CN109009033A - A kind of blood pressure prediction technique based on wavelet analysis and echo state network - Google Patents
A kind of blood pressure prediction technique based on wavelet analysis and echo state network Download PDFInfo
- Publication number
- CN109009033A CN109009033A CN201810749008.6A CN201810749008A CN109009033A CN 109009033 A CN109009033 A CN 109009033A CN 201810749008 A CN201810749008 A CN 201810749008A CN 109009033 A CN109009033 A CN 109009033A
- Authority
- CN
- China
- Prior art keywords
- blood pressure
- low frequency
- prediction
- wavelet
- frequency sequence
- 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
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/02—Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
- A61B5/021—Measuring pressure in heart or blood vessels
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7253—Details of waveform analysis characterised by using transforms
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7271—Specific aspects of physiological measurement analysis
- A61B5/7275—Determining trends in physiological measurement data; Predicting development of a medical condition based on physiological measurements, e.g. determining a risk factor
Landscapes
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Heart & Thoracic Surgery (AREA)
- Public Health (AREA)
- Physiology (AREA)
- Biophysics (AREA)
- Pathology (AREA)
- Veterinary Medicine (AREA)
- Biomedical Technology (AREA)
- Physics & Mathematics (AREA)
- Medical Informatics (AREA)
- Molecular Biology (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Psychiatry (AREA)
- Signal Processing (AREA)
- Cardiology (AREA)
- Vascular Medicine (AREA)
- Measuring Pulse, Heart Rate, Blood Pressure Or Blood Flow (AREA)
Abstract
The blood pressure prediction technique based on wavelet analysis and echo state network that the invention discloses a kind of, include the following steps: that original blood pressure data is decomposed into high frequency series part and low frequency Sequence first with wavelet-decomposing method, then echo state network model is constructed, carry out the training of echo state network model to the training sample of low frequency Sequence and high frequency series part respectively again, the echo state network model obtained followed by training carries out Single-step Prediction and multi-step prediction to the test sample of high frequency series part and low frequency Sequence respectively, wavelet reconstruction finally is carried out to the prediction result of high frequency series part and low frequency Sequence, obtain Single-step Prediction blood pressure data and multi-step prediction blood pressure data.The method of the present invention can predict blood pressure trend well in the Single-step Prediction and multi-step prediction of blood pressure time series, and accurately sketch out the variation form of blood pressure, have important clinical significance to the promptly and accurately prediction of blood pressure time series.
Description
Technical field
The invention belongs to medical treatment & health, machine learning, artificial intelligence and medical signals process fields, are related to blood pressure time sequence
The analysis and prediction of column, in particular to a kind of blood pressure prediction technique based on wavelet analysis and echo state network.
Background technique
Blood pressure is the important indicator for reflecting human health status.Dysarteriotony can threaten the life security of human body, such as suffer from
The morbidity of person's postoperative acute low blood pressure and the morbidity of senile hypertension.The generation of dysarteriotony needs to carry out timely and effectively
Therapeutic intervention, time seem most valuable and urgent.Therefore, it is important can timely and accurately to predict that the trend of blood pressure has
Clinical meaning.
Human blood-pressure is by various influences, such as time change, seasonal variations, age growth, weight, drink round the clock
Food, mood etc., therefore there is stronger randomness and complexity, it is a kind of time series of nonlinear and nonstationary.
Currently, the prediction technique of blood pressure time series mainly has neural network method, BP neural network and wavelet decomposition group
Close prediction technique, AR model and support vector machines combination forecasting method, wavelet transformation and support vector machines combination forecasting method, double
Spectrum analysis-BP neural network combination forecasting method, EMD and GEP combination forecasting method, wavelet analysis and Gaussian regression combination
Prediction technique etc..
Practice have shown that single-mode is difficult to obtain ideal precision of prediction to the blood pressure time series of nonlinear and nonstationary；
Existing combination forecasting method is mainly used in the Single-step Prediction of blood pressure time series, in the multi-step prediction of blood pressure time series
In, still remain the problems such as precision of prediction is not high, and prediction step is too short.
