CN110427924A - A kind of heart impact signal based on LSTM more classifying identification methods automatically - Google Patents
A kind of heart impact signal based on LSTM more classifying identification methods automatically Download PDFInfo
- Publication number
- CN110427924A CN110427924A CN201910846195.4A CN201910846195A CN110427924A CN 110427924 A CN110427924 A CN 110427924A CN 201910846195 A CN201910846195 A CN 201910846195A CN 110427924 A CN110427924 A CN 110427924A
- Authority
- CN
- China
- Prior art keywords
- lstm
- heart
- point
- wave
- impact signal
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 41
- 238000012549 training Methods 0.000 claims abstract description 19
- 239000011159 matrix material Substances 0.000 claims abstract description 14
- 238000005457 optimization Methods 0.000 claims abstract description 8
- 230000003044 adaptive effect Effects 0.000 claims abstract description 5
- 230000008569 process Effects 0.000 claims abstract description 5
- 230000006870 function Effects 0.000 claims description 9
- 238000000354 decomposition reaction Methods 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000013461 design Methods 0.000 claims description 6
- 210000002569 neuron Anatomy 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 210000004027 cell Anatomy 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 4
- 239000002033 PVDF binder Substances 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims description 3
- 239000012528 membrane Substances 0.000 claims description 3
- 229920002981 polyvinylidene fluoride Polymers 0.000 claims description 3
- 238000013139 quantization Methods 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 238000013459 approach Methods 0.000 claims 1
- 238000007796 conventional method Methods 0.000 abstract description 2
- 230000001419 dependent effect Effects 0.000 abstract description 2
- 208000024172 Cardiovascular disease Diseases 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 238000009610 ballistocardiography Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000036541 health Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 208000037998 chronic venous disease Diseases 0.000 description 2
- 230000008034 disappearance Effects 0.000 description 2
- 201000010099 disease Diseases 0.000 description 2
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 2
- 239000003814 drug Substances 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 208000019622 heart disease Diseases 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 230000002123 temporal effect Effects 0.000 description 2
- 239000000853 adhesive Substances 0.000 description 1
- 230000001070 adhesive effect Effects 0.000 description 1
- 238000010009 beating Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000001010 compromised effect Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000006806 disease prevention Effects 0.000 description 1
- 230000002526 effect on cardiovascular system Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- JEIPFZHSYJVQDO-UHFFFAOYSA-N ferric oxide Chemical compound O=[Fe]O[Fe]=O JEIPFZHSYJVQDO-UHFFFAOYSA-N 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000002401 inhibitory effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
- A61B5/346—Analysis of electrocardiograms
- A61B5/349—Detecting specific parameters of the electrocardiograph cycle
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
- G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Abstract
The invention discloses a kind of heart impact signal based on LSTM more classifying identification methods automatically, step are as follows: obtain data set training sample；The data set training sample of acquisition is pre-processed using the method for wavelet threshold denoising, obtains pure heart impact signal；The positioning of IJK wave is completed to obtained heart impact signal cooperation adaptive threshold, the heart for obtaining heart impact signal claps interception；LSTM network model is constructed, the heart of obtained heart impact signal is clapped into input data of the interception as LSTM model, network model is trained and is tested；Training process carries out right-value optimization using LSTM network model of the back-propagation algorithm to building, makes network convergence to global optimum；LSTM network model exports discrimination, obtains classification accuracy, according to correcting errors for confusion matrix classification, calculates kapaa coefficient, assesses the nicety of grading of the model.This method is able to solve Long-range dependence and conventional method existing for RNN and is overly dependent upon the not high problem of engineer and nicety of grading, obtains preferable classifying quality.
Description
Technical field
The present invention relates to the physiological signal of deep learning classification fields, and in particular to a kind of heart impact signal based on LSTM
Automatic more classifying identification methods.
Background technique
Cardiovascular disease (Cardiovascular Disease, CVD), also known as circulation system disease, are a series of relate to
And the disease of the circulatory system, have the characteristics that high illness rate, high disability rate and high mortality, referred to as endangers human health
" number one killer ".In China, cardiovascular disease seriously threatens the health of the people, and 2017 " Chinese cardiovascular disease report " is reported,
CVD patient is about 2.9 hundred million in the whole country at present, just has 1 people to suffer from cardiovascular disease in average every 5 people.Therefore, cardiovascular
The prevention of disease, diagnosing and treating have become one of the medicine project of relationship China national health.
Acquisition for heart rate, ECG signal are using mature and extensive.But first by the acquisition mode of adhesive electrode
It is contact, the discomfort on body can be brought to patient, secondly using the electrocardiograph of profession, this had both needed certain medicine
Knowledge, and it is expensive, long-term use can not be also carried out at home.BCG signal, that is, ballistocardiography mode becomes as a result,
Preferable alternative.Ballistocardiography refers to that heart can give human contact's close supporter pressure during beating
Power, and it is exactly ballistocardiography signal that this pressure change, which is collected after being converted into electric signal, this signal contains heart rate letter
Breath, due to its contactless design, is very suitable to middle use of being in, does not also need excessive medical knowledge.It is examined using BCG signal
The technology for surveying heart disease is also rapidly developing, this patient's also important in inhibiting for heart disease.
