CN108827634A - Manifold merges empirical mode decomposition method - Google Patents

Abstract

The present invention relates to a kind of manifolds to merge empirical mode decomposition method, including：Analysis signal in be added mean value be 0, the random white noise that standard deviation is σ, obtain noisy signal；EMD processing is carried out to the noisy signal, obtains the IMF comprising fault message, i.e. failure modalities component；Change the value of σ, repeat the above steps n times, obtains N number of failure modalities component with different noise intensities, wherein N is positive integer；N number of failure modalities component is merged according to given manifold learning, obtains the inherent manifold structure of higher-dimension failure modalities component, i.e. failure transient ingredient.Above-mentioned manifold merges empirical mode decomposition method, takes different values to the standard deviation of each random white noise being added in analysis signal, using the outstanding feature mining ability of manifold learning, extracts the transient components with rock-steady structure from higher-dimension failure modalities component.

Description

Manifold Fusion Empirical Mode Decomposition Method

technical field

The invention relates to fault diagnosis of mechanical equipment, in particular to a manifold fusion empirical mode decomposition method.

Background technique

Rotating mechanical equipment is developing towards large-scale, precision and automation, which puts forward stricter requirements for the manufacture, installation and daily maintenance of each component in the entire equipment system. A slight damage or vibration of any component Misalignment may affect the normal operation of the entire system and even cause major accidents. When a rotating component fails, there is a transient shock response component in its vibration signal, and the successful detection of the transient component in the signal is a common method for effective fault diagnosis of rotating machinery. However, due to the complexity of the working environment, the vibration signals of rotating machinery often appear as non-stationary, and contain a variety of frequency components, including a large amount of environmental noise, resulting in the transient components related to faults appearing very weak in the signal, thus giving Fault diagnosis is inconvenient.

The Empirical Mode Decomposition (EMD) method is a non-stationary signal processing method, which can adaptively decompose complex signals into a finite number of Intrinsic Mode Functions (IMF). Each IMF satisfies two conditions: a) the function has the same number of local extremum points and zero-crossing points in the entire time range, or a difference of at most one; b) at any time point, the envelope of the local maximum (upper envelope Envelope) and the envelope of the local minimum (lower envelope) have an average value of zero. Each IMF obtained by the EMD method contains local feature information of different time scales of the original signal. Because the vibration transient components excited by rotating machinery faults meet the conditions of IMF, EMD has been widely used in the detection of transient components of mechanical vibration signals. However, the EMD method has the problem of mode aliasing, that is, one time-scale sequence is distributed in two IMFs, or there are multiple time-scale sequences in one IMF. The main reason for the modal aliasing phenomenon is the discontinuity of the signal, for example, the signal is doped with noise, shock pulse and intermittent signal. The modal mixing problem leads to incomplete transient component information obtained by the EMD method or mixed with interference components, which is not conducive to fault identification.

In order to solve the mode mixing problem in the EMD method, the existing technology is mainly Ensemble Empirical Mode Decomposition (EEMD) method. The principle of this method is to add white noise to the original signal, and use the uniform distribution characteristics of the white noise spectrum to make the signal have continuity on different time scales, so that the EMD processing of the noise-added signal can avoid the problem of modal aliasing . The noise added to the signal will be distributed in each IMFs. In order to remove these introduced noises, the method adopted by EEMD is to perform EMD processing after multiple noise additions, and then calculate the average of multiple IMFs with similar frequency bands. The zero-mean property of noise removes the introduced noise. The implementation steps of EEMD are as follows: first, add random white noise with a mean value of 0 and a standard deviation of σ to the analysis signal; then perform EMD processing on the noise-added signal to obtain a sub-signal group composed of n IMFs; then repeat the above steps M times, M sub-signal groups are obtained; finally, the average value of IMFs with similar frequency band ranges in all sub-signal groups is calculated, and n IMFs that remove the introduced noise are obtained.