Summary of the invention
It is an object of the invention to overcome shortcoming and deficiency in the prior art, provide a kind of based on wavelet analysis and echo
It is inaccurate to the blood pressure time series forecasting of nonlinear and nonstationary to improve existing prediction model for the blood pressure prediction technique of state network
Really, prediction step is too short, training algorithm is excessively complicated, is easily trapped into the problems such as local optimum.Especially in blood pressure time series
Multi-step prediction in, prediction effect is obviously improved.
In order to achieve the above object, the present invention adopts the following technical scheme that:
A kind of blood pressure prediction technique based on wavelet analysis and echo state network, includes the following steps:
S1, small echo is carried out using the T/F local analysis method of wavelet transformation to original blood pressure time series data
It decomposes, is low frequency Sequence and high frequency series part by original blood pressure Time Series；
S2, the reserve pool parameter for determining echo state network model；
S3, the instruction for carrying out echo state network model to the training sample of low frequency Sequence and high frequency series part respectively
Practice, obtains the state equation of reserve pool output prediction；
S4, the echo shape for the echo state network model and high frequency series part for training obtained low frequency Sequence is utilized
State network model carries out Single-step Prediction and multi-step prediction to the test sample of low frequency Sequence and high frequency series part respectively；
S5, by the single step of the Single-step Prediction data of low frequency Sequence and multi-step prediction prediction data and high frequency series part
Prediction data and multi-step prediction prediction data carry out wavelet reconstruction respectively, obtain Single-step Prediction blood pressure data and multi-step prediction blood pressure
Data.
As a preferred technical solution, step S1 specifically include the following steps:
S11, the original blood pressure time series of acquisition, and it is classified as training sample and test sample；
S12, wavelet decomposition is carried out to the original blood pressure time series, obtains the low frequency sequence of original blood pressure time series
Part and high frequency series part；
The continuous wavelet transform formula of wavelet decomposition process are as follows:
Signal u (t) ∈ L in formula2(R), L2(R) indicate be defined on real axis, measurable quadractically integrable function space；Indicate that the continuous wavelet generated by wavelet mother function Ψ (t), a indicate contraction-expansion factor, b table
Show shift factor；
Contraction-expansion factor a and the corresponding discrete wavelet function of shift factor b discretization are as follows:
Wherein, a=2-s, b=2-sK, s and k belong to set of integers Z；
With the Mallet fast algorithm of orthogonal wavelet transformation, by signal u (t) rectangular projection to SPACE VjAnd WjOn, it is corresponding
Obtain resolution ratio 2jThe approximation signal a of lower signal u (t)j(t) and discrete signal dj(t), the approximation signal ajIt (t) is decomposition
Low frequency signal, the discrete signal djIt (t) is the high-frequency signal decomposed, signal u (t) obtains high frequency sequence after Multiresolution Decomposition
The signal of part and the signal of low frequency Sequence are arranged, is indicated are as follows:N is positive integer,
Indicate the wavelet decomposition number of plies；I is positive integer；T is the moment.
As a preferred technical solution, in step S2, the performance of the echo state network model by reserve pool parameter
It determines, the parameter of the reserve pool includes: connection weight spectral radius SR, reserve pool scale N, reserve pool input unit inside reserve pool
The scale IS and sparse degree SD of reserve pool.
As a preferred technical solution, step S3 specifically include the following steps:
S31, the connection weight matrix for initializing echo state network model, including input connection weight matrix Win, deposit
The connection weight matrix W of the inside connection weight matrix W in pond and output layer to reserve poolback, wherein WinIt is randomly generated with W
, set Wback=0；
S32, echo state network model is carried out to the training sample of low frequency Sequence and high frequency series part respectively
Training, specifically: it send low frequency Sequence or the training sample of high frequency series part as input neuron into initialization
In echo state network model, reserve pool intrinsic nerve member state renewal equation is as follows:
X (t+1)=f (Winuin(t+1)+Wx(t)+Wbacky(t))
F is intrinsic nerve member activation primitive in formula, specifically uses hyperbolic tangent function；When t=0, reserve pool intrinsic nerve
First state vector x (t)=0；The reserve pool intrinsic nerve member state vector x (t+1) at t+1 moment is by currently inputting neuron uin(t
+ 1), excitation generates the reserve pool intrinsic nerve member state vector x (t) and desired output y (t) of last moment jointly；With each
The state vector x (t) at moment is that row forms matrix X, and corresponding desired output y (t) constitutes a column vector Y；
The state equation of reserve pool output prediction are as follows:
W in formulaoutTo export connection weight matrix, the output connection weight matrix is solved using least square method:
Wout=XTY=(XTX)-1XTY
Wherein, XTIt is the pseudoinverse of X.