Existing sorting technique can be roughly divided into two classes, i.e., classification method based on conventional machines study and be based on depth
The classification method of habit, the former mainly uses the methods of support vector machines, k nearest neighbor and FUZZY NETWORK identification.The latter is now artificial
The research hotspot of smart field, is the technology to be grown up based on ANN, and deep learning passes through regulating networks parameter, iteration
Study has powerful feature extraction and data mining capability to search out optimal feature representation model.Constantly to depth
Neural network is trained, and data is transformed into deeper sample space from original sample space layer-by-layer, so that in advance
It surveys or the accuracy of classification task is greatly improved.
In deep learning algorithm, most representative and most commonly used algorithm is exactly convolutional neural networks
(Convolution Neural Network, CNN) and Recognition with Recurrent Neural Network RNN.RNN can be regarded as to be uploaded in a time
The neural network passed, the mode being unfolded in temporal sequence are more applicable for the processing of the time series data in BCG signal, but it can be
On time shaft occur " gradient disappearances ", for t moment, it generate gradient on a timeline to history propagation it is several layers of after
Just disappear.This is in place of the deficiencies in the prior art.
Therefore, for the Long-range dependence problem of prior art RNN, it is automatic to propose a kind of heart impact signal based on LSTM
More classifying identification methods are necessary with solving drawbacks described above in the prior art.
Summary of the invention
It is an object of the invention to be directed to the time series data of BCG signal, existing defect problem is handled with RNN, and is provided
A kind of heart impact signal based on LSTM more classifying identification methods automatically, this method provides one kind to be able to solve existing for RNN
Long-range dependence problem, and solve the problems, such as that conventional method is overly dependent upon engineer and nicety of grading is not high, it obtains preferable
Classifying quality method.
Realizing the technical solution of the object of the invention is:
A kind of heart impact signal based on LSTM more classifying identification methods automatically, include the following steps:
S1, data set training sample is obtained；
S2, the data set training sample that step S1 is obtained is pre-processed using the method for wavelet threshold denoising, is obtained
Pure heart impact signal；
S3, the positioning that IJK wave is completed to the heart impact signal cooperation adaptive threshold that step S2 is obtained, obtain heart impact letter
Number the heart clap interception；
The heart of the obtained heart impact signal of step S3 is clapped interception as LSTM model by S4, building LSTM network model
Input data is trained network model and tests；
S5, training process carry out right-value optimization using LSTM network model of the back-propagation algorithm to building, receive network
Hold back global optimum；
S6, LSTM network model export discrimination, obtain classification accuracy, according to correcting errors for confusion matrix classification, calculate
Kapaa coefficient assesses the nicety of grading of the model.
In step S1, the data set training sample is the heart impact signal hardware Acquisition Circuit acquisition with design,
Sensor of the PVDF piezoelectric membrane developed using U.S. MEAS as acquisition BCG signal.
In step S2, the pretreatment includes the following steps:
S2-1, multi-scale wavelet decompose: suitable wavelet basis and Decomposition order being selected to carry out small wavelength-division to original BCG signal
Solution, obtains the correspondence wavelet coefficient of different decomposition layer；
S2-2, wavelet threshold processing: suitable threshold function table and threshold value are selected, quantifies corresponding wavelet coefficient, by setting
The wavelet coefficient for being less than threshold value is carried out zero setting, achievees the purpose that BCG signal denoising by fixed reasonable threshold value；
S2-3, wavelet inverse transformation is carried out to the wavelet coefficient after quantization, the BCG signal after being denoised.
In step S3, the heart claps interception, is the IJK wave positioned in BCG signal, includes the following steps:
S3-1, positioning J wave point: J wave is usually amplitude highest point in BCG signal signal period, is easier to confirm, is used
The position n of the legal position J wave of local maximum is designed heartbeat interval minimum 0.4 second, then the J-J of BCG signal is set to
0.4*fs sampled point, wherein fs is sample frequency；
S3-2, positioning I wave point: using J wave point position n as origin, Look-ahead, when finding first minimum of J wavefront
When to be judged, after each Look-ahead to minimum point, centered on the point, judge 200 points before the point value whether
The both greater than value of the point is then left I wave, otherwise abandons if more than the value of the point, continues Look-ahead；
S3-3, positioning K wave point: it using J wave as origin, searches backward, first minimum point after finding J wave, with the point
Centered on, judge whether the value of 200 points after the point is both greater than the value of the point, carry out threshold decision, meet threshold decision i.e.
For K wave；
S3-4, BCG signal heart clap interception: on the basis of the position of IJK wave, if doing respectively forwardly, backward, by this
The interception of number of segment strong point is clapped as the heart.
