CN105260990B - Contaminate the denoising method for infrared spectroscopy signals of making an uproar - Google Patents

Abstract

The present invention relates to the technical field of denoising methods for infrared spectral signals, and relates to a denoising method for noise-stained infrared spectral signals, which is carried out according to the following steps: the first step is to conduct empirical After modal decomposition, j eigenmode components are obtained, and the weighted average eigenmode components and residuals are obtained after j eigenmode components are weighted and averaged; the second step is to perform signal reconstruction on the noise-stained infrared spectrum signal After that, the reconstructed signal is obtained. The denoising method of the noise-stained infrared spectrum signal described in the present invention has a better denoising effect on the noise-stained infrared spectrum signal, and the infrared spectrum signal after denoising according to the denoising method of the noise-stained infrared spectrum signal of the present invention The infrared spectrum signal that is closer to the unstained noise can provide a better basis for the diagnosis and treatment of breast cancer and other diseases.

Description

Denoising Method of Noise-stained Infrared Spectral Signal

technical field

The invention relates to the technical field of denoising methods for infrared spectrum signals, and relates to a denoising method for noise-stained infrared spectrum signals.

Background technique

Spectral detection technology is a new technology for diagnosing breast cancer. Infrared spectroscopy is widely used as a fast, simple and non-destructive qualitative and quantitative analysis method. The measured infrared spectrum often contains a lot of interference information (noise signal, etc.), and the noise mainly comes from three aspects: detector noise, electronic circuit noise and environmental noise. Therefore, noise reduction is extremely important in spectral analysis and processing.

At present, the commonly used noise reduction methods mainly include wavelet, empirical mode decomposition (EMD) and EEMD, etc. Among them, the noise reduction method based on wavelet transform needs prior knowledge, that is, the selection of wavelet base, wavelet layer number and threshold and so on. EMD is a new adaptive signal processing method, which is suitable for nonlinear and non-stationary signals. However, an important defect of the EMD method is modal aliasing, which distorts the signal after noise reduction. In addition, the LMS adaptive filter is a learning algorithm based on error correction learning rules. Because its algorithm is simple and does not require prior knowledge, it has been widely used and has become a standard algorithm for adaptive filtering. However, there is an inevitable contradiction between the steady-state error and the convergence rate of the LMS algorithm, and the noise reduction effect is sometimes good or bad. Wu(Wu Z H, Huang NE.Ensemble empirical mode decomposition: a noise assisted data analysis method[J].Advances in Adaptive DataAnalysis) proposed EEMD on the basis of the research on the mode mixing phenomenon encountered in EMD decomposition Methods. After a non-stationary signal is decomposed by EEMD, several stationary intrinsic mode functions (IMFs) can be obtained. The IMF obtained by this method effectively overcomes the problem of mode aliasing in EMD decomposition, but encounters low signal-to-noise ratio signals , under the influence of abnormal events, the high-frequency eigenmode functions of the noise-stained signal decomposed by EEMD appear white noise pollution to varying degrees.

Contents of the invention

The present invention provides a kind of noise-removing method of dyed-noise infrared spectrum signal, has overcome above-mentioned deficiencies in the prior art, and the de-noise method of dye-and-noise infrared spectrum signal of the present invention adopts EEMD combined with VS-LMS (variable step) for the first time Long least mean square adaptive filter) denoises the noise-stained infrared spectrum signal, and the denoising method of the noise-stained infrared spectrum signal of the present invention has better performance than the existing infrared spectrum noise reduction method The noise reduction effect can provide a better basis for the diagnosis and treatment of breast cancer and other diseases.

