CN108113665B - Automatic noise reduction method for electrocardiosignal - Google Patents

Abstract

The invention discloses an automatic noise reduction algorithm for electrocardiosignals, which comprises the following processing procedures: A) establishing an echo state network and initializing; B) acquiring electrocardiosignals of a human body and constructing a training set on the basis of the electrocardiosignals; C) training an echo state network by using an online training algorithm based on a recursive least square method to obtain a trained echo state network; D) and constructing a test set, and inputting the test set into the trained echo state network to obtain a clean electrocardiosignal. By the processing of the method, the denoised clean electrocardiosignal not only effectively filters noise, but also recovers the low-frequency characteristic wave of the electrocardiosignal and retains the effective information of the electrocardiosignal.

Description

Automatic noise reduction method for electrocardiosignal

Technical Field

The invention relates to an automatic electrocardiosignal detection and analysis method, in particular to an automatic electrocardiosignal noise reduction method.

Background

Cardiovascular diseases become the healthy first killer of people, and have high morbidity, high disability rate and high mortality. The low-frequency component of the electrocardiosignal is very important for the intelligent diagnosis of cardiovascular diseases, for example, the ST section which is one of the low-frequency components represents the potential change condition in a period of time after the ventricular depolarization is finished and after the ventricular repolarization is started, and the ST section is an important index for diagnosing myocardial ischemia. At the same time, however, the electrocardiographic signal itself has the commonalities of bioelectric signals: the electrocardiosignal acquisition device has the characteristics of weak amplitude, low frequency, large impedance, randomness and the like, and the electrocardiosignal is very easy to be interfered by various noises in the acquisition process. Particularly, in the current remote medical background, the noise during the acquisition of the electrocardiosignals presents more complex characteristics, so that the low-frequency part of the electrocardiosignals is easily submerged by the complex noise, and the phenomenon of effective information loss occurs.

There are three methods commonly used today: first, Empirical Mode Decomposition (EMD) is to extract the Intrinsic Mode Function (IMF) to reduce the noise of a noisy ECG signal, resulting in the removal of 95% white gaussian noise. However, EMD does not separate out signals that are similar in frequency to ECG signals. Therefore, this noise reduction method may filter out P-waves and T-waves from the signal, resulting in misdiagnosis. Second, Principal Component Analysis (PCA) and Independent Component Analysis (ICA) are blind source separation techniques used to find independent source signals from non-gaussian noise signals, with significant advantages for in-band filtering. However, the model is very sensitive to slight changes of the noisy ECG signal, and is not suitable for the case where the ECG signal contains more complicated noise. Thirdly, the wavelet denoising method achieves good results in terms of removing high-frequency noise, but the selection of wavelet basis and the setting of threshold value need to be determined through a large amount of experiments, and has direct influence on the final filtering effect. Therefore, the existing algorithm is easy to lose effective low-frequency information of the electrocardiosignals in the noise reduction process, and the judgment of a doctor is influenced.

Disclosure of Invention

The invention aims to provide an automatic noise reduction algorithm for an electric signal, which is used for realizing noise reduction of the electrocardiosignal and simultaneously reserving low-frequency characteristic waves and solving the problem that the effective low-frequency information of the electrocardiosignal is easily lost in the noise reduction process of the existing algorithm.

The purpose of the invention is realized as follows:

an automatic noise reduction algorithm for electrocardiosignals comprises the following steps:

A) establishing an echo state network and initializing;

B) acquiring electrocardiosignals of a human body and constructing a training set on the basis of the electrocardiosignals;

C) training an echo state network by using an online training algorithm based on a recursive least square method, and obtaining and storing the trained echo state network;

D) and constructing a test set by using the electrocardiosignals next to the training set, and inputting the test set into the trained echo state network to obtain clean electrocardiosignals.