In step S4, the building LSTM network model, which includes input layer, hidden layer and output layer, LSTM
The concept of " input gate, out gate forget door " is introduced, interact the memory mould built in LSTM network jointly between them
Block solves the problems, such as Long-range dependence existing for RNN；
The forgetting door determines the location mode C of last momentt-1How many remains into current time Ct, mathematical table
Up to formula are as follows:
ft=σ (Wf[xt,ht-1,Ct-1]+bf)*Ct-1
Wherein xtFor list entries, ht-1For the output of a preceding memory block, Ct-1For cell state before, WfFor weight
Vector, bfFor bias vector, σ is sigmoid function, if sigmoid function is exported close to 0, then the information stored before
Ct-1It will be by " forgetting "；
The input gate determines the input x of current time networktHow many is saved in location mode Ct, first part is defeated
It is out it, using sigmoid activation primitive, second part output is Ct, use tanh activation primitive；
The out gate is used for control unit state CtHow many is output to the current output value h of LSTMt；
The relationship of input gate and out gate is as follows:
it=σ (Wi*[xt,ht-1,Ct-1]+bi)
Ct=ft+it*tanh(Wc*[xt,ht-1,Ct-1]+bc)
ot=σ (Wo*[xt,ht-1,Ct]+bo)
ht=tanh (Ct)*ot
Wherein itFor input gate, xtFor list entries, WiFor the weight vector of input gate, biFor the bias vector of input gate,
CtFor the output state of input gate, WcFor the weight vector of input gate output state, bcFor being biased towards for input gate output state
Amount, otFor out gate, WoFor the weight vector of out gate, boFor the bias vector of out gate, htFor the output of out gate.
In step S5, the right-value optimization is included the following steps: using the back-propagation algorithm of LSTM
(1) output valve of each neuron of forward calculation, i.e. it、Ct、otAnd htTotally 4 variables；
(2) the error entry value of each neuron of retrospectively calculate, the backpropagation of LSTM error term include both direction: one
It is the backpropagation along the time, i.e., since current t moment, calculates the error term at each moment；One is that error term is upward
One Es-region propagations；
(3) according to corresponding error term, the gradient of each weight is calculated；
If indicating that error amount, W represent the weight matrix in network with E, then in LSTM backpropagation, the output of t moment
Error gradient δhtWith location mode CtError gradient δctIt can be derived by the gradient conditions at t+1 moment:
According to gradient, then the right value update for forgeing door, input gate and out gate is respectively as follows:
In step S6, the classification accuracy is to pass through calculating according to discrimination with Kappa coefficient appraisal procedure
The size of kappa coefficient determines that the recognition performance of LSTM network model, kappa coefficient really divide particular by all earth's surfaces
Pixel in class total (N) then subtracts certain true pixel sum of one kind earth's surface and quilt multiplied by sum of confusion matrix diagonal line (Xkk)
Accidentally be divided into the product of such pixel sum to all categories sum as a result, again divided by total pixel number with square subtracting in certain one kind
The result that the true pixel sum of table sums to all categories with the product for being accidentally divided into such pixel sum in such is obtained；
The value of kappa coefficient is higher, then the classification accuracy that representative model is realized is higher, the calculation formula of classification accuracy
It is as follows:
Kappa coefficient is the model evaluation parameter of being calculated based on confusion matrix, and the value of this coefficient is higher, then generation
The classification accuracy that table model is realized is higher, Kappa coefficient formulas are as follows:
Wherein, p0It is expressed as total classification accuracy, peIt is expressed asaiStatement i-th
Class authentic specimen number, biIt states the i-th class and predicts the number of samples come.
The utility model has the advantages that a kind of heart impact signal based on LSTM provided by the invention more classifying identification methods automatically, just like
Lower advantage:
First, the filtering processing of heart impact signal is completed using the method for wavelet threshold, is obtained more pure signal, is subtracted
Interference of other noises to accuracy rate when classification is lacked；
Second, the application of deep learning reduces a large amount of manual features and extracts work, is compared to CNN and utilizes convolution kernel
The method to extract feature is translated in data to be processed, the mode that RNN is unfolded in temporal sequence is more applicable for BCG
The processing of signal, but the memory of RNN can only be related with several sequences in front, and gradient disappearance or quick-fried is easy to produce in training
Broken phenomenon；And LSTM type neural network the problem of having well solved RNN, it can really efficiently use BCG signal characteristic
Timing information, to improve the discrimination of BCG signal.
Detailed description of the invention
Fig. 1 is flow chart of the invention；
Fig. 2 is the hardware circuit diagram that data sample obtains；
Fig. 3 is Wavelet Denoising Method flow chart；
Fig. 4 is denoising effect picture；
Fig. 5 is the I wave point labelling method program flow diagram with threshold decision；
Fig. 6 is the application schematic diagram of LSTM network model；
Fig. 7 is the basic and unfolding assumption diagram of RNN；
Fig. 8 is the unit detail view of RNN and LSTM, wherein (a) is RNN, it (b) is LSTM.
Specific embodiment
The present invention is further elaborated with reference to the accompanying drawings and examples, but is not limitation of the invention.
Embodiment:
As shown in Figure 1, a kind of heart impact signal based on LSTM more classifying identification methods automatically, include the following steps:
S1, data set training sample is obtained；
S2, the data set training sample that step S1 is obtained is pre-processed using the method for wavelet threshold denoising, is obtained
Pure heart impact signal；
S3, the positioning that IJK wave is completed to the heart impact signal cooperation adaptive threshold that step S2 is obtained, obtain heart impact letter
Number the heart clap interception；
The heart of the obtained heart impact signal of step S3 is clapped interception as LSTM model by S4, building LSTM network model
Input data is trained network model and tests；
S5, training process carry out right-value optimization using LSTM network model of the back-propagation algorithm to building, receive network
Hold back global optimum；
S6, LSTM network model export discrimination, obtain accurate true rate of classifying, and according to correcting errors for confusion matrix classification, count
Kapaa coefficient is calculated, the nicety of grading of the model is assessed.
