Processing method of arteriovenous fistula blood flow sound signal and electronic equipment

Abstract

The application discloses an arteriovenous fistula blood flow sound signal processing method and electronic equipment, and relates to the technical field of artificial intelligence medical information processing. The method comprises the following steps: performing periodic segmentation positioning on the blood flow sound signal at the arteriovenous anastomosis; calculating the logarithmic mel spectrum of the blood flow sound signal; calculating the energy entropy of the periodic segment corresponding to each frequency of the logarithmic mel spectrum; participating in the calculation of the proxy attention mechanism by using the query agent matrix obtained by the energy entropy calculation to obtain an aggregated feature matrix. The application greatly reduces the complexity of attention mechanism calculation and obtains blood flow sound deep features that can more reflect arteriovenous fistula stenosis.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology based on artificial intelligence, and in particular to a method and electronic device for processing sound signals of blood flow in arteriovenous fistulas. Background Technology

[0002] Currently, there are two main methods for detecting arteriovenous fistula stenosis using blood flow sounds:

[0003] One approach involves extracting statistical features of blood flow sound signals in the time or frequency domain, such as the high-to-low frequency energy ratio and root mean square energy, and then comparing the differences in statistical quantitative indicators between narrow and unobstructed conditions to make a judgment. However, this type of method is affected by the quality of the collected data and is not very robust.

[0004] Another approach combines the time-frequency domain features of audio with deep neural networks such as Vision Transformer for classification. However, when arteriovenous fistulas are stenotic, the sound at the anastomosis may exhibit a "whistling" sound, with a sudden change in the time-frequency domain, making this method insufficient for capturing the unique morphology of blood flow sound signals under stenotic conditions. Furthermore, while the Vision-Transformer network model comprehensively considers the local characteristics of each patch and the interrelationships between patches, it also considers the relationship between each patch and all other patches when calculating attention. This means that the relationships between different patches are selectively emphasized, and not all patches are necessarily related; moreover, this design can also impose a computational burden.

[0005] Therefore, there is an urgent need for a processing technique for the blood flow sound signal of arteriovenous fistula, which can extract the depth features of the blood flow sound signal that can more clearly reflect the stenosis of arteriovenous fistula and reduce the matrix dimension of the depth features, so as to solve the above-mentioned technical problems in the existing technology. Summary of the Invention

[0006] In view of the above-mentioned defects or deficiencies in the prior art, the present invention provides a method and electronic device for processing blood flow sound signals of arteriovenous fistula. It utilizes the fragment-level energy entropy in the frequency dimension as the proxy query vector of the attention mechanism, integrates the information of the original time-frequency features and local energy entropy, and works together to train the deep neural network. This results in an improved attention mechanism with two Softmax attention modules, which are equivalent to a generalized linear attention. This achieves the fusion of high-performance Softmax attention and efficient linear attention, and has the advantages of low computational complexity and strong model expressive power.

[0007] One aspect of the present invention provides a method for processing the sound signal of blood flow in an arteriovenous fistula, comprising:

[0008] The blood flow sound signal at the arteriovenous anastomosis is periodically segmented and located to pinpoint the start and end points of each cycle.

[0009] The log-Mel spectrum of the blood flow sound signal is calculated, and the start and end positions of the cycle are mapped onto the time dimension of the log-Mel spectrum according to a preset window length and a preset frame shift, thereby dividing the log-Mel spectrum into time dimensions. Each periodic segment is used to calculate the position of each periodic segment in the log-Mel spectrum; where each periodic segment corresponds to several frequencies and time frames.

[0010] Calculate the energy entropy of the periodic segment corresponding to each frequency of the log-Mel spectrum, and construct... An energy entropy matrix of dimension, in which This represents the number of frequencies in the logarithmic Mel spectrum. This indicates the number of periodic segments in the logarithmic Mel spectrum;

[0011] The query proxy matrix of the proxy attention mechanism is calculated based on the energy entropy matrix. The query matrix, key matrix, and value matrix of the proxy attention mechanism are calculated based on the log-Mel spectrum. A first attention score is calculated using the query proxy matrix and the key matrix. A first aggregated feature matrix of the log-Mel spectrum is calculated based on the first attention score and the value matrix. A second attention score is calculated using the query matrix and the query proxy matrix. A second aggregated feature matrix of the proxy attention mechanism is calculated based on the second attention score and the first aggregated feature matrix. The second aggregated feature matrix is ​​used for training the classification layer of the deep neural network.

