CN111582132B - Improved EEMD and PCNN-based gas leakage signal noise reduction method - Google Patents

Abstract

The invention discloses a gas leakage sound wave signal noise reduction method based on improved EEMD and PCNN, and belongs to the field of signal processing and computer simulation application. And decomposing the sampling signal into IMF components of several orders by means of the unique decomposition advantage of the EEMD method on the non-stationary nonlinear signal. And then, further reducing the noise of each noise component by using a neural network signal transmission mode of the improved PCNN model. And finally, recombining the noise reduction components into a signal. Compared with the traditional noise reduction algorithm, the method overcomes the defect of mode aliasing, overcomes the defect of difficulty in threshold selection, has higher signal-to-noise ratio of processed signals and smaller root mean square error, and has the advantages of strong noise reduction capability, strong effective signal retention capability and small defects of the traditional method.

Description

Improved EEMD and PCNN-based gas leakage signal noise reduction method

Technical Field

The invention belongs to the field of one-dimensional signal processing, is applied to preprocessing of gas leakage sound wave signals, and particularly relates to a gas leakage signal noise reduction method based on improved EEMD and PCNN.

Background

Toxic, harmful, inflammable and explosive gases are the biggest potential safety hazards in production and transportation of chemical enterprises, and effective monitoring of hazardous gases is one of the main means for reducing occurrence of serious disastrous events. The premise of timely repairing the gas leakage is that the accident occurrence is judged according to the collected signal characteristics, so that the research on the signal noise reduction algorithm with robustness and reliability has great significance for monitoring and remedying the dangerous gas leakage accident.

The gas leak signal generates a primary frequency signal of about 20Hz, and the main sources of gas leak signal noise include mechanical noise and environmental noise. The ultrasonic signal is different from the surrounding noise, but in order to effectively monitor the accurate gas leakage signal in the environment, the acquired sound wave signal needs to be subjected to noise reduction processing.

Signal preprocessing is crucial to subsequent signal feature extraction and analysis, in order to research the problem of gas leakage signal noise reduction, noise reduction algorithms are continuously developed in recent years, for example, a wavelet semi-soft threshold noise reduction method is proposed by Wanglong, and detail coefficients and approximation coefficients of various scales are formulated according to signal correlation coefficients obtained by decomposing signals, so that the deviation of a wavelet soft threshold function and a hard threshold function in the signal reconstruction process is reduced; zhang Xingli proposes an improved threshold function denoising algorithm combining variation modal decomposition and wavelet energy entropy, after VMD decomposition is carried out on an original signal, thresholds of all scale layers are calculated by the wavelet energy entropy to carry out denoising, a threshold selection method is improved, and the signal-to-noise ratio of the signal is improved; the weekly construction provides a self-adaptive wavelet packet threshold function denoising algorithm based on Shannon entropy, a sparse punishment mechanism is formulated, the difficulty caused by threshold selection is avoided, and a better denoising effect is achieved; the method proposes a sequencing data denoising algorithm based on bilinear regression, and is effective in single cell data processing. Therefore, a novel noise reduction method is proposed, and a new idea is provided for reducing the noise of the hazardous gas leakage signal.

Disclosure of Invention

The purpose of the invention is as follows: aiming at the problem of limitation of the traditional signal noise reduction algorithm in the aspect of threshold selection, the invention provides a gas leakage signal noise reduction method based on improved integrated Empirical Mode Decomposition (EEMD) and Pulse Coupled Neural Network (PCNN). The method is suitable for removing the noise of the non-stationary nonlinear signal.

The technical scheme is as follows: in order to realize the purpose of the invention, the technical scheme adopted by the invention is as follows: a gas leakage signal noise reduction method based on improved EEMD and PCNN comprises the following steps:

step 1, collecting an original gas leakage signal, constructing a sample signal, performing integrated empirical mode decomposition (EEMD) on the sample signal to obtain a plurality of IMF components, and solving a correlation coefficient between each IMF component and the original signal; the correlation coefficient refers to a coefficient describing the degree of correlation between each order of component and the original signal after decomposition;

step 2, judging whether each IMF component is a noise component, executing step 3 to further reduce noise for the IMF component judged as the noise component, and reserving the IMF components which are not judged as the noise component to wait for signal reconstruction;

step 3, adjusting internal activity items and neuron attenuation coefficients of the PCNN model according to correlation coefficients of the IMF components and the original signal, and performing PCNN noise reduction on the IMF components respectively;

and 4, performing signal reconstruction on the IMF component subjected to PCNN noise reduction and the IMF component which is not determined as the noise component in the step 2, and completing the noise reduction of the original signal.