A kind of heart impact signal based on LSTM more classifying identification methods automatically, include the following steps:
S1, data set training sample is obtained；
S2, the data set training sample that step S1 is obtained is pre-processed using the method for wavelet threshold denoising, is obtained
Pure heart impact signal；
S3, the positioning that IJK wave is completed to the heart impact signal cooperation adaptive threshold that step S2 is obtained, obtain heart impact letter
Number the heart clap interception；
The heart of the obtained heart impact signal of step S3 is clapped interception as LSTM model by S4, building LSTM network model
Input data is trained network model and tests；
S5, training process carry out right-value optimization using LSTM network model of the back-propagation algorithm to building, receive network
Hold back global optimum；
S6, LSTM network model export discrimination, obtain accurate true rate of classifying, and according to correcting errors for confusion matrix classification, count
Kapaa coefficient is calculated, the nicety of grading of the model is assessed.
In step S1, the data set training sample is the heart impact signal hardware Acquisition Circuit acquisition with design,
The hardware circuit diagram that data sample obtains is as shown in Fig. 2, the PVDF piezoelectric membrane developed using U.S. MEAS is believed as acquisition BCG
Number sensor.
In step S2, the pretreatment, as shown in figure 3, including the following steps:
S2-1, multi-scale wavelet decompose: suitable wavelet basis and Decomposition order being selected to carry out small wavelength-division to original BCG signal
Solution, obtains the correspondence wavelet coefficient of different decomposition layer；After wavelet decomposition, the effective information of signal is concentrated mainly on amplitude
Wavelet coefficient on, and noise signal is then distributed in each multi-scale wavelet domain, is concentrated mainly on the wavelet coefficient compared with small magnitude.
S2-2, wavelet threshold processing: suitable threshold function table and threshold value are selected, quantifies corresponding wavelet coefficient, by setting
The wavelet coefficient for being less than threshold value is carried out zero setting, achievees the purpose that BCG signal denoising by fixed reasonable threshold value；
S2-3, wavelet inverse transformation is carried out to the wavelet coefficient after quantization, the BCG signal after being denoised denoises effect picture
As shown in Figure 4.
In step S3, the heart claps interception, is the IJK wave positioned in BCG signal, includes the following steps:
S3-1, positioning J wave point: J wave is usually amplitude highest point in BCG signal signal period, is easier to confirm, is used
The position n of the legal position J wave of local maximum is designed heartbeat interval minimum 0.4 second, then the J-J of BCG signal is set to
0.4*fs sampled point, wherein fs is sample frequency；
S3-2, positioning I wave point: using J wave point position n as origin, Look-ahead, when finding first minimum of J wavefront
When to be judged, after each Look-ahead to minimum point, centered on the point, judge 200 points before the point value whether
The both greater than value of the point is then left I wave, otherwise abandons if more than the value of the point, continues Look-ahead；
S3-3, positioning K wave point: it using J wave as origin, searches backward, first minimum point after finding J wave, with the point
Centered on, judge whether the value of 200 points after the point is both greater than the value of the point, carry out threshold decision, meet threshold decision i.e.
For K wave；
S3-4, BCG signal heart clap interception: on the basis of the position of IJK wave, if doing respectively forwardly, backward, by this
The interception of number of segment strong point is clapped as the heart.No matter IJK wave location algorithm design it is how perfect, all necessarily have false retrieval and missing inspection.The present invention
The means compromised using one: the heart for ignoring missing inspection is clapped, and false retrieval and the heart correctly detected are clapped all interceptions and come out.Since IJK is fixed
Position algorithm, performance it is more perfect, so interception come out the heart clap only have small part be the false retrieval heart clap.And the deep learning classified
Algorithm all has certain robustness, and less false retrieval heart umber of beats will not have much impact to result.
As shown in fig. 6, in step S4, the building LSTM network model, the model includes input layer, hidden layer and defeated
Layer out, LSTM introduce the concept of " input gate, out gate forget door ", interact between them and build LSTM network jointly
In memory module, solve the problems, such as Long-range dependence existing for RNN.From figure 7 it can be seen that by past output and working as in RNN
Preceding input connects together, the output both controlled by tanh, it only considers the state at nearest moment, Fig. 8 be RNN and
The unit detail view of LSTM compares, and as can be seen from the figure LSTM increases on the basis of RNN in order to remember long-term state
It input and exports all the way all the way, increased this be exactly cell state all the way；
The forgetting door determines the location mode C of last momentt-1How many remains into current time Ct, mathematical table
Up to formula are as follows:
ft=σ (Wf[xt,ht-1,Ct-1]+bf)*Ct-1
Wherein xtFor list entries, ht-1For the output of a preceding memory block, Ct-1For cell state before, WfFor weight
Vector, bfFor bias vector, σ is sigmoid function, if sigmoid function is exported close to 0, then the information stored before
Ct-1It will be by " forgetting "；
The input gate determines the input x of current time networktHow many is saved in location mode Ct, first part is defeated
It is out it, using sigmoid activation primitive, second part output is Ct, use tanh activation primitive；
The out gate is used for control unit state CtHow many is output to the current output value h of LSTMt；
The relationship of input gate and out gate is as follows:
it=σ (Wi*[xt,ht-1,Ct-1]+bi)
Ct=ft+it*tanh(Wc*[xt,ht-1,Ct-1]+bc)
ot=σ (Wo* [xt,ht-1,Ct]+bo)
ht=tanh (Ct)*ot
Wherein itFor input gate, xtFor list entries, WiFor the weight vector of input gate, biFor the bias vector of input gate,
CtFor the output state of input gate, WcFor the weight vector of input gate output state, bcFor being biased towards for input gate output state
Amount, otFor out gate, WoFor the weight vector of out gate, boFor the bias vector of out gate, htFor the output of out gate.
