CN110123367A - Computer equipment, recognition of heart sound device, method, model training apparatus and storage medium - Google Patents
Computer equipment, recognition of heart sound device, method, model training apparatus and storage medium Download PDFInfo
- Publication number
- CN110123367A CN110123367A CN201910268611.7A CN201910268611A CN110123367A CN 110123367 A CN110123367 A CN 110123367A CN 201910268611 A CN201910268611 A CN 201910268611A CN 110123367 A CN110123367 A CN 110123367A
- Authority
- CN
- China
- Prior art keywords
- data
- heart sound
- cardiechema signals
- diastole
- sound data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B7/00—Instruments for auscultation
- A61B7/02—Stethoscopes
- A61B7/04—Electric stethoscopes
Landscapes
- Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Heart & Thoracic Surgery (AREA)
- Medical Informatics (AREA)
- Molecular Biology (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- General Health & Medical Sciences (AREA)
- Public Health (AREA)
- Veterinary Medicine (AREA)
- Measuring Pulse, Heart Rate, Blood Pressure Or Blood Flow (AREA)
Abstract
This application involves the global characteristics in intelligent decision field, the local feature and cardiechema signals that combine heart sound cycle stages to identify to cardiechema signals, and the accuracy of identification is higher.Specifically disclose computer equipment, recognition of heart sound device, method, model training apparatus and storage medium；Computer equipment includes memory and processor, and processor is used to execute the computer program of memory storage and the realization when executing computer program: pre-processing to the cardiechema signals of acquisition；First heart sound data, systole phase data, second heart sound data and diastole data are obtained from pretreated cardiechema signals；First heart sound data, systole phase data, second heart sound data and diastole data are handled to obtain local feature according to processing rule；The global characteristics of pretreated cardiechema signals are obtained according to trained neural network model；Based on trained recognition of heart sound model, the type of cardiechema signals is identified according to local feature and global characteristics.
Description
Technical field
This application involves health detection technical field more particularly to a kind of computer equipment, recognition of heart sound device, method,
Model training apparatus and storage medium.
Background technique
Heart sound is one kind of voice signal, contains the important information in relation to human health status, this by extracting
Information and progress, which effectively identify, can be realized objectively digitlization cardiophony, tie so as to provide reliably to diagnose for patient
Fruit；Thus caardiophonogram has unique meaning in the diagnosis of cardiac determination and certain cardiovascular diseases.
Existing recognition of heart sound or classification method are roughly divided into two classes, based on conventional machines study classification method and be based on
The classification method of neural network learning；Wherein, the former is the characteristic value study based on manual extraction, cannot be true and comprehensively anti-
Heart sound data substantive characteristics is reflected, and entire heart sound periodicity extraction characteristic value cannot be learnt to accurate targetedly part
Information；And the classification method based on neural network learning, the difficulty of study deep layer network is not only faced, but also need a large amount of instruction
Practicing data can just accomplish that the performance than shallow-layer framework improves to some extent, and the accuracy of identification is lower.
Summary of the invention
The embodiment of the present application provides a kind of computer equipment, recognition of heart sound device, method, model training apparatus and storage and is situated between
Matter, combines the local feature of heart sound cycle stages and the global characteristics of cardiechema signals, and the accuracy of identification is higher.
In a first aspect, the computer equipment includes memory and processor this application provides a kind of computer equipment；
The memory is for storing computer program；
The processor, for executing the computer program and the realization when executing the computer program:
The cardiechema signals of acquisition are pre-processed；
First heart sound data, systole phase data, second heart sound data and diastole are obtained from pretreated cardiechema signals
Data；
According to default processing rule to the first heart sound data, systole phase data, second heart sound data and diastole issue
Local feature is obtained according to being handled；
The global characteristics of the pretreated cardiechema signals are obtained according to trained neural network model；
Based on trained recognition of heart sound model, the heart sound is believed according to the local feature and the global characteristics
Number type identified.
Second aspect, this application provides a kind of computer equipments, which is characterized in that the computer equipment includes storage
Device and processor；The memory is for storing computer program；
The processor, for executing the computer program and the realization when executing the computer program:
Obtain training sample set, the training sample set includes several cardiechema signals as training sample and each described
The labeled data of cardiechema signals；
First heart sound data, systole phase data, second heart sound data and diastole data are obtained from the cardiechema signals；
According to default processing rule to the first heart sound data, systole phase data, second heart sound data and diastole issue
Local feature is obtained according to being handled；
The global characteristics of the cardiechema signals are obtained according to trained neural network model；
Based on the recognition of heart sound model, according to the local feature and the global characteristics to the cardiechema signals
Type is identified to obtain the prediction data of the cardiechema signals；
The parameter of the recognition of heart sound model is adjusted according to the prediction data of the cardiechema signals and labeled data.
The third aspect, this application provides a kind of recognition of heart sound device, described device includes:
Preprocessing module, for being pre-processed to the cardiechema signals of acquisition；
First obtains module, for obtaining first heart sound data, systole phase data, second from pretreated cardiechema signals
Heart sound data and diastole data；
Second obtains module, for the default processing rule of basis to the first heart sound data, systole phase data, second heart
Sound data and diastole data are handled to obtain local feature；
Third obtains module, for obtaining the pretreated cardiechema signals according to trained neural network model
Global characteristics；
Identification module, for being based on trained recognition of heart sound model, according to the local feature and the global spy
Sign identifies the type of the cardiechema signals.
Fourth aspect, this application provides a kind of recognition of heart sound model training apparatus, comprising:
4th obtains module, and for obtaining training sample set, the training sample set includes several as training sample
The labeled data of cardiechema signals and each cardiechema signals；
5th obtains module, for obtaining first heart sound data, systole phase data, second heart sound number from the cardiechema signals
According to diastole data；
6th obtains module, for the default processing rule of basis to the first heart sound data, systole phase data, second heart
Sound data and diastole data are handled to obtain local feature；
7th obtains module, for obtaining the global characteristics of the cardiechema signals according to trained neural network model；
Identification module, for being based on the recognition of heart sound model, according to the local feature and the global characteristics pair
The type of the cardiechema signals is identified to obtain the prediction data of the cardiechema signals；
Module is adjusted, for adjusting the recognition of heart sound model according to the prediction data and labeled data of the cardiechema signals
Parameter.
5th aspect, this application provides a kind of computer readable storage medium, the computer readable storage medium is deposited
Contain computer program, which is characterized in that if the computer program is executed by processor, realize:
The cardiechema signals of acquisition are pre-processed；
First heart sound data, systole phase data, second heart sound data and diastole are obtained from pretreated cardiechema signals
Data；
According to default processing rule to the first heart sound data, systole phase data, second heart sound data and diastole issue
Local feature is obtained according to being handled；
The global characteristics of the pretreated cardiechema signals are obtained according to trained neural network model；
Based on trained recognition of heart sound model, the heart sound is believed according to the local feature and the global characteristics
Number type identified.
6th aspect, this application provides a kind of recognition of heart sound methods, comprising:
The cardiechema signals of acquisition are pre-processed；
First heart sound data, systole phase data, second heart sound data and diastole are obtained from pretreated cardiechema signals
Data；
According to default processing rule to the first heart sound data, systole phase data, second heart sound data and diastole issue
Local feature is obtained according to being handled；
The global characteristics of the pretreated cardiechema signals are obtained according to trained neural network model；
Based on trained recognition of heart sound model, the heart sound is believed according to the local feature and the global characteristics
Number type identified.
This application discloses a kind of computer equipment, recognition of heart sound device, method, model training apparatus and storage medium,
By according to default processing rule to the first heart sound data, systole phase data, second heart sound data and diastole data into
Row processing obtains the pretreated cardiechema signals to obtain local feature, and according to trained neural network model
Global characteristics realize the part spy not only according to first heart sound data, systole phase data, second heart sound data and diastole data
Sign, identifies cardiechema signals also according to the global characteristics of cardiechema signals；Therefore the identification process of recognition of heart sound model is comprehensive
The details feature of heart sound cycle stages and the global characteristics of cardiechema signals, the accuracy of identification are higher.
Detailed description of the invention
Technical solution in ord to more clearly illustrate embodiments of the present application, below will be to required use in embodiment description
Attached drawing be briefly described, it should be apparent that, the accompanying drawings in the following description is some embodiments of the present application, for this field
For those of ordinary skill, without creative efforts, it is also possible to obtain other drawings based on these drawings.