Further, in step 1, the sample signal is constructed and EEMD is decomposed by the following method:

step 1.1, adding isometric random white gaussian noise G (t) with an amplitude coefficient of k into an acquired original gas leakage signal X (t) to form a sample signal N (t), namely N (t) = X (t) + G (t);

step 1.2, preliminarily eliminating noise in the sample signal by utilizing the zero-mean characteristic of Gaussian white noise and calculating the mean value of the sample signal, and carrying out EEMD decomposition on the denoised signal to obtain a plurality of IMF components;

and step 1.3, screening the IMF components according to the constraint conditions to obtain IMF components of the characterization signals, and calling a corrcoef function by utilizing matlab to obtain correlation coefficients between the IMF components and the original signals.

Further, the constraint conditions include:

a) The absolute value of the difference between the zero point and the extreme point is less than or equal to 1;

b) The mean of the upper and lower envelope lines of the sample signal N (t) is zero.

Further, in the step 3, adjusting an internal activity item and a neuron attenuation coefficient of the PCNN model, and performing PCNN noise reduction on each IMF component respectively; the method comprises the following steps:

step 3.1, changing the dimensionality of two parallel channels of neuron feedback input and coupling connection input in a receiving domain of a traditional pulse coupling neural network PCNN model from m & n to 1 & n, and simultaneously changing a connection matrix and a weight matrix into a matrix of 1 & n, so that the weight obtained by performing Euclidean square inverse on the weight matrix is suitable for a one-dimensional vector, and the transfer mechanism of the pulse coupling neural network PCNN is applied to one-dimensional signal noise reduction;

the principle of filtering noise in a signal by using the PCNN model is as follows: when a signal enters a neural network as an input signal, each number in a signal sequence is a neuron, and the relation between an activity item in a PCNN model representing the activity state of the neuron and a threshold value is determined to be output;

step 3.2, initializing the PCNN model, inputting each IMF component as a feedback input signal into a receiving domain of the PCNN model, and forming a feedback input signal F _i [n]＝S _i In which S is _i Representing a neuron excitation; the coupled connection input signal calculation formula of the PCNN model is as follows:

wherein L is _i [n]Representing a coupled input signal, V _L Denotes the connection amplitude constant, W _i Representing a coupling connection matrix, Y _j Representing the neuron ignition state in the jth ignition process of the PCNN, i representing the time domain serial number of a certain numerical value in the signal sequence, j representing the counting times in the ignition process, and n representing the iteration times;

3.3, capturing neurons by using the improved PCNN model and determining whether to ignite to generate pulses, changing the size of a threshold value if the pulses are generated by ignition, namely, performing signal denoising processing by using a dynamic threshold value, generating the pulses by ignition, changing the size of the dynamic threshold value and filtering signals with ineffective amplitude values;

capture satisfaction

The neurons jointly determine the state of an internal activity item of the PCNN model and determine whether the neurons ignite or not; />

The principle of pulse generation is as follows:

wherein T is _i [n]Is a dynamic threshold value, alpha _E Is a threshold attenuation coefficient, V, determining a threshold attenuation speed _E Is the lifting factor of the dynamic threshold, E _i [n-1]Is a link input L _i [n]Last moment state of (Y) _i [n-1]Is the last time state of the neuron iteration n times pulse output;

when alpha is _E Enlarged and internal activity item U _i [n]>T _i [n]When the neuron meets the ignition condition, the ith neuron is ignited; y is _i Taking the amplitude of the effective signal at the ith point as an output value; when alpha is _E When reduced, neurons do not meet firing conditions, Y _i Output 0, expressed as follows:

step 3.4, calculating a weight matrix of the model, and adjusting the neuron internal salient connection coefficient and the neuron attenuation coefficient in the internal activity item of the PCNN model according to the correlation coefficient of each IMF component and the original signal, so that the PCNN model is suitable for each IMF component; internal activity item U _i [n]Comprises the following steps:

U _i [n]＝F _i [n](1+βL _i [n])

wherein F _i [n]Represents the feedback input obtained after the neuron iterates for n times, beta represents the neuron internal prominent connection coefficient, and L _i [n]Representing the coupled connection input after iteration for n times;

step 3.5, judging whether the termination condition is met, namely continuous k times of iteration Y _i Outputting 0, if the termination condition is satisfied, iteratingEnding, outputting the current signal sequence; otherwise, returning to the step 3.2, re-initializing the model, and re-inputting the current signal sequence into the model for the next iteration until the termination condition is met.

Has the beneficial effects that: compared with the prior art, the technical scheme of the invention has the following beneficial technical effects:

the invention changes the dimensionality of two parallel channels of feedback input and coupling connection input of a neuron in a receiving domain of the traditional pulse coupling neural network from m & n to 1 & n, and simultaneously changes a connection matrix and a weight matrix into a matrix of 1 & n. Therefore, the weight obtained by making the weight matrix into the inverse of the Euclidean square is suitable for the one-dimensional vector, and the neural network transmission mechanism of the PCNN excellence is transplanted into the one-dimensional signal processing. Compared with a common modal decomposition noise reduction method, the PCNN noise reduction model is added, and the noise reduction is performed on the signals from different angles, so that the noise reduction capability of the method is stronger. In addition, on the basis of the traditional threshold noise reduction concept, the defect that the threshold is difficult to determine is overcome. Therefore, the noise reduction method has the advantages of strong noise reduction capability, strong effective signal retention capability and small defects of the traditional method.

Drawings

FIG. 1 is a flow chart of a method implementation of the present invention;

FIG. 2 is a flow chart of the PCNN algorithm;

FIG. 3 is a graph of a raw signal for gas leakage;

FIG. 4 is an exploded view of the EEMD;

FIG. 5 is a comparison graph before and after IMF1 denoising;

FIG. 6 is a comparison graph before and after IMF2 denoising;

FIG. 7 is a graph of the original signal spectrum;

FIG. 8 is a graph of the spectrum of the original signal after EEMD-PCNN denoising;

FIG. 9 is a signal diagram of the original signal after EEMD-PCNN denoising;

FIG. 10 is a front-to-back comparison of a signal plus white Gaussian noise with a signal-to-noise ratio of 50 dB;

FIG. 11 is a graph of a spectrum of a noisy signal;

fig. 12 is a signal spectrum diagram after noise reduction.

Detailed Description

The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.

The invention relates to a gas leakage signal noise reduction method based on improved EEMD and PCNN, which is implemented by the process shown in figure 1, wherein a sample signal is decomposed by EEMD, IMF components which are determined as noise components are subjected to PCNN noise reduction continuously, and the rest components are used for signal reconstruction subsequently, namely the IMF components which are not determined as the noise components and the noise components subjected to the subsequent PCNN noise reduction are subjected to signal reconstruction together. The specific implementation comprises the following steps:

step 1, collecting an original gas leakage signal, constructing a sample signal, performing integrated empirical mode decomposition (EEMD) on the sample signal to obtain a plurality of IMF components, and solving a correlation coefficient between each IMF component and the original signal; the correlation coefficient is a coefficient describing the degree of correlation between each order component and the original signal after decomposition.

Constructing a sample signal and EEMD decomposition, wherein the method comprises the following steps:

step 1.1, adding isometric random white gaussian noise G (t) with an amplitude coefficient k of 0.3 into an acquired original gas leakage signal X (t) to form a sample signal N (t), namely N (t) = X (t) + G (t);

step 1.2, utilizing the zero-mean characteristic of Gaussian white noise, preliminarily eliminating noise in a sample signal by averaging the sample signal, and carrying out EEMD decomposition on the denoised signal to obtain a plurality of IMF components;

and step 1.3, screening the IMF components according to the constraint conditions to obtain IMF components of the characterization signals, and calling a corrcoef function by utilizing matlab to obtain the correlation coefficient between each IMF component and the original signal.

