CN113158769A - CEEMDAN and FastICA-based electromechanical device bearing vibration signal denoising method - Google Patents

Abstract

The invention relates to the technical field of complete empirical mode decomposition CEEMDAN and fixed point algorithm FastICA of self-adaptive noise, and discloses a method for removing noise of a bearing vibration signal of an electromechanical device based on CEEMDAN and FastICA. The specific process is as follows: acquiring a device vibration signal as a raw data sample through a sensor placed on the electromechanical device; decomposing the acquired data samples by using a complete empirical mode decomposition method of CEEMDAN adaptive noise to obtain a plurality of intrinsic mode functions IMF which are used as observation signals X (t); then obtaining source signal estimation S (t) by a FastICA algorithm; then, obtaining a new IMF by ICA inverse transformation, and directly accumulating the IMF to reach a required reconstruction signal; and (4) the deep learning network model is used for classification and identification, so that the bearing state is diagnosed. Through the technical scheme of the invention, the signal extraction effect is obviously improved, and the identification accuracy is obviously improved.

Description

CEEMDAN and FastICA-based electromechanical device bearing vibration signal denoising method

Technical Field

The invention belongs to the technical field of electromechanical fault signal processing, and particularly relates to a CEEMDAN and FastICA-based fault diagnosis method for a bearing used in a mine machinery for signal denoising.

Background

Most of mine electromechanical equipment is large mechanical equipment, and most of the machines work in the open air due to the complex environment around the mine, the mining field often receives various interferences, stones and the like easily enter the interior of the machines during mining, so that the vibration is abnormal, and mechanical faults are generated. The bearing is an important part in mechanical equipment and is often the cause of most faults, and the fault diagnosis of the bearing is one of the gravity centers of the diagnostic mechanical equipment. At present, measures such as manual periodic inspection, manual experience evaluation and repair after failure are adopted to perform early warning and maintenance work on mechanical equipment, but the problems of missing of optimal maintenance time, poor real-time performance, high maintenance cost and the like easily occur, and once a large-scale machine fails, serious consequences such as property loss, casualty events and the like can be brought. The specific vibration signal generated by the bearing operation can represent abundant mechanical state information, so that the initial processing of the vibration fault signal is very important.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a CEEMDAN and FastICA-based electromechanical device bearing vibration signal denoising method.

The design fully considers the characteristic information of a sensor for collecting vibration data samples, carefully knows the advantages of signal denoising of CEEMDAN and FastICA, provides clearer original signals for a subsequent deep learning network, and is favorable for better optimizing a model. The diagnosis accuracy and efficiency are improved.

The technical scheme adopted by the invention for solving the technical problems is as follows: a CEEMDAN and FastICA-based electromechanical device bearing vibration signal denoising method comprises the following steps:

step one, decomposing vibration data acquired by a sensor by using a complete empirical mode decomposition method of CEEMDAN adaptive noise to obtain a plurality of intrinsic mode functions IMF which are used as observation signals X (t).

And step two, denoising the eigen function obtained by decomposition through a FastICA algorithm, and obtaining a new IMF' by utilizing ICA inverse transformation.

And step three, directly accumulating and restoring the new IMF' to obtain a reconstructed signal, extracting a characteristic vector, and reducing the dimension by using LLE.

And step four, the feature vector is used as the input of the deep learning network, and the diagnosis result is obtained through output.

In the first step, the method comprises the following steps:

firstly, solving a first-order modal component, adding positive-negative pair Gaussian white noise which obeys standard normal distribution into an original data signal, taking m as a coefficient, epsilon as an amplitude, ni (t) as an ith added white noise sequence, and i as an auxiliary noise frequency, namely:

wherein

2. Performing EMD on the obtained new signal to obtain a plurality of IMF components;

3. obtaining a first-order final component by averaging a plurality of IMFs

And a first order residual component r1(t);

the decomposition formula is:

wherein the first order final decomposition component is:

=

the first order residual component is:

solving the second-order modal component, adding positive and negative Gaussian white noise into the first-order residual component r1(t), and carrying out N-time decomposition on the formed new signal to obtain the second-order component

And a residual component r2(t);

the decomposition formula is:

+

=

wherein the second order final decomposition component is:

=

the second order residual component is:

5. repeating the steps until the residual signal is inseparable, and then expressing the original signal as:

the observation signals x (t) = { x1(t), x2(t), …, xn (t) } are finally obtained

In the second step, the method comprises the following steps:

on the basis of the ICA independent component analysis method, FastICA was changed as follows,