Fig. 1 is a kind of structural schematic diagram for computer equipment that one embodiment of the application provides；
Fig. 2 is the flow diagram of processing step when computer equipment realizes recognition of heart sound method；
Fig. 3 is the spectrogram schematic diagram of cardiechema signals；
Fig. 4 is the waveform diagram of the cardiechema signals before bandpass filtering；
Fig. 5 is the waveform diagram of the cardiechema signals after bandpass filtering；
Fig. 6 is the flow diagram that computer equipment realizes to cardiechema signals progress processing step when stage by stage；
Fig. 7 is the flow diagram of processing step when computer equipment realizes acquisition local feature；
The flow diagram of processing step when Fig. 8 is the training of computer equipment realization neural network model；
Fig. 9 is a kind of structural schematic diagram of neural network model；
Figure 10 is the flow diagram of processing step when computer equipment realizes acquisition global characteristics；
Figure 11 is the flow diagram of processing step when computer equipment realizes recognition of heart sound model training method；
Figure 12 is the structural schematic diagram of the recognition of heart sound device of one embodiment of the application；
Figure 13 is the structural schematic diagram of the recognition of heart sound model training apparatus of one embodiment of the application.
Specific embodiment
Below in conjunction with the attached drawing in the embodiment of the present application, technical solutions in the embodiments of the present application carries out clear, complete
Site preparation description, it is clear that described embodiment is some embodiments of the present application, instead of all the embodiments.Based on this Shen
Please in embodiment, every other implementation obtained by those of ordinary skill in the art without making creative efforts
Example, shall fall in the protection scope of this application.
Flow chart shown in the drawings only illustrates, it is not necessary to including all content and operation/step, also not
It is that must be executed by described sequence.For example, some operation/steps can also decompose, combine or partially merge, therefore practical
The sequence of execution is possible to change according to the actual situation.In addition, though the division of functional module has been carried out in schematic device,
But in some cases, it can be divided with the module being different from schematic device.
With reference to the accompanying drawing, it elaborates to some embodiments of the application.In the absence of conflict, following
Feature in embodiment and embodiment can be combined with each other.
Referring to Fig. 1, Fig. 1 is a kind of structural schematic diagram for computer equipment that embodiments herein provides.The calculating
Machine equipment can be server or terminal.
Refering to fig. 1, which includes processor, memory and the network interface connected by system bus,
In, memory may include non-volatile memory medium and built-in storage.
Processor supports the operation of entire computer equipment for providing calculating and control ability.
Built-in storage provides environment for the operation of the computer program in non-volatile memory medium, the computer program quilt
When processor executes, processor may make to execute any one recognition of heart sound method.
The network interface such as sends the task dispatching of distribution for carrying out network communication.It will be understood by those skilled in the art that
Structure shown in Fig. 1, only the block diagram of part-structure relevant to application scheme, is not constituted to application scheme institute
The restriction for the computer equipment being applied thereon, specific computer equipment may include than more or fewer portions as shown in the figure
Part perhaps combines certain components or with different component layouts.
It should be understood that processor can be central processing unit (Central Processing Unit, CPU), it should
Processor can also be other general processors, digital signal processor (Digital Signal Processor, DSP), specially
With integrated circuit (Application Specific Integrated Circuit, ASIC), field programmable gate array
(Field-Programmable Gate Array, FPGA) either other programmable logic device, discrete gate or transistor are patrolled
Collect device, discrete hardware components etc..Wherein, general processor can be microprocessor or the processor be also possible to it is any often
The processor etc. of rule.
Non-volatile memory medium can storage program area and computer program, which includes program instruction；
The processor is for executing the computer program.
In some embodiments, when which is executed by processor, processor may make to realize a kind of recognition of heart sound
Method.
In one embodiment, referring to Fig. 2, Fig. 2 is processor described in the embodiment of the present application for run memory
The flow diagram of processing step when middle computer program is to realize recognition of heart sound method.Processor implements following steps:
Step S110, the cardiechema signals of acquisition are pre-processed.
In some embodiments, cardiechema signals are acquired by the digital stethoscope of profession, in further embodiments, by just
Take the mobile terminal connection heart sound transducer acquisition cardiechema signals of formula.
Illustratively, the cardiechema signals got are the ecg signal data of number format, such as caardiophonogram
(phonocardiogram, PCG).Electrocardiosignal can be intercepted by preset duration, such as one section of maximum time of interception is not more than 60
The electrocardiosignal of second is as identification object.
Be illustrated in figure 3 the spectrogram of one section of cardiechema signals, the frequency of cardiechema signals about concentrate on 400hz hereinafter, and
In some embodiments, the sample frequency of the cardiechema signals acquired from caardiophonogram is 2000Hz；According to Nyquist
(Nyquist) sample frequency known to sampling thheorem should can retain enough signal letters greater than twice of cardiechema signals frequency
Breath,；It can be the original of 2000Hz or so by sample frequency in order to which calculation amount can be reduced while meeting sampling thheorem
1000Hz is arrived in the down-sampled processing of cardiechema signals.
Since during heart sound signal acquisition some interference signals, Ke Yitong can be carried due to friction, ambient noise etc.
Bandpass filtering is crossed to remove the spiking introduced in collection process, obtains pretreated cardiechema signals.It is illustrated in figure 4 band
Cardiechema signals before pass filter, Fig. 5 show the cardiechema signals after bandpass filtering, it can be seen that the point of cardiechema signals section start
Peak noise is effectively filtered out, and low-frequency disturbance and high-frequency noise are also effectively filtered out.
Step S120, first heart sound data, systole phase data, second heart sound data are obtained from pretreated cardiechema signals
With diastole data.
Cardiechema signals include one or more heart sound periods, correspond to one or many heartbeats.The heart sound period includes every time
First heart sound (S1), systole phase (Systole), second heart sound (S2) and diastole (Diastole)；Wherein first heart sound occurs
At the beginning of the systole, indicate the beginning of ventricular contraction, second heart sound occurred at the beginning of diastole, indicated ventricle
The beginning of diastole.The heart sound period of cardiechema signals is divided into four-stage, more acurrate can be identified cardiechema signals, also
Targetedly the abnormal heart sound in each stage can be identified.
In some embodiments, as shown in fig. 6, the processor is executing step S120 from pretreated cardiechema signals
When obtaining first heart sound data, systole phase data, second heart sound data and diastole data, for realizing step S121, step
S122:
Step S121, according to the R wave of the corresponding electrocardiogram of the cardiechema signals and T wave by the cardiechema signals at least one
Heart sound period divisions are first heart sound, systole phase, second heart sound, diastole four-stage.
In the present embodiment, the dividing method proposed using Springer divides by pretreated cardiechema signals
It cuts, by heart sound period divisions are as follows: first heart sound, systole phase, second heart sound and diastole.Specifically, can be in cardiechema signals
Some or all of the heart sound period be split.
Illustratively, it collects and records the electrocardiogram with the cardiechema signals same period and extracts the spy of the electrocardio in electrocardiogram
Sign, ecg characteristics include the information of R wave and T wave, according to the information of R wave and T wave by least one heart sound of the cardiechema signals
Period divisions are first heart sound, systole phase, second heart sound, diastole four-stage.
Step S122, the cardiechema signals are obtained in the first heart sound, systole phase, second heart sound, four ranks of diastole
The data of section.
Specifically, the data by cardiechema signals after segmentation in each stage, i.e. first heart sound data, systole phase data, second
Heart sound data and diastole data store respectively.
Step S130, according to default processing rule to the first heart sound data, systole phase data, second heart sound data and
Diastole data are handled to obtain local feature.
In some embodiments, as shown in fig. 7, the processor is executing step S130 according to default processing rule to institute
When stating first heart sound data, systole phase data, second heart sound data and diastole data and being handled to obtain local feature, use
At least one step in realization step S131- step S133:
Step S131, according to preset time-domain analysis rule to the first heart sound data, systole phase data, second heart sound
Data and diastole data are handled to obtain temporal signatures.
In some embodiments, step S131 is according to preset time-domain analysis rule to the first heart sound data, contraction
Issue evidence, second heart sound data and diastole data are handled to obtain temporal signatures, are specifically included: to the first heart sound
Data, systole phase data, second heart sound data and diastole data carry out Time Domain Processing to obtain and can characterize heart sound rhythm and pace of moving things spy
The heart sound interval data of sign is as temporal signatures.