The automatic noise reduction algorithm for the electrocardiosignals comprises the following steps of A) initializing a weight matrix and initializing and setting network parameters, wherein:

the process of initializing the weight matrix is as follows:

weight matrix W connecting input unit with the interior of reserve tank_inRandomly initializing a weight matrix W connected with neurons in the reserve pool to ensure that the weights are uniformly distributed;

connecting weight matrix W of output unit and reserve pool_backThe reserve pool is connected with the output unit to form a weight matrix W_outInitializing to a zero matrix;

the initialization setting process of the network parameters comprises the following steps:

the number of input neurons is set to be 4, the number of output neurons is set to be 1, and the number of neurons in a reserve pool is set to be N which is 1000 +/-300;

the sparsity degree SD of the reserve pool is m/N, wherein m is the number of the neurons which are connected with each other in the reserve pool, and the value range of SD is 1-5%;

debugging the value of the spectrum radius SR of the reserve pool according to the step length of 0.1 within the range of 0-1, and taking the SR value corresponding to the optimal result as a final value, wherein the SR is max { abs (characteristic value of W) };

the value of the input unit dimension IS of the reserve pool IS debugged in a range of-1 to 1 by taking 0.01 as a step length, and the corresponding IS value when the result IS optimal IS taken as a final value.

The automatic noise reduction algorithm for the electrocardiosignal comprises the following specific steps in the step B):

firstly, a plurality of continuous heart beats x (n) are intercepted from the acquired human body electrocardiosignal data, wherein n is 1,2, …, M is 55 +/-2, and then the noise-containing electrocardiosignal x (n) at the current moment, the noise-containing electrocardiosignal x (n-1) at the previous moment, the first derivative x '(n) of the noise-containing electrocardiosignal at the current moment and the second derivative x' (n) of the noise-containing electrocardiosignal at the current moment form an input signal x (n) of an echo state network_r(n) is represented by

Defining clean heart beat data corresponding to the electrocardiosignals x (n) containing noise at the current moment as d (n);

from x_r(n) and d (n) form a training set (x)_r(n),d(n),n＝1,2,…,M)；

Definition of-

Wherein g (-) is an activation function of the output neuron, and is selected as an identity function, i.e. the output neuron is a linear neuron, and the matrix v (n) in the function is defined as

Wherein r (n) is a reserve pool state variable;

defining the error e (n) ═ d (n) -y (n).

The automatic noise reduction algorithm for the electrocardiosignal comprises the following specific steps in the step C):

selecting a network, setting the initial value of a reserve pool state variable to be 0, namely, setting r (0) to be 0, and training set (x)_rSample x in (n), d (n), n ═ 1,2, …, M)_r(n) passing through W_inD (n) through W_backAdding into a storage pool respectively, and using the formula r (n) ═ (1-alpha) [ f (W. r (n-1) + W_in·u(n)+W_back·d(n))]+ alpha r (n-1) updating the reserve pool state variable r (n);

wherein f (-) is a reservoir neuron activation function selected from a non-linear tanh (-) function;

u (n) ═ ψ (n-1) · v (n), ψ (n) is a correlation coefficient matrix of v (n) and is formed by the following formula

To perform the update, wherein,

lambda is a forgetting factor, the value of the lambda is adjusted in the range of 0.95-1 by taking 0.0001 as a step length, the lambda value is the final value when the result is optimal, psi (0) ═ I, I is an identity matrix, the lambda value is a very small positive number, and the lambda value is in the range of 0-0.001Debugging by taking 0.00001 as a step length, and taking the result optimal value as a final value;

alpha is forgetting rate, the value of the alpha is debugged in a step length of 0.1 within the range of 0-1, and the alpha value with the optimal result is taken as the final value;

② according to the formula W_out(n)＝W_out(n-1) + k (n) e (n) to output the weight W_outUpdating of (1);

③ by updating W_out(n) and r (n) approximating y (n) to a clean ECG signal d (n) and allowing the error e (n) to approximate 0, resulting in a trained echo state network.

The automatic noise reduction algorithm for the electrocardiosignal comprises the following specific steps in the step D):

constructing a test set (X) by using the T heart beats next to the M heart beats intercepted during the construction of the training set_r(n), n ═ 1,2, …, T ═ 200 ± 10), where,

namely, the electrocardiosignal X (n) at the current moment, the last electrocardiosignal X (n-1), the first derivative X '(n) and the second derivative X' (n) thereof are included;

test set (X)_rX in (n), n-1, 2, …, T)_r(n) inputting the echo state network after training, and outputting a clean electrocardiosignal Y (n) by the network.