As a preferred technical solution, in step S4, to the test specimens one's duty of low frequency Sequence and high frequency series part
Not carry out Single-step Prediction and multi-step prediction, specifically include the following steps:
S41, the test sample of low frequency Sequence and high frequency series part is separately input to low frequency sequence echo state
Network model and high frequency series echo state network model, obtain the Single-step Prediction data of low frequency SequenceWith
The Single-step Prediction data of high frequency series part
S42, the multi-step prediction data for obtaining low frequency Sequence, detailed process is as follows:
The Single-step Prediction data of S421, low frequency SequenceFor the expectation of t+1 moment low frequency Sequence
The desired output is returned to network inputs part, replaces the test sample of the low frequency Sequence at t+2 moment and conduct by output
It inputs neuron and inputs network, obtain the prediction data of t+2 moment low frequency Sequence；
S422, step S421 is repeated until obtaining s step prediction dataWherein s is positive integer,For the prediction data of t+s moment low frequency Sequence；
S43, according to the identical principle of step S42, obtain the multi-step prediction data of high frequency series part.
As a preferred technical solution, in step S5, the composite formula of the wavelet reconstruction are as follows:
WhereinFor the prediction data of low frequency Sequence,For the prediction data of high frequency series part, n
For the wavelet decomposition number of plies.
The present invention has the following advantages compared with the existing technology and effect:
Blood pressure prediction technique provided by the invention based on wavelet analysis and echo state network, can calculate to a nicety blood
Press variation tendency.Wavelet analysis technology is combined with echo state network method, wavelet analysis in processing non-stationary signal
The good time-frequency multiresolution that has and localization the advantages that and echo state network in processing nonlinear properties training algorithm
It is relatively simple, be not easy the advantages that falling into local optimum and combine, improve existing prediction model to the blood of nonlinear and nonstationary
Press the problems such as time series forecasting precision is not high, prediction step is too short.
The experimental results showed that the root-mean-square error (RMSE) of this method is about traditional base when multi-step prediction step-length is 15
In the 1/3 of the prediction technique of ESN.The present invention can be pre- well in the Single-step Prediction and multi-step prediction of blood pressure time series
Measuring blood pressure variation tendency, and the variation form of blood pressure is accurately sketched out, to the promptly and accurately pre- of blood pressure time series
Measuring tool has important clinical significance.
Detailed description of the invention
Fig. 1 is the method flow diagram of the invention based on the prediction of the blood pressure of wavelet analysis and echo state network；
Fig. 2 is the original blood pressure time series chart in the present embodiment；
Fig. 3 is the echo state network model in the present embodiment；
Fig. 4 be in the present embodiment the blood pressure Single-step Prediction result of the prediction technique of the method for the present invention and tradition ESN with really
Value compares figure；
Fig. 5 (a)~Fig. 5 (d) is the blood pressure multi-step prediction knot of the method for the present invention and the prediction technique of tradition ESN in embodiment
Fruit figure compared with true value；Fig. 5 (a), Fig. 5 (b), Fig. 5 (c), Fig. 5 (d) be respectively step-length be 3,5,9,15 when prediction data
Comparison diagram.
Specific embodiment
In order to which the purpose of the present invention, technical solution and advantage is more clearly understood, with reference to the accompanying drawings and embodiments,
The present invention is further described in detail.It should be appreciated that described herein, the specific embodiments are only for explaining the present invention,
It is not limited to the present invention.