Illustratively, during heart sound interval data includes R wave spacing, first heart sound interval, second heart sound interval, contraction
Every, diastole interval, systole phase and the ratio of eartbeat interval, the ratio of diastole and eartbeat interval, systole phase and diastole
The mean value and/or variance of at least one of ratio.
It may include one or more heart sound periods in a cardiechema signals, or can use in a cardiechema signals
One or more heart sound periods；Therefore each interval in sound interval data, each ratio can take mean value and/or variance；It is exemplary
, a cardiechema signals include 60 heart sound periods, and each interval in heart sound interval data, each ratio took in 60 heart sound periods
Corresponding interval, the mean value of ratio and/or variance.
For example, temporal signatures include totally 16 temporal signatures values in heart sound interval data.
In some embodiments, step S131 is according to preset time-domain analysis rule to the first heart sound data, contraction
Issue evidence, second heart sound data and diastole data are handled to obtain temporal signatures, are specifically included: to the first heart sound
Data, systole phase data, second heart sound data and diastole data carry out Time Domain Processing to obtain and can characterize heart sound amplitude spy
The heart sound amplitude data of sign is as temporal signatures.
Illustratively, heart sound amplitude data include systole phase averaged amplitude value and first heart sound averaged amplitude value ratio,
Diastole averaged amplitude value and the ratio of second heart sound averaged amplitude value, the amplitude degree of bias of first heart sound, the amplitude in systole phase are inclined
Spend the amplitude kurtosis of (skewness), the amplitude degree of bias of second heart sound, the amplitude degree of bias of diastole, first heart sound
(kurtosis), at least one of the amplitude kurtosis of the amplitude kurtosis, second heart sound in systole phase, amplitude kurtosis of diastole is equal
Value and/or variance.
Illustratively, a cardiechema signals include 60 heart sound periods, and each ratio, amplitude in heart sound amplitude data are inclined
Degree, amplitude kurtosis can take corresponding ratio, the amplitude degree of bias, the mean value and/or variance of amplitude kurtosis in this 60 heart sound periods.
For example, temporal signatures include totally 20 temporal signatures values in heart sound amplitude data.
From two aspect of the amplitude of heart sound and the rhythm and pace of moving things, normal person and cardiac's heart can be analyzed to a certain extent
The difference of sound.
Step S132, according to preset frequency-domain analysis rule to the first heart sound data, systole phase data, second heart sound
Data and diastole data are handled to obtain frequency domain character.
Illustratively, pass through the heart sound data to each heart sound stage, i.e. first heart sound data, systole phase data, second
Heart sound data and diastole data do frequency-domain analysis, obtain frequency domain character.
In some embodiments, step S132 is according to preset frequency-domain analysis rule to the first heart sound data, contraction
Issue evidence, second heart sound data and diastole data are handled to obtain frequency domain character and specifically include: to the first heart sound
Data, systole phase data, second heart sound data and diastole data carry out frequency-domain analysis to obtain each multiple and different frequency bands of leisure
On spectrum amplitude intermediate value as frequency domain character.
Specifically, multiplying hamming to first heart sound data, systole phase data, second heart sound data and diastole data respectively
(Hamming) window obtains Energy distribution of each phase data on frequency spectrum using Fast Fourier Transform (FFT), seeks spectrum amplitude
Frequency domain character of the intermediate value (i.e. median) as cardiechema signals respective stage.
According to the research of human auditory system mechanism, human ear has different auditory sensitivities to the sound wave of different frequency.Bass holds
Easily masking high pitch, and high pitch masking bass is more difficult, so the critical bandwidth higher-frequency of the sound mask at low frequency wants small；Consider
To this characteristic, when filter is set from low to high in this section of frequency band by the size of critical bandwidth by close to dredging one group of arrangement
Bandpass filter is filtered input signal.Illustratively, be arranged nine frequency band 25-45,45-65,65-85,85-105,
105-125,125-150,150-200,200-300,300-400, unit Hz, and nine bandpass filters are set accordingly.
Each stage heart sound data is inputted into each bandpass filter, the signal energy of each bandpass filter output is as the substantially special of signal
Sign, Energy distribution of each phase data on frequency spectrum, the intermediate value (i.e. median) for seeking spectrum amplitude are corresponding as cardiechema signals
The frequency domain character in stage.
When nine frequency bands are arranged, four-stage shares 4 × 9=36 spectrum amplitude intermediate value, i.e. frequency domain character value.
Step S133, rule is handled to the first heart sound data, systole phase data, second heart sound according to preset cepstrum
Data and diastole data are handled to obtain mel cepstrum feature.
Meier (Mel) cepstrum is often applied in sound signal processing.The auditory system of people is a special nonlinear system,
The sensitivity that it responds different frequency signals is different.Mel-frequency cepstrum coefficient (Mel Frequency Cepstrum
Coefficient, MFCC) aural signature that considers the mankind, it is non-that linear spectral is first mapped to the Mel based on Auditory Perception
In linear spectral, then switch on cepstrum.
The shape of sound channel is shown in the envelope that voice short-time rating is composed, and mel-frequency cepstrum feature, such as Meier
Frequency cepstral coefficient is exactly one kind can be with a kind of feature of this envelope of accurate description.Mel cepstrum is for sound signal processing ratio
Cepstrum is closer to human ear to the analytical characteristics of sound.
In some embodiments, to the first heart sound data, systole phase data, second heart sound data and diastole data
Preemphasis processing, framing plus the processing of hamming window and mel cepstrum analysis are carried out to obtain respective multiple mel-frequency cepstrums
Coefficient.Illustratively, the process for extracting the data mel-frequency cepstrum coefficient in a certain stage in cardiechema signals includes: first to this
Data carry out preemphasis processing, framing and add the processing of hamming window, and wherein the purpose of preemphasis is for the high frequency section to sound
It is aggravated, increases the high frequency resolution of sound；Sub-frame processing is to divide that data into some short sections to be handled, and such as will
Every second data is divided into 33-100 frame, and framing for example can be using the method that the window of moveable finite length is weighted come real
It is existing；Add the processing of hamming window that the data overall situation can be made more continuous, avoid the occurrence of Gibbs' effect, and after adding hamming window,
Originally the Partial Feature of periodic function is showed without periodic data-signal.Then to adding hamming window treated each
Short-time analysis window carries out Fourier transformation and obtains signal spectrum, then this linear signal spectrum is reflected by Mel filter group
It is mapped in the Mel non-linear spectrum based on Auditory Perception and obtains Mel frequency spectrum；Mel cepstrum point is carried out on Mel frequency spectrum later
Analysis, specifically includes and takes logarithm, DCT discrete cosine transform, to obtain mel-frequency cepstrum coefficient.
In some embodiments, the first heart sound data, systole phase data, second heart sound data and diastole data
Mel cepstrum feature includes that the first heart sound data, systole phase data, second heart sound data and diastole data are respective more
A mel-frequency cepstrum coefficient.
Illustratively, 13 are respectively extracted to first heart sound data, systole phase data, second heart sound data and diastole data
A such as mel-frequency cepstrum coefficient, the mel cepstrum feature of a cardiechema signals includes 4 × 13=52 mel-frequency cepstrum altogether
Coefficient.
First heart sound data, systole phase data, second heart sound data and the diastole obtained according to step S131- step S133
The temporal signatures of issue evidence, frequency domain character and mel cepstrum feature these Discrete Eigenvalues, can be used as and tentatively judge each rank
Section is with the presence or absence of abnormal heart sound, such as the R wave spacing and the R wave spacing of cardiac etc. of normal person, is also used as training
Foundation in sample set to demarcate labeled data as the cardiechema signals of training sample.
Step S140, the overall situation for obtaining the pretreated cardiechema signals according to trained neural network model is special
Sign.
In some possible embodiments, as shown in figure 8, the processor is in the training step for executing neural network model
When, implement following steps S101- step S103:
Step S101, training sample set is obtained, the training sample set includes several cardiechema signals as training sample
And the labeled data of each cardiechema signals.
Illustratively, collecting part does not have the cardiechema signals of heart sound abnormal conditions and a part there are heart sound exception feelings
The cardiechema signals of condition are as training sample；Each cardiechema signals calibration in training sample has labeled data, and labeled data can wrap
Heart sound classification is included if any without exception, first heart sound exception, diastole exception etc.；Labeled data further includes the heart in another embodiment
Sound signal is the probability value of each heart sound classification.