The technical solution of the present invention is realized by the following measures: a denoising method of a noise-stained infrared spectrum signal is carried out according to the following steps: the first step is to empirically model the noise-stained infrared spectrum signal after adding Gaussian white noise After state decomposition, j eigenmode components are obtained, j is a natural number, and the weighted average eigenmode components and residuals are obtained after j eigenmode components are weighted and averaged; in the second step, the noise-stained infrared spectrum signal is After the signal is reconstructed, the reconstructed signal is obtained; the third step is to input the reconstructed signal into the variable step-size minimum mean square adaptive filter, and initialize the weight and set the step size in the variable step-size minimum mean square adaptive filter. The initial value of the long factor, the set order and the set number of runs, the variable step size minimum mean square adaptive filter updates the step size factor during the running process, and when the curve of the mean square error is a convergent curve, it can be The variable step size minimum mean square adaptive filter output signal is a noise reduction signal; specifically,

In the first step, several Gaussian white _noises with a mean value of 0 and a constant standard deviation are added to the noise-dyed infrared spectrum signal. (n) represents, adding the Gaussian white noise dyed noise infrared spectrum signal is represented by x _i (n), x _i (n)=x(n)+n _i (n), will add the Gaussian white noise dyed noise infrared spectrum After the signal is subjected to empirical mode decomposition, j eigenmode components are obtained, and the eigenmode components are denoted by c _ij (n), and the j eigenmode components are divided into The weighted average operation eliminates the noisy eigenmode components. After the weighted average operation, the weighted average eigenmode components and residuals are obtained. The weighted average eigenmode components are denoted by c _j and the residuals are denoted by r(n) , M is a natural number,

In the second step, the reconstructed signal is obtained after signal reconstruction of the noise-stained infrared spectrum signal, and the reconstructed signal is represented by x(n), x(n)=∑ _j c _j (n)+r(n), c _j (n) is the jth eigenmode component obtained after performing empirical mode decomposition on the noise-stained infrared spectrum signal after adding Gaussian white noise;

In the third step, the sequence of inputting the reconstructed signal into the variable-step minimum mean square adaptive filter is denoted by X(n),

X(n)=[x(n),x(n-1),...,x(n-M+1)],

The weight vector of the variable step least mean square adaptive filter is represented by W(n),

W(n)=[W _n1 ,W _n2 ,W _n3 ,...,W _nM ] ^T ,

The output signal of the variable step least mean square adaptive filter is represented by y(n),

The error of y(n) relative to the expected signal d(n) is represented by e(n),

e(n)=d(n)-W(n) ^T X(n),

The relationship between the error and the step factor satisfies:

W(n+1)=W(n)+2mue(n)X(n),

mu is the step size factor, mu is a constant, used to control the convergence speed, e(n) is the error of the signal, in order to ensure convergence after iteration, mu must satisfy: 0<mu<λ _max , λ _max is the input sequence X(n ) the maximum eigenvalue of the autocorrelation matrix R _xx , the iterative formula of the step size factor is:

mu ₀ is the initial value of the step factor, all n are natural numbers, and in the iterative formula Among them, n is the number of iterations of the signal point, the number of iterations is a natural number, mu is the nth step factor, and the step factor is updated according to the iterative formula of the step factor. During the update process, when the mean square error When the curve of is the convergence curve, the output signal of the variable step least mean square adaptive filter is the noise reduction signal.

The denoising method of the noise-stained infrared spectrum signal described in the present invention has a better denoising effect on the noise-stained infrared spectrum signal, and the infrared spectrum signal after denoising according to the denoising method of the noise-stained infrared spectrum signal of the present invention The infrared spectrum signal that is closer to the unstained noise can provide a better basis for the diagnosis and treatment of breast cancer and other diseases.

Description of drawings

Accompanying drawing 1 is adopting EQUINOX55 type Fourier transform infrared spectrometer (Germany Bruker company) to collect the mid-infrared spectrogram (unstained noise infrared spectrogram) of pure breast cancer section.

Accompanying drawing 2 is the infrared spectrogram of breast cancer slice staining noise.

Accompanying drawing 3 is the learning curve of the noise-stained infrared spectrum signal of breast cancer slices by the noise-stained infrared spectral signal denoising method according to the present invention.

Accompanying drawing 4 is the infrared spectrogram after processing the infrared spectrum signal of the breast cancer slice stained and noised according to the denoising method of the infrared spectrum signal of the stained noise.