1. changing the original cumulative distribution function sigmoid function into

The probability distribution function is the derivative of the cumulative distribution function as:

wherein often a1=1

The iterative formula of W in the original ICA is as follows:

the iterative formula for W in FastICA is changed to:

finally, the original signal estimation S (t) = { s1(t), s2(t), …, sm (t) }is obtained

And subjecting to reverse ICA conversion to obtain novel IMF'

In the fourth step, the method comprises the following steps:

performing dimensionality reduction operation on the feature extraction vector through an LLE technology to enable the data feature to meet the dimensionality suitable for network processing;

in consideration of the characteristics of sample data of mechanical equipment, an unsupervised domain self-adaptive transfer learning model RTN is adopted, fault type judgment is carried out by utilizing spectral clustering, then the Bayesian network based on maximum likelihood estimation is utilized to analyze the fault degree, and finally, a network diagnosis output result is obtained, wherein the network diagnosis output result comprises real-time feedback of fault components, fault types and fault early warning degrees.

Compared with the prior art, the invention has the following beneficial effects:

1. on the basis of EEMD, white noise is added in a self-adaptive manner through the CEEMDAN method, so that the problems that decomposition loses completeness and reconstruction errors are generated after white noise is added in other modal decomposition such as EEMD and the like are solved, and the accuracy of signal decomposition is improved.

2. The functions of removing noise and purifying signals are achieved by introducing FastICA, more accurate signal reconstruction is achieved, original signals with the reduction degree as high as possible are provided for a subsequent deep learning network, network model optimization is improved, and identification accuracy is enhanced.

3. By means of the transfer learning spectral clustering algorithm and the Bayesian network of the maximum likelihood estimation, the problems that large-scale electromechanical equipment is few in fault data samples and high in data acquisition cost are solved, the clustering processing capacity for judging fault information types is improved, the interpretability is enhanced, and the fault degree of a certain fault category is visualized.

Drawings

FIG. 1 is a flow chart illustrating the steps of a CEEMDAN and FastICA-based method for denoising vibration signals of bearings of electromechanical devices according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a CEEMDAN EMD;

FIG. 3 is a detailed operation flow of the deep learning module in the first step;

FIG. 4 is a chart of the results of a CEEMDAN decomposition of the university of Kaiser West reservoir 105 vibration data set experiment based on the present invention;

FIG. 5 is a schematic diagram of the experiment of noise-removing effect of FastICA signal separation.

Detailed Description

The technical solutions in the embodiments of the present invention will be described below in detail and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all embodiments of the present invention.

Referring to fig. 1, a method for denoising electromechanical device bearing vibration signals based on CEEMDAN and FastICA is shown according to an embodiment of the present invention. The implementation process can be divided into four steps:

firstly, decomposing vibration data acquired by a sensor by using a complete empirical mode decomposition method of CEEMDAN adaptive noise to obtain a plurality of intrinsic mode functions IMF, and taking the intrinsic mode functions IMF as observation signals X (t) = { x1(t), x2(t), …, xn (t) }.

Please refer to fig. 2, which illustrates the principle process of the empirical mode decomposition of CEEMDAN

Please refer to fig. 4, which is a diagram illustrating the experimental results of the CEEMDAN decomposition of the kaiser university 105 vibration data set according to the present invention

In the first step, the method comprises the following steps:

1. firstly, solving a first-order modal component, adding positive-negative pair Gaussian white noise which obeys standard normal distribution into an original data signal, taking m as a coefficient, epsilon as an amplitude, ni (t) as an ith added white noise sequence, and i as an auxiliary noise frequency, namely:

wherein

2. Performing EMD on the obtained new signal to obtain a plurality of IMF components;

3. obtaining a first-order final component by averaging a plurality of IMFs

And a first order residual component r1(t);

the decomposition formula is:

wherein the first order final decomposition component is:

=

the first order residual component is:

solving the second-order modal component, adding positive and negative Gaussian white noise into the first-order residual component r1(t), and carrying out N-time decomposition on the formed new signal to obtain the second-order component

And a residual component r2(t);

the decomposition formula is:

+

=

wherein the second order final decomposition component is:

=

the second order residual component is:

repeating the steps until the residual signal is inseparable, and then expressing the original signal as:

finally obtaining observed signals X (t) = { x1(t), x2(t), …, xn (t)

And step two, denoising the eigen function obtained by decomposition through a FastICA algorithm, and obtaining a new IMF' by utilizing ICA inverse transformation.