[0012] In another aspect, the present invention provides an electronic device comprising: one or more processors; a storage device for storing one or more computer programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors are configured to implement the above-described method for processing the blood flow sound signal of an arteriovenous fistula.

[0013] The advantages of the method and electronic device for processing blood flow sound signals in arteriovenous fistulas provided by this invention are as follows:

[0014] (1) In this invention, fragment energy entropy is used as the query proxy matrix in attention calculation. Fragment energy entropy can measure the local change characteristics of sound when arteriovenous fistula is stenotic. Using it as the query proxy matrix can make the depth features of the generated blood flow sound signal more discriminative.

[0015] (2) The number of periodic segments of the fragment energy entropy of the present invention is significantly reduced compared with the number of time frames in the original time-frequency features, which can significantly save computational costs when calculating attention.

[0016] (3) The algorithm of the present invention is simple to implement, has good scalability, and is easy to integrate into subsequent convolutional neural network architectures or Transformer architectures. Attached Figure Description

[0017] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0018] Figure 1 This is a flowchart of a method for processing the sound signal of blood flow in an arteriovenous fistula according to an embodiment of this application. Figure 1 ;

[0019] Figure 2 This is a flowchart of a method for processing the sound signal of blood flow in an arteriovenous fistula according to an embodiment of this application. Figure 2 ;

[0020] Figure 3 This is a waveform diagram illustrating the periodic localization result of blood flow sound signals provided in one embodiment of this application;

[0021] Figure 4 This is a spectrum of the log-Mel spectrum of a blood flow sound signal provided in one embodiment of this application;

[0022] Figure 5 This is a schematic diagram illustrating the calculation of agent attention using fragment energy entropy according to an embodiment of this application;

[0023] Figure 6 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0026] It should be understood that although the terms first, second, third, etc., may be used to describe the acquisition modules in the embodiments of the present invention, these acquisition modules should not be limited to these terms. These terms are only used to distinguish the acquisition modules from each other.

[0027] See Figure 1 One embodiment of this application provides a method for processing blood flow sound signals from arteriovenous fistulas, mainly comprising four stages: localization of periodic segments of the blood flow sound signal, calculation of log-Mel spectrum, calculation of segment energy entropy and energy entropy matrix, surrogate attention calculation based on segment energy entropy, and training of a deep neural network (optional). This method can run as a computer program on dedicated medical equipment, a PC, a smart terminal, or a server.

[0028] See Figure 2 The method for processing the sound signal of blood flow in an arteriovenous fistula according to this embodiment includes the following steps:

[0029] Step S101: Periodically segment and locate the blood flow sound signal at the arteriovenous anastomosis site to determine the start and end positions of each period.

[0030] Blood flow turbulence is most pronounced at the arteriovenous anastomosis site, and abnormal sounds generated by arteriovenous fistula stenosis, such as high-frequency murmurs, are most easily detected at the anastomosis site. Therefore, this method collects blood flow sound signals from the arteriovenous anastomosis site. Due to cardiac fluctuations, the blood flow sound signals at the anastomosis site exhibit a certain quasi-periodicity. In this step, a blood flow sound signal periodic segmentation module is used to locate the start and end positions of each period, so as to facilitate the localization of each subsequent period at the feature level.

[0031] Specifically, the periodic segmentation and localization of blood flow sound signals includes the following process:

[0032] Step S1011: Eliminate DC offset introduced during signal acquisition.

[0033] Because the acquired blood flow sound signal may include natural background noise or bias, this background noise may not be caused by the blood flow itself, causing the entire blood flow sound signal to drift above or below the zero line. In this embodiment, the average value of all sound data from the acquired blood flow sound signal is taken. This average value represents the constant background noise or bias. Subtracting this average value from the sound value of each point in the blood flow sound signal brings the entire sound waveform back to the position centered on the zero line. At this point, the fluctuations in the waveform truly represent the changes in the blood flow sound itself. Specifically, this step calculates the average amplitude of all sampling points of the original blood flow sound signal and subtracts this average amplitude from the original blood flow sound signal to eliminate DC offsets such as device baseline drift introduced by signal acquisition, obtaining the first signal.