Step S102, the cardiechema signals as training sample are identified to obtain according to the neural network model
Take the prediction data of the cardiechema signals.
Illustratively, prediction data includes the heart sound type predicted, can also include that the heart sound type predicted corresponds to
Probability.
Step S103, according to the parameter of the prediction data of the cardiechema signals and labeled data adjustment neural network model.
Illustratively, random optimization can be carried out to neural network model using Adam, selects cross entropy as loss letter
Number is minimized.Neural network model is trained by multiple samples that training sample is concentrated, when prediction data and mark
The deviation of note data terminates to train when being less than preset threshold value, obtains trained neural network model.
In some embodiments, the processor is executing step S140 according to trained neural network model acquisition institute
When stating the global characteristics of pretreated cardiechema signals, it is specifically used for realizing:
Convolution operation, Chi Hua are at least carried out to the pretreated cardiechema signals according to trained neural network model
Operation, full attended operation are to export recessive character as global characteristics.
Neural network generally comprises input layer, output layer and the middle layer between input layer and output layer, also cries hidden
Hide layer；Hidden layer can provide recessive character；What it is due to input layer input is cardiechema signals, rather than first heart sound number above-mentioned
According to, systole phase data, second heart sound data and diastole data, therefore, these recessive characters are can to embody cardiechema signals spy
The global characteristics of property.
In some embodiments, the structure of neural network model is as shown in Figure 9.As shown in Figure 10, the processor is being held
Row at least carries out convolution operation to the pretreated cardiechema signals according to trained neural network model, pondization operates,
When full attended operation is to export recessive character as global characteristics, it is specifically used for realization step S1411- step S1415:
Step S1411, from multiple heart sound Wave datas of the pretreated multiple and different frequency bands of heart sound signal extraction.
Illustratively, Inupt1, Input2, Input3, Input4 are respectively each heartbeat of cardiechema signals by 4 frequencies
Band, such as the filtered heart sound Wave data of 25-45,45-80,80-200,200-400 (Hz), the heart sound waveform number of each frequency band
It is 2500 according to length.
Step S1412, the first eigenvector of each heart sound Wave data is extracted.
It is also 2500 that each heart sound Wave data, which passes through the length exported after input layer, and dimension is 1, is expressed as (2500,1).
Step S1413, convolution operation is carried out to each first eigenvector and pondization operates, with output and each described the
The corresponding second feature vector of one feature vector.
Each first eigenvector of different frequency bands respectively passes through first time convolution conv, for the first time maximum pond max_
Pooling, second of convolution conv, second of maximum pond max_pooling and flatten operation processing obtain accordingly
Second feature vector.Wherein the tensor (Tensor) of higher-dimension is processed into an one-dimensional tensor (vector) by flatten operation.
The first eigenvector of (2500,1) is (2496,8) by the output of first time convolution conv, i.e., length is
2496, dimension is 8；Output of the vector of (2496,8) through maximum pond max_pooling for the first time is (1248,8), (1248,
8) output of the vector through second of convolution conv is (1244,4), and the vector of (1244,4) is through second of maximum pond max_
The output of pooling is (622,4)；The output that the vector of (622,4) is operated through flatten is one-dimensional that length is 2488
Two feature vectors.
Step S1414, all second feature vectors are integrated into a third feature vector.
As shown in figure 9, flatten:Concatenate layers are integrated four second feature vectors for corresponding to different frequency bands
For a third feature vector, the length of the third feature vector is 9952, and dimension is 1.
Step S1415, full attended operation is carried out to export recessive character as global characteristics to the third feature vector.
Illustratively, length 9952, dimension be 1 third feature vector first through dropout layers handle, the length of output
It is 9952；It then is being vector that length is 20 through Dense layers of processing；The characteristic value that the vector that the length is 20 includes can be made
For the recessive character.
It is 20 to the length by another dropout layers and another Dense layers when being trained to neural network model
Vector handled, obtain the prediction data of the cardiechema signals.
Cardiechema signals application multi-channel filter is divided into different frequency bands, training neural network model can in deep learning
With by by the temporal signatures of first heart sound data, systole phase data, second heart sound data and diastole data, frequency domain character and
These Discrete Eigenvalues of mel cepstrum feature are integrated into training process as one group of constraint condition by backpropagation, enhancing mind
Through network model feature representation ability, prediction accuracy and global characteristics are promoted, such as the ability to express of recessive character.
Step S150, it is based on trained recognition of heart sound model, according to the local feature and the global characteristics pair
The type of the cardiechema signals is identified.
Illustratively, local feature include step S131 obtain heart sound interval data in totally 16 temporal signatures values,
The 52 of 36 frequency domain character values, step S133 acquisition that totally 20 temporal signatures values, step 132 in heart sound amplitude data obtain
A mel-frequency cepstrum coefficient shares 124 local features；Global characteristics include 20 recessive characters, i.e. global characteristics；Heart sound
Identification model identifies the type of the cardiechema signals according to 124 local features and 20 global characteristics, with identification
The type of the cardiechema signals is normal cardiac sound or abnormal heart sound out.
In some embodiments, the recognition of heart sound model includes CatBoost model.
Boosting algorithm has it in scenes such as training sample amount is limited, the required training time is shorter, shortage tune ginseng knowledge
Indispensable advantage.CatBoost is that a kind of gradient promotes library, be Gradient Boosting (gradient promotions) technology and
The combination of Categorical Features (classification type feature) technology is also based on the machine learning frame that gradient promotes decision tree
Frame.Robustness is high, reduces the demand to many hyper parameter tunings, and reduce the chance of overfitting.
The embodiment of the invention also provides a kind of recognition of heart sound methods, illustratively, as shown in Fig. 2, recognition of heart sound method
Include:
Step S110, the cardiechema signals of acquisition are pre-processed；
Step S120, first heart sound data, systole phase data, second heart sound data are obtained from pretreated cardiechema signals
With diastole data；
Step S130, according to default processing rule to the first heart sound data, systole phase data, second heart sound data and
Diastole data are handled to obtain local feature；
Step S140, the overall situation for obtaining the pretreated cardiechema signals according to trained neural network model is special
Sign；
Step S150, it is based on trained recognition of heart sound model, according to the local feature and the global characteristics pair
The type of the cardiechema signals is identified.
The processor of computer equipment provided by the above embodiment and recognition of heart sound method, computer equipment executes memory
Recognition of heart sound method is realized when middle computer program, by regular to the first heart sound data, systole phase according to default processing
Data, second heart sound data and diastole data are handled to obtain local feature, and according to trained neural network
Model obtain the pretreated cardiechema signals global characteristics realize not only according to first heart sound data, systole phase data,
The local feature of second heart sound data and diastole data knows cardiechema signals also according to the global characteristics of cardiechema signals
Not；Therefore the identification process of recognition of heart sound model combines the details feature of heart sound cycle stages and the overall situation of cardiechema signals
The accuracy of feature, identification is higher.
And since the heart sound period is divided into multiple stages, so that certain a kind of heart disease can be identified with specific aim.
In other feasible embodiments, non-volatile memory medium can storage program area and computer program, should
Computer program includes program instruction；The processor is for executing the computer program.The program instruction is held by processor
When row, processor may make to realize a kind of recognition of heart sound model training method.
In one embodiment, Figure 11 is please referred to, Figure 11 is processor described in the embodiment of the present application for running storage
The flow diagram of processing step when computer program is in device to realize recognition of heart sound model training method.Processor specific implementation
Following steps:
Step S210, training sample set is obtained, the training sample set includes several cardiechema signals as training sample
And the labeled data of each cardiechema signals.
Illustratively, collecting part does not have the cardiechema signals of heart sound abnormal conditions and a part there are heart sound exception feelings
The cardiechema signals of condition are as training sample；Each cardiechema signals calibration in training sample has labeled data；Labeled data can wrap
Heart sound classification is included if any without exception, first heart sound exception, diastole exception etc.；Labeled data further includes the heart in another embodiment
Sound signal is the probability value of each heart sound classification.
In some embodiments, the cardiechema signals that training sample is concentrated have already been through pretreatment.
Step S220, first heart sound data, systole phase data, second heart sound data and diastole are obtained from the cardiechema signals
Issue evidence.