By the processing of the method, the denoised clean electrocardiosignal not only effectively filters noise, but also recovers the low-frequency characteristic wave of the electrocardiosignal and retains the effective information of the electrocardiosignal.

Drawings

FIG. 1 is a flow chart of the process of carrying out the method of the present invention.

Fig. 2 is a schematic diagram of an echo state network structure.

Fig. 3 is a schematic diagram of the waveform structure of human electrocardiosignals.

Fig. 4 is a graph of a noisy cardiac signal prior to filtering.

FIG. 5 is a graph of a filtered clean cardiac signal.

Detailed Description

The embodiment is realized in a computer with an Intel Xeon CPU E5-2697@2.70GHz and a 128.00GB Win7 internal memory 64-bit operating system, and the whole electrocardiosignal automatic classification algorithm is realized by adopting Matlab language.

With reference to fig. 1, the implementation of the present invention is as follows:

A) establishing an Echo State Network (ESN):

firstly, initializing a weight matrix:

weight matrix W for random initialization input unit and internal connection of reserve tank_inConnecting the weight matrix W with neurons in the reserve pool, and enabling the weight matrix W and the neurons to be uniformly distributed;

connection weight matrix W of output unit and reserve pool_backThe reserve pool is connected with the output unit to form a weight matrix W_outInitializing to a zero matrix;

secondly, setting parameters:

the number of input neurons of ESN is set to 4, the number of output neurons is set to 1, and the reserve pool size N, which is the number of neurons in the reserve pool, is 1000;

the reservoir sparsity degree SD of the ESN is calculated according to the formula SD ═ m/N, where m is the number of neurons connected to each other in the reservoir, and in this example, m is taken as m ═ 20, and SD is 2%;

the spectrum radius SR of the reserve pool is defined as SR ═ max { abs (characteristic value of W) }, because ESN has the property of echo state when SR is less than 1, in this example, the SR value is debugged by taking 0.1 as the step length according to the range of 0-1, and the SR ═ 0.4 is taken when the result is optimal;

definition of W_SFor a randomly generated sparse matrix with uniformly distributed sparsity degree SD, the neuron connection weight matrix W in the reserve pool is SR.W_SEnsuring the spectrum radius of W as SR;

the input unit scale IS of the reserve pool IS a scale factor which needs to be multiplied before an input signal IS connected to the reserve pool, the input signal IS multiplied by the input factor, so that the input signal IS transformed into an input range of a neuron activation function of the reserve pool, the input range of the neuron activation function IS [ -1, 1], in the embodiment, the value of IS IS debugged in a range of-1 to 1 by taking 0.01 as a step length, and the IS IS [0.01 ] when the result IS optimal; 0.01; 0.9; 0.8 ];

B) collecting human body electrocardio original signals at a collection frequency of 250Hz, wherein the waveform structure is shown in figure 3 and is stored in a data form of a TXT document, reading the electrocardio original signal data stored in the TXT document into a computer by Matlab software to be used as original electrocardiosignals, and constructing a training set as shown in figure 4:

definition of x_r(n) is an input signal of ESN, which is an input signal obtained by extracting, from an original electrocardiographic signal, a continuous sequence of M ═ 55 heart beats x (n), where n ═ 1,2, …,55, and is composed of a noisy electrocardiographic signal x (n) at the current time, a noisy electrocardiographic signal x (n-1) at the previous time, a first derivative x' (n) of the noisy electrocardiographic signal at the current time, and a second derivative x ″ (n) of the noisy electrocardiographic signal at the current time, and is expressed as

Is an Nx 4 matrix; defining clean heart beat data corresponding to the current heart beat as d (n); from x_r(n) and d (n) form a training set (x)_r(n),d(n),n＝1,2,…,55)。

Definition of

Wherein g (-) is an activation function of the output neuron, and is selected as an identity function, i.e. the output neuron is a linear neuron, and the matrix v (n) in the function is defined as

Wherein r (n) IS the state variable of the reserve pool, and diag (IS) represents that a diagonal matrix IS generated by taking IS as a diagonal;

defining an error e (n) ═ d (n) -y (n);