[0034] Step S1012: Scale the amplitude of the first signal to the range of [-1, 1] to facilitate subsequent processing steps.

[0035] Because the acquired blood flow sound signal may have a very small amplitude (low sound) or a very large amplitude (high sound), different volume levels can lead to significant differences in the range of values ​​calculated subsequently. This makes it inconvenient to process uniformly and can easily make some calculations inaccurate or complex. Therefore, this step detects the maximum and minimum amplitude values ​​in the entire blood flow sound signal, and then scales the amplitude of each point in the blood flow sound signal proportionally to these maximum and minimum values ​​to a value between -1 and 1, thus obtaining the standard value of the first signal.

[0036] Step S1013: Divide the first signal into three segments, perform autocorrelation operation on each segment, ignore the global maximum value at zero delay, take the first secondary peak position as the estimated period value of each segment, and take the median of the estimated period value of each segment as the final period length of the first signal.

[0037] The next step is to find the period T of the blood flow sound signal. In this embodiment, the preferred approach is as follows:

[0038] First, the first signal is divided into multiple segments, such as three segments. This is because the blood flow sound signal is not completely uniform, and the sound quality of a certain segment may be poor and mixed with noise. Multiple segments can avoid a bad spot affecting the overall judgment, thereby improving reliability.

[0039] Next, autocorrelation is performed on each segment of the signal individually to obtain the autocorrelation curve for each segment. The principle is: by detecting the similarity between a signal segment and itself after a certain time interval, it is possible to determine whether the signal segment has periodicity. This is because if a signal has no periodicity at all, it will not resemble itself at any time interval; if a signal has periodicity, it will be most similar to itself when the time interval is exactly equal to one period T.

[0040] Next, observe each autocorrelation curve. The curve has a zero time offset, meaning the signal has no delay. This is equivalent to judging the correlation between two identical signals. There must be a global maximum peak. This maximum peak means 100% similarity. However, this peak is meaningless, so the maximum value at zero delay is ignored.

[0041] Next, skipping the maximum peak with zero delay and detecting to the right along the time axis, we encounter the first significant secondary peak position after the aforementioned global maximum peak. This position represents the time required for the sound pattern to repeat once in this signal segment, thus allowing each segment to estimate the global period length T1, T2, T3, ... of a blood flow sound signal.

[0042] Finally, the estimated cycle lengths T1, T2, T3, etc., may not be exactly the same because blood flow sound signals will always have slight fluctuations or noise. In order to obtain the most reliable final cycle T that is least affected by individual outliers, we take the median of all estimated cycle lengths as the final cycle length, because the median is more robust to interference than the average.

[0043] Step S1014: Locate the start and end positions of each cycle.

[0044] First, the local peak points of the first signal are detected to obtain a sequence of local peak points. Several candidate peak points are selected that have an adjacent peak time interval greater than or equal to 0.7T (to avoid detecting nearby noise peaks) and a local peak value greater than the threshold 0.7 (because in the blood flow sound signal, the maximum amplitude corresponds to the peak value of blood ejection during cardiac systole, which is the most reliable physiological marker point for period segmentation). The candidate peak points are used as the starting point of the segmented periodic segments.

[0045] Next, calculate whether the difference between the starting point of the current period and the maximum value at the previous time point is stable within the preset range and maintained for the predetermined duration.

[0046] Specifically, calculating the difference between the starting point and the signal to the left of the starting point (previous time points) is to ensure that the starting point itself is not affected by residual interference from the previous cycle. If there are still large fluctuations to the left of the starting point, it indicates that the signal has not yet fully stabilized from the fluctuations of the previous cycle and entered the new cycle starting point, requiring further verification. Only when the left side of the starting point is relatively stable and continues to stabilize to the right can the true cycle endpoint be confirmed. In other words, if there is significant left-side interference near the starting point (the difference between the starting point and the previously predetermined time points is consistently below the standard), it indicates that the current starting point may be contaminated by residual fluctuations from the previous cycle, requiring the current starting point to be discarded and the candidate peak point sequence from the previous step to be returned, skipping the current starting point and selecting the next candidate peak point as the new starting point. Preferably, the above-mentioned preset amplitude range is... The time period is 0.1 seconds.

[0047] Next, if the difference between the starting point of the current period and the maximum value at a previous time point remains stable within a preset range and for a predetermined duration, then the search proceeds backward from the starting point of the current period. When the difference between the starting point and the maximum value at a subsequent time point remains stable within the preset range and for the predetermined duration, the first point to reach the maximum value will be reached. The time point within is taken as the end point of the current period segment. Preferably, when searching to the right (subsequent time points), the above-mentioned preset amplitude range is also [missing information]. The time period is also scheduled for 0.1 seconds.

[0048] Figure 3This is a waveform representation of the periodic localization result of the blood flow sound signal.

[0049] Step S102: Calculate the log-Mel spectrum of the blood flow sound signal, and map the start and end positions of the cycle onto the time dimension of the log-Mel spectrum according to a preset window length and a preset frame shift, thereby dividing the log-Mel spectrum into time dimensions. Each periodic segment is assigned a position in the log-Mel spectrum; where each periodic segment corresponds to several frequencies and time frames.

[0050] The log-Mel spectrum of the blood flow sound signal is calculated using the following steps:

[0051] Step S1021 involves pre-emphasizing the blood flow sound signal at the anastomosis site using a first-order FIR high-pass filter to increase the weight of high-frequency components. The preferred system function of the first-order FIR high-pass filter is:

[0052]

[0053] in, This represents the complex variable of the Z-transform, i.e., the time shift operator in a discrete-time system.

[0054] Step S1022: The pre-aggravated blood flow sound signal is framed using the Hanning window.

[0055] The purpose of using the Hanning window for framing in this step is to divide the continuous signal into short-duration segments to reduce spectral leakage. The subsequent Short-Time Fourier Transform (STFT) also requires a finite-length input signal, which also necessitates framing in this step. Considering the time-frequency resolution requirements, the preferred setting for this step is a Hanning window length of 0.5s and a frame shift (window shift) of 0.25s.

[0056] Step S1023: Perform a short-time Fourier transform on each frame of the segmented signal to obtain a time-frequency graph. Take the modulus of each frame in the obtained time-frequency graph to obtain the power spectrum. Considering that the Mel scale is more consistent with human auditory characteristics, the power spectrum is transformed from the linear domain to the Mel domain. Then, logarithmic processing is performed on the characteristics of the Mel domain to obtain the logarithmic Mel spectrum. The logarithmic Mel spectrum is a two-dimensional matrix containing frequency and time dimensions. Dimension D represents frequency, typically 64, and dimension K represents the time dimension, indicating the number of time frames. See the spectral diagram for a log-Mel spectrum. Figure 4 .

[0057] Step S1024: Map the positioning positions of each cycle obtained in step S101 onto the log-Mel spectrum according to a window length of 0.5s and a window shift of 0.25s, thereby dividing the log-Mel spectrum into periodic segments in the time dimension. There are several periodic segments, each corresponding to several frequencies and time frames (the log-Mel spectrum in the time-frequency domain has the time dimension on the horizontal axis and the frequency dimension on the vertical axis). The position index of each periodic segment in the log-Mel spectrum is calculated using the following formula. :

[0058] in, Indicates the index of the segmentation point of the periodic segment. Sampling frequency, This indicates the preset window length, which is 0.5s in this embodiment. This indicates the preset frame shift, which is 0.25s in this embodiment.

[0059] Step S103: Calculate the energy entropy of the periodic segment corresponding to each frequency of the log-Mel spectrum, and construct... An energy entropy matrix of dimension, in which This represents the number of frequencies in the logarithmic Mel spectrum. This indicates the number of periodic segments in the logarithmic Mel spectrum.

[0060] Energy entropy reflects the uncertainty of a signal. This step calculates the energy entropy of each periodic segment at each frequency based on the previously determined log-Mel spectrum. The resulting segment energy entropy reflects the local changes in sound characteristics during arteriovenous fistula stenosis. The energy entropies of different periodic segments constitute a... A two-dimensional matrix, where This represents the number of frequencies in the logarithmic Mel spectrum. This indicates the number of periodic segments in the logarithmic Mel spectrum.

[0061] like Figure 4 The aforementioned log-Mel spectrum has 64 frequency dimensions and 50 time frames. Figure 1 After locating the periodic segments, the spectrum can be roughly divided into 12 periodic segments. Calculating the energy entropy of each periodic segment at each frequency dimension yields a... Two-dimensional spectrum.