Step S230, according to default processing rule to the first heart sound data, systole phase data, second heart sound data and
Diastole data are handled to obtain local feature.
In some embodiments, the processor is executing step S230 according to default processing rule to the first heart sound
When data, systole phase data, second heart sound data and diastole data are handled to obtain local feature, for realizing:
It to the first heart sound data, systole phase data, second heart sound data and is relaxed according to preset time-domain analysis rule
It opens issue and obtains temporal signatures according to being handled；And/or
It to the first heart sound data, systole phase data, second heart sound data and is relaxed according to preset frequency-domain analysis rule
It opens issue and obtains frequency domain character according to being handled；And/or
Rule is handled to the first heart sound data, systole phase data, second heart sound data according to preset cepstrum and is relaxed
It opens issue and obtains mel cepstrum feature according to being handled.
In some embodiments, the processor execute it is described according to preset time-domain analysis rule to first heart
When sound data, systole phase data, second heart sound data and diastole data are handled to obtain temporal signatures, for realizing:
To the first heart sound data, systole phase data, second heart sound data and diastole data carry out Time Domain Processing with
Obtain heart sound interval data and heart sound amplitude data, wherein the heart sound interval data is for characterizing heart sound prosodic feature
Data, the heart sound amplitude data are the heart sound amplitude data for characterizing heart sound amplitude characteristic.
The processor execute it is described according to preset frequency-domain analysis rule to the first heart sound data, shrink issue
When being handled according to, second heart sound data and diastole data to obtain frequency domain character, for realizing:
To the first heart sound data, systole phase data, second heart sound data and diastole data carry out frequency-domain analysis with
The intermediate value of the spectrum amplitude on each multiple and different frequency bands of leisure is obtained as frequency domain character；
The processor is described regular to the first heart sound data, contraction issue according to the processing of preset cepstrum in execution
When being handled according to, second heart sound data and diastole data to obtain mel cepstrum feature, for realizing:
To the first heart sound data, systole phase data, second heart sound data and diastole data carry out preemphasis processing,
Framing plus the processing of hamming window and mel cepstrum analysis are to obtain respective multiple mel-frequency cepstrum coefficients.
Step S240, the global characteristics of the cardiechema signals are obtained according to trained neural network model.
In some embodiments, the processor is executing step S240 according to trained neural network model acquisition institute
When stating the global characteristics of pretreated cardiechema signals, for realizing:
According to trained neural network model cardiechema signals are at least carried out with convolution operation, pondization operation, full connection behaviour
Make to export recessive character as global characteristics.
In some embodiments, the processor execute according to trained neural network model to cardiechema signals at least
Carry out convolution operation, when pondization operates, full attended operation is to export recessive character as global characteristics, for realizing:
From multiple heart sound Wave datas of the multiple and different frequency bands of the heart sound signal extraction；Extract each heart sound waveform number
According to first eigenvector；Convolution operation and pondization operation are carried out to each first eigenvector, with output and each described the
The corresponding second feature vector of one feature vector；All second feature vectors are integrated into a third feature vector；
Full attended operation is carried out to export recessive character as global characteristics to the third feature vector.
Step S250, it is based on the recognition of heart sound model, according to the local feature and the global characteristics to described
The type of cardiechema signals is identified to obtain the prediction data of the cardiechema signals.
Specifically, based on the recognition of heart sound model in training, according to the local feature and the global characteristics to institute
The type for stating cardiechema signals is identified to obtain the prediction data of cardiechema signals.Illustratively, prediction data includes predicting
Heart sound type can also include the corresponding probability of heart sound type predicted.
In some embodiments, recognition of heart sound model includes CatBoost model.
In some embodiments, it is based on same training sample set, neural network model can be trained, to be used for
The global characteristics of the pretreated cardiechema signals are obtained according to trained neural network model.
Step S260, the ginseng of the recognition of heart sound model is adjusted according to the prediction data of the cardiechema signals and labeled data
Number.
In some embodiments, it concentrates 80% training sample to be input to recognition of heart sound model training sample to instruct
Practice, then is tested with the result that remaining 20% training sample exports recognition of heart sound model.If prediction data, i.e., test is tied
Fruit is greater than preset threshold, then is designated as 1；If test result is less than preset threshold, then 0 is designated as；Illustratively, preset threshold is
0.5；The labeled data of test result and training sample is compared again；Illustratively, labeled data is 1 or 0, wherein 1 indicates
Heart sound is abnormal, and 0 indicates that no heart sound is abnormal；If prediction data is identical with labeled data, then test result is correct；If predicting number
According to different with labeled data, then test result mistake.
Record above-mentioned test, as a result, the correct quantity of such as test result is m, the quantity of test result mistake is n；It is logical
It crosses ACC check and effect assessment is trained to test result, for example, calculating the value of m ÷ (m+n), preset if the value is less than
Value, such as 2%, then can be with deconditioning；If the value is greater than preset value, continue to train, adjusts the parameter of recognition of heart sound model；
Preferably to judge cardiechema signals with the presence or absence of correctness abnormal, and that abnormal recognition of heart sound can be improved.
Computer equipment provided by the above embodiment, processor realize recognition of heart sound when executing computer program in memory
Model training method；By the first heart sound data, the systole phase data, second heart sound that obtain the cardiechema signals as training sample
The local feature of data and diastole data, and the global characteristics of the acquisition cardiechema signals carry out prediction knowledge to cardiechema signals
Not, according to the parameter of the labeled data of the prediction data of prediction and training sample adjustment recognition of heart sound model；It can be with less
Heart sound sample complete training, the obtained recognition of heart sound model of training can also be made not only according to first heart sound data, systole phase
The local feature of data, second heart sound data and diastole data, also according to cardiechema signals global characteristics to cardiechema signals into
Row identification, the accuracy of identification are higher.
In some embodiments, recognition of heart sound method, recognition of heart sound model training method can be applied to terminal or service
In device, it is therefore desirable to which trained model is stored in terminal or server.Wherein, which can be mobile phone, plate electricity
The electronic equipments such as brain, laptop, desktop computer, personal digital assistant and wearable device；Server can be independent
Server, or server cluster.
If it is applied in terminal, in order to guarantee the normal operation and quickly and effectively identification classification of the terminal, go back
It needs the recognition of heart sound model obtained to training, neural network model to carry out compression processing, the model after compression processing is saved
In terminal.
Wherein, which specifically includes at recognition of heart sound model, neural network model progress beta pruning processing, quantization
Reason and Huffman encoding processing etc., to reduce the size of recognition of heart sound model, neural network model, and then are conveniently stored in capacity
In lesser terminal.
A kind of structural schematic diagram of the recognition of heart sound device provided as shown in figure 12 for one embodiment of the application, the heart sound are known
Other device can be configured in server or terminal, for executing recognition of heart sound method above-mentioned.
As shown in figure 12, the recognition of heart sound device, comprising:
Preprocessing module 110, for being pre-processed to the cardiechema signals of acquisition.
First obtain module 120, for from pretreated cardiechema signals obtain first heart sound data, systole phase data,
Second heart sound data and diastole data.
Second obtains module 130, for according to default processing rule to the first heart sound data, systole phase data, the
Two heart sound datas and diastole data are handled to obtain local feature.
Specifically, the second acquisition module 130 includes
First acquisition unit, according to preset time-domain analysis rule to the first heart sound data, systole phase data, second
Heart sound data and diastole data are handled to obtain temporal signatures；And/or
Second acquisition unit, for according to preset frequency-domain analysis rule to the first heart sound data, systole phase data,
Second heart sound data and diastole data are handled to obtain frequency domain character；And/or
Third acquiring unit handles rule to the first heart sound data, systole phase data, second according to preset cepstrum
Heart sound data and diastole data are handled to obtain mel cepstrum feature.
Specifically, first acquisition unit is used for the first heart sound data, systole phase data, second heart sound data and relaxes
Issue is according to carrying out Time Domain Processing to obtain heart sound interval data and heart sound amplitude data, wherein the heart sound interval data is
For characterizing the data of heart sound prosodic feature, the heart sound amplitude data is the heart sound amplitude number for characterizing heart sound amplitude characteristic
According to.