The traditional technology has the following technical problems:

The EEMD method solves the modal aliasing problem by adding white noise. The aggregation number M and the standard deviation σ of the white noise are two parameters that the EEMD method needs to be set artificially. Different parameters will have a certain impact on the signal decomposition results. Increasing the value of M can reduce (but not completely eliminate) the content of noise introduced in the final IMFs, but at the same time it will also increase the amount of calculation, and the self-noise in the respective frequency bands of IMFs, that is, the in-band self-noise, cannot It is eliminated due to the increase of M; if σ is too small, the mode aliasing problem cannot be suppressed, and if it is too large, it will not only increase the number of decomposed IMFs and increase the amount of calculation, but also make it difficult to decompose the high-frequency components in the signal and the IMFs in the IMFs. Introduce the problem of excessive noise residual. Therefore, there are two main problems in existing EEMD methods: a) the limited ensemble averaging cannot completely eliminate the introduced noise in each IMF, let alone the in-band self-noise in each IMF; b) the standard deviation of the introduced white noise It is difficult to determine an appropriate value for σ.

Contents of the invention

Based on this, it is necessary to provide a manifold fusion empirical mode decomposition method for the above technical problems. This method is aimed at the problem that the limited ensemble average in the EEMD method cannot effectively eliminate the introduced noise and self-noise, and the standard deviation of the introduced noise For problems that are difficult to determine, the method of manifold learning is used to fuse multiple IMFs with different noise intensities containing fault information, so that there is no need to determine the standard deviation of the introduced noise, and the introduced noise and self-noise in the band can be better eliminated , to effectively detect the transient components in the signal.

A manifold fusion empirical mode decomposition method, comprising:

Add random white noise with a mean of 0 and a standard deviation of σ to the analysis signal to obtain a noise-added signal;

EMD processing is performed on the noise-added signal to obtain an IMF containing fault information, that is, a fault modal component;

Change the value of σ and repeat the above steps N times to obtain N fault modal components with different noise intensities, where N is a positive integer;

According to a given manifold learning method, the N fault modal components are fused to obtain the intrinsic manifold structure of the high-dimensional fault modal components, that is, the fault transient components.

The above-mentioned manifold fusion empirical mode decomposition method takes different values for the standard deviation of the random white noise added to the analysis signal each time, and uses the excellent feature mining ability of manifold learning to extract fault modal components with The transient component of the stable structure is removed, the in-band noise and inherent noise without a stable structure are removed, and the non-fault component caused by the modal aliasing problem caused by the improper noise intensity is removed, and the fault transient component in the signal is realized. effective detection. This technical method has at least the following advantages: there is no need to determine the standard deviation of the introduced noise, there is no modal aliasing problem, the in-band self-noise can be removed, and a higher signal-to-noise ratio can be obtained.

In another embodiment, "add random white noise with mean value 0 and standard deviation σ to the analysis signal to obtain a noise-added signal;", the standard deviation σ is 0.01 to 0.01 to 0.01% of the standard deviation of the analysis signal 1 times.

In another embodiment, "the EMD processing is performed on the noise-added signal, and an IMF containing fault information is obtained, that is, a failure mode component;" the failure mode component is obtained from a plurality of IMFs obtained by EMD processing Selected according to the given failure mode determination method.

In another embodiment, the method for determining a given failure mode includes but is not limited to using kurtosis, smoothness factor, sparse value, correlation coefficient, energy and their combination to select from multiple IMFs obtained by EMD The method of extracting the failure mode components.

In another embodiment, "the N fault modal components are fused according to a given manifold learning method to obtain the intrinsic manifold structure of the high-dimensional fault modal components, that is, the fault transient components.", The given manifold learning method includes but not limited to local tangent space alignment algorithm, isometric mapping algorithm, local linear embedding algorithm, Laplacian feature mapping algorithm or locality-preserving projection algorithm.

A computer device includes a memory, a processor, and a computer program stored on the memory and operable on the processor, and the steps of the method are implemented when the processor executes the program.