In step S5, the right-value optimization is included the following steps: using the back-propagation algorithm of LSTM
(2) output valve of each neuron of forward calculation, i.e. it、Ct、otAnd htTotally 4 variables；
(2) the error entry value of each neuron of retrospectively calculate, the backpropagation of LSTM error term include both direction: one
It is the backpropagation along the time, i.e., since current t moment, calculates the error term at each moment；One is that error term is upward
One Es-region propagations；
(3) according to corresponding error term, the gradient of each weight is calculated；
If indicating that error amount, W represent the weight matrix in network with E, then in LSTM backpropagation, the output of t moment
Error gradient δhtWith location mode CtError gradient δctIt can be derived by the gradient conditions at t+1 moment:
According to gradient, then the right value update for forgeing door, input gate and out gate is respectively as follows:
In step S6, the classification accuracy is to pass through calculating according to discrimination with Kappa coefficient appraisal procedure
The size of kappa coefficient determines that the recognition performance of LSTM network model, kappa coefficient really divide particular by all earth's surfaces
Pixel in class total (N) then subtracts certain true pixel sum of one kind earth's surface and quilt multiplied by sum of confusion matrix diagonal line (Xkk)
Accidentally be divided into the product of such pixel sum to all categories sum as a result, again divided by total pixel number with square subtracting in certain one kind
The result that the true pixel sum of table sums to all categories with the product for being accidentally divided into such pixel sum in such is obtained；
Kappa calculated result is -1~1, but usually kappa is fallen between 0~1, can be divided into five groups to indicate different stage
Consistency: 0.0~0.20 extremely low consistency (slight), 0.21~0.40 general consistency (fair), 0.41~
0.60 medium consistency (moderate), 0.61~0.80 high consistency (substantial) and 0.81~1 are almost complete
Complete consistent (almost perfect).
The value of kappa coefficient is higher, then the classification accuracy that representative model is realized is higher, the calculation formula of classification accuracy
It is as follows:
Kappa coefficient is the model evaluation parameter of being calculated based on confusion matrix, and the value of this coefficient is higher, then generation
The classification accuracy that table model is realized is higher, Kappa coefficient formulas are as follows:
Wherein, p0 is expressed as total classification accuracy, peIt is expressed asaiStatement i-th
Class authentic specimen number, bi state the i-th class and predict the number of samples come.
Claims (7)
1. a kind of heart impact signal based on LSTM more classifying identification methods automatically, which comprises the steps of:
S1, data set training sample is obtained；
S2, the data set training sample that step S1 is obtained is pre-processed using the method for wavelet threshold denoising, is obtained pure
Heart impact signal；
S3, the positioning that IJK wave is completed to the heart impact signal cooperation adaptive threshold that step S2 is obtained, obtain heart impact signal
The heart claps interception；
The heart of the obtained heart impact signal of step S3 is clapped input of the interception as LSTM model by S4, building LSTM network model
Data are trained network model and test；
S5, training process carry out right-value optimization using LSTM network model of the back-propagation algorithm to building, arrive network convergence
Global optimum；
S6, LSTM network model export discrimination, obtain classification accuracy, according to correcting errors for confusion matrix classification, calculate kapaa
Coefficient assesses the nicety of grading of the model.
2. more classifying identification methods, feature exist a kind of heart impact signal based on LSTM according to claim 1 automatically
In in step S1, the data set training sample is the heart impact signal hardware Acquisition Circuit acquisition with design, uses
Sensor of the PVDF piezoelectric membrane that U.S. MEAS is developed as acquisition BCG signal.
3. more classifying identification methods, feature exist a kind of heart impact signal based on LSTM according to claim 1 automatically
In in step S2, the pretreatment includes the following steps:
S2-1, multi-scale wavelet decompose: selecting suitable wavelet basis and Decomposition order to carry out wavelet decomposition to original BCG signal, obtain
To the correspondence wavelet coefficient of different decomposition layer；
S2-2, wavelet threshold processing: suitable threshold function table and threshold value are selected, corresponding wavelet coefficient is quantified, is closed by setting
The wavelet coefficient for being less than threshold value is carried out zero setting, achievees the purpose that BCG signal denoising by the threshold value of reason；
S2-3, wavelet inverse transformation is carried out to the wavelet coefficient after quantization, the BCG signal after being denoised.