Embodiment
As shown in Figure 1, a kind of blood pressure prediction technique based on wavelet analysis and echo state network, includes the following steps:
S1, small echo is carried out using the T/F local analysis method of wavelet transformation to original blood pressure time series data
It decomposes, is low frequency Sequence and high frequency series part by original blood pressure Time Series；
S11, the original blood pressure time series of acquisition, as shown in Fig. 2, original blood pressure time series totally 1283 sampled points, preceding
860 parts are used as training, and latter 423 parts are tested；
S12, wavelet decomposition is carried out to the original blood pressure time series, obtains the low frequency sequence of original blood pressure time series
Part and high frequency series part；
The continuous wavelet transform formula of wavelet decomposition process are as follows:
Signal u (t) ∈ L in formula2(R), L2(R) indicate be defined on real axis, measurable quadractically integrable function space；Indicate that the continuous wavelet generated by wavelet mother function Ψ (t), a indicate contraction-expansion factor, b table
Show shift factor；
Contraction-expansion factor a and the corresponding discrete wavelet function of shift factor b discretization are as follows:
Wherein, a=2-s, b=2-sK, s and k belong to set of integers Z；
With the Mallet fast algorithm of orthogonal wavelet transformation, by signal u (t) rectangular projection to SPACE VjAnd WjOn, it is corresponding
Obtain resolution ratio 2jApproximation signal (low frequency signal) a of lower signal u (t)j(t) and discrete signal (high-frequency signal) dj(t), it enables and dividing
Resolution 2jIncrease step by step, obtain the realization step by step of signal decomposition, every level of decomposition the result is that last time decomposition is obtained low
Frequency signal is further decomposed into low frequency and high-frequency signal, and the high-frequency signal is not considered；Signal u (t) is obtained after Multiresolution Decomposition
To the signal of low frequency Sequence and the signal of high frequency series part, indicate are as follows:N is
Positive integer indicates the wavelet decomposition number of plies；I is positive integer；T is the moment.
S2, the reserve pool parameter for determining echo state network model；
As shown in figure 3, echo state network model is a kind of completely new recurrent neural network, part connection weight is only trained,
It overcomes that the intrinsic training algorithm of conventional recursive neural network is excessively complicated, is easily trapped into the problems such as local optimum, is increasingly becoming
One of time series analysis and the main tool of prediction.
The final performance of echo state network model is determined by the parameters of reserve pool, comprising: inside reserve pool
Connection weight spectral radius SR, reserve pool scale N, reserve pool input unit scale IS, the sparse degree SD of reserve pool.
In the present embodiment, connection weight spectral radius SR inside reserve pool is the inside connection weight matrix W of reserve pool
The characteristic value of maximum absolute value, is denoted as λmax, λmax< 1 is the necessary condition for guaranteeing network stabilization；
λmax=max { abs (characteristic value of W) }
Only work as λmaxWhen < 1, echo state network just has echo status attribute.
Reserve pool scale N is the number of neuron in reserve pool, and the scale selection of reserve pool is related with number of samples,
Very big on network performance influence, reserve pool scale is bigger, and echo state network is more accurate to the description of given dynamical system, still
Overfitting problem can be brought.It is 7 that N is taken in embodiment.
Reserve pool input unit scale IS is to need before the input signal of reserve pool is connected to reserve pool intrinsic nerve member
The scale factor to be multiplied, i.e., carry out certain scaling to input signal.The object for generally requiring processing is non-linear stronger,
IS is bigger.It is 0.7 that IS is taken in embodiment.
The sparse degree SD of reserve pool indicates the connection in reserve pool between neuron, is not institute in reserve pool
Have between neuron and all there is connection, SD indicates that neuron population interconnected in reserve pool accounts for the percentage of total neuron N
Than value is bigger, and None-linear approximation ability is stronger.It is 10% that SD is taken in embodiment.Wherein n is to be connected with each other nerve
First number, N are total neuron number.
In the present embodiment, for specific echo state network model, the parameters of reserve pool are rule of thumb set.