Accompanying drawing 5 is the infrared spectrogram after processing the infrared spectrum signal of the breast cancer slice stained with wavelet noise.

Accompanying drawing 6 is the infrared spectrogram after processing the infrared spectrum signal of EEMD decomposition to the noise of breast cancer slices.

Accompanying drawing 7 is the infrared spectrogram after the LMS adaptive filter processes the infrared spectral signal of the breast cancer slice stained with noise.

Detailed ways

The present invention is not limited by the following examples, and specific implementation methods can be determined according to the technical solutions of the present invention and actual conditions.

The present invention will be further described below in conjunction with embodiment:

Embodiment 1: the denoising method of this noise-stained infrared spectrum signal is carried out according to the following steps: the first step, after adding Gaussian white noise, the noise-stained infrared spectrum signal is subjected to empirical mode (EMD) decomposition to obtain j basic eigenmode component (IMF), j is a natural number, and the weighted average eigenmode component and residual are obtained after j eigenmode components are weighted and averaged; in the second step, after signal reconstruction of the noise-stained infrared spectrum signal The reconstructed signal is obtained; the third step is to input the reconstructed signal into the variable step-size least mean square adaptive filter (VS-LMS), and initialize the weights and set The initial value of the step size factor, the set order and the set number of operations, the variable step size least mean square adaptive filter is in the process of running, update the step size factor, when the mean square error (e ² (n)) When the curve of is the convergence curve, the output signal of the variable step least mean square adaptive filter is the noise reduction signal; specifically,

In the first step, several Gaussian white _noises with a mean value of 0 and a constant standard deviation are added to the noise-dyed infrared spectrum signal. (n) represents, adding the Gaussian white noise dyed noise infrared spectrum signal is represented by x _i (n), x _i (n)=x(n)+n _i (n), will add the Gaussian white noise dyed noise infrared spectrum After the signal is subjected to empirical mode decomposition, j eigenmode components are obtained, and the eigenmode components are denoted by c _ij (n), and the j eigenmode components are divided into The weighted average operation eliminates the noisy eigenmode components. After the weighted average operation, the weighted average eigenmode components and residuals are obtained. The weighted average eigenmode components are denoted by c _j and the residuals are denoted by r(n) , M is a natural number,

In the second step, the reconstructed signal is obtained after signal reconstruction of the noise-stained infrared spectrum signal, and the reconstructed signal is represented by x(n), x(n)=∑ _j c _j (n)+r(n), c _j (n) is the jth eigenmode component obtained after performing empirical mode decomposition on the noise-stained infrared spectrum signal after adding Gaussian white noise;

In the third step, the sequence of inputting the reconstructed signal into the variable-step minimum mean square adaptive filter is denoted by X(n),

X(n)=[x(n),x(n-1),...,x(n-M+1)],

The weight vector of the variable step least mean square adaptive filter is represented by W(n),

W(n)=[W _n1 ,W _n2 ,W _n3 ,...,W _nM ] ^T ,

The output signal of the variable step least mean square adaptive filter is represented by y(n),

The error of y(n) relative to the expected signal d(n) is represented by e(n),

e(n)=d(n)-W(n) ^T X(n),

The relationship between the error and the step factor satisfies:

W(n+1)=W(n)+2mue(n)X(n),

mu is the step size factor, mu is a constant, used to control the convergence speed, e(n) is the error of the signal, in order to ensure convergence after iteration, mu must satisfy: 0<mu<λ _max , λ _max is the input sequence X(n ) the largest eigenvalue of the autocorrelation matrix R _xx ,

The iterative formula for the step factor is:

mu ₀ is the initial value of the step factor, all n are natural numbers, and in the iterative formula Among them, n is the number of iterations of the signal point, the number of iterations is a natural number, mu is the nth step factor, and the step factor is updated according to the iterative formula of the step factor. During the update process, when the mean square error When the curve of is the convergence curve, the output signal of the variable step least mean square adaptive filter is the noise reduction signal.

According to the denoising method of the noise-stained infrared spectrum signal of the present invention, the infrared spectrum signal of breast cancer slices stained with noise (with noise) is processed, and the variable step size minimum mean square adaptive filter order is set as K= 12. The number of operations M=30, the initial step size factor mu ₀ =0.044, n=1 to 1744 in the step size factor iteration formula, initialize the weight of the forward filter (initialization weight):

W=[0.1642, 0.1341, 0.0529, -0.0624, -0.1586, -0.1932, -0.1555, -0.0599, 0.0584, 0.1229, 0.1106],

The curve (learning curve) of the mean square error e ² (n) in the process of processing the noise-stained infrared spectrum signal of breast cancer slices according to the method for denoising the infrared spectrum signal of the dyed noise of the present invention is shown in Figure 3, According to the denoising method of the noise-stained infrared spectrum signal of the present invention, the infrared spectrum diagram after the processing of the infrared spectrum signal of the breast cancer slice stained noise is shown in Figure 4, using wavelet, EEMD decomposition and LMS adaptive filter to respectively The infrared spectrum of breast cancer slices is processed with wavelet noise, and the infrared spectrum after wavelet processing is shown in Figure 5. The infrared spectrum of breast cancer slices with EEMD decomposition is shown in Figure 5. As shown in Fig. 6, the infrared spectrogram after LMS adaptive filter is processed to the infrared spectrum signal of breast cancer slice dyeing noise is shown in Fig. 7, compares the denoising method, wavelet, wavelet, EEMD decomposition and the denoising effect of the LMS adaptive filter, the denoising effect is reflected by the three indicators of SNR, RMSE and ρ, the denoising method of the noise-stained infrared spectrum signal of the present invention (EEMD combined with VS-LMS)) , wavelet, EEMD decomposition (EEMD) and LMS adaptive filter (LMS) SNR, RMSE and ρ three indicators are shown in Table 1. SNR (signal-to-noise ratio) indicates the energy ratio of the real signal to noise, so the larger the SNR, the better, RMSE (root mean square error) reflects the error between the noise reduction signal and the original signal, so the smaller the RMSE, the better, ρ (correlation Coefficient) indicates the degree of waveform correlation before and after noise reduction, so the closer ρ is to 1, the better.

From Figure 1 to Figure 7, the resolution of the infrared spectrum in Figure 5 has been improved, but the noise reduction ability of the wavelet is insufficient. Compared with the unstained infrared spectrum in Figure 1, the details in Figure 5 are submerged in the noise And added a lot of peaks and troughs, making the filtered peak characteristics deviate from the original image (Figure 1); from Figure 6, we can see that Figure 6 well retains the characteristics of the peak and has high stability, but the noise reduction Insufficient ability, the peak resolution of the spectrum after noise reduction is insufficient; from Figure 7, it can be seen that the infrared spectrum image processed by the LMS adaptive filter to the noise-stained infrared spectrum signal of breast cancer slices still has obvious noise signals; from Figure 7 3 It can be seen that the learning curve is a convergence curve, and Figure 4 retains the characteristics of the spectrum well, and the main peak is closer to the main peak in Figure 1, that is, the denoising method of the noise-stained infrared spectral signal according to the present invention is adopted The infrared spectrum of breast cancer after noise reduction can better reflect the characteristics of the sample.

It can be seen from Table 1 that the SNR of EEMD combined with VS-LMS is greater than the SNR of the other three methods in Table 1, and the RMSE of EEMD combined with VS-LMS is smaller than that of the other three methods in Table 1. EEMD combined with VS-LMS Compared with the ρ of the other three methods in Table 1, the ρ is closer to 1. It can be seen that the denoising method of the noise-stained infrared spectrum signal of the present invention can reduce the noise of the infrared spectrum signal of the breast cancer section. The effect is better.

In summary, according to the denoising method of the noise-stained infrared spectrum signal of the present invention, the denoising effect of the noise-stained infrared spectrum signal is better, and according to the de-noising method of the noise-stained infrared spectrum signal of the present invention, the denoising effect is reduced. The noised infrared spectrum signal is closer to the unstained infrared spectrum signal, which can provide a better basis for the diagnosis and treatment of breast cancer and other diseases.