4. more classifying identification methods, feature exist a kind of heart impact signal based on LSTM according to claim 1 automatically
In in step S3, the heart claps interception, is the IJK wave positioned in BCG signal, includes the following steps:
S3-1, positioning J wave point: J wave is usually amplitude highest point in BCG signal signal period, is easier to confirm, using part
Maximum Approach positions the position n of J wave, designs heartbeat interval minimum 0.4 second, then the J-J of BCG signal is set to 0.4*fs
A sampled point, wherein fs is sample frequency；
S3-2, positioning I wave point: using J wave point position n as origin, Look-ahead is wanted when finding first minimum of J wavefront
Judged after each Look-ahead to minimum point, centered on the point, judge whether the value of 200 points before the point is all big
In the value of the point, if more than the value of the point, then it is left I wave, is otherwise abandoned, continues Look-ahead；
S3-3, positioning K wave point: using J wave as origin, searching, first minimum point after finding J wave backward, is with the point
The heart, judges whether the value of 200 points after the point is both greater than the value of the point, carries out threshold decision, and that meet threshold decision is K
Wave；
S3-4, BCG signal heart clap interception: on the basis of the position of IJK wave, if doing respectively forwardly, backward, by this number of segment
Strong point interception is clapped as the heart.
5. more classifying identification methods, feature exist a kind of heart impact signal based on LSTM according to claim 1 automatically
In in step S4, the building LSTM network model, the model includes input layer, hidden layer and output layer, and LSTM is introduced
The concept of " input gate, forgets door at out gate ", interact the memory module built in LSTM network jointly between them；
The forgetting door determines the location mode C of last momentt-1How many remains into current time Ct, mathematic(al) representation
Are as follows:
ft=σ (Wf[xt,ht-1,Ct-1]+bf)*Ct-1
Wherein xtFor list entries, ht-1For the output of a preceding memory block, Ct-1For cell state before, WfFor weight vector,
bfFor bias vector, σ is sigmoid function, if sigmoid function is exported close to 0, then the information C stored beforet-1It will
It can be by " forgetting "；
The input gate determines the input x of current time networktHow many is saved in location mode Ct, first part, which exports, is
it, using sigmoid activation primitive, second part output is Ct, use tanh activation primitive；
The out gate is used for control unit state CtHow many is output to the current output value h of LSTMt；
The relationship of input gate and out gate is as follows:
it=σ (Wi*[xt,ht-1,Ct-1]+bi)
Ct=ft+it*tanh(Wc*[xt,ht-1,Ct-1]+bc)
ot=σ (Wo*[xt,ht-1,Ct]+bo)
ht=tanh (Ct)*ot
Wherein itFor input gate, xtFor list entries, WiFor the weight vector of input gate, biFor the bias vector of input gate, CtFor
The output state of input gate, WcFor the weight vector of input gate output state, bcFor the bias vector of input gate output state, ot
For out gate, WoFor the weight vector of out gate, boFor the bias vector of out gate, htFor the output of out gate.
6. more classifying identification methods, feature exist a kind of heart impact signal based on LSTM according to claim 1 automatically
In in step S5, the right-value optimization is included the following steps: using the back-propagation algorithm of LSTM
(1) output valve of each neuron of forward calculation, i.e. it、Ct、otAnd htTotally 4 variables；
(2) the error entry value of each neuron of retrospectively calculate, the backpropagation of LSTM error term include both direction: one is edge
The backpropagation of time calculates the error term at each moment that is, since current t moment；One is by error term upper layer
It propagates；
(3) according to corresponding error term, the gradient of each weight is calculated；
If indicating that error amount, W represent the weight matrix in network with E, then in LSTM backpropagation, the output error of t moment
Gradient δhtWith location mode CtError gradient δctIt can be derived by the gradient conditions at t+1 moment:
According to gradient, then the right value update for forgeing door, input gate and out gate is respectively as follows:
7. more classifying identification methods, feature exist a kind of heart impact signal based on LSTM according to claim 1 automatically
In in step S6, the classification accuracy is to pass through calculating kappa system according to discrimination with Kappa coefficient appraisal procedure
Several sizes determines the recognition performance of LSTM network model, kappa coefficient particular by all earth's surfaces are really classified in
Pixel sum (N) then subtracts certain one kind and the true pixel sum of earth's surface and is accidentally divided into multiplied by sum of confusion matrix diagonal line (Xkk)
The product of such pixel sum to all categories sum as a result, again divided by total pixel number square to subtract earth's surface in certain one kind true
Pixel sum and the result that the product for being accidentally divided into such pixel sum in such sums to all categories are obtained；