Illustratively, during the temporal signatures include R wave spacing, first heart sound interval, second heart sound interval, contraction
Every, diastole interval, systole phase and the ratio of eartbeat interval, the ratio of diastole and eartbeat interval, systole phase and diastole
Ratio, the ratio of systole phase averaged amplitude value and first heart sound averaged amplitude value, diastole averaged amplitude value and second heart sound are flat
The equal ratio of range value, the amplitude degree of bias of first heart sound, the amplitude degree of bias, the amplitude degree of bias of second heart sound, diastole in systole phase
The amplitude degree of bias, the amplitude kurtosis of first heart sound, the amplitude kurtosis in systole phase, second heart sound amplitude kurtosis, the width of diastole
Spend the mean value and/or variance of at least one of kurtosis.
Specifically, second acquisition unit is used for the first heart sound data, systole phase data, second heart sound data and relaxes
Issue is according to carrying out frequency-domain analysis to obtain the intermediate value of the spectrum amplitude on each multiple and different frequency bands of leisure as frequency domain character.
The frequency domain character includes the first heart sound data, systole phase data, second heart sound data and diastole data
Spectrum amplitude intermediate value on each multiple and different frequency bands of leisure.
Specifically, third acquiring unit is used for the first heart sound data, systole phase data, second heart sound data and relaxes
Opening issue, preemphasis processing, framing plus hamming window are handled and mel cepstrum analysis is to obtain respective multiple Meiers according to carrying out
Frequency cepstral coefficient.
The mel cepstrum feature packet of the first heart sound data, systole phase data, second heart sound data and diastole data
Include the first heart sound data, systole phase data, second heart sound data and the respective multiple mel-frequency cepstrums of diastole data
Coefficient.
Third obtains module 140, believes for obtaining the pretreated heart sound according to trained neural network model
Number global characteristics.
It is used for according to trained neural network model specifically, third obtains module 140 to the pretreated heart
Sound signal at least carries out convolution operation, pondization operation, full attended operation to export recessive character as global characteristics.
Specifically, third acquisition module 140 includes:
First extraction unit, for multiple heart sound waves from the pretreated multiple and different frequency bands of heart sound signal extraction
Graphic data；
Second extraction unit, for extracting the first eigenvector of each heart sound Wave data；
First operating unit, for carrying out convolution operation to each first eigenvector and pondization operates, with output and
The corresponding second feature vector of each first eigenvector；
First integral unit, for all second feature vectors to be integrated into a third feature vector；
First full connection unit, for the third feature vector carry out full attended operation using export recessive character as
Global characteristics.
Identification module 150, for being based on trained recognition of heart sound model, according to the local feature and the overall situation
Feature identifies the type of the cardiechema signals.
Specifically, the recognition of heart sound model includes CatBoost model.
A kind of structural schematic diagram of the recognition of heart sound model training apparatus provided as shown in figure 13 for one embodiment of the application,
The recognition of heart sound model training apparatus can be configured in server or terminal, for executing recognition of heart sound model training above-mentioned
Method.
As shown in figure 13, the recognition of heart sound model training apparatus, comprising:
4th obtains module 210, and for obtaining training sample set, the training sample set includes several as training sample
Cardiechema signals and each cardiechema signals labeled data.
5th obtains module 220, for obtaining first heart sound data, systole phase data, second heart from the cardiechema signals
Sound data and diastole data.
6th obtains module 230, for according to default processing rule to the first heart sound data, systole phase data, the
Two heart sound datas and diastole data are handled to obtain local feature.
Specifically, the 6th acquisition module 230 includes:
4th acquiring unit, for according to preset time-domain analysis rule to the first heart sound data, systole phase data,
Second heart sound data and diastole data are handled to obtain temporal signatures；And/or
5th acquiring unit, for according to preset frequency-domain analysis rule to the first heart sound data, systole phase data,
Second heart sound data and diastole data are handled to obtain frequency domain character；And/or
6th acquiring unit, for according to preset cepstrum handle rule to the first heart sound data, systole phase data,
Second heart sound data and diastole data are handled to obtain mel cepstrum feature.
7th obtains module 240, and the overall situation for obtaining the cardiechema signals according to trained neural network model is special
Sign.
Specifically, the 7th acquisition module 240 includes:
Third extraction unit, for multiple heart sound Wave datas from the multiple and different frequency bands of the heart sound signal extraction；
4th extraction unit, for extracting the first eigenvector of each heart sound Wave data；
Second operating unit, for carrying out convolution operation to each first eigenvector and pondization operates, with output and
The corresponding second feature vector of each first eigenvector；
Second integral unit, for all second feature vectors to be integrated into a third feature vector；
Second full connection unit, for the third feature vector carry out full attended operation using export recessive character as
Global characteristics.
Identification module 250, for being based on the recognition of heart sound model, according to the local feature and the global characteristics
The type of the cardiechema signals is identified to obtain the prediction data of the cardiechema signals.
Module 260 is adjusted, for adjusting the recognition of heart sound according to the prediction data and labeled data of the cardiechema signals
The parameter of model.
It should be noted that it is apparent to those skilled in the art that, for convenience of description and succinctly,
The device of foregoing description and each module, the specific work process of unit, can be with reference to pair in aforementioned computer apparatus embodiments
Process is answered, details are not described herein.
Computer equipment can be realized in numerous general or special purpose computing system environments or configuration.Such as: individual calculus
Machine, server computer, handheld device or portable device, multicomputer system, based on microprocessor are at laptop device
System, programmable consumer-elcetronics devices, network PC, minicomputer, mainframe computer, including any of the above system or equipment
Distributed computing environment etc..
As seen through the above description of the embodiments, those skilled in the art can be understood that the application can
It realizes by means of software and necessary general hardware platform.Based on this understanding, the technical solution essence of the application
On in other words the part that contributes to existing technology can be embodied in the form of software products, the computer software product
It can store in storage medium, such as ROM/RAM, magnetic disk, CD, including some instructions are used so that a computer equipment
(can be personal computer, server or the network equipment etc.) executes the certain of each embodiment of the application or embodiment
Method described in part, such as:
A kind of computer readable storage medium, the computer-readable recording medium storage have computer program, the meter
It include program instruction in calculation machine program, the processor executes described program instruction, realizes provided by the embodiments of the present application any
Item recognition of heart sound method or any one recognition of heart sound model training method provided by the embodiments of the present application.
In some embodiments, the computer-readable recording medium storage has computer program, if the computer journey
Sequence is executed by processor, and is realized:
The cardiechema signals of acquisition are pre-processed；
First heart sound data, systole phase data, second heart sound data and diastole are obtained from pretreated cardiechema signals
Data；
According to default processing rule to the first heart sound data, systole phase data, second heart sound data and diastole issue
Local feature is obtained according to being handled；
The global characteristics of the pretreated cardiechema signals are obtained according to trained neural network model；
Based on trained recognition of heart sound model, the heart sound is believed according to the local feature and the global characteristics
Number type identified.
Wherein, the computer readable storage medium can be the storage inside of computer equipment described in previous embodiment
Unit, such as the hard disk or memory of the computer equipment.The computer readable storage medium is also possible to the computer
The plug-in type hard disk being equipped on the External memory equipment of equipment, such as the computer equipment, intelligent memory card (Smart
Media Card, SMC), secure digital (Secure Digital, SD) card, flash card (Flash Card) etc..
The above, the only specific embodiment of the application, but the protection scope of the application is not limited thereto, it is any
Those familiar with the art within the technical scope of the present application, can readily occur in various equivalent modifications or replace
It changes, these modifications or substitutions should all cover within the scope of protection of this application.Therefore, the protection scope of the application should be with right
It is required that protection scope subject to.
Claims (10)
1. a kind of computer equipment, which is characterized in that the computer equipment includes memory and processor；
The memory is for storing computer program；
The processor, for executing the computer program and the realization when executing the computer program:
The cardiechema signals of acquisition are pre-processed；
First heart sound data, systole phase data, second heart sound data and diastole data are obtained from pretreated cardiechema signals；
According to default processing rule to the first heart sound data, systole phase data, second heart sound data and diastole data into
Row processing is to obtain local feature；
The global characteristics of the pretreated cardiechema signals are obtained according to trained neural network model；
Based on trained recognition of heart sound model, according to the local feature and the global characteristics to the cardiechema signals
Type is identified.