[0062] Furthermore, the specific calculation process for the fragment energy entropy is as follows:

[0063] Calculate the first according to the following formula. The period segment corresponding to the frequency of the first Energy per time frame :

[0064]

[0065] in, Indicates the first The period segment corresponding to the frequency of the first The amplitude of each time frame;

[0066] According to the following formula, the first The energy of each time frame of the periodic segment corresponding to each frequency is normalized:

[0067]

[0068] in, This represents the total number of frames in the periodic segment. Indicates the first The period segment corresponding to the frequency of the first The normalized probability of each time frame in this periodic segment;

[0069] Calculate the first according to the following formula. The energy entropy of a periodic segment corresponding to a frequency :

[0070]

[0071] in, express The natural logarithm of .

[0072] Finally, based on the energy entropy of each period segment Build A dimensional energy entropy matrix.

[0073] Step S104: Calculate the query proxy matrix of the proxy attention mechanism based on the energy entropy matrix; calculate the query matrix, key matrix, and value matrix of the proxy attention mechanism based on the log-Mel spectrum; calculate the first attention score using the query proxy matrix and the key matrix; calculate the first aggregated feature matrix of the log-Mel spectrum based on the first attention score and the value matrix; calculate the second attention score using the query matrix and the query proxy matrix; calculate the second aggregated feature matrix of the proxy attention mechanism based on the second attention score and the first aggregated feature matrix; the second aggregated feature matrix is ​​used for training the classification layer of the deep neural network.

[0074] Because arteriovenous fistula stenosis causes a sudden energy shift in the blood flow acoustic signal at this location, leading to an increase in entropy, energy entropy can mark pathologically sensitive areas. Furthermore, the dimension of the energy entropy matrix is ​​lower than that of the log-Mel spectrum (the original log-Mel spectrum has a dimension from...). Compressing each time frame into an energy entropy matrix A periodic segment, in which Therefore, this step performs a linear transformation on the energy entropy matrix to generate a low-dimensional query surrogate matrix. This query surrogate matrix replaces the high-dimensional query matrix, which should be generated by the log-Mel spectrum, in the attention mechanism for attention calculation. This improves the ability to express pathological features by capturing local sound mutations (such as "whistling") caused by narrowing through energy entropy, and significantly reduces the computational complexity of the attention mechanism by utilizing the significant compression of the time dimension of the energy entropy matrix.

[0075] See Figure 5 The calculation process of the attention mechanism is as follows (the step numbers are only used to distinguish each step and do not restrict the order of the steps):

[0076] (1) Calculate the query proxy matrix according to the following formula. :

[0077]

[0078] in, Represents the energy entropy matrix. This represents the transpose of the energy entropy matrix. This represents a linear transformation matrix whose elements are automatically adjusted as the model is trained.

[0079] (2) Calculate the query matrix, key matrix, and value matrix (the three matrices used by the attention mechanism) of the agent attention mechanism based on the log-Mel spectrum, including:

[0080] ;

[0081] ;

[0082] ;

[0083] in, Represents the logarithmic Mel spectrum. express transpose, Represents the query matrix. Represents the key matrix. Represents a value matrix, , and Both represent linear transformation matrices whose elements are automatically adjusted as the model is trained.

[0084] (3) Calculate the first aggregate feature matrix

[0085] The first attention score is calculated using the following formula. :

[0086]

[0087] in, () represents the normalized exponential function. Indicates querying the proxy matrix. Key matrix transpose, Represents matrix multiplication;

[0088] The first aggregation feature matrix is ​​calculated using the following formula. :

[0089]

[0090] in, Represents a value matrix.

[0091] (4) Calculate the second aggregation feature matrix

[0092] The second attention score is calculated using the following formula. :

[0093]

[0094] in, () represents the normalized exponential function. Represents the query matrix. Indicates querying the proxy matrix. Represents matrix multiplication;

[0095] The second aggregation feature matrix is ​​calculated using the following formula. :

[0096]

[0097] in, This represents the first feature output matrix. This represents matrix multiplication.