The value of kappa coefficient is higher, then the classification accuracy that representative model is realized is higher, and the calculation formula of classification accuracy is such as
Under:
Kappa coefficient is the model evaluation parameter of being calculated based on confusion matrix, and the value of this coefficient is higher, then represents mould
The classification accuracy that type is realized is higher, Kappa coefficient formulas are as follows:
Wherein, p0It is expressed as total classification accuracy, peIt is expressed asaiIt is true to state the i-th class
Real number of samples, biIt states the i-th class and predicts the number of samples come.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910846195.4A CN110427924A (en) | 2019-09-09 | 2019-09-09 | A kind of heart impact signal based on LSTM more classifying identification methods automatically |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910846195.4A CN110427924A (en) | 2019-09-09 | 2019-09-09 | A kind of heart impact signal based on LSTM more classifying identification methods automatically |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110427924A true CN110427924A (en) | 2019-11-08 |
Family
ID=68417881
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910846195.4A Pending CN110427924A (en) | 2019-09-09 | 2019-09-09 | A kind of heart impact signal based on LSTM more classifying identification methods automatically |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110427924A (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111044617A (en) * | 2019-11-28 | 2020-04-21 | 东南大学 | Inhaul cable damage identification method based on deep learning and acoustic emission technology |
| CN111134662A (en) * | 2020-02-17 | 2020-05-12 | 武汉大学 | Electrocardio abnormal signal identification method and device based on transfer learning and confidence degree selection |
| CN111141412A (en) * | 2019-12-25 | 2020-05-12 | 深圳供电局有限公司 | Cable temperature and anti-theft dual-monitoring method and system and readable storage medium |
| CN111387968A (en) * | 2020-03-18 | 2020-07-10 | 南京润楠医疗电子研究院有限公司 | Accurate heart attack signal beat-by-beat heart rate calculation device and method |
| CN111476205A (en) * | 2020-02-26 | 2020-07-31 | 安徽建筑大学 | Personnel counting method and device based on L STM model |
| CN111772669A (en) * | 2020-08-18 | 2020-10-16 | 中国科学院合肥物质科学研究院 | Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network |
| CN112197983A (en) * | 2020-09-02 | 2021-01-08 | 中车长春轨道客车股份有限公司 | Train service performance identification method |
| CN112294264A (en) * | 2020-10-31 | 2021-02-02 | 无锡中物云信息科技有限公司 | Sleep staging method based on BCG and blood oxygen saturation rate |
| CN113647925A (en) * | 2021-07-05 | 2021-11-16 | 新绎健康科技有限公司 | Heart rate determination method and device based on heart attack signal |
| CN113970697A (en) * | 2021-09-09 | 2022-01-25 | 北京无线电计量测试研究所 | Analog circuit state evaluation method and device |
| CN114004342A (en) * | 2021-11-29 | 2022-02-01 | 中国科学院光电技术研究所 | Laser communication system distortion wavefront prediction method based on LSTM network |
| CN114010186A (en) * | 2022-01-11 | 2022-02-08 | 华南师范大学 | Ballistocardiogram signal positioning method and computer equipment |
| CN114176548A (en) * | 2021-12-03 | 2022-03-15 | 新绎健康科技有限公司 | Heart attack signal heart rate calculation method and system based on template matching |
| CN114376521A (en) * | 2021-12-27 | 2022-04-22 | 天翼云科技有限公司 | Sleep state recognition model training and sleep staging method and device |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108354612A (en) * | 2018-01-19 | 2018-08-03 | 深圳和而泰数据资源与云技术有限公司 | A kind of signal processing method and device |
| CN109171707A (en) * | 2018-10-24 | 2019-01-11 | 杭州电子科技大学 | A kind of intelligent cardiac figure classification method |
| CN109614885A (en) * | 2018-11-21 | 2019-04-12 | 齐鲁工业大学 | A kind of EEG signals Fast Classification recognition methods based on LSTM |
-
2019
- 2019-09-09 CN CN201910846195.4A patent/CN110427924A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108354612A (en) * | 2018-01-19 | 2018-08-03 | 深圳和而泰数据资源与云技术有限公司 | A kind of signal processing method and device |
| CN109171707A (en) * | 2018-10-24 | 2019-01-11 | 杭州电子科技大学 | A kind of intelligent cardiac figure classification method |
| CN109614885A (en) * | 2018-11-21 | 2019-04-12 | 齐鲁工业大学 | A kind of EEG signals Fast Classification recognition methods based on LSTM |
Non-Patent Citations (2)
| Title |
|---|
| 潘周光: "基于PVDF压电薄膜的非接触式人体生理信号检测系统研究", 《中国优秀硕士学位论文全文数据库 工程科技I辑》 * |
| 王曼: "基于高阶统计量的BCG信号特征分析与识别", 《中国优秀硕士学位论文全文数据库 医药卫生科技辑》 * |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111044617B (en) * | 2019-11-28 | 2022-08-02 | 东南大学 | Inhaul cable damage identification method based on deep learning and acoustic emission technology |
| CN111044617A (en) * | 2019-11-28 | 2020-04-21 | 东南大学 | Inhaul cable damage identification method based on deep learning and acoustic emission technology |
| CN111141412A (en) * | 2019-12-25 | 2020-05-12 | 深圳供电局有限公司 | Cable temperature and anti-theft dual-monitoring method and system and readable storage medium |
| CN111134662A (en) * | 2020-02-17 | 2020-05-12 | 武汉大学 | Electrocardio abnormal signal identification method and device based on transfer learning and confidence degree selection |
| CN111134662B (en) * | 2020-02-17 | 2021-04-16 | 武汉大学 | Electrocardio abnormal signal identification method and device based on transfer learning and confidence degree selection |