FIG. 5 is a schematic diagram of the noise-removing effect of FastICA signal separation

In the second step, the method comprises the following steps:

on the basis of the ICA independent component analysis method, FastICA was changed as follows,

1. changing the original cumulative distribution function sigmoid function into

The probability distribution function is the derivative of the cumulative distribution function as: p(s) = g(s)' = tanh (a1, y)

Wherein often a1=1

The iterative formula of W in the original ICA is as follows:

the iterative formula for W in FastICA is changed to:

finally, the source signal estimation S (t) = { s1(t), s2(t), …, sm (t) }is obtained

And subjecting to reverse ICA conversion to obtain novel IMF'

And step three, directly accumulating and restoring the new IMF' to obtain a reconstructed signal, extracting a characteristic vector, and reducing the dimension by using LLE.

And step four, the feature vector is used as the input of the deep learning network, and the diagnosis result is obtained through output.

In the fourth step, the method comprises the following steps:

performing dimensionality reduction operation on the feature extraction vector through an LLE technology to enable the data feature to meet the dimensionality suitable for network processing;

please refer to fig. 3, which illustrates a specific operation process of the deep learning module in the first step

In consideration of the characteristics of sample data of mechanical equipment, an unsupervised domain self-adaptive transfer learning model RTN is adopted, fault type judgment is carried out by utilizing spectral clustering, then the Bayesian network based on maximum likelihood estimation is utilized to analyze the fault degree, and finally, a network diagnosis output result is obtained, wherein the network diagnosis output result comprises real-time feedback of fault components, fault types and fault early warning degrees.

The above-described embodiments, objects, technical principles, decision schemes and advantages of the present invention are described in more detail, it should be understood that the above-described embodiments are merely exemplary embodiments of the present invention, and are not intended to limit the scope of the present invention, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (5)

1. The electromechanical device bearing vibration signal denoising method based on CEEMDAN and FastICA is characterized by comprising the following steps:

decomposing vibration data acquired by a sensor by using a complete empirical mode decomposition method of CEEMDAN adaptive noise to obtain a plurality of intrinsic mode functions IMF which are used as observation signals X (t);

denoising the eigen function obtained by decomposition through a FastICA algorithm, and obtaining a new IMF' by utilizing ICA inverse transformation;

step three, directly accumulating and restoring the new IMF' to obtain a reconstructed signal, extracting a characteristic vector, and reducing the dimension by using LLE;

and step four, the feature vector is used as the input of the deep learning network, and the diagnosis result is obtained through output.

2. The CEEMDAN and FastICA-based electromechanical device bearing vibration signal denoising method of claim 1, wherein the first step comprises: the device vibration signal is collected as a raw data sample by a sensor placed on the electromechanical device.

3. The CEEMDAN and FastICA-based electromechanical device bearing vibration signal denoising method of claim 1, wherein the first step comprises:

performing complete empirical mode decomposition on CEEMDAN self-adaptive noise on an original vibration data sample, wherein the complete empirical mode decomposition comprises the following steps:

firstly, solving a first-order modal component, and adding positive-negative pair Gaussian white noise which follows standard normal distribution into an original data signal;

performing EMD on the obtained new signal to obtain a plurality of IMF components;

obtaining a first-order final component by averaging a plurality of IMFs

And a first order residual component r1(t);

then solving the second order modal component to obtain the second order component

And a residual component r2(t);

repeating the steps until the residual signal is inseparable, and then expressing the original signal as:

。

4. the CEEMDAN and FastICA-based method of denoising electromechanical device bearing vibration signals as in claim 1, wherein the second step comprises:

the method has the advantages that the traditional ICA independent component analysis method is improved, the FastICA algorithm is adopted, the operation speed is higher, the effect is better, the iterative formula of the accumulative distribution function and the confusion matrix is changed on the basis of the traditional ICA independent component analysis method, and the better denoising performance is obtained;

and then ICA inverse change is carried out on the original signal estimation S (t) obtained by FastICA to obtain new IMF', and accumulation reconstruction, feature extraction and dimension reduction by LLE are carried out in the next step.

5. The CEEMDAN and FastICA-based method of denoising electromechanical device bearing vibration signals in accordance with claim 1, wherein the fourth step comprises:

extracting features, namely reducing dimensions by using LLE, and taking the obtained feature vector as the input of a deep learning network;

in consideration of the characteristics of sample data of mechanical equipment, an unsupervised domain self-adaptive transfer learning model RTN is adopted, fault type judgment is carried out by utilizing spectral clustering, then the Bayesian network based on maximum likelihood estimation is utilized to analyze the fault degree, and finally, a network diagnosis output result is obtained, wherein the network diagnosis output result comprises real-time feedback of fault components, fault types and fault early warning degrees.

Images