The above technical features constitute the embodiment of the present invention, which has strong adaptability and implementation effect, and non-essential technical features can be increased or decreased according to actual needs to meet the needs of different situations.

Claims (1)

1. The denoising method of infrared spectroscopy signals 1. a kind of dye is made an uproar, it is characterised in that carry out in the steps below：The first step is high by being added Dye after this white noise makes an uproar after infrared spectroscopy signals carry out empirical mode decomposition and obtains j intrinsic modal components, and j is natural number, The intrinsic modal components of weighted average and residual volume will be obtained after j intrinsic modal components weighted averages；Second step makes an uproar to dye infrared Spectral signal obtains reconstruction signal after carrying out signal reconstruction；Third walks, and reconstruction signal input variable step size lowest mean square is adaptive Filter is answered, weights, the initial value for setting step factor, setting are initialized in variable step size minimum mean square self-adaption filter Exponent number and setting number of run, variable step size minimum mean square self-adaption filter in the process of running, update step factor, when When the curve of mean square error is convergence curve, variable step size minimum mean square self-adaption filter output signal is de-noising signal；Tool Body,

   In the first step, it is the white Gaussian noise that 0, standard deviation is constant that mean value several times is added in dye makes an uproar infrared spectroscopy signals, White Gaussian noise n_i(n) it indicates, i is natural number, contaminates infrared spectroscopy signals x (n) expressions of making an uproar, the dye of white Gaussian noise is added Infrared spectroscopy signals of making an uproar x_i(n) it indicates, x_i(n)=x (n)+n_i(n), the dye that white Gaussian noise is added is made an uproar infrared spectroscopy signals J intrinsic modal components, intrinsic modal components c are obtained after carrying out empirical mode decomposition_ij(n) indicate, using it is incoherent with J intrinsic modal components weighted mean operations are eliminated noisy intrinsic mode point by the characteristic that machine sequence statistic mean value is 0 It measures, the intrinsic modal components of weighted average and residual volume, the intrinsic modal components c of weighted average is obtained after weighted mean operation_jTable Showing, residual volume indicates that M is natural number with r (n),

   In second step, to make an uproar after infrared spectroscopy signals carry out signal reconstruction to dye and obtain reconstruction signal, reconstruction signal is indicated with x (n), X (n)=∑_jc_j(n)+r (n), c_j(n) it is to carry out empirical modal point to the infrared spectroscopy signals of making an uproar of the dye after white Gaussian noise are added J-th of the intrinsic modal components obtained after solution；

   In third step, the sequence X (n) that reconstruction signal inputs in variable step size minimum mean square self-adaption filter is indicated,

   X (n)=[x (n), x (n-1) ..., x (n-M+1)],

   W (n) expressions of the weight vectors of variable step size minimum mean square self-adaption filter,

   W (n)=[W_n1,W_n2,W_n3,...,W_nM]^T,

   Y (n) expressions of the output signal of variable step size minimum mean square self-adaption filter,

   Error e (n) expressions of the y (n) relative to desired signal d (n),

   E (n)=d (n)-W (n)^TX (n),

   Error and the relationship of step factor meet：

   W (n+1)=W (n)+2mue (n) X (n),

   Mu is step factor, and mu is constant, is used for control convergence speed, and e (n) is the error of signal, in order to receive after ensureing iteration It holds back, mu must satisfy：0 ＜ mu ＜ λ_max, λ_maxFor list entries X (n) autocorrelation matrixes R_xxMaximum eigenvalue,

   The iterative formula of step factor is：

   mu₀For the initial value of step factor, all n are natural number, and in iterative formulaIn, n is letter The iterations of number point, iterations are natural number, and mu is n-th of step factor, and step factor is according to the iteration of step factor Formula is updated, and for step factor during newer, when the curve of mean square error is convergence curve, variable step size is minimum Square sef-adapting filter output signal is de-noising signal.