| CN111476205B (en) * | 2020-02-26 | 2023-10-17 | 安徽建筑大学 | Personnel counting method and device based on LSTM model |
| CN111476205A (en) * | 2020-02-26 | 2020-07-31 | 安徽建筑大学 | Personnel counting method and device based on L STM model |
| CN111387968A (en) * | 2020-03-18 | 2020-07-10 | 南京润楠医疗电子研究院有限公司 | Accurate heart attack signal beat-by-beat heart rate calculation device and method |
| CN111387968B (en) * | 2020-03-18 | 2023-09-26 | 南京润楠医疗电子研究院有限公司 | Accurate heart beat-by-beat heart rate calculation device and method for ballistocardiogram signals |
| CN111772669A (en) * | 2020-08-18 | 2020-10-16 | 中国科学院合肥物质科学研究院 | Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network |
| CN111772669B (en) * | 2020-08-18 | 2022-08-19 | 中国科学院合肥物质科学研究院 | Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network |
| CN112197983A (en) * | 2020-09-02 | 2021-01-08 | 中车长春轨道客车股份有限公司 | Train service performance identification method |
| CN112294264A (en) * | 2020-10-31 | 2021-02-02 | 无锡中物云信息科技有限公司 | Sleep staging method based on BCG and blood oxygen saturation rate |
| CN113647925A (en) * | 2021-07-05 | 2021-11-16 | 新绎健康科技有限公司 | Heart rate determination method and device based on heart attack signal |
| CN113970697A (en) * | 2021-09-09 | 2022-01-25 | 北京无线电计量测试研究所 | Analog circuit state evaluation method and device |
| CN113970697B (en) * | 2021-09-09 | 2023-06-13 | 北京无线电计量测试研究所 | Analog circuit state evaluation method and device |
| CN114004342B (en) * | 2021-11-29 | 2023-05-30 | 中国科学院光电技术研究所 | Laser communication system distortion wavefront prediction method based on LSTM network |
| CN114004342A (en) * | 2021-11-29 | 2022-02-01 | 中国科学院光电技术研究所 | Laser communication system distortion wavefront prediction method based on LSTM network |
| CN114176548A (en) * | 2021-12-03 | 2022-03-15 | 新绎健康科技有限公司 | Heart attack signal heart rate calculation method and system based on template matching |
| CN114376521A (en) * | 2021-12-27 | 2022-04-22 | 天翼云科技有限公司 | Sleep state recognition model training and sleep staging method and device |
| CN114010186B (en) * | 2022-01-11 | 2022-03-18 | 华南师范大学 | Ballistocardiogram signal positioning method and computer equipment |
| CN114010186A (en) * | 2022-01-11 | 2022-02-08 | 华南师范大学 | Ballistocardiogram signal positioning method and computer equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110427924A (en) | A kind of heart impact signal based on LSTM more classifying identification methods automatically | |
| CN109948647B (en) | Electrocardiogram classification method and system based on depth residual error network | |
| Yang et al. | Spatiotemporal differentiation of myocardial infarctions | |
| Javadi et al. | Improving ECG classification accuracy using an ensemble of neural network modules | |
| CN108549841A (en) | A kind of recognition methods of the Falls Among Old People behavior based on deep learning | |
| CN108511055B (en) | Ventricular premature beat recognition system and method based on classifier fusion and diagnosis rules | |
| CN105427296A (en) | Ultrasonic image low-rank analysis based thyroid lesion image identification method | |
| Odeh | Using an Adaptive Neuro-Fuzzy Inference System(AnFis) Algorithm for Automatic Diagnosis of Skin Cancer | |
| CN112396001B (en) | Rope skipping number statistical method based on human body posture estimation and TPA (tissue placement model) attention mechanism | |
| CN110313894A (en) | Arrhythmia cordis sorting algorithm based on convolutional neural networks | |
| CN114469124B (en) | Method for identifying abnormal electrocardiosignals in movement process | |
| CN111626114A (en) | Electrocardiosignal arrhythmia classification system based on convolutional neural network | |
| Yao et al. | Interpretation of electrocardiogram heartbeat by CNN and GRU | |
| CN111053552A (en) | QRS wave detection method based on deep learning | |
| CN113705645A (en) | Adaptive joint model semi-supervised learning classification method for electroencephalogram signals | |
| CN113705507A (en) | Mixed reality open set human body posture recognition method based on deep learning | |
| CN111144167A (en) | Gait information identification optimization method, system and storage medium | |
| Priyadharsini et al. | An Efficient method for the removal of ECG artifact from measured EEG Signal using PSO algorithm | |
| Hossain et al. | Detection & classification of tumor cells from bone mr imagery using connected component analysis & neural network | |
| CN114004258B (en) | Semi-supervised electrocardiographic abnormality detection method | |
| CN107092895A (en) | A kind of multi-modal emotion identification method based on depth belief network | |
| Gangadharan et al. | Detection and classification of various pest attacks and infection on plants using RBPN with GA based PSO algorithm | |
| CN116595465A (en) | High-dimensional sparse data outlier detection method and system based on self-encoder and data enhancement | |
| CN115601833A (en) | Myoelectric gesture recognition memory network method and system integrating double-layer attention and multi-stream convolution | |
| CN111563411A (en) | Method for classifying electrocardiosignals by using optimized AdaBoost weighting mode and weak classifier |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191108 |