The constraint conditions include:

a) The absolute value of the difference between the zero point and the extreme point is less than or equal to 1;

b) The mean of the upper and lower envelope lines of the sample signal N (t) is zero.

And 2, judging whether each IMF component is a noise component, executing step 3 to further reduce noise for the IMF component judged as the noise component, and reserving the IMF components which are not judged as the noise component to wait for signal reconstruction.

And 3, adjusting internal activity items and neuron attenuation coefficients of the PCNN model according to the correlation coefficients of the IMF components and the original signal, and performing PCNN noise reduction on the IMF components respectively. The method comprises the following steps:

step 3.1, changing the dimensionality of two parallel channels of neuron feedback input and coupling connection input in a receiving domain of a traditional pulse coupling neural network PCNN model from m & n to 1 & n, and simultaneously changing a connection matrix and a weight matrix into a matrix of 1 & n, so that the weight obtained by performing Euclidean square inverse on the weight matrix is suitable for a one-dimensional vector, and the transfer mechanism of the pulse coupling neural network PCNN is applied to one-dimensional signal noise reduction;

the principle of filtering noise in a signal by using the PCNN model is as follows: when a signal enters a neural network as an input signal, each number in a signal sequence is a neuron, and the relation between an activity item in a PCNN model representing the activity state of the neuron and a threshold value determines the output.

Step 3.2, initializing the PCNN model, inputting each IMF component as a feedback input signal into a receiving domain of the PCNN model, and forming a feedback input signal F _i [n]＝S _i In which S is _i Representing a neuron excitation; the coupled connection input signal calculation formula of the PCNN model is as follows:

/>

wherein L is _i [n]Representing a coupled input signal, V _L Denotes the connection amplitude constant, W _i Representing a coupling connection matrix, Y _j The method is characterized in that the method represents the neuron ignition state in the jth ignition process of the PCNN, i represents the time domain serial number of a certain numerical value in a signal sequence, j represents the counting times of the ignition process, and n is the iteration times.

3.3, capturing neurons by using the improved PCNN model and determining whether to ignite to generate pulses, changing the size of a threshold value if the pulses are generated by ignition, wherein the threshold value is a dynamic threshold value, mapping the pulses generated by each ignition, and actually representing the strength of effective signals in signal components, namely, performing signal denoising processing by using the dynamic threshold value, generating the pulses by ignition, so as to change the size of a dynamic threshold value and filter signals with ineffective amplitude values;

as shown in FIG. 2, the improved PCNN modifies the dimensions of the neuron channel, capturing the satisfaction

The neurons jointly determine the state of an activity item in the PCNN model and determine whether the neurons fire or not;

the principle of pulse generation is as follows:

wherein T is _i [n]Is a dynamic threshold value, alpha _E Is a threshold attenuation coefficient which determines a threshold attenuation speed when alpha is _E Increase, T _i [n]With a consequent decrease in V _E Is the coefficient of rise of the dynamic threshold, the variation of which is in a positive relationship with the variation of the dynamic threshold, E _i [n-1]Is a link input L _i [n]Last moment state of (Y) _i [n-1]Is the last time state of the neuron iteration n times of pulse output;

when alpha is _E Augmented and internal activity item U _i [n]>T _i [n]When the neuron meets the ignition condition, the ith neuron is ignited; y is _i The amplitude of the effective signal at the ith point is taken as an output value, and thus, the generation of the pulse is facilitated; the effective signal is a signal which can be used for extracting features besides the noise signal;

when alpha is _E When decreasing, T _i [n]Rapidly increase, neuron not meeting ignition condition, Y _i Output 0, generation of pulses is suppressed, and the pulse output state is expressed as follows:

step 3.4, calculating a weight matrix of the model, and adjusting the neuron internal salient connection coefficient and the neuron attenuation coefficient in the internal activity item of the PCNN model according to the correlation coefficient of each IMF component and the original signal, so that the PCNN model is suitable for each IMF component; internal activity item U _i [n]Comprises the following steps:

U _i [n]＝F _i [n](1+βL _i [n])

wherein F _i [n]Representing the feedback input obtained after n iterations of the neuron, beta representing the coefficient of prominent connections within the neuron, L _i [n]Representing the coupled connection input after iteration for n times;

in this embodiment, the PCNN parameter is adjusted before the IMF9 component is processed, and a value with the best effect, such as an attenuation coefficient, is obtained, the ability to detect an effective signal is low when the attenuation coefficient is 0.3, and it is difficult to separate noise, and the effective signal is filtered together with noise when the attenuation coefficient is 0.6, and then a proper value is obtained by using a bisection method.

Step 3.5, judging whether the termination condition is met, namely continuous k times of iteration Y _i Outputting 0, if the termination condition is met, finishing the iteration, and outputting the current signal sequence; otherwise, returning to the step 3.2, re-initializing the model, and re-inputting the current signal sequence into the model for the next iteration until the termination condition is met. In this embodiment, k is 50.

And 4, performing signal reconstruction on the IMF component subjected to noise reduction by the pulse coupled neural network model (PCNN) and the IMF component which is not determined as the noise component in the step 2, and completing noise reduction on the original signal.

As shown in fig. 3, the original signal of the gas leakage is doped with a large amount of noise in the environment. The experiment of this example used a cDAQ data acquisition card developed by NI instruments with an ultrasonic transducer to acquire the simulated gas leak signal. The sampling frequency is 100kHZ, the sampling distance is 2m, and 2000 continuous sampling points are selected from the experimental data.

As shown in fig. 4, the original signal is screened and decomposed into 9 IMF components meeting the conditions:

the signal W (t) to be detected is obtained by subtracting the local average value from the original signal, N (t) is a sample signal, and the upper envelope H (t) and the lower envelope L (t) of the original signal are obtained by cubic spline interpolation.

As shown in fig. 5 and 6, the comparison between the original IMF1 and IMF2 components and the components subjected to PCNN noise reduction is shown, where the signal-to-noise ratio of the noise components is significantly improved. As shown in fig. 7 and 8, after the noise reduction is completed, the unnecessary noise components are filtered out from the signal, and the main frequency characteristics of the effective signal are relatively more obvious. As shown in FIG. 9, the signal is noise reduced by EEMD-PCNN to improve the signal-to-noise ratio and reduce the root mean square error. As shown in FIG. 10, gaussian white noise having a signal-to-noise ratio (SNR) of 50dB is added to the signal. The above experiments all verify that EEMD-PCNN noise reduction has good effect of filtering unknown noise, but in order to verify that the method can effectively filter noise instead of effective signals, gaussian white noise with known signal-to-noise ratio is added to the signals, and EEMD-PCNN noise reduction is carried out. As shown in fig. 11, the signal spectrum has a high frequency component due to the addition of white gaussian noise. As shown in fig. 12, the high frequency components are filtered out and the effective signal is retained after noise reduction, which confirms that the method can filter out noise while retaining the effective signal.

The above is a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, it is possible to make various improvements and modifications without departing from the technical principle of the present invention, and those improvements and modifications should be considered as the protection scope of the present invention.

Claims (3)

1. A gas leakage signal noise reduction method based on improved EEMD and PCNN is characterized in that: the method comprises the following steps:

step 1, collecting an original gas leakage signal, constructing a sample signal, carrying out integrated empirical mode decomposition (EEMD) on the sample signal to obtain a plurality of IMF components, and solving a correlation coefficient between each IMF component and the original signal;

step 2, judging whether each IMF component is a noise component, executing step 3 to further reduce noise for the IMF component judged as the noise component, and reserving the IMF components which are not judged as the noise component to wait for signal reconstruction;

step 3, adjusting internal activity items and neuron attenuation coefficients of the PCNN model according to correlation coefficients of the IMF components and the original signal, and performing PCNN noise reduction on the IMF components respectively;