S3, the instruction for carrying out echo state network model to the training sample of low frequency Sequence and high frequency series part respectively
Practice, obtains the state equation of reserve pool output prediction；
S31, the connection weight matrix for initializing echo state network model, including input connection weight matrix Win, deposit
The connection weight matrix W of the inside connection weight matrix W in pond and output layer to reserve poolbackWherein WinIt is randomly generated with W,
Set Wback=0；
S32, echo state network model is carried out to the training sample of low frequency Sequence and high frequency series part respectively
Training, specifically: it send low frequency Sequence or the training sample of high frequency series part as input neuron into initialization
In echo state network model, reserve pool intrinsic nerve member state renewal equation is as follows:
X (t+1)=f (Winuin(t+1)+Wx(t)+Wbacky(t))
F is intrinsic nerve member activation primitive in formula, uses hyperbolic tangent functionWhen t=0, deposit
Pond intrinsic nerve member state vector x (t)=0；The reserve pool intrinsic nerve member state vector x (t+1) at t+1 moment is by currently inputting
Neuron uin(t+1), the reserve pool intrinsic nerve member state vector x (t) and desired output y (t) of last moment is excited jointly
It generates；It is that row forms matrix X with the state vector x (t) at each moment, corresponding desired output y (t) constitutes a column vector
Y；
The state equation of reserve pool output prediction are as follows:
W in formulaoutTo export connection weight matrix, the output connection weight matrix is solved using least square method:
Wout=XTY=(XTX)-1XTY
Wherein, XTIt is the pseudoinverse of X.
S4, the echo state network model of low frequency Sequence obtained using step S3 training and high frequency series part
Echo state network model carries out Single-step Prediction and multistep to the test sample of low frequency Sequence and high frequency series part respectively
Prediction；
S41, the test sample of low frequency Sequence and high frequency series part is separately input to corresponding echo state network
Network model obtains the Single-step Prediction data of low frequency SequenceWith the Single-step Prediction data of high frequency series part
S42, the multi-step prediction data for obtaining low frequency Sequence, detailed process is as follows:
The Single-step Prediction data of S421, low frequency SequenceExpectation for t+1 moment low frequency Sequence is defeated
Out, which is returned into network inputs part, replaces the test sample of the low frequency Sequence at t+2 momentAnd network is inputted as input neuron, obtain the prediction of t+2 moment low frequency Sequence
Data；
S422, step S421 is repeated until obtaining s step prediction dataWherein s is positive integer,For the prediction data of t+s moment low frequency Sequence；
S43, according to the identical principle of step S42, obtain the multi-step prediction data of high frequency series part.
S5, by the single step of the Single-step Prediction data of low frequency Sequence and multi-step prediction prediction data and high frequency series part
Prediction data and multi-step prediction prediction data carry out wavelet reconstruction respectively, obtain Single-step Prediction blood pressure data and multi-step prediction blood pressure
Data；
The composite formula of the wavelet reconstruction are as follows:
WhereinFor the prediction data of low frequency Sequence,For the prediction data of high frequency series part, n
For the wavelet decomposition number of plies.
In the present embodiment, by the Single-step Prediction data of the Single-step Prediction data of low frequency Sequence and high frequency series part
Wavelet reconstruction is carried out, obtains Single-step Prediction blood pressure data, as shown in Figure 4；
The multi-step prediction data of low frequency Sequence and the multi-step prediction data of high frequency series part are subjected to wavelet reconstruction,
Multi-step prediction blood pressure data is obtained, as Fig. 5 (a)~Fig. 5 (d) show the blood pressure time series predicting model of the present embodiment and passes
The multi-step prediction blood pressure data comparison diagram that system is obtained based on the prediction model of echo state network (ESN), wherein Fig. 5 (a) is step
The comparison diagram of a length of 3 prediction blood pressure data；Fig. 5 (b) is the comparison diagram for the prediction blood pressure data that step-length is 5；Fig. 5 (c) is step
The comparison diagram of a length of 9 prediction blood pressure data；Fig. 5 (d) is the comparison diagram for the prediction blood pressure data that step-length is 15.