[0098] In summary, the second aggregated feature matrix output by the proxy attention module... In fact, it integrates the global time-frequency information of the original logarithmic Mel spectrum with the local pathological features extracted from energy entropy, therefore the second aggregated feature matrix is ​​used. Training the classification layer of a subsequent deep neural network can yield a deep neural network model with lower computational complexity and better performance.

[0099] The method described in this embodiment uses fragment energy entropy as the query proxy matrix in attention calculation. Fragment energy entropy measures the local changes in sound during arteriovenous fistula stenosis, and using it as the query proxy matrix makes the generated deep features more discriminative. Since the number of periodic fragments in fragment energy entropy is significantly reduced compared to the number of time frames in the original time-frequency features, the computational complexity can be greatly reduced in subsequent attention calculations, significantly saving computational costs. This invention is simple to implement, highly scalable, and easy to integrate into subsequent convolutional neural network architectures or Transformer architectures.

[0100] Another embodiment of the present invention provides an electronic device, including: one or more processors; a storage device for storing one or more computer programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method for processing arteriovenous fistula blood flow sound signals in the above method embodiment.

[0101] See Figure 6 Electronic device 100 may include, but is not limited to, specialized medical equipment, PCs, industrial computers, smart terminals, or servers. Figure 6 The illustrated electronic device 100 is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention. Exemplarily, the electronic device 100 may include a processing device (e.g., a central processing unit, graphics processor, etc.) 101, which may perform various appropriate actions and processes to implement the methods described in the embodiments of the present invention, based on a program stored in a read-only memory (ROM) 102 or a program loaded from a storage device 108 into a random access memory (RAM) 103. Various programs and data required for the operation of the electronic device 100 are also stored in the RAM 103. The processing device 101, ROM 102, and RAM 103 are interconnected via a bus 105. An input / output (I / O) interface 104 is also connected to the bus 105.

[0102] Typically, the following devices can be connected to I / O interface 104: input devices 106 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 107 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 108 including, for example, magnetic tapes, hard disks, etc.; and communication devices 109. Communication device 109 allows electronic device 100 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 An electronic device 100 with various components is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0103] The above description is merely a preferred embodiment of the present invention. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to the specific combination of the above-described technical features, but should also cover other technical solutions formed by any combination of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A method for processing the sound signal of blood flow in an arteriovenous fistula, characterized in that, include: The blood flow sound signal at the arteriovenous anastomosis is periodically segmented and located to pinpoint the start and end points of each cycle. The log-Mel spectrum of the blood flow sound signal is calculated, and the start and end positions of the cycle are mapped onto the time dimension of the log-Mel spectrum according to a preset window length and a preset frame shift, thereby dividing the log-Mel spectrum into time dimensions. Each periodic segment is used to calculate the position of each periodic segment in the log-Mel spectrum; where each periodic segment corresponds to several frequencies and time frames. Calculate the energy entropy of the periodic segment corresponding to each frequency of the log-Mel spectrum, and construct... An energy entropy matrix of dimension, where This represents the number of frequencies in the logarithmic Mel spectrum. This indicates the number of periodic segments in the logarithmic Mel spectrum; The query proxy matrix of the proxy attention mechanism is calculated based on the energy entropy matrix. The query matrix, key matrix, and value matrix of the proxy attention mechanism are calculated based on the log-Mel spectrum. A first attention score is calculated using the query proxy matrix and the key matrix. A first aggregated feature matrix of the log-Mel spectrum is calculated based on the first attention score and the value matrix. A second attention score is calculated using the query matrix and the query proxy matrix. A second aggregated feature matrix of the proxy attention mechanism is calculated based on the second attention score and the first aggregated feature matrix. The second aggregated feature matrix is ​​used for training the classification layer of the deep neural network.

2. The method for processing the sound signal of blood flow in an arteriovenous fistula according to claim 1, characterized in that, The periodic segmentation and localization of the blood flow sound signal at the arteriovenous anastomosis includes: Calculate the mean amplitude of all sampling points of the original blood flow sound signal, and subtract the mean amplitude from the original blood flow sound signal to eliminate the DC offset introduced by signal acquisition, and obtain the first signal; The first signal is divided into multiple segments. Autocorrelation is performed on each segment. After ignoring the global maximum value at zero delay, the position of the first secondary peak is taken as the estimated period value of each segment. The median of the estimated period value of each segment is taken as the final period length of the first signal.