step 3.1, changing the dimensionality of neuron feedback input and coupling connection input in a PCNN model receiving domain from m & n to 1 & n, and simultaneously changing a connection matrix and a weight matrix to a 1 & n matrix;

step 3.2, initializing the PCNN model, inputting each IMF component as a feedback input signal into a receiving domain of the PCNN model, and forming a feedback input signal F _i [n]＝S _i In which S is _i Representing a neuron excitation; the coupled connection input signal calculation formula of the PCNN model is as follows:

wherein L is _i [n]Representing a coupled input signal, V _L Denotes the connection amplitude constant, W _i Representing a coupling connection matrix, Y _j Representing the neuron ignition state in the jth ignition process of the PCNN, i representing the time domain serial number of a certain numerical value in the signal sequence, j representing the counting times in the ignition process, and n representing the iteration times;

3.3, capturing neurons by using the improved PCNN model and determining whether to ignite to generate pulses, changing the size of a threshold value if the pulses are generated by ignition, namely, performing signal denoising processing by using a dynamic threshold value, generating the pulses by ignition, changing the size of the dynamic threshold value and filtering signals with ineffective amplitude values;

capture satisfaction

The neurons jointly determine the state of an internal activity item of the PCNN model and determine whether the neurons ignite or not;

pulse generation is represented as follows:

wherein T is _i [n]Is a dynamic threshold value, alpha _E Is a threshold attenuation coefficient, V, determining a threshold attenuation speed _E Is the coefficient of rise of the dynamic threshold, E _i [n-1]Is a link input L _i [n]Last moment state of (A), Y _i [n-1]Is the last time state of the neuron iteration n times of pulse output;

when alpha is _E Enlarged and internal activity item U _i [n]>T _i [n]When the neuron meets the ignition condition, the ith neuron is ignited; y is _i Taking the amplitude of the effective signal at the ith point as an output value; when alpha is _E When reduced, neurons do not meet firing conditions, Y _i Output 0, expressed as follows:

step 3.4, calculating a weight matrix of the model, and adjusting neuron internal salient connection coefficients and neuron attenuation coefficients in internal activity items of the PCNN model according to correlation coefficients of each IMF component and an original signal; internal activity item U _i [n]Comprises the following steps:

U _i [n]＝F _i [n](1+βL _i [n])

wherein F _i [n]Represents the feedback input obtained after the neuron iterates for n times, beta represents the neuron internal prominent connection coefficient, and L _i [n]Representing the coupling input after n iterations;

step 3.5, judging whether the termination condition is met, namely continuous k times of iteration Y _i Outputting 0, if the termination condition is met, iteratingEnding, outputting the current signal sequence; otherwise, returning to the step 3.2, re-initializing the model, and re-inputting the current signal sequence into the model for next iteration until the termination condition is met;

and 4, performing signal reconstruction on the IMF component subjected to PCNN noise reduction and the IMF component which is not determined as the noise component in the step 2, and completing noise reduction on the original signal.

2. The method of claim 1, wherein the method comprises the steps of: in the step 1, a sample signal and EEMD decomposition are constructed, and the method comprises the following steps:

step 1.1, adding isometric random white gaussian noise G (t) with an amplitude coefficient of k into an acquired original gas leakage signal X (t) to form a sample signal N (t), namely N (t) = X (t) + G (t);

step 1.2, utilizing the zero-mean characteristic of Gaussian white noise, preliminarily eliminating noise in a sample signal by averaging the sample signal, and carrying out EEMD decomposition on the denoised sample signal to obtain a plurality of IMF components;

and step 1.3, screening the IMF components according to the constraint conditions to obtain IMF components of the characterization signals, and calling a corrcoef function by utilizing matlab to obtain correlation coefficients between the IMF components and the original signals.

3. The method of claim 2, wherein the noise reduction method for the gas leakage signal based on EEMD and PCNN comprises: the constraint conditions include:

a) The absolute value of the difference between the zero point and the extreme point is less than or equal to 1;

b) The mean of the upper and lower envelope lines of the sample signal N (t) is zero.

Images