In this embodiment, it for the prediction effect of quantitative analysis blood pressure time series predicting model, needs to carry out error
Analysis generallys use average relative error (MSE), root-mean-square error (RMSE), standard root-mean-square error (NRMSE), is averaged absolutely
The indexs such as error (MAD), average absolute very error (MAPE) are evaluated.
It, will be using prediction technique of the tradition based on ESN and of the invention in embodiment in order to illustrate superiority of the invention
The results are shown in Table 1 for combination forecasting method acquisition.
As can be known from Table 1, in Single-step Prediction, prediction technique and this implementation of the tradition based on echo state network (ESN)
Combination forecasting method prediction effect of the example based on wavelet analysis and echo state network is good.But in multi-step prediction,
The combination forecasting method of the present embodiment to be substantially better than tradition be based on echo state network (ESN) prediction technique, and with
The increase of prediction step, advantage are more obvious.Such as prediction step, when being 15, the root-mean-square error (RMSE) of the present embodiment about passes
The 1/3 of the prediction technique based on ESN of uniting.
Table 1
The embodiments described above only express several embodiments of the present invention, and the description thereof is more specific and detailed, but simultaneously
Limitations on the scope of the patent of the present invention therefore cannot be interpreted as.It should be pointed out that for those of ordinary skill in the art
For, without departing from the inventive concept of the premise, various modifications and improvements can be made, these belong to guarantor of the invention
Protect range.Therefore, the scope of protection of the patent of the present invention should subject to the claims.
Claims (6)
1. a kind of blood pressure prediction technique based on wavelet analysis and echo state network, which is characterized in that include the following steps:
S1, wavelet decomposition is carried out using the T/F local analysis method of wavelet transformation to original blood pressure time series data,
It is low frequency Sequence and high frequency series part by original blood pressure Time Series；
S2, the reserve pool parameter for determining echo state network model；
S3, the training for carrying out echo state network model to the training sample of low frequency Sequence and high frequency series part respectively,
Obtain the state equation of reserve pool output prediction；
S4, the echo state network for the echo state network model and high frequency series part for training obtained low frequency Sequence is utilized
Network model carries out Single-step Prediction and multi-step prediction to the test sample of low frequency Sequence and high frequency series part respectively；
S5, by the Single-step Prediction of the Single-step Prediction data of low frequency Sequence and multi-step prediction prediction data and high frequency series part
Data and multi-step prediction prediction data carry out wavelet reconstruction respectively, obtain Single-step Prediction blood pressure data and multi-step prediction blood pressure number
According to.
2. the blood pressure prediction technique according to claim 1 based on wavelet analysis and echo state network, which is characterized in that
Step S1 specifically include the following steps:
S11, the original blood pressure time series of acquisition, and it is classified as training sample and test sample；
S12, wavelet decomposition is carried out to the original blood pressure time series, obtains the low frequency Sequence of original blood pressure time series
With high frequency series part；
The continuous wavelet transform formula of wavelet decomposition process are as follows:
Signal u (t) ∈ L in formula2(R), L2(R) indicate be defined on real axis, measurable quadractically integrable function space；Indicate that the continuous wavelet generated by wavelet mother function Ψ (t), a indicate contraction-expansion factor, b table
Show shift factor；
Contraction-expansion factor a and the corresponding discrete wavelet function of shift factor b discretization are as follows:
Wherein, a=2-s, b=2-sK, s and k belong to set of integers Z；
With the Mallet fast algorithm of orthogonal wavelet transformation, by signal u (t) rectangular projection to SPACE VjAnd WjOn, correspondence obtains
Resolution ratio 2jThe approximation signal a of lower signal u (t)j(t) and discrete signal dj(t), the approximation signal aj(t) low frequency to decompose
Signal, the discrete signal djIt (t) is the high-frequency signal decomposed, signal u (t) obtains high frequency series portion after Multiresolution Decomposition
The signal of the signal and low frequency Sequence that divide indicates are as follows:N is positive integer, is indicated
The wavelet decomposition number of plies；I is positive integer；T is the moment.