A computer-readable storage medium stores a computer program thereon, and implements the steps of the method when the program is executed by a processor.

A processor, the processor is used to run a program, wherein the method is executed when the program runs.

Description of drawings

FIG. 1 is a flow chart of a manifold fusion empirical mode decomposition method disclosed in an embodiment of the present invention.

Fig. 2 is a time-domain waveform diagram of a bearing sound vibration signal provided by an embodiment of the present invention.

Fig. 3 is the fault transient component obtained after processing the signal described in Fig. 2 by using the EEMD method.

FIG. 4 shows the fault transient components obtained after processing the signal shown in FIG. 2 using the technology disclosed in the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

A manifold fusion empirical mode decomposition method, comprising:

Add random white noise with a mean of 0 and a standard deviation of σ to the analysis signal to obtain a noise-added signal;

EMD processing is performed on the noise-added signal to obtain an IMF containing fault information, that is, a fault modal component;

Change the value of σ and repeat the above steps N times to obtain N failure mode components with different noise intensities, where N is a positive integer;

According to a given manifold learning method, the N fault modal components are fused to obtain the intrinsic manifold structure of the high-dimensional fault modal components, that is, the fault transient components.

The above-mentioned manifold fusion empirical mode decomposition method takes different values for the standard deviation of the random white noise added to the analysis signal each time, and uses the excellent feature mining ability of manifold learning to extract fault modal components with The transient component of the stable structure is removed, the in-band noise and inherent noise without a stable structure are removed, and the non-fault component caused by the modal aliasing problem caused by the improper noise intensity is removed, and the fault transient component in the signal is realized. effective detection. This technical method has at least the following advantages: there is no need to determine the standard deviation of the introduced noise, there is no modal aliasing problem, the in-band self-noise can be removed, and a higher signal-to-noise ratio can be obtained.

In another embodiment, "add random white noise with mean value 0 and standard deviation σ to the analysis signal to obtain a noise-added signal;", the standard deviation σ is 0.01 to 0.01 to 0.01% of the standard deviation of the analysis signal 1 times.

In another embodiment, "the EMD processing is performed on the noise-added signal, and an IMF containing fault information is obtained, that is, a failure mode component;" the failure mode component is obtained from a plurality of IMFs obtained by EMD processing Selected according to the given failure mode determination method.

In another embodiment, the method for determining a given failure mode includes but is not limited to using kurtosis, smoothness factor, sparse value, correlation coefficient, energy and their combination to select from multiple IMFs obtained by EMD The method of extracting the failure mode components.

In another embodiment, "the N fault modal components are fused according to a given manifold learning method to obtain the intrinsic manifold structure of the high-dimensional fault modal components, that is, the fault transient components.", The given manifold learning method includes but not limited to local tangent space alignment algorithm, isometric mapping algorithm, local linear embedding algorithm, Laplacian feature mapping algorithm or locality-preserving projection algorithm.

It can be understood that the given manifold learning method may also be other methods with a dimension reduction function.

A computer device includes a memory, a processor, and a computer program stored on the memory and operable on the processor, and the steps of the method are implemented when the processor executes the program.

A computer-readable storage medium stores a computer program thereon, and implements the steps of the method when the program is executed by a processor.

A processor, the processor is used to run a program, wherein the method is executed when the program runs.

A specific application scenario of the present invention is introduced below:

It can be seen from the background art that the existing EEMD method uses ensemble averaging to remove the introduced noise in IMFs. However, finite ensemble averaging cannot completely remove the introduced noise, nor can it remove the in-band self-noise, let alone solve the problem of modal aliasing caused by improper noise intensity.

Therefore, the invention discloses a manifold fusion empirical mode decomposition method and a signal transient component detection device based on the method. The method adopts manifold learning to fuse multiple fault mode components with different noise intensities, and extracts fault transient components in the signal from them. Due to the excellent feature mining ability of manifold learning, it can extract the inherent manifold structure of fault transient components, thereby removing the in-band noise, self-noise, and modal aliasing caused by improper noise intensity. The non-fault component of the signal realizes the effective detection of the fault transient component in the signal.

According to the above-mentioned content of the invention and the flow chart of the manifold fusion empirical mode decomposition method and the signal transient component detection device based on the method of accompanying drawing 1, the technology specifically includes:

Step 101: adding random white noise with a mean value of 0 and a standard deviation of σ to the analysis signal to obtain a noise-added signal;

Neither too large nor too small a value of σ can well solve the mode aliasing problem of EMD, and the optimal value of σ will change due to the difference of the analysis signal. In order to avoid discussing the optimal value of σ, the method of the present invention takes different values for σ, the range of which is 0.01 to 1 times the standard deviation of the analysis signal, and σ takes a uniform value within this range.

Step 102: Perform EMD processing on the noise-added signal to obtain an IMF containing fault information, that is, a fault modal component;

After EMD processing, the fault transient component mainly exists in the fault mode component, so the method of the present invention only performs further processing on the fault mode component. The failure mode component is selected from a plurality of IMFs obtained by EMD processing according to a given failure mode determination method.

The method for determining a given failure mode includes but is not limited to using kurtosis, smoothness factor, sparse value, correlation coefficient, energy and their combination to select the failure mode component from a plurality of IMFs obtained by EMD method.

Step 103: Change the value of σ, repeat the above steps N times, and obtain N failure mode components with different noise intensities;

Since the method of the present invention does not use statistical properties to remove introduced noise, the number of repetitions N does not need to be large. In order to reduce the amount of calculation, the value of N is generally 10.

Step 104: Fusing the N fault modal components according to a given manifold learning method to obtain the intrinsic manifold structure of the high-dimensional fault modal components, ie, the fault transient components.

Manifold learning is a dimensionality reduction and data mining method that can be used to extract the intrinsic low-dimensional manifold structure embedded in high-dimensional data. The N failure mode components may form N-dimensional data. Among them, the fault transient component exists in each dimensional data, has a stable structure, can be regarded as the manifold structure of N-dimensional data, and will be preserved in the manifold learning results; while other components, including the introduction of noise, Self-noise and non-fault components caused by modal aliasing caused by improper noise intensity are different in each dimensional data, do not have a stable structure, and will be eliminated in the manifold learning results. Therefore, after the high-dimensional fault modal components are fused by a given manifold learning method, a fault transient component with a high signal-to-noise ratio can be obtained.

The given manifold learning method includes but not limited to local tangent space alignment algorithm, isometric mapping algorithm, local linear embedding algorithm, Laplacian feature mapping algorithm, local preserving projection algorithm and other methods with dimension reduction function.

In order to understand the technical solution of the present invention and its effect more clearly, a specific embodiment will be described in detail below.

Take the early weak fault detection of the bearing as an example. The bearing model is N306E, and the inner ring of the bearing is driven by a motor at a speed of 1464.6rpm. A sound pressure sensor is fixed near the bearing to collect the sound vibration signal of the bearing, and the sampling frequency is 20kHz.

First, the main fault characteristic period is calculated according to the rotation speed of the inner ring of the bearing and the geometric dimensions of the bearing, and the results are shown in Table 1.

Table 1: Bearing fault characteristic cycle

Inner ring fault characteristic cycle T _i =0.0068s Outer ring fault characteristic cycle T _o =0.0102s Rolling element failure characteristic cycle Tb = _0.0085s

Referring to accompanying drawing 2, Fig. 2 is the time-domain waveform diagram of the bearing sound vibration signal provided by the embodiment of the present invention. Some transient pulse components can be observed from the waveform diagram, but there are also some transient pulse components submerged in noise that are difficult to identify, so a bearing fault characteristic period cannot be found from the diagram.

The EEMD method is used to process the signal described in Figure 2, and the standard deviation of the introduced noise is taken as a common empirical value, that is, 0.2 times the standard deviation of the original signal, and the set average number is 200. Figure 3 shows the failure modal component, the fault transient component, in the processing results. Periodic transient pulse components can be observed from this figure, but there is still a large amount of noise that has not been removed between each pulse, so the signal-to-noise ratio of the fault transient components obtained is not high.

The technology disclosed in the present invention is used to process the signal described in Figure 2, the given fault mode determination method is a method combining the time domain and frequency domain kurtosis values of each modal component, and the given manifold learning method is a local tangent space permutation algorithm. Figure 4 shows the processing results. The noise between the transient pulses in the figure is almost completely removed, making the periodicity of the pulses more obvious. The average interval of the pulses is calculated to be 0.0068s, which is the same as the characteristic period of the bearing inner ring failure in Table 1, so it can be determined that the test bearing The inner ring is defective. In fact, before carrying out the experiment, a crack defect has been artificially set in the inner ring of the test bearing with a defect width of 0.5 mm. Therefore, the technology disclosed in the present invention can accurately detect the transient components of bearing faults from the noise-containing vibration signals of the bearings.

To sum up, by introducing noises of different intensities to the bearing sound vibration signal, and then performing EMD processing separately, and finally using manifold learning to fuse the obtained high-dimensional fault modal components, various fault modal components in the fault modal components can be removed. Noise components, so as to effectively detect the transient components of bearing faults. This method overcomes the problems that the existing EEMD technology is difficult to determine the intensity of the introduced noise and remove the inherent noise in the band, and can extract the transient components submerged by the noise, which is of great significance to the detection of the transient components of the signal under the background of strong noise.

The technical features of the above-mentioned embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, should be considered as within the scope of this specification.

The above-mentioned embodiments only express several implementation modes of the present invention, and the descriptions thereof are relatively specific and detailed, but should not be construed as limiting the patent scope of the invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention should be based on the appended claims.

Claims (8)

1. A manifold fusion empirical mode decomposition method, characterized in that, comprising: Adding random white noise with a mean value of 0 and a standard deviation of σ to the analysis signal to obtain a noise-added signal; EMD processing is performed on the noise-added signal to obtain an IMF containing fault information, that is, a fault modal component; Change the value of σ and repeat the above steps N times to obtain N failure mode components with different noise intensities, where N is a positive integer; According to a given manifold learning method, the N fault modal components are fused to obtain the intrinsic manifold structure of the high-dimensional fault modal components, that is, the fault transient components. 2. The manifold fusion empirical mode decomposition method according to claim 1 is characterized in that, "in the analysis signal, add random white noise with a mean value of 0 and a standard deviation of σ to obtain a noise-added signal;" in, the The standard deviation σ is 0.01 to 1 times the standard deviation of the analysis signal. 3. manifold fusion empirical mode decomposition method according to claim 1, is characterized in that, " carry out EMD processing to described adding noise signal, obtain an IMF that contains fault information, i.e. fault modal component; " described The failure mode components are selected from multiple IMFs obtained by EMD processing according to a given failure mode determination method. 4. The manifold fusion empirical mode decomposition method according to claim 3, wherein said given fault mode determination method includes but not limited to using kurtosis, smooth factor, sparse value, correlation coefficient, energy and Their combination and the like can select the failure mode components from multiple IMFs obtained by EMD. 5. manifold fusion empirical mode decomposition method according to claim 1, is characterized in that, "according to given manifold learning method, described N fault mode components are fused, obtain the high-dimensional fault mode component Intrinsic manifold structure, that is, fault transient components." In, the given manifold learning method includes but not limited to local tangent space arrangement algorithm, isometric mapping algorithm, local linear embedding algorithm, Laplacian eigenmap algorithm or Locality-preserving projection algorithm. 6. A computer device, comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, characterized in that, when the processor executes the program, any one of claims 1 to 5 is realized The steps of the method. 7. A computer-readable storage medium, on which a computer program is stored, wherein, when the program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are realized. 8. A processor, wherein the processor is used to run a program, wherein the method according to any one of claims 1 to 5 is executed when the program runs.