3. The method for processing the sound signal of blood flow in an arteriovenous fistula according to claim 2, characterized in that, The method of locating the start and end positions of each cycle includes: Detect local peak points of the first signal to obtain a local peak point sequence, filter out a number of candidate peak points whose adjacent peak time interval is greater than or equal to a first preset threshold and whose local peak is greater than a second preset threshold, and use the candidate peak points as the starting point of the segmented periodic segment; Calculate whether the difference between the starting point of the current period and the maximum value at a previous time point is stable within a preset range and maintained for a predetermined duration. If so, search backward from the starting point of the current period. When the difference between the starting point and the maximum value at a subsequent time point is stable within the preset range and maintained for the predetermined duration, it will be the first to reach the target. The time point within is taken as the end point of the current period.

4. The method for processing the sound signal of blood flow in an arteriovenous fistula according to claim 1, characterized in that, The calculation of the log-Melb spectrum of the blood flow sound signal includes: The blood flow sound signal is pre-emphasized using a first-order FIR high-pass filter to increase the weight of high-frequency components; The pre-emphasis blood flow sound signal was framed using the Hanning window. Perform a short-time Fourier transform on each frame of the signal after framing to obtain a time-frequency diagram; The power spectrum is obtained by taking the modulus of each frame in the obtained time-frequency graph; The power spectrum is transformed from the linear domain to the Mel domain, and the characteristics of the Mel domain are logarithmically processed to obtain the logarithmic Mel spectrum.

5. The method for processing the sound signal of blood flow in an arteriovenous fistula according to claim 1, characterized in that, The calculation of the energy entropy of each frequency segment corresponding to the log-Mel spectrum includes: Calculate the first according to the following formula. The period segment corresponding to the frequency of the first Energy per time frame : ； in, Indicates the first The period segment corresponding to the frequency of the first The amplitude of each time frame; According to the following formula, the first The energy of each time frame of the periodic segment corresponding to each frequency is normalized: ； in, This represents the total number of frames in the periodic segment. Indicates the first The period segment corresponding to the frequency of the first The normalized probability of each time frame in this periodic segment; Calculate the first according to the following formula. The energy entropy of a periodic segment corresponding to a frequency : ； in, express The natural logarithm of .

6. The method for processing the sound signal of blood flow in an arteriovenous fistula according to claim 5, characterized in that, Calculate the query proxy matrix according to the following formula. : ； in, Represents the energy entropy matrix. This represents the transpose of the energy entropy matrix. This represents a linear transformation matrix whose elements are automatically adjusted as the model is trained.

7. The method for processing the sound signal of blood flow in an arteriovenous fistula according to claim 6, characterized in that, The calculation of the query matrix, key matrix, and value matrix of the agent attention mechanism based on the log-Mel spectrum includes: ； ； ； in, Represents the logarithmic Mel spectrum. express transpose, Represents the query matrix. Represents the key matrix, Represents a value matrix, , and Both represent linear transformation matrices whose elements are automatically adjusted as the model is trained.

8. The method for processing the sound signal of blood flow in an arteriovenous fistula according to claim 7, characterized in that, Also includes: The first attention score is calculated using the following formula. : ； in, () represents the normalized exponential function. Indicates querying the proxy matrix. Key matrix transpose, Represents matrix multiplication; The first aggregation feature matrix is ​​calculated using the following formula. : ； in, Represents a value matrix; The second attention score is calculated using the following formula. : ； in, () represents the normalized exponential function. Represents the query matrix. Indicates querying the proxy matrix. Represents matrix multiplication; The second aggregation feature matrix is ​​calculated using the following formula. : ； in, This represents the first feature output matrix. This represents matrix multiplication.

9. The method for processing the sound signal of blood flow in an arteriovenous fistula according to claim 1, characterized in that, The calculation of the position of each periodic segment in the log-Mel spectrum includes: The position index of each periodic segment in the log-Mel spectrum is calculated using the following formula. : ; in, Indicates the index of the segmentation point of the periodic segment. Sampling frequency, Preset window length, This indicates the preset frame shift.

10. An electronic device, characterized in that, include: One or more processors; 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 configured to implement the method for processing the blood flow sound signal of an arteriovenous fistula as described in any one of claims 1-9.

Images