3. the blood pressure prediction technique according to claim 1 based on wavelet analysis and echo state network, which is characterized in that
In step S2, the performance of the echo state network model is determined by the parameter of reserve pool, and the parameter of the reserve pool includes: storage
Connection weight spectral radius SR, reserve pool scale N, reserve pool input unit scale IS and the sparse degree SD of reserve pool inside standby pond.
4. the blood pressure prediction technique according to claim 1 based on wavelet analysis and echo state network, which is characterized in that
Step S3 specifically include the following steps:
S31, the connection weight matrix for initializing echo state network model, including input connection weight matrix Win, reserve pool
The connection weight matrix W of internal connection weight matrix W and output layer to reserve poolback, wherein WinIt is randomly generated with W, if
Determine Wback=0；
S32, the training for carrying out echo state network model to the training sample of low frequency Sequence and high frequency series part respectively,
Specifically: the echo shape of initialization is sent into using low frequency Sequence or the training sample of high frequency series part as input neuron
In state network model, reserve pool intrinsic nerve member state renewal equation is as follows:
X (t+1)=f (Winuin(t+1)+Wx(t)+Wbacky(t))
F is intrinsic nerve member activation primitive in formula, specifically uses hyperbolic tangent function；When t=0, reserve pool intrinsic nerve member shape
State vector x (t)=0；The reserve pool intrinsic nerve member state vector x (t+1) at t+1 moment is by currently inputting neuron uin(t+1)、
Excitation generates the reserve pool intrinsic nerve member state vector x (t) and desired output y (t) of last moment jointly；With each moment
State vector x (t) be that row forms matrix X, corresponding desired output y (t) constitutes a column vector Y；
The state equation of reserve pool output prediction are as follows:
W in formulaoutTo export connection weight matrix, the output connection weight matrix is solved using least square method:
Wout=XTY=(XTX)-1XTY
Wherein, XTIt is the pseudoinverse of X.
5. the blood pressure prediction technique according to claim 1 based on wavelet analysis and echo state network, which is characterized in that
In step S4, Single-step Prediction and multi-step prediction are carried out to the test sample of low frequency Sequence and high frequency series part respectively, had
Body includes the following steps:
S41, the test sample of low frequency Sequence and high frequency series part is separately input to low frequency sequence echo state network
Model and high frequency series echo state network model, obtain the Single-step Prediction data of low frequency SequenceAnd high frequency
The Single-step Prediction data of Sequence
S42, the multi-step prediction data for obtaining low frequency Sequence, detailed process is as follows:
The Single-step Prediction data of S421, low frequency SequenceFor the desired output of t+1 moment low frequency Sequence,
The desired output is returned into network inputs part, replaces the test sample of the low frequency Sequence at t+2 moment and as input
Neuron inputs network, obtains the prediction data of t+2 moment low frequency Sequence；
S422, step S421 is repeated until obtaining s step prediction dataWherein s is positive integer,For
The prediction data of t+s moment low frequency Sequence；
S43, according to the identical principle of step S42, obtain the multi-step prediction data of high frequency series part.
