosis is performed, and the start and end positions of each period are located; The log mel spectrum of the blood flow sound signal is calculated, the periodic segmentation position is mapped to the log mel spectrum, the highest and lowest amplitude ratios of each frequency within the log mel spectrum in the time-frequency domain are calculated, and a local energy ratio spectrum diagram is generated; Convolution transformation is performed on the log mel spectrum and the local energy ratio spectrum diagram respectively, two depth features with the same dimensions are obtained, and a cosine similarity matrix of the two depth features with the same dimensions is calculated; The two depth features with the same dimensions are fused in a channel, and then input into a to-be-trained deep neural network, wherein a loss function of the to-be-trained deep neural network comprises two parts of feature alignment loss and classification loss, and the feature alignment loss is determined according to the cosine similarity matrix.

2. The method for feature extraction of blood flow sound signals of an arteriovenous fistula according to claim 1, characterized in that, The periodic segmentation positioning of the blood flow sound signal at the arteriovenous anastomosis comprises the following steps: The amplitude mean of all sampling points of the original blood flow sound signal is calculated, and the amplitude mean is subtracted from the original blood flow sound signal to eliminate the direct current offset introduced by signal acquisition, and a first signal is obtained; The first signal is divided into multiple segments, autocorrelation operation is performed on each segment of the signal, after the global maximum value at zero delay is ignored, the first secondary peak position is taken as the estimated period value of each segmented signal, and the median of the estimated period value of each segmented signal is taken as the final period length of the first signal.

3. The method for AVF blood flow sound signal feature extraction according to claim 2, characterized in that, The start and end positions of each period are located, comprising the following steps: Local peak points of the first signal are detected to obtain a local peak point sequence, a plurality of candidate peak points with a time interval between adjacent peak values greater than or equal to a first preset threshold and a local peak value greater than a second preset threshold are screened out, and the candidate peak points are taken as the starting points of the segmented period segments; It is judged whether the difference between the starting point of the current period segment and the maximum value of the previous time point is stable within a preset amplitude range and maintains a predetermined time length, and if so, the starting point is searched backward, and when the difference between the starting point and the maximum value of the subsequent time point is stable within the preset amplitude range and maintains the predetermined time length, the time point that first reaches the preset amplitude range is taken as the end point of the current period segment.

4. The method for feature extraction of blood flow sound signals of an arteriovenous fistula according to claim 1, characterized in that, The log mel spectrum of the blood flow sound signal is calculated, comprising the following steps: The blood flow sound signal is pre-emphasized by a first-order FIR high-pass filter to increase the weight of high-frequency components; The pre-emphasized blood flow sound signal is framed by using a Hanning window; Short-time Fourier transform is performed on each framed signal to obtain a time-frequency diagram; Each frame of the obtained time-frequency diagram is taken modulo to obtain a power spectrum; The power spectrum is converted from a linear domain to a mel domain, and the features in the mel domain are logarithmically processed to obtain a log mel spectrum.

5. The method for feature extraction of blood flow sound signals of an arteriovenous fistula according to claim 1, characterized in that, a loss function of the deep neural network to be trained is represented by the following equation: ； ； ； wherein, is a hyperparameter that measures the importance of the entire loss calculation; represents a unit diagonal matrix; represents the calculation of the cross-entropy loss of ; represents the feature alignment loss, is a cross-entropy loss function of and ; represents the calculation of the cross-entropy loss of ; represents the classification loss, is a cross-entropy loss function of and ; represents a cosine similarity matrix of the two dimensions of the deep features, represents a label value of a corresponding blood flow sound signal sample, represents a predicted value output by the model.

6. The method for AVF blood flow sound signal feature extraction according to claim 4, characterized in that, The window length of the Hanning window is 0.5s, and the window shift is 0.25s.

7. The method for AVF blood flow sound signal feature extraction according to claim 4, characterized in that, The system function of the first-order FIR high-pass filter is: wherein denotes the complex variable of the Z-transform, i.e. the time-shift operator in discrete-time systems.

8. The method for feature extraction of blood flow sound signals of an arteriovenous internal fistula according to claim 3, characterized in that, The first preset threshold is 0.7 times the final period length of the first signal; and the second preset threshold is 0.7 times the amplitude of the first signal.

9. The method for feature extraction of blood flow sound signals of an arteriovenous internal fistula according to claim 3, characterized in that, The preset amplitude range is 0.7-0.9 times the amplitude of the first signal, and the predetermined time length is 0.1s.

10. An electronic device, comprising: The method comprises the following steps: One or more processors; a storage device for storing one or more computer programs; when the one or more computer programs are executed by the one or more processors, the one or more processors are caused to implement the method for extracting features of an arteriovenous fistula blood flow sound signal according to any one of claims 1-9.

Images