C) training an echo state network by using an online training algorithm based on a recursive least square method:

the initial value of the state variable of the reserve pool of the selected network is 0, namely r (0) is 0, and the training set (x)_rSample x in (n), d (n), n ═ 1,2, …,55)_r(n) passing through W_inD (n) through W_backRespectively added into a reserve pool and utilizes a maleFormula r (n) ((1- α) [ f (W · r (n-1) + W-_in·u(n)+W_back·d(n))]+ α · r (n-1) updates the pool status r (n):

wherein f (-) is a reserve pool neuron activation function, and is selected as a nonlinear tanh (-) function due to the fact that strong nonlinearity is needed when the network is modeled;

u (n) ═ ψ in the function^-1(n-1). v (n), ψ (n) is the inverse of the correlation coefficient matrix of v (n) and is updated by the following expression:

wherein,

λ is forgetting factor, defined as a constant with a range of 0.95 < λ < 1, and its value is adjusted by using 0.0001 as step length, and when the result is optimum, the value of λ is taken as final value, in this example, λ is 0.9999, and ψ (0) is equal to ψ (0)^-1I, I is an identity matrix, is a very small positive number, has a value range of 0 to 0.001, is debugged with 0.00001 as a step length, and has a value of final value, which is 0.00172, when the result is optimal (that is, when input signals in a training set are put into a network again, and the obtained error between y (n) and d (n) is close to 0);

in the function, α is a forgetting rate, defined as a positive number smaller than 1, and has a value range of 0 to 1, and the function is debugged with a step size of 0.1, and the value α when the result is optimal is taken as a final value, in this example, the value α is 0.8.

② according to the formula W_out(n)＝W_out(n-1) + k (n) e (n) to output the weight W_outAnd (4) updating.

③ by updating W_out(n) and r (n) approximating y (n) to a clean ECG signal d (n) and allowing the error e (n) to approximate 0, resulting in and preserving a well-trained echo state network.

When y (n) is close to d (n), namely the error e (n) is close to 0, the network is in a stable state, and the trained state-returning network can be stored and used for denoising the electrocardiosignals.

D) And (3) verification:

constructing a test set by using the immediately adjacent T (200) heartbeats after the 55 heartbeats intercepted in the process of constructing the training set, wherein the test set is a test set (X)_r(n), n is 1,2, …,200), wherein,

namely, the electrocardiosignal X (n) at the current moment, the last electrocardiosignal X (n-1), the first derivative X '(n) and the second derivative X' (n) thereof are included;

test set (X)_rX in (n), n is 1,2, …,200)_r(n) inputting the trained echo state network, and outputting a clean electrocardiosignal Y (n) by the network, as shown in FIG. 5.

From the above processing results, it can be seen that the low-frequency characteristic waves of the electrocardiosignals T and P are seriously distorted under the interference of noise when the electrocardiosignals containing the noise (as shown in FIG. 4) are compared with the waveforms of the normal electrocardiosignals (as shown in FIG. 2), and the clean electrocardiosignals (as shown in FIG. 5) after being denoised not only effectively filter the noise, but also recover the low-frequency characteristic waves of the electrocardiosignals and retain the effective information of the electrocardiosignals after being processed by the method of the present invention.

Claims (1)

1. An automatic noise reduction algorithm for electrocardiosignals is characterized by comprising the following steps:

A) establishing an echo state network and initializing;

B) acquiring electrocardiosignals of a human body and constructing a training set on the basis of the electrocardiosignals;

C) training an echo state network by using an online training algorithm based on a recursive least square method, and obtaining and storing the trained echo state network;