2. computer equipment as described in claim 1, which is characterized in that the processor is executing the default processing of the basis
Rule handles to obtain part the first heart sound data, systole phase data, second heart sound data and diastole data
When feature, for realizing:
According to preset time-domain analysis rule to the first heart sound data, systole phase data, second heart sound data and diastole
Data are handled to obtain temporal signatures；And/or
According to preset frequency-domain analysis rule to the first heart sound data, systole phase data, second heart sound data and diastole
Data are handled to obtain frequency domain character；And/or
Rule is handled to the first heart sound data, systole phase data, second heart sound data and diastole according to preset cepstrum
Data are handled to obtain mel cepstrum feature.
3. computer equipment as claimed in claim 2, which is characterized in that the processor execute it is described according to it is preset when
Domain analysis rule handles to obtain the first heart sound data, systole phase data, second heart sound data and diastole data
When taking temporal signatures, for realizing:
Time Domain Processing is carried out to obtain to the first heart sound data, systole phase data, second heart sound data and diastole data
Heart sound interval data and heart sound amplitude data, wherein the heart sound interval data is the data for characterizing heart sound prosodic feature,
The heart sound amplitude data is the heart sound amplitude data for characterizing heart sound amplitude characteristic；
The processor execute it is described according to preset frequency-domain analysis rule to the first heart sound data, systole phase data,
When second heart sound data and diastole data are handled to obtain frequency domain character, for realizing:
Frequency-domain analysis is carried out to obtain to the first heart sound data, systole phase data, second heart sound data and diastole data
The intermediate value of spectrum amplitude on each multiple and different frequency bands of leisure is as frequency domain character；
The processor execute it is described according to preset cepstrum handle rule to the first heart sound data, systole phase data,
When second heart sound data and diastole data are handled to obtain mel cepstrum feature, for realizing:
Preemphasis processing is carried out to the first heart sound data, systole phase data, second heart sound data and diastole data, is divided
Frame plus the processing of hamming window and mel cepstrum analysis are to obtain respective multiple mel-frequency cepstrum coefficients.
4. computer equipment as described in claim 1, which is characterized in that the processor is described according to trained in execution
When neural network model obtains the global characteristics of the pretreated cardiechema signals, for realizing:
Convolution operation, Chi Huacao are at least carried out to the pretreated cardiechema signals according to trained neural network model
Make, full attended operation is to export recessive character as global characteristics.
5. computer equipment as claimed in claim 4, which is characterized in that the processor is described according to trained in execution
Neural network model carries out convolution operation, pondization operation, full attended operation at least to the pretreated cardiechema signals with defeated
When recessive character is as global characteristics out, for realizing:
From multiple heart sound Wave datas of the pretreated multiple and different frequency bands of heart sound signal extraction；
Extract the first eigenvector of each heart sound Wave data；
Convolution operation and pondization operation are carried out to each first eigenvector, it is corresponding with each first eigenvector to export
Second feature vector；
All second feature vectors are integrated into a third feature vector；
Full attended operation is carried out to export recessive character as global characteristics to the third feature vector.
6. a kind of computer equipment, which is characterized in that the computer equipment includes memory and processor；The memory is used
In storage computer program；
The processor, for executing the computer program and the realization when executing the computer program:
Training sample set is obtained, the training sample set includes several cardiechema signals as training sample and each heart sound
The labeled data of signal；
First heart sound data, systole phase data, second heart sound data and diastole data are obtained from the cardiechema signals；
According to default processing rule to the first heart sound data, systole phase data, second heart sound data and diastole data into
Row processing is to obtain local feature；
The global characteristics of the cardiechema signals are obtained according to trained neural network model；
Based on the recognition of heart sound model, according to the local feature and the global characteristics to the type of the cardiechema signals
It is identified to obtain the prediction data of the cardiechema signals；
The parameter of the recognition of heart sound model is adjusted according to the prediction data of the cardiechema signals and labeled data.
7. a kind of recognition of heart sound device characterized by comprising
Preprocessing module, for being pre-processed to the cardiechema signals of acquisition；
First obtains module, for obtaining first heart sound data, systole phase data, second heart sound from pretreated cardiechema signals
Data and diastole data；
Second obtains module, for the default processing rule of basis to the first heart sound data, systole phase data, second heart sound number
It is handled according to diastole data to obtain local feature；
Third obtains module, for obtaining the overall situation of the pretreated cardiechema signals according to trained neural network model
Feature；
Identification module, for being based on trained recognition of heart sound model, according to the local feature and the global characteristics pair
The type of the cardiechema signals is identified.
8. a kind of recognition of heart sound model training apparatus characterized by comprising
4th obtains module, and for obtaining training sample set, the training sample set includes several heart sound as training sample
The labeled data of signal and each cardiechema signals；
5th obtain module, for from the cardiechema signals obtain first heart sound data, systole phase data, second heart sound data and
Diastole data；
6th obtains module, for the default processing rule of basis to the first heart sound data, systole phase data, second heart sound number
It is handled according to diastole data to obtain local feature；
7th obtains module, for obtaining the global characteristics of the cardiechema signals according to trained neural network model；
Identification module, for being based on the recognition of heart sound model, according to the local feature and the global characteristics to described
The type of cardiechema signals is identified to obtain the prediction data of the cardiechema signals；
Module is adjusted, the ginseng of the recognition of heart sound model is adjusted for the prediction data and labeled data according to the cardiechema signals
Number.
9. a kind of computer readable storage medium, the computer-readable recording medium storage has computer program, and feature exists
In, if the computer program is executed by processor, realization:
The cardiechema signals of acquisition are pre-processed；
First heart sound data, systole phase data, second heart sound data and diastole data are obtained from pretreated cardiechema signals；
According to default processing rule to the first heart sound data, systole phase data, second heart sound data and diastole data into
Row processing is to obtain local feature；
The global characteristics of the pretreated cardiechema signals are obtained according to trained neural network model；
Based on trained recognition of heart sound model, according to the local feature and the global characteristics to the cardiechema signals
Type is identified.
10. a kind of recognition of heart sound method characterized by comprising
The cardiechema signals of acquisition are pre-processed；
First heart sound data, systole phase data, second heart sound data and diastole data are obtained from pretreated cardiechema signals；
According to default processing rule to the first heart sound data, systole phase data, second heart sound data and diastole data into
Row processing is to obtain local feature；
The global characteristics of the pretreated cardiechema signals are obtained according to trained neural network model；
Based on trained recognition of heart sound model, according to the local feature and the global characteristics to the cardiechema signals
Type is identified.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910268611.7A CN110123367B (en) | 2019-04-04 | 2019-04-04 | Computer device, heart sound recognition method, model training device, and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910268611.7A CN110123367B (en) | 2019-04-04 | 2019-04-04 | Computer device, heart sound recognition method, model training device, and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110123367A true CN110123367A (en) | 2019-08-16 |
| CN110123367B CN110123367B (en) | 2022-11-15 |
Family
ID=67569233