6. the blood pressure prediction technique according to claim 1 based on wavelet analysis and echo state network, which is characterized in that
In step S5, the composite formula of the wavelet reconstruction are as follows:
WhereinFor the prediction data of low frequency Sequence,For the prediction data of high frequency series part, n is small echo
Decomposition order.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810749008.6A CN109009033A (en) | 2018-07-10 | 2018-07-10 | A kind of blood pressure prediction technique based on wavelet analysis and echo state network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810749008.6A CN109009033A (en) | 2018-07-10 | 2018-07-10 | A kind of blood pressure prediction technique based on wavelet analysis and echo state network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109009033A true CN109009033A (en) | 2018-12-18 |
Family
ID=64640791
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810749008.6A Pending CN109009033A (en) | 2018-07-10 | 2018-07-10 | A kind of blood pressure prediction technique based on wavelet analysis and echo state network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109009033A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109640351A (en) * | 2019-01-25 | 2019-04-16 | 南京邮电大学 | A kind of unified prediction of base station flow |
| CN111554398A (en) * | 2020-05-11 | 2020-08-18 | 济南浪潮高新科技投资发展有限公司 | Remote vital sign evaluation method and system based on 5G |
| WO2021164349A1 (en) * | 2020-02-21 | 2021-08-26 | 乐普（北京）医疗器械股份有限公司 | Blood pressure prediction method and apparatus based on photoplethysmography signal |
| WO2021205581A1 (en) * | 2020-04-08 | 2021-10-14 | 富士通株式会社 | Information processing system, information processing device, information processing method, and information processing program |
-
2018
- 2018-07-10 CN CN201810749008.6A patent/CN109009033A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109640351A (en) * | 2019-01-25 | 2019-04-16 | 南京邮电大学 | A kind of unified prediction of base station flow |
| WO2021164349A1 (en) * | 2020-02-21 | 2021-08-26 | 乐普（北京）医疗器械股份有限公司 | Blood pressure prediction method and apparatus based on photoplethysmography signal |
| WO2021205581A1 (en) * | 2020-04-08 | 2021-10-14 | 富士通株式会社 | Information processing system, information processing device, information processing method, and information processing program |
| JP7376832B2 (en) | 2020-04-08 | 2023-11-09 | 富士通株式会社 | Information processing system, information processing device, information processing method, and information processing program |
| CN111554398A (en) * | 2020-05-11 | 2020-08-18 | 济南浪潮高新科技投资发展有限公司 | Remote vital sign evaluation method and system based on 5G |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109009033A (en) | A kind of blood pressure prediction technique based on wavelet analysis and echo state network | |
| Tao et al. | Unsupervised cross-domain rolling bearing fault diagnosis based on time-frequency information fusion | |
| CN111772619B (en) | Heart beat identification method based on deep learning, terminal equipment and storage medium | |
| Omi et al. | Hawkes process model with a time-dependent background rate and its application to high-frequency financial data | |
| CN107550499B (en) | A kind of method and system detecting human-body fatigue degree | |
| CN105913066B (en) | A kind of digitlization lungs sound feature dimension reduction method based on Method Using Relevance Vector Machine | |
| CN109480835A (en) | A kind of mental fatigue detection method based on shot and long term Memory Neural Networks | |
| CN110313894A (en) | Arrhythmia cordis sorting algorithm based on convolutional neural networks | |
| CN109620210A (en) | A kind of electrocardiosignal classification method of the CNN based on from coding mode in conjunction with GRU | |
| Cresswell et al. | CaloMan: Fast generation of calorimeter showers with density estimation on learned manifolds | |
| CN108647807A (en) | The prediction technique of river discharge | |
| Ozturk et al. | Comparison of neuro-fuzzy systems for classification of transcranial Doppler signals with their chaotic invariant measures | |
| Resta | Hurst exponent and its applications in time-series analysis | |
| CN113486925A (en) | Model training method, fundus image generation method, model evaluation method and device | |
| Nie et al. | Skin Cancer Classification based on Cosine Cyclical Learning Rate with Deep Learning | |
| Paetz | Metric rule generation with septic shock patient data | |
| CN108510076A (en) | A kind of algorithm of metric sequence confusion degree | |
| Yan et al. | Analyzing Wrist Pulse Signals Measured with Polyvinylidene Fluoride Film for Hypertension Identification. | |
| CN112686091B (en) | Two-step arrhythmia classification method based on deep neural network | |
| Kaur et al. | Hybrid of fuzzy logic and random walker method for medical image segmentation | |
| George et al. | Cancer tissues recognition system using box counting method and artificial neural network | |
| Shi et al. | BreastNet: entropy-regularized transferable multi-task learning for classification with limited breast data | |
| CN117582185B (en) | Pulse force level prediction method based on CLLSR mixed model | |
| Demidova et al. | Application of the Phase Analysis of Time Series for the Identification of Macroeconomic Cycles Based on the Dynamics of the Exchange Rates | |
| Guo et al. | Detection and analysis of wheat storage year based on electronic tongue and DWT-IPSO-LSSVM algorithm |
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 |
Application publication date: 20181218 |
|
| RJ01 | Rejection of invention patent application after publication |