D) constructing a test set by using the electrocardiosignals next to the training set, and inputting the test set into the trained echo state network to obtain clean electrocardiosignals;

the step A) comprises initializing a weight matrix and initializing and setting network parameters, wherein:

the process of initializing the weight matrix is as follows:

weight matrix W connecting input unit with the interior of reserve tank_inConnecting weight moment with neuron in reserve poolRandomly initializing the array W to ensure that the array W is uniformly distributed;

connecting weight matrix W of output unit and reserve pool_backThe reserve pool is connected with the output unit to form a weight matrix W_outInitializing to a zero matrix;

the initialization setting process of the network parameters comprises the following steps:

the number of input neurons is set to be 4, the number of output neurons is set to be 1, and the number of neurons in a reserve pool is set to be N which is 1000 +/-300;

the sparsity degree SD of the reserve pool is m/N, wherein m is the number of the neurons which are connected with each other in the reserve pool, and the value range of SD is 1-5%;

debugging the value of the spectrum radius SR of the reserve pool according to the step length of 0.1 within the range of 0-1, and taking the SR value corresponding to the optimal result as a final value, wherein the SR is max { abs (characteristic value of W) };

the value of the input unit dimension IS of the reserve pool IS debugged in a range of-1 to 1 by taking 0.01 as a step length, and the corresponding IS value when the result IS optimal IS taken as a final value;

the specific process of the step B) is as follows:

firstly, a plurality of continuous heart beats x (n) are intercepted from the acquired human body electrocardiosignal data, wherein n is 1,2, …, M is 55 +/-2, and then the noise-containing electrocardiosignal x (n) at the current moment, the noise-containing electrocardiosignal x (n-1) at the previous moment, the first derivative x '(n) of the noise-containing electrocardiosignal at the current moment and the second derivative x' (n) of the noise-containing electrocardiosignal at the current moment form an input signal x (n) of an echo state network_r(n) is represented by

Defining clean heart beat data corresponding to the electrocardiosignals x (n) containing noise at the current moment as d (n);

from x_r(n) and d (n) form a training set (x)_r(n),d(n),n＝1,2,…,M)；

Definition of-

Wherein g (-) is an activation function of the output neuron, and is selected as an identity function, i.e. the output neuron is a linear neuron, and the matrix v (n) in the function is defined as

Wherein r (n) IS a reserve pool state variable, and diag (IS) represents that a diagonal matrix IS generated by taking IS as a diagonal;

(iii) defining an error e (n) ═ d (n) -y (n);

the specific process of the step C) is as follows:

selecting a network, setting the initial value of a reserve pool state variable to be 0, namely, setting r (0) to be 0, and training set (x)_rSample x in (n), d (n), n ═ 1,2, …, M)_r(n) passing through W_inD (n) through W_backAdding into a storage pool respectively, and using the formula r (n) ═ (1-alpha) [ f (W. r (n-1) + W_in·u(n)+W_back·d(n))]+ alpha r (n-1) updating the reserve pool state variable r (n);

wherein f (-) is a reservoir neuron activation function selected from a non-linear tanh (-) function;

u (n) ═ ψ (n-1) · v (n), ψ (n) is a correlation coefficient matrix of v (n) and is formed by the following formula

To perform the update, wherein,

lambda is forgetting factor, the value of which is adjusted in the range of 0.95-1 by taking 0.0001 as step length, the lambda value is the final value when the result is optimal, psi (0) ═ I, I is an identity matrix, which is a tiny positive number, and the positive number is takenDebugging the value within the range of 0-0.001 by taking 0.00001 as a step length, and taking the optimal value of the result as a final value;

alpha is forgetting rate, the value of the alpha is debugged in a step length of 0.1 within the range of 0-1, and the alpha value with the optimal result is taken as the final value;

② according to the formula W_out(n)＝W_out(n-1) + k (n) e (n) to output the weight W_outUpdating of (1);

③ by updating W_out(n) and r (n) approximating y (n) to a clean ECG signal d (n) and allowing the error e (n) to approximate 0, resulting in a trained echo state network;

the specific process of the step D) is as follows:

constructing a test set (X) by using the T heart beats next to the M heart beats intercepted during the construction of the training set_r(n), n ═ 1,2, …, T ═ 200 ± 10), where,

namely, the electrocardiosignal X (n) at the current moment, the last electrocardiosignal X (n-1), the first derivative X '(n) and the second derivative X' (n) thereof are included;

test set (X)_rX in (n), n-1, 2, …, T)_r(n) inputting the echo state network after training, and outputting a clean electrocardiosignal Y (n) by the network.

Images