1. Introduction
The rolling bearing is one of the most common components in rotating machinery, and it is also one of the components with the highest failure rate. The running condition of bearings is directly related to the safe and stable operation of the whole rotating machinery system. Extracting fault feature information from bearing vibration signals under strong background noise is a hot issue in the field of mechanical fault diagnosis. The process of bearing fault diagnosis is divided into two parts: vibration parameter (displacement, velocity and acceleration) acquisition and vibration data processing. Adamczak et al. analyzed the existing bearing fault testing system. They strived for more accurate bearing fault measurement equipment and fault measurement parameters by comparing and analyzing existing measuring devices [
1]. Meanwhile, a large number of vibration signal processing methods have been put forward, among which the kurtogram method is able to adaptively recognize the resonance frequency band of the vibration signal. By filtering the resonance frequency band and integrating the spectrum analysis of filtered signals, the fault characteristics of vibration signals can be effectively extracted.
Kurtosis is the 4th-order cumulant of random variables. This index cannot reflect the change of specific signals, and is not suitable for condition monitoring under strong noise environments. In order to overcome the shortage of kurtosis in engineering applications, Antoni made a formal definition of kurtosis [
2]. Subsequently, Antoni further proposed the kurtogram. The core idea of the algorithm is that the spectral kurtosis index of the unstable signal is maximized by reasonable selection of frequency resolution [
3]. In order to reduce the operation time of the kurtogram and enable it to be widely used in engineering practice, Antoni further proposed a fast spectral kurtosis (FSK) calculation method based on a fast algorithm for computing the kurtogram over a grid. FSK can simplify the process of selecting the optimal band
[
4]. In the FSK algorithm, the horizontal axis represents the frequency, the vertical axis represents the number of decomposition layers, and the depth of the color indicates the spectral kurtosis value of each sub band.
Since then, many algorithms based on the kurtogram have been proposed. Lei et al. proposed an improved kurtogram method [
5], which uses wavelet packet transform (WPT) instead of the short time Fourier transform (STFT) and FIR filters used in the traditional kurtogram. A filter based on WPT can accurately divide the frequency bands and control the noise effectively [
5]. Then, on the basis of Lei et al., Wang et al. made further improvements. The kurtosis calculation method in the traditional kurtogram was changed to calculate the power spectrum of the envelope of different nodes of WPT. The power spectrum of the envelope signal reflects the sparsity of the signal and helps to capture the resonance band in the signal [
6]. Based on the same idea, Tse and Wang proposed a new method called sparsogram. The power spectrum of the envelope was applied to improve the kurtosis index in the kurtogram, so that it reflected the sparsity of signals [
6,
7]. Gu proposed a method for analyzing kurtograms in the frequency domain [
8]. The correlated kurtosis of the envelope was used to replace the traditional calculating method in the kurtogram, which makes the calculation index sensitive to periodic impulses and has better robustness. To obtain the correct relationship between the node and frequency band in WPT, a vital process called frequency ordering is conducted to solve the frequency folding problem due to down-sampling. Tian et al. proposed the modulation signal bispectrum (MSB)-based robust detector to reduce the noise and accurately identify the resonance frequency band [
9]. The high-magnitude features that result from the use of the MSB also enhance the modulation effects of a bearing fault. In gearboxes, the kurtogram method would fail when gears and bearings partially malfunctioned at the same time. In view of this, Wang proposed an improved method by combining kurtogram and menshing resonance (MRgram) [
10]. Random impulses often occur in the frequency bands of vibration signals of bearings, making the selection of kurtogram resonance frequency bands inaccurate. Xu et al. proposed a periodicity-based kurtogram to deal with random impulse resistance [
11]. The periodic component to aperiodic component ratio (PAR) was utilized in this method to differentiate the types of impulses. Li et al. improved the kurtogram based on an impulse step dictionary and a reweighted minimizing nonconvex penalty Lq regular for rolling bearing fault diagnosis [
12].
Entropy was originally a thermodynamic concept describing the regularity of information. Because entropy has some statistical properties superior to kurtosis in some respects, the application scope of entropy was gradually extended to the field of mechanical fault diagnosis. Antoni introduced Shannon entropy into the kurtogram algorithm and proposed infogram [
13]. Infogram contains three parts, the envelope (SE) infogram, squared envelope spectrum (SES) infogram, and SE
1/2/SES
1/2 infogram. The calculation index of the SE infogram is the negative entropy of the square envelope, and the calculation index of the SES infogram is the negative entropy of the square envelope spectrum. On the basis of infogram, the multiscale clustering grey infogram (MCGI) was proposed by Li et al. [
14]. Hemmati et al. proposed an index that combines kurtosis and Shannon entropy, and the index was used to obtain the optimal band pass filter utilizing wavelet packet transform (WPT) and envelope detection [
15]. The MCGI combined the negentropies of the time and frequency domains in a grey fashion using multiscale clustering. Chen et al. proposed a fault classification method combining time-spectral kurtosis, entropy and support vector machines (SVM). Combining the advantages of time-spectral kurtosis (T-SK) and entropy, this method makes fault extraction more obvious [
16].
To sum up, the improvement of kurtogram has mainly focused on two aspects. On the one hand, the decomposition methods (STFT, FIR) of the traditional kurtogram algorithm have been improved, so that the frequency band segmentation will have better results. On the other hand, the calculation index of each sub-band has been improved. The aim is to make the calculation index sensitive to the periodic impact caused by the fault, and to have strong robustness to the interference caused by noise or accidental impact. These improvements have made a significant contribution in some respects. In order to further improve the accuracy and efficiency of the algorithm in determining the center frequency and bandwidth, a new adaptive method called TEERgram is proposed in this article.
2. The Index of Teager Energy Entropy
In 2016, Antoni proposed the infogram method. Infogram theories include the squared envelope (SE) infogram, the squared envelope spectrum (SES) infogram, and the SE
1/2/SES
1/2 infogram. The SE and SES were defined as Equations (1) and (2).
where
,
is the frequency, and
is the frequency resolution.
and
are entropy values in the time domain and entropy in the frequency domain respectively. That is, they were used to represent the impact characteristics and cyclostationarity characteristics of