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910268611.7A Active CN110123367B (en) | 2019-04-04 | 2019-04-04 | Computer device, heart sound recognition method, model training device, and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110123367B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110558944A (en) * | 2019-09-09 | 2019-12-13 | 成都智能迭迦科技合伙企业(有限合伙) | Heart sound processing method and device, electronic equipment and computer readable storage medium |
| CN111345815A (en) * | 2020-02-11 | 2020-06-30 | 广州视源电子科技股份有限公司 | Method, device, equipment and storage medium for detecting QRS wave in electrocardiosignal |
| CN111603161A (en) * | 2020-05-28 | 2020-09-01 | 苏州小蓝医疗科技有限公司 | Electroencephalogram classification method |
| WO2021051559A1 (en) * | 2019-09-18 | 2021-03-25 | 平安科技（深圳）有限公司 | Categorizing method for heart sound signal, device, equipment, and storage medium |
| CN112690823A (en) * | 2020-12-22 | 2021-04-23 | 海南力维科贸有限公司 | Method and system for identifying physiological sounds of lungs |
| CN112971802A (en) * | 2021-02-08 | 2021-06-18 | 中北大学 | Heart sound signal detection method and system based on deep learning model |
| CN113855065A (en) * | 2021-09-28 | 2021-12-31 | 平安科技（深圳）有限公司 | Heart sound identification method based on fusion of shallow learning and deep learning and related device |
| CN114010220A (en) * | 2021-10-29 | 2022-02-08 | 平安科技（深圳）有限公司 | Heart sound signal processing method, computer device and storage medium |
| CN116807435A (en) * | 2023-08-28 | 2023-09-29 | 嘉兴学院G60科创走廊产业与创新研究院 | Portable heart sound sensing device and heart rate monitoring system |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102697520A (en) * | 2012-05-08 | 2012-10-03 | 天津沃康科技有限公司 | Electronic stethoscope based on intelligent distinguishing function |
| CN106548159A (en) * | 2016-11-08 | 2017-03-29 | 中国科学院自动化研究所 | Reticulate pattern facial image recognition method and device based on full convolutional neural networks |
| CN107122826A (en) * | 2017-05-08 | 2017-09-01 | 京东方科技集团股份有限公司 | Processing method and system and storage medium for convolutional neural networks |
| CN107126206A (en) * | 2017-06-16 | 2017-09-05 | 重庆睿凡科技发展有限公司 | A kind of heart sound automatic analysis system and method |
| CN107137072A (en) * | 2017-04-28 | 2017-09-08 | 北京科技大学 | A kind of ventricular ectopic beating detection method based on 1D convolutional neural networks |
| CN107154021A (en) * | 2017-05-03 | 2017-09-12 | 西安电子科技大学 | Image super-resolution method based on deep layer thresholding convolutional neural networks |
| CN107529645A (en) * | 2017-06-29 | 2018-01-02 | 重庆邮电大学 | A kind of heart sound intelligent diagnosis system and method based on deep learning |
| CN108143407A (en) * | 2017-12-25 | 2018-06-12 | 四川大学 | A kind of heart sound segmentation method for automatically extracting heart sound envelope characteristic |
| CN108323158A (en) * | 2018-01-18 | 2018-07-24 | 深圳前海达闼云端智能科技有限公司 | Heart sound identification method and cloud system |
| CN109044396A (en) * | 2018-06-25 | 2018-12-21 | 广东工业大学 | A kind of intelligent recognition of heart sound method based on two-way length Memory Neural Networks in short-term |
-
2019
- 2019-04-04 CN CN201910268611.7A patent/CN110123367B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102697520A (en) * | 2012-05-08 | 2012-10-03 | 天津沃康科技有限公司 | Electronic stethoscope based on intelligent distinguishing function |
| CN106548159A (en) * | 2016-11-08 | 2017-03-29 | 中国科学院自动化研究所 | Reticulate pattern facial image recognition method and device based on full convolutional neural networks |
| CN107137072A (en) * | 2017-04-28 | 2017-09-08 | 北京科技大学 | A kind of ventricular ectopic beating detection method based on 1D convolutional neural networks |
| CN107154021A (en) * | 2017-05-03 | 2017-09-12 | 西安电子科技大学 | Image super-resolution method based on deep layer thresholding convolutional neural networks |
| CN107122826A (en) * | 2017-05-08 | 2017-09-01 | 京东方科技集团股份有限公司 | Processing method and system and storage medium for convolutional neural networks |
| CN107126206A (en) * | 2017-06-16 | 2017-09-05 | 重庆睿凡科技发展有限公司 | A kind of heart sound automatic analysis system and method |
| CN107529645A (en) * | 2017-06-29 | 2018-01-02 | 重庆邮电大学 | A kind of heart sound intelligent diagnosis system and method based on deep learning |
| CN108143407A (en) * | 2017-12-25 | 2018-06-12 | 四川大学 | A kind of heart sound segmentation method for automatically extracting heart sound envelope characteristic |
| CN108323158A (en) * | 2018-01-18 | 2018-07-24 | 深圳前海达闼云端智能科技有限公司 | Heart sound identification method and cloud system |
| CN109044396A (en) * | 2018-06-25 | 2018-12-21 | 广东工业大学 | A kind of intelligent recognition of heart sound method based on two-way length Memory Neural Networks in short-term |
Non-Patent Citations (2)
| Title |
|---|
| 周克良等: "心音信号特征分析与识别方法研究", 《广西师范大学学报(自然科学版)》 * |
| 庞春颖等: "HHT和提升小波包相结合的心音信号识别研究", 《信号处理》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110558944A (en) * | 2019-09-09 | 2019-12-13 | 成都智能迭迦科技合伙企业(有限合伙) | Heart sound processing method and device, electronic equipment and computer readable storage medium |
| WO2021051559A1 (en) * | 2019-09-18 | 2021-03-25 | 平安科技（深圳）有限公司 | Categorizing method for heart sound signal, device, equipment, and storage medium |
| CN111345815A (en) * | 2020-02-11 | 2020-06-30 | 广州视源电子科技股份有限公司 | Method, device, equipment and storage medium for detecting QRS wave in electrocardiosignal |
| CN111345815B (en) * | 2020-02-11 | 2023-05-02 | 广州视源电子科技股份有限公司 | Method, device, equipment and storage medium for detecting QRS wave in electrocardiosignal |
| CN111603161A (en) * | 2020-05-28 | 2020-09-01 | 苏州小蓝医疗科技有限公司 | Electroencephalogram classification method |
| CN112690823A (en) * | 2020-12-22 | 2021-04-23 | 海南力维科贸有限公司 | Method and system for identifying physiological sounds of lungs |
| CN112971802A (en) * | 2021-02-08 | 2021-06-18 | 中北大学 | Heart sound signal detection method and system based on deep learning model |
| CN113855065A (en) * | 2021-09-28 | 2021-12-31 | 平安科技（深圳）有限公司 | Heart sound identification method based on fusion of shallow learning and deep learning and related device |
| CN113855065B (en) * | 2021-09-28 | 2023-09-22 | 平安科技（深圳）有限公司 | Heart sound identification method and related device based on shallow learning and deep learning fusion |
| CN114010220A (en) * | 2021-10-29 | 2022-02-08 | 平安科技（深圳）有限公司 | Heart sound signal processing method, computer device and storage medium |
| CN116807435A (en) * | 2023-08-28 | 2023-09-29 | 嘉兴学院G60科创走廊产业与创新研究院 | Portable heart sound sensing device and heart rate monitoring system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110123367B (en) | 2022-11-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110123367A (en) | Computer equipment, recognition of heart sound device, method, model training apparatus and storage medium | |
| Chowdhury et al. | Time-frequency analysis, denoising, compression, segmentation, and classification of PCG signals | |
| Deng et al. | Heart sound classification based on improved MFCC features and convolutional recurrent neural networks | |
| Patidar et al. | Classification of cardiac sound signals using constrained tunable-Q wavelet transform | |
| Karar et al. | Automated diagnosis of heart sounds using rule-based classification tree | |
| US11139048B2 (en) | Discovering novel features to use in machine learning techniques, such as machine learning techniques for diagnosing medical conditions | |
| CN109961017A (en) | A kind of cardiechema signals classification method based on convolution loop neural network | |
| US11062792B2 (en) | Discovering genomes to use in machine learning techniques | |
| US7610324B2 (en) | System for detection and estimation of periodic patterns in a noisy signal | |
| Vepa | Classification of heart murmurs using cepstral features and support vector machines | |
| CN109497997A (en) | Based on majority according to the seizure detection and early warning system of acquisition | |
| Das et al. | Supervised model for Cochleagram feature based fundamental heart sound identification | |
| Kucharski et al. | A deep learning approach for valve defect recognition in heart acoustic signal | |
| Usman et al. | Heart rate detection and classification from speech spectral features using machine learning | |
| Shankar et al. | Seizure types classification by generating input images with in-depth features from decomposed EEG signals for deep learning pipeline | |
| CN114305484A (en) | Heart disease heart sound intelligent classification method, device and medium based on deep learning | |
| Banerjee et al. | Multi-class heart sounds classification using 2D-convolutional neural network | |
| Ankışhan | Blood pressure prediction from speech recordings | |
| CN110415824B (en) | Cerebral apoplexy disease risk assessment device and equipment | |
| CN115486865A (en) | Heart sound classification method based on convolution cyclic neural network | |
| Mustafa et al. | Detection of heartbeat sounds arrhythmia using automatic spectral methods and cardiac auscultatory | |
| Malik et al. | Accurate classification of heart sound signals for cardiovascular disease diagnosis by wavelet analysis and convolutional neural network: preliminary results | |
| Taşkıran et al. | A deep learning based decision support system for diagnosis of Temporomandibular joint disorder | |
| Abbas et al. | Artificial intelligence framework for heart disease classification from audio signals | |
| Wu et al. | Influence of sampling rate on voice analysis for assessment of Parkinson's disease |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |