CN110123367B

CN110123367B - Computer device, heart sound recognition method, model training device, and storage medium

Info

Publication number: CN110123367B
Application number: CN201910268611.7A
Authority: CN
Inventors: 康延妮; 李响; 贾晓雨; 绳立淼
Original assignee: Ping An Technology Shenzhen Co Ltd
Current assignee: Ping An Technology Shenzhen Co Ltd
Priority date: 2019-04-04
Filing date: 2019-04-04
Publication date: 2022-11-15
Anticipated expiration: 2039-04-04
Also published as: CN110123367A

Abstract

The application relates to the field of intelligent decision, and the local features of all stages of a heart sound period and the global features of a heart sound signal are integrated to identify the heart sound signal, so that the identification accuracy is higher. Specifically disclosed are a computer device, a heart sound recognition method, a model training device, and a storage medium; the computer device comprises a memory and a processor for executing a computer program stored by the memory and realizing, when executing the computer program: preprocessing the acquired heart sound signals; acquiring first heart sound data, systolic phase data, second heart sound data and diastolic phase data from the preprocessed heart sound signals; processing the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a processing rule to obtain local features; acquiring global features of the preprocessed heart sound signals according to the trained neural network model; and based on the trained heart sound recognition model, recognizing the type of the heart sound signal according to the local features and the global features.

Description

Computer device, heart sound recognition method, model training device, and storage medium

Technical Field

The present application relates to the field of health detection technologies, and in particular, to a computer device, a heart sound recognition apparatus, a heart sound recognition method, a model training apparatus, and a storage medium.

Background

The heart sound is one of sound signals, which contains important information about the health condition of a human body, and objective digital heart sound auscultation can be realized by extracting the information and carrying out effective identification, so that a reliable diagnosis result can be provided for a patient; therefore, the heart sound diagram has unique significance in the measurement of the cardiac function and the diagnosis of certain cardiovascular diseases.

The existing heart sound identification or classification methods are roughly divided into two categories, namely a classification method based on traditional machine learning and a classification method based on neural network learning; the former is based on manually extracted characteristic value learning, which can not truly and comprehensively reflect the essential characteristics of heart sound data, and the characteristic value extracted for the whole heart sound period can not learn accurate and targeted local information; the classification method based on neural network learning not only faces the difficulty of learning deep networks, but also needs a large amount of training data to improve the performance of the deep networks compared with the performance of a shallow architecture, and the recognition accuracy is low.

Disclosure of Invention

The embodiment of the application provides computer equipment, a heart sound recognition device, a heart sound recognition method, a model training device and a storage medium, local features of all phases of a heart sound period and global features of heart sound signals are integrated, and recognition accuracy is higher.

In a first aspect, the present application provides a computer device comprising a memory and a processor;

the memory is used for storing a computer program;

the processor is used for executing the computer program and realizing the following when the computer program is executed:

preprocessing the acquired heart sound signals;

acquiring first heart sound data, systolic phase data, second heart sound data and diastolic phase data from the preprocessed heart sound signals;

processing the first heart sound data, the systolic phase data, the second heart sound data and the diastolic phase data according to a preset processing rule to obtain local characteristics;

acquiring the global characteristics of the preprocessed heart sound signals according to the trained neural network model;

and identifying the type of the heart sound signal according to the local features and the global features based on the trained heart sound identification model.

In a second aspect, the present application provides a computer device comprising a memory and a processor; the memory is used for storing a computer program;

acquiring a training sample set, wherein the training sample set comprises a plurality of heart sound signals serving as training samples and labeling data of each heart sound signal;

acquiring first heart sound data, systolic phase data, second heart sound data and diastolic phase data from the heart sound signal;

acquiring global features of the heart sound signals according to the trained neural network model;

based on the heart sound identification model, identifying the type of the heart sound signal according to the local features and the global features to obtain prediction data of the heart sound signal;

and adjusting parameters of the heart sound identification model according to the prediction data and the labeling data of the heart sound signal.

In a third aspect, the present application provides a heart sound recognition apparatus, the apparatus comprising:

the preprocessing module is used for preprocessing the acquired heart sound signals;

the first acquisition module is used for acquiring first heart sound data, systolic data, second heart sound data and diastolic data from the preprocessed heart sound signals;

the second acquisition module is used for processing the first heart sound data, the systolic phase data, the second heart sound data and the diastolic phase data according to a preset processing rule to acquire local characteristics;

a third obtaining module, configured to obtain a global feature of the preprocessed heart sound signal according to the trained neural network model;

and the recognition module is used for recognizing the type of the heart sound signal according to the local features and the global features based on the trained heart sound recognition model.

In a fourth aspect, the present application provides a heart sound recognition model training device, including:

the fourth acquisition module is used for acquiring a training sample set, wherein the training sample set comprises a plurality of heart sound signals serving as training samples and labeling data of each heart sound signal;

a fifth obtaining module, configured to obtain the first heart sound data, the systolic data, the second heart sound data, and the diastolic data from the heart sound signal;

a sixth obtaining module, configured to process the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset processing rule to obtain a local feature;

a seventh obtaining module, configured to obtain global features of the heart sound signal according to the trained neural network model;

the identification module is used for identifying the type of the heart sound signal according to the local features and the global features based on the heart sound identification model so as to obtain the prediction data of the heart sound signal;

and the adjusting module is used for adjusting the parameters of the heart sound identification model according to the predicted data and the labeled data of the heart sound signal.

In a fifth aspect, the present application provides a computer-readable storage medium storing a computer program, where the computer program is executed by a processor to implement:

preprocessing the acquired heart sound signals;

In a sixth aspect, the present application provides a method for recognizing heart sounds, including:

preprocessing the acquired heart sound signals;

The application discloses a computer device, a heart sound recognition method, a model training device and a storage medium, wherein local features are obtained by processing first heart sound data, systolic phase data, second heart sound data and diastolic phase data according to preset processing rules, and global features of preprocessed heart sound signals are obtained according to a trained neural network model, so that the heart sound signals are recognized according to the local features of the first heart sound data, the systolic phase data, the second heart sound data and the diastolic phase data and the global features of the heart sound signals; therefore, the recognition process of the heart sound recognition model integrates the detail characteristics of each phase of the heart sound period and the global characteristics of the heart sound signal, and the recognition accuracy is higher.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the description below are some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.

Fig. 1 is a schematic structural diagram of a computer device according to an embodiment of the present application;

FIG. 2 is a flow chart illustrating the processing steps of a computer device implementing the heart sound identification method;

FIG. 3 is a schematic diagram of a frequency spectrum of a heart sound signal;

FIG. 4 is a waveform diagram of a heart sound signal before band-pass filtering;

FIG. 5 is a schematic waveform diagram of a band-pass filtered heart sound signal;

FIG. 6 is a schematic flow chart of the processing steps implemented by the computer device to stage the heart sound signal;

FIG. 7 is a flowchart illustrating processing steps performed by a computer device to obtain local features;

FIG. 8 is a flow chart illustrating the processing steps of a computer device in implementing the training of the neural network model;

FIG. 9 is a schematic diagram of a neural network model;

FIG. 10 is a flowchart illustrating processing steps implemented by a computer device in acquiring global features;

FIG. 11 is a flowchart illustrating processing steps of a computer device implementing the method for training a heart sound recognition model;

fig. 12 is a schematic structural diagram of a heart sound recognition apparatus according to an embodiment of the present application;

fig. 13 is a schematic structural diagram of a heart sound recognition model training device according to an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present application without making any creative effort belong to the protection scope of the present application.

The flowcharts shown in the figures are illustrative only and do not necessarily include all of the contents and operations/steps, nor do they necessarily have to be performed in the order described. For example, some operations/steps may be decomposed, combined or partially combined, so that the actual execution sequence may be changed according to the actual situation. In addition, although the division of the functional blocks is made in the device diagram, it may be divided in different blocks from that in the device diagram in some cases.

Some embodiments of the present application will be described in detail below with reference to the accompanying drawings. The embodiments described below and the features of the embodiments can be combined with each other without conflict.

Referring to fig. 1, fig. 1 is a schematic structural diagram of a computer device according to an embodiment of the present application. The computer device may be a server or a terminal.

Referring to fig. 1, the computer device includes a processor, a memory and a network interface connected by a system bus, wherein the memory may include a nonvolatile storage medium and an internal memory.

The processor is used for providing calculation and control capability and supporting the operation of the whole computer equipment.

The internal memory provides an environment for execution of a computer program in a non-volatile storage medium, which when executed by the processor, causes the processor to perform any one of the heart sound recognition methods.

The network interface is used for network communication, such as sending assigned tasks and the like. Those skilled in the art will appreciate that the architecture shown in fig. 1 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.

It should be understood that the Processor may be a Central Processing Unit (CPU), and the Processor may be other general purpose Processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Programmable logic device, discrete Gate or transistor logic, discrete hardware components, etc. Wherein a general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The non-volatile storage medium may store an operating system and a computer program, the computer program comprising program instructions; the processor is configured to execute the computer program.

In some embodiments, the program instructions, when executed by a processor, may cause the processor to implement a method of heart sound recognition.

In one embodiment, please refer to fig. 2, fig. 2 is a flowchart illustrating processing steps of a processor for executing a computer program in a memory to implement a heart sound identification method according to an embodiment of the present application. The processor specifically implements the following steps:

and step S110, preprocessing the acquired heart sound signals.

In some embodiments, the heart sound signals are collected by a professional digital stethoscope, and in other embodiments, the heart sound signals are collected by a portable mobile terminal connected with a heart sound sensor.

Illustratively, the acquired heart sound signals are electrocardiographic signal data in a digital format, such as Phonocardiogram (PCG). The electrocardiosignal can be intercepted according to a preset duration, for example, the electrocardiosignal with the maximum duration not more than 60 seconds is intercepted as an identification object.

FIG. 3 shows a spectrogram of a segment of a heart sound signal having a frequency centered below about 400Hz, and in some embodiments, a sampling frequency of 2000Hz of the heart sound signal obtained from the spectrogram; according to the Nyquist sampling theorem, it is known that enough signal information can be reserved if the sampling frequency is more than twice of the heart sound signal frequency; in order to satisfy the sampling theorem and reduce the calculation amount, the original heart sound signal with the sampling frequency of about 2000Hz can be subjected to down-sampling processing to 1000Hz.

Because some interference signals can be carried by friction, environmental noise and the like in the heart sound signal acquisition process, peak signals introduced in the acquisition process can be removed through band-pass filtering, and the preprocessed heart sound signals are obtained. As shown in fig. 4, which shows the heart sound signal before the band-pass filtering, and in fig. 5, which shows the heart sound signal after the band-pass filtering, it can be seen that the spike noise at the start of the heart sound signal is effectively filtered, and the low frequency interference and the high frequency noise are also effectively filtered.

Step S120, obtaining first heart sound data, systolic data, second heart sound data, and diastolic data from the preprocessed heart sound signal.

The heart sound signal comprises one or more heart sound cycles corresponding to one or more heart beats. Each heart sound cycle comprises a first heart sound (S1), a Systole (Systole), a second heart sound (S2) and a Diastole (Diastole); wherein the first heart sound occurs at the onset of systole, marking the onset of ventricular systole, and the second heart sound occurs at the onset of diastole, marking the onset of ventricular diastole. The heart sound period of the heart sound signal is divided into four stages, so that the heart sound signal can be identified more accurately, and abnormal heart sound in each stage can be identified in a targeted manner.

In some embodiments, as shown in fig. 6, when the processor performs step S120 to acquire the first heart sound data, the systolic data, the second heart sound data and the diastolic data from the preprocessed heart sound signal, the processor is configured to implement steps S121, S122:

step S121, dividing at least one heart sound period of the heart sound signal into a first heart sound stage, a systolic stage, a second heart sound stage and a diastolic stage according to R waves and T waves of the corresponding electrocardiogram of the heart sound signal.

In this embodiment, the preprocessed heart sound signal is segmented by the segmentation method proposed by Springer, and the heart sound period is segmented into: a first heart sound, a systolic phase, a second heart sound, and a diastolic phase. In particular, some or all of the heart sound periods in the heart sound signal may be segmented.

Illustratively, an electrocardiogram in the same time period as the heart sound signal is acquired and recorded, electrocardiogram features in the electrocardiogram are extracted, the electrocardiogram features comprise information of R waves and T waves, and at least one heart sound cycle of the heart sound signal is divided into four stages, namely a first heart sound stage, a systolic stage, a second heart sound stage and a diastolic stage according to the information of the R waves and the T waves.

And S122, acquiring data of the heart sound signal in four stages of the first heart sound, the systolic period, the second heart sound and the diastolic period.

Specifically, the data of the divided heart sound signal at each stage, i.e., the first heart sound data, the systolic data, the second heart sound data, and the diastolic data are stored.

Step S130, processing the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset processing rule to obtain a local feature.

In some embodiments, as shown in fig. 7, when the processor performs step S130 to process the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset processing rule to obtain the local feature, the processor is configured to implement at least one of steps S131 to S133:

step S131, processing the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset time domain analysis rule to obtain time domain characteristics.

In some embodiments, the step S131 processes the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset time domain analysis rule to obtain the time domain feature, specifically including: and performing time domain processing on the first heart sound data, the systolic period data, the second heart sound data and the diastolic period data to acquire heart sound interval data which can represent heart sound rhythm characteristics as time domain characteristics.

Illustratively, the heart sound interval data includes a mean and/or variance of at least one of an R-wave interval, a first heart sound interval, a second heart sound interval, a systolic interval, a diastolic interval, a ratio of systolic and heartbeat intervals, a ratio of diastolic and heartbeat intervals, and a ratio of systolic and diastolic.

One or more heart sound periods may be included in a volume of heart sound signals, or one or more heart sound periods in a volume of heart sound signals may be utilized; thus, intervals, ratios in the tone interval data may be averaged and/or squared; illustratively, a heart sound signal comprises 60 heart sound periods, and each interval and each ratio in the heart sound interval data takes a mean and/or a variance of the corresponding interval and ratio in the 60 heart sound periods.

For example, the time-domain feature includes a total of 16 time-domain feature values in the heart sound interval data.

In some embodiments, the step S131 processes the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset time domain analysis rule to obtain the time domain feature, specifically including: and performing time domain processing on the first heart sound data, the systolic data, the second heart sound data and the diastolic data to acquire heart sound amplitude data which can represent the heart sound amplitude characteristics and serve as time domain characteristics.

Illustratively, the heart sound amplitude data includes a mean and/or variance of at least one of a ratio of the systolic average amplitude value to the first heart sound average amplitude value, a ratio of the diastolic average amplitude value to the second heart sound average amplitude value, an amplitude skewness of the first heart sound, an amplitude skewness of the systolic phase (skewness), an amplitude skewness of the second heart sound, an amplitude skewness of the diastolic phase (kurtosis), an amplitude skewness of the first heart sound, an amplitude skewness of the systolic phase (kurtosis), an amplitude skewness of the second heart sound, and an amplitude skewness of the diastolic phase (kurtosis).

Illustratively, a volume of heart sound signals includes 60 heart sound periods, and the ratios, amplitude skewness, and amplitude kurtosis in the heart sound amplitude data may be the mean and/or variance of the corresponding ratios, amplitude skewness, and amplitude kurtosis in the 60 heart sound periods.

For example, the time-domain feature includes a total of 20 time-domain feature values in the heart sound amplitude data.

The difference between the heart sounds of a normal person and a heart disease patient can be analyzed to a certain extent from the aspects of the amplitude and the rhythm of the heart sounds.

Step S132, processing the first heart sound data, the systolic phase data, the second heart sound data and the diastolic phase data according to a preset frequency domain analysis rule to obtain frequency domain characteristics.

Illustratively, the frequency domain features are obtained by performing frequency domain analysis on the heart sound data of each heart sound stage, i.e., the first heart sound data, the systolic data, the second heart sound data, and the diastolic data.

In some embodiments, the step S132 of processing the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset frequency domain analysis rule to obtain the frequency domain feature specifically includes: and performing frequency domain analysis on the first heart sound data, the systolic data, the second heart sound data and the diastolic data to acquire a median value of spectral amplitudes of the first heart sound data, the systolic data, the second heart sound data and the diastolic data on a plurality of different frequency bands respectively as frequency domain characteristics.

Specifically, the first heart sound data, the systolic period data, the second heart sound data and the diastolic period data are multiplied by a Hamming window respectively, energy distribution of data of each phase on a frequency spectrum is obtained through fast Fourier transform, and a median (namely a median) of the frequency spectrum amplitude is obtained to be used as a frequency domain characteristic of a corresponding phase of the heart sound signal.

According to the research of human auditory mechanism, human ears have different auditory sensitivities to sound waves with different frequencies. Bass is easy to mask treble, while treble is more difficult to mask bass, so the critical bandwidth of sound masking at low frequencies is smaller than at high frequencies; in consideration of this characteristic, the filter is set so that a group of band-pass filters are arranged from dense to sparse in the band from low frequency to high frequency to filter the input signal. Illustratively, nine frequency bands 25-45, 45-65, 65-85, 85-105, 105-125, 125-150, 150-200, 200-300, 300-400 are set in Hz and nine bandpass filters are set accordingly. Inputting the heart sound data of each stage into each band-pass filter, taking the signal energy output by each band-pass filter as the basic characteristic of the signal, and calculating the median (namely the median) of the spectrum amplitude as the frequency domain characteristic of the corresponding stage of the heart sound signal according to the energy distribution of the data of each stage on the spectrum.

When nine bands are set, there are 4 × 9=36 spectral median amplitudes, i.e., frequency-domain eigenvalues, in all of the four stages.

And S133, processing the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset cepstrum processing rule to acquire Mel cepstrum characteristics.

Mel (Mel) cepstrum is often applied in sound signal processing. The human auditory system is a special nonlinear system whose sensitivity to signals of different frequencies is different. The Mel Frequency Cepstrum Coefficient (MFCC) takes human auditory features into consideration, and first maps a linear spectrum into a Mel nonlinear spectrum based on auditory perception, and then converts the Mel nonlinear spectrum onto a Cepstrum.

The shape of the vocal tract is shown in the envelope of the speech short-time power spectrum, and mel-frequency cepstrum features, such as mel-frequency cepstrum coefficients, are a feature that can accurately describe the envelope. Mel cepstrum is closer to the analysis characteristic of human ear to sound than cepstrum to sound signal processing.

In some embodiments, the first heart sound data, the systolic data, the second heart sound data, and the diastolic data are pre-emphasized, framed, hamming windowed, and mel-cepstral analyzed to obtain respective pluralities of mel-frequency cepstral coefficients. Illustratively, the process of extracting data mel-frequency cepstrum coefficients of a certain stage in the heart sound signal comprises: firstly, carrying out pre-emphasis processing, framing and Hamming window processing on the data, wherein the pre-emphasis is used for emphasizing the high-frequency part of the sound and increasing the high-frequency resolution of the sound; the framing processing is to divide the data into short segments for processing, such as dividing the data per second into 33-100 frames, and the framing can be implemented by weighting a movable window with a limited length, for example; the Haiming window adding process can enable data to be global and continuous, the Gibbs effect is avoided, and after the Haiming window is added, no periodic data signal originally presents partial characteristics of a periodic function. Then, fourier transform is carried out on each short-time analysis window processed by the Haiming window to obtain a signal spectrum, and the linear signal spectrum is mapped into a Mel nonlinear spectrum based on auditory perception through a Mel filter bank to obtain a Mel spectrum; and then carrying out Mel cepstrum analysis on the Mel frequency spectrum, specifically comprising logarithm taking and DCT discrete cosine transform to obtain Mel frequency cepstrum coefficients.

In some embodiments, the mel-cepstral features of the first, systolic, second and diastolic heart sound data comprise a plurality of mel-frequency cepstral coefficients for each of the first, systolic, second and diastolic heart sound data.

Illustratively, 13 mel-frequency cepstrum coefficients are extracted for each of the first heart sound data, the systolic data, the second heart sound data and the diastolic data, and the mel-frequency cepstrum features of one heart sound signal totally include 4 × 13=52 mel-frequency cepstrum coefficients.

According to the discrete characteristic values of the time domain characteristic, the frequency domain characteristic and the mel cepstrum characteristic of the first heart sound data, the systolic data, the second heart sound data and the diastolic data obtained in the steps S131 to S133, the discrete characteristic values can be used for preliminarily judging whether abnormal heart sounds exist in each stage, such as the R wave interval of a normal person and the R wave interval of a heart disease patient, and can also be used as a basis for calibrating and labeling data for heart sound signals serving as training samples in a training sample set.

And S140, acquiring the global characteristics of the preprocessed heart sound signals according to the trained neural network model.

In some possible embodiments, as shown in fig. 8, when executing the training step of the neural network model, the processor implements the following steps S101 to S103:

step S101, a training sample set is obtained, wherein the training sample set comprises a plurality of heart sound signals serving as training samples and labeling data of each heart sound signal.

Illustratively, part of the heart sound signals without the heart sound abnormal condition and part of the heart sound signals with the heart sound abnormal condition are collected to be used as training samples; labeling data are marked on each heart sound signal in the training sample, wherein the labeling data can comprise the existence of abnormality of the heart sound type, first heart sound abnormality, diastole abnormality and the like; in another embodiment the annotation data further comprises probability values for the heart sound signal for the respective heart sound category.

Step S102, the heart sound signal serving as the training sample is identified according to the neural network model so as to obtain prediction data of the heart sound signal.

Illustratively, the prediction data includes a predicted heart sound type and may further include a probability corresponding to the predicted heart sound type.

And S103, adjusting parameters of a neural network model according to the prediction data and the labeling data of the heart sound signals.

Illustratively, adam can be used to randomly optimize the neural network model, and cross entropy is selected as a loss function to minimize. And training the neural network model through a plurality of samples in the training sample set, and finishing the training when the deviation between the predicted data and the labeled data is smaller than a preset threshold value to obtain the trained neural network model.

In some embodiments, when the processor performs step S140 to obtain the global features of the preprocessed heart sound signals according to the trained neural network model, the processor is specifically configured to implement:

and performing convolution operation, pooling operation and full-connection operation on the preprocessed heart sound signals according to the trained neural network model to output implicit characteristics as global characteristics.

A neural network generally comprises an input layer, an output layer and an intermediate layer, also called a hidden layer, between the input layer and the output layer; the hidden layer may provide a covert feature; since the input layer inputs the heart sound signal instead of the aforementioned first heart sound data, systolic data, second heart sound data and diastolic data, these implicit features are global features that can characterize the heart sound signal.

In some embodiments, the neural network model is structured as shown in fig. 9. As shown in fig. 10, when the processor performs at least convolution operation, pooling operation, and full-link operation on the preprocessed heart sound signal according to the trained neural network model to output an implicit feature as a global feature, the processor is specifically configured to implement steps S1411 to S1415:

step S1411, a plurality of heart sound waveform data of a plurality of different frequency bands are extracted from the preprocessed heart sound signal.

Illustratively, inupt1, input2, input3 and Input4 are heart sound waveform data after 4 frequency bands, such as 25-45, 45-80, 80-200 and 200-400 (Hz) filtering, are respectively passed by a heart sound signal every heartbeat, and the length of the heart sound waveform data of each frequency band is 2500.

In step S1412, the first feature vector of each piece of the heart sound waveform data is extracted.

The length of each heart sound waveform data output after passing through the input layer is also 2500, and the dimension is 1, and is represented by (2500,1).

Step S1413, performing convolution operation and pooling operation on each of the first feature vectors to output a second feature vector corresponding to each of the first feature vectors.

And processing each first feature vector of different frequency bands by a first convolution conv, a first maximum pooling max _ pooling, a second convolution conv, a second maximum pooling max _ pooling and a scatter operation to obtain a corresponding second feature vector. Wherein the flatten operation processes the higher-dimensional Tensor (Tensor) into a one-dimensional Tensor (vector).

The output of the first eigenvector of (2500, 1) after the first convolution conv is (2496, 8), i.e. the length is 2496 and the dimension is 8; the output of the vector of (2496, 8) for the first maximum pooling max _ pooling is (1248, 8), (1248, 8) for the second convolution conv is (1244, 4), (1244, 4) for the second maximum pooling max _ pooling is (622, 4); the output of the vector of (622, 4) via the flatten operation is a one-dimensional second feature vector of length 2488.

Step S1414, integrating all the second feature vectors into a third feature vector.

As shown in fig. 9, flatten: the concatemate layer integrates four second eigenvectors corresponding to different frequency bands into one third eigenvector, which has a length of 9952 and a dimension of 1.

And step S1415, performing full connection operation on the third feature vector to output a recessive feature as a global feature.

Illustratively, the length is 9952, the third eigenvector with the dimension of 1 is processed by a dropout layer, and the output length is 9952; then processed into a vector with the length of 20 by a Dense layer; the length-20 vector may include a feature value as the implicit feature.

When the neural network model is trained, the vector with the length of 20 is processed through another dropout layer and another Dense layer, and prediction data of the heart sound signal are obtained.

The heart sound signals are divided into different frequency bands by applying multi-channel filtering, a neural network model is trained, discrete characteristic values of time domain characteristics, frequency domain characteristics and Mel cepstrum characteristics of first heart sound data, systolic phase data, second heart sound data and diastolic phase data are integrated into a training process through back propagation by taking the discrete characteristic values as a group of constraint conditions in deep learning, the characteristic expression capability of the neural network model is enhanced, and the prediction accuracy and the global characteristics, such as the expression capability of recessive characteristics, are improved.

And S150, identifying the type of the heart sound signal according to the local features and the global features based on the trained heart sound identification model.

Illustratively, the local features include 16 time domain feature values in the heart sound interval data obtained in step S131, 20 time domain feature values in the heart sound amplitude data, 36 frequency domain feature values obtained in step S132, and 52 mel-frequency cepstrum coefficients obtained in step S133, and there are 124 local features in total; the global features include 20 implicit features, namely global features; the heart sound identification model identifies the type of the heart sound signal according to the 124 local features and the 20 global features so as to identify that the type of the heart sound signal is a normal heart sound or an abnormal heart sound.

In some embodiments, the heart sound recognition model comprises a CatBoost model.

The Boosting algorithm has the indispensable advantages in the scenes of limited training sample size, short required training time, lack of parameter adjusting knowledge and the like. The Catboost is a Gradient boost library, is a combination of a Gradient Boosting technology and a category Features technology, and is also a machine learning framework based on a Gradient boost decision tree. The robustness is high, the requirement for tuning many hyper-parameters is reduced, and the chance of over-fitting is reduced.

An embodiment of the present invention further provides a method for recognizing a heart sound, which includes, as shown in fig. 2:

step S110, preprocessing the acquired heart sound signals;

step S120, acquiring first heart sound data, systolic phase data, second heart sound data and diastolic phase data from the preprocessed heart sound signals;

step S130, processing the first heart sound data, the systolic period data, the second heart sound data and the diastolic period data according to a preset processing rule to obtain local characteristics;

step S140, acquiring the global characteristics of the preprocessed heart sound signals according to the trained neural network model;

In the computer device and the heart sound recognition method provided in the above embodiments, when a processor of the computer device executes a computer program in a memory, the heart sound recognition method is implemented, by processing the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset processing rule to obtain local features, and obtaining global features of the preprocessed heart sound signals according to a trained neural network model, the heart sound signals are recognized according to not only the local features of the first heart sound data, the systolic data, the second heart sound data, and the diastolic data, but also the global features of the heart sound signals; therefore, the recognition process of the heart sound recognition model integrates the detail characteristics of each phase of the heart sound period and the global characteristics of the heart sound signal, and the recognition accuracy is higher.

Moreover, because the heart sound period is divided into a plurality of phases, a certain type of heart diseases can be identified in a targeted mode.

In other possible embodiments, the non-volatile storage medium may store an operating system and a computer program, the computer program comprising program instructions; the processor is configured to execute the computer program. The program instructions, when executed by the processor, may cause the processor to implement a method of training a heart sound recognition model.

In one embodiment, please refer to fig. 11, fig. 11 is a flowchart illustrating processing steps of a processor for executing a computer program in a memory to implement a method for training a heart sound recognition model according to an embodiment of the present application. The processor specifically implements the following steps:

step S210, a training sample set is obtained, where the training sample set includes a plurality of heart sound signals as training samples and labeling data of each heart sound signal.

Illustratively, part of the heart sound signals without heart sound abnormal conditions and part of the heart sound signals with the heart sound abnormal conditions are collected as training samples; marking labeled data on each heart sound signal in the training sample; the annotation data may include the heart sound category, such as whether there is an abnormality, a first heart sound abnormality, a diastolic abnormality, etc.; in another embodiment the annotation data further comprises probability values for the heart sound signal for the respective heart sound category.

In some embodiments, the heart sound signals in the training sample set have been pre-processed.

Step S220, obtaining first heart sound data, systolic phase data, second heart sound data and diastolic phase data from the heart sound signal.

And step S230, processing the first heart sound data, the systolic period data, the second heart sound data and the diastolic period data according to a preset processing rule to obtain local characteristics.

In some embodiments, when executing step S230, the processor is configured to process the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset processing rule to obtain the local feature, to implement:

processing the first heart sound data, the systolic phase data, the second heart sound data and the diastolic phase data according to a preset time domain analysis rule to obtain a time domain characteristic; and/or

Processing the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset frequency domain analysis rule to obtain frequency domain characteristics; and/or

And processing the first heart sound data, the systolic period data, the second heart sound data and the diastolic period data according to a preset cepstrum processing rule to obtain Mel cepstrum characteristics.

In some embodiments, the processor, when executing the processing of the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset time domain analysis rule to obtain a time domain feature, is configured to:

and performing time domain processing on the first heart sound data, the systolic phase data, the second heart sound data and the diastolic phase data to acquire heart sound interval data and heart sound amplitude data, wherein the heart sound interval data are used for representing heart sound rhythm characteristics, and the heart sound amplitude data are used for representing heart sound amplitude characteristics.

The processor is configured to implement, when performing the processing of the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset frequency domain analysis rule to obtain frequency domain characteristics:

performing frequency domain analysis on the first heart sound data, the systolic data, the second heart sound data and the diastolic data to acquire a median value of spectral amplitudes of the first heart sound data, the systolic data, the second heart sound data and the diastolic data on a plurality of different frequency bands respectively as frequency domain characteristics;

the processor is configured to, when executing the processing of the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset cepstrum processing rule to obtain mel cepstrum features, implement:

pre-emphasis processing, framing, haiming window processing, and Mel cepstral analysis are performed on the first heart sound data, systolic data, second heart sound data, and diastolic data to obtain respective multiple Mel frequency cepstral coefficients.

And S240, acquiring the global characteristics of the heart sound signals according to the trained neural network model.

In some embodiments, when the processor performs step S240 to obtain the global features of the preprocessed heart sound signals according to the trained neural network model, the processor is configured to:

and performing convolution operation, pooling operation and full-connection operation on the heart sound signal according to the trained neural network model to output implicit features as global features.

In some embodiments, the processor, when executing at least a convolution operation, a pooling operation, and a full-link operation on the heart sound signal according to the trained neural network model to output an implicit feature as the global feature, is configured to:

extracting a plurality of heart sound waveform data of a plurality of different frequency bands from the heart sound signal; extracting a first feature vector of each piece of heart sound waveform data; performing convolution operation and pooling operation on each first feature vector to output a second feature vector corresponding to each first feature vector; integrating all the second feature vectors into a third feature vector; and performing full-connection operation on the third feature vector to output an implicit feature as a global feature.

And S250, identifying the type of the heart sound signal according to the local feature and the global feature based on the heart sound identification model to obtain the prediction data of the heart sound signal.

Specifically, based on a trained heart sound recognition model, the type of the heart sound signal is recognized according to the local features and the global features to obtain prediction data of the heart sound signal. Illustratively, the prediction data includes a predicted heart sound type and may further include a probability corresponding to the predicted heart sound type.

In some embodiments, based on the same training sample set, a neural network model may be trained for obtaining global features of the preprocessed heart sound signals according to the trained neural network model.

And S260, adjusting parameters of the heart sound identification model according to the prediction data and the labeling data of the heart sound signal.

In some embodiments, 80% of the training samples in the training sample set are input into the heart sound recognition model for training, and the remaining 20% of the training samples are used for testing the result output by the heart sound recognition model. If the data is predicted, namely the test result is greater than a preset threshold value, marking as 1; if the test result is smaller than the preset threshold value, marking the test result as 0; illustratively, the preset threshold is 0.5; then comparing the test result with the labeled data of the training sample; illustratively, the label data is 1 or 0, where 1 represents a heart sound abnormality and 0 represents a heart sound abnormality; if the predicted data and the marked data are the same, the test result is correct; and if the predicted data and the marked data are different, the test result is wrong.

Recording the test results, for example, the number of correct test results is m, and the number of errors in test results is n; the training effect evaluation of the test result is checked by the ACC check, for example, a value of m ÷ (m + n) is calculated, and if the value is less than a preset value, such as 2%, the training can be stopped; if the value is larger than the preset value, continuing training and adjusting parameters of the heart sound recognition model; so as to better judge whether the heart sound signal is abnormal or not and improve the accuracy of abnormal heart sound identification.

In the computer device provided by the above embodiment, the processor implements the heart sound recognition model training method when executing the computer program in the memory; the method comprises the steps of predicting and identifying a heart sound signal by acquiring local characteristics of first heart sound data, systolic data, second heart sound data and diastolic data of the heart sound signal serving as a training sample and acquiring global characteristics of the heart sound signal, so as to adjust parameters of a heart sound identification model according to predicted prediction data and labeled data of the training sample; the training can be completed by a small number of heart sound samples, and the heart sound recognition model obtained by training can recognize the heart sound signals according to the local characteristics of the first heart sound data, the systolic data, the second heart sound data and the diastolic data and the global characteristics of the heart sound signals, so that the recognition accuracy is higher.

In some embodiments, the heart sound recognition method and the heart sound recognition model training method may be applied to a terminal or a server, and therefore, the trained models need to be stored in the terminal or the server. The terminal can be an electronic device such as a mobile phone, a tablet computer, a notebook computer, a desktop computer, a personal digital assistant and a wearable device; the servers may be independent servers or server clusters.

If the method is applied to the terminal, in order to ensure the normal operation of the terminal and the rapid and effective identification and classification, the trained heart sound identification model and the trained neural network model need to be compressed, and the compressed model is stored in the terminal.

The compression processing specifically comprises pruning processing, quantization processing, huffman coding processing and the like on the heart sound recognition model and the neural network model so as to reduce the sizes of the heart sound recognition model and the neural network model and further facilitate storage in a terminal with smaller capacity.

Fig. 12 is a schematic structural diagram of a heart sound recognition apparatus according to an embodiment of the present application, which can be configured in a server or a terminal for executing the aforementioned heart sound recognition method.

As shown in fig. 12, the heart sound recognition apparatus includes:

and a preprocessing module 110, configured to preprocess the acquired heart sound signal.

A first obtaining module 120, configured to obtain first heart sound data, systolic data, second heart sound data, and diastolic data from the preprocessed heart sound signal.

A second obtaining module 130, configured to process the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset processing rule to obtain a local feature.

Specifically, the second obtaining module 130 includes

The first acquisition unit is used for processing the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset time domain analysis rule to acquire a time domain feature; and/or

The second acquisition unit is used for processing the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset frequency domain analysis rule to acquire frequency domain characteristics; and/or

And the third acquisition unit is used for processing the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset cepstrum processing rule so as to acquire Mel cepstrum characteristics.

Specifically, the first obtaining unit is configured to perform time domain processing on the first heart sound data, the systolic data, the second heart sound data, and the diastolic data to obtain heart sound interval data and heart sound amplitude data, where the heart sound interval data is data used for representing a rhythm characteristic of a heart sound, and the heart sound amplitude data is heart sound amplitude data used for representing a heart sound amplitude characteristic.

Illustratively, the time-domain features include a mean and/or variance of at least one of an R-wave interval, a first heart sound interval, a second heart sound interval, a systolic interval, a diastolic interval, a ratio of systolic and heartbeat intervals, a ratio of diastolic and heartbeat intervals, a ratio of systolic and diastolic, a ratio of systolic mean magnitude value to first heart sound mean magnitude value, a ratio of diastolic mean magnitude value to second heart sound mean magnitude value, a magnitude skewness of the first heart sound, a magnitude skewness of the systolic, a magnitude skewness of the second heart sound, a magnitude skewness of the diastolic, a magnitude kurtosis of the first heart sound, a magnitude kurtosis of the systolic, a magnitude kurtosis of the second heart sound, a magnitude kurtosis of the diastolic.

Specifically, the second obtaining unit is configured to perform frequency domain analysis on the first heart sound data, the systolic data, the second heart sound data, and the diastolic data to obtain a median of spectral amplitudes of the first heart sound data, the systolic data, the second heart sound data, and the diastolic data in a plurality of different frequency bands, respectively, as a frequency domain feature.

The frequency domain features include spectral median amplitudes of the first heart sound data, the systolic data, the second heart sound data, and the diastolic data over a plurality of different frequency bands, respectively.

Specifically, the third obtaining unit is configured to perform pre-emphasis processing, framing, haiming window adding processing, and mel-frequency cepstrum analysis on the first heart sound data, the systolic data, the second heart sound data, and the diastolic data to obtain a plurality of mel-frequency cepstrum coefficients.

The mel cepstral features of the first, systolic, second and diastolic heart sound data comprise a plurality of mel-frequency cepstral coefficients for each of the first, systolic, second and diastolic heart sound data.

A third obtaining module 140, configured to obtain global features of the preprocessed heart sound signals according to the trained neural network model.

Specifically, the third obtaining module 140 is configured to perform at least convolution operation, pooling operation, and full connection operation on the preprocessed heart sound signal according to the trained neural network model, so as to output a hidden feature as a global feature.

Specifically, the third obtaining module 140 includes:

a first extraction unit configured to extract a plurality of heart sound waveform data of a plurality of different frequency bands from the preprocessed heart sound signal;

a second extraction unit configured to extract a first feature vector of each of the heart sound waveform data;

a first operation unit, configured to perform convolution operation and pooling operation on each of the first feature vectors to output a second feature vector corresponding to each of the first feature vectors;

a first integration unit, configured to integrate all the second feature vectors into a third feature vector;

and the first full connection unit is used for performing full connection operation on the third feature vector so as to output a recessive feature as a global feature.

And the recognition module 150 is configured to recognize the type of the heart sound signal according to the local features and the global features based on the trained heart sound recognition model.

Specifically, the heart sound identification model includes a Catboost model.

Fig. 13 is a schematic structural diagram of a heart sound recognition model training device according to an embodiment of the present invention, which can be configured in a server or a terminal for executing the aforementioned heart sound recognition model training method.

As shown in fig. 13, the heart sound recognition model training device includes:

a fourth obtaining module 210, configured to obtain a training sample set, where the training sample set includes a plurality of heart sound signals serving as training samples and labeling data of each of the heart sound signals.

A fifth obtaining module 220, configured to obtain the first heart sound data, the systolic data, the second heart sound data, and the diastolic data from the heart sound signal.

A sixth obtaining module 230, configured to process the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset processing rule to obtain a local feature.

Specifically, the sixth obtaining module 230 includes:

the fourth acquisition unit is used for processing the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset time domain analysis rule to acquire time domain characteristics; and/or

A fifth obtaining unit, configured to process the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset frequency domain analysis rule to obtain a frequency domain feature; and/or

A sixth obtaining unit, configured to process the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset cepstrum processing rule to obtain a mel cepstrum feature.

And a seventh obtaining module 240, configured to obtain the global features of the heart sound signal according to the trained neural network model.

Specifically, the seventh obtaining module 240 includes:

a third extraction unit configured to extract a plurality of pieces of heart sound waveform data of a plurality of different frequency bands from the heart sound signal;

a fourth extraction unit configured to extract a first feature vector of each of the cardioid data;

a second operation unit, configured to perform convolution operation and pooling operation on each of the first feature vectors to output a second feature vector corresponding to each of the first feature vectors;

a second integration unit, configured to integrate all the second feature vectors into a third feature vector;

and the second full-connection unit is used for performing full-connection operation on the third feature vector to output a recessive feature as a global feature.

An identifying module 250, configured to identify a type of the heart sound signal according to the local feature and the global feature based on the heart sound identification model to obtain prediction data of the heart sound signal.

An adjusting module 260, configured to adjust parameters of the heart sound identification model according to the prediction data and the labeling data of the heart sound signal.

It should be noted that, as will be clear to those skilled in the art, for convenience and brevity of description, the specific working processes of the apparatus, the modules and the units described above may refer to the corresponding processes in the foregoing computer device embodiment, and are not described herein again.

The computing device may be implemented in numerous general purpose or special purpose computing system environments or configurations. For example: personal computers, server computers, hand-held or portable devices, tablet-type devices, multiprocessor systems, microprocessor-based systems, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments that include any of the above systems or devices, and the like.

From the above description of the embodiments, it is clear to those skilled in the art that the present application can be implemented by software plus necessary general hardware platform. Based on such understanding, the technical solution of the present application may be substantially implemented or contributed to by the prior art in the form of a software product, which may be stored in a storage medium, such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method according to the embodiments or some parts of the embodiments of the present application, such as:

a computer-readable storage medium, where a computer program is stored, where the computer program includes program instructions, and a processor executes the program instructions to implement any heart sound identification method provided in this application embodiment or any heart sound identification model training method provided in this application embodiment.

In some embodiments, the computer readable storage medium stores a computer program that, if executed by a processor, implements:

preprocessing the acquired heart sound signals;

The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiment, for example, a hard disk or a memory of the computer device. The computer readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card), and the like, provided on the computer device.

While the invention has been described with reference to specific embodiments, the scope of the invention is not limited thereto, and those skilled in the art can easily conceive various equivalent modifications or substitutions within the technical scope of the invention. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A computer device, wherein the computer device comprises a memory and a processor;

the memory is used for storing a computer program;

preprocessing the acquired heart sound signals;

processing the first heart sound data, the systolic data, the second heart sound data and the diastolic data according to a preset processing rule to obtain local features;

based on a trained heart sound recognition model, recognizing the type of the heart sound signal according to the local features and the global features;

wherein, when the processor executes the processing of the first heart sound data, the systolic phase data, the second heart sound data and the diastolic phase data according to the preset processing rule to obtain the local features, the processor is configured to:

and performing frequency domain analysis on the first heart sound data, the systolic data, the second heart sound data and the diastolic data to acquire a median of spectral amplitudes of the first heart sound data, the systolic data, the second heart sound data and the diastolic data respectively on a plurality of different frequency bands as frequency domain features, wherein critical bandwidth sizes of all band-pass filters for acquiring the frequency domain features are arranged from dense to sparse in a range from low frequency to high frequency.

2. The computer device of claim 1, wherein the processor, when executing the processing of the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to preset processing rules to obtain local features, is further configured to:

3. The computer device of claim 2, wherein the processor, when executing the processing of the first heart sound data, the systolic phase data, the second heart sound data and the diastolic phase data according to preset temporal analysis rules to obtain temporal features, is configured to:

performing time domain processing on the first heart sound data, the systolic phase data, the second heart sound data and the diastolic phase data to acquire heart sound interval data and heart sound amplitude data, wherein the heart sound interval data are used for representing heart sound rhythm characteristics, and the heart sound amplitude data are used for representing heart sound amplitude characteristics;

pre-emphasis processing, framing, hamming window processing, and mel-frequency cepstral analysis are performed on the first heart sound data, the systolic data, the second heart sound data, and the diastolic data to obtain respective pluralities of mel-frequency cepstral coefficients.

4. The computer device of claim 1, wherein the processor, when executing the obtaining global features of the preprocessed heart sound signals according to the trained neural network model, is configured to:

5. The computer device of claim 4, wherein the processor, when executing the convolution operation, pooling operation, and full join operation on the preprocessed heart sound signals according to the trained neural network model to output implicit features as global features, is configured to:

extracting a plurality of heart sound waveform data of a plurality of different frequency bands from the preprocessed heart sound signal;

extracting a first feature vector of each piece of heart sound waveform data;

performing convolution operation and pooling operation on each first feature vector to output a second feature vector corresponding to each first feature vector;

integrating all the second feature vectors into a third feature vector;

and performing full-connection operation on the third feature vector to output an implicit feature as a global feature.

6. A computer device, wherein the computer device comprises a memory and a processor; the memory is used for storing a computer program;

acquiring global characteristics of the heart sound signals according to the trained neural network model;

adjusting parameters of the heart sound identification model according to the prediction data and the labeling data of the heart sound signals;

wherein, the processor is configured to perform the processing on the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset processing rule to obtain a local feature, and is configured to:

7. A heart sound identification apparatus, comprising:

the first acquisition module is used for acquiring first heart sound data, systolic phase data, second heart sound data and diastolic phase data from the preprocessed heart sound signals;

the recognition module is used for recognizing the type of the heart sound signal according to the local features and the global features based on a trained heart sound recognition model;

the second obtaining module is configured to, when processing the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset processing rule to obtain a local feature:

8. A heart sound recognition model training device, comprising:

a fifth obtaining module, configured to obtain first heart sound data, systolic data, second heart sound data, and diastolic data from the heart sound signal;

the adjusting module is used for adjusting parameters of the heart sound identification model according to the prediction data and the labeling data of the heart sound signals;

the sixth obtaining module is configured to, when processing the first heart sound data, the systolic data, the second heart sound data, and the diastolic data according to a preset processing rule to obtain a local feature:

9. A computer-readable storage medium storing a computer program, wherein if the computer program is executed by a processor, the computer program implements:

preprocessing the acquired heart sound signals;

acquiring first heart sound data, systolic data, second heart sound data and diastolic data from the preprocessed heart sound signals;

wherein, the processing the first heart sound data, the systolic period data, the second heart sound data and the diastolic period data according to a preset processing rule to obtain local features comprises:

10. A method for recognizing a heart sound, comprising:

preprocessing the acquired heart sound signals;

performing frequency domain analysis on the first heart sound data, the systolic data, the second heart sound data and the diastolic data to acquire a median of spectral amplitudes of the first heart sound data, the systolic data, the second heart sound data and the diastolic data respectively on a plurality of different frequency bands as a frequency domain feature, wherein critical bandwidth sizes of the band-pass filters for acquiring the frequency domain feature are arranged from dense to sparse in a range from low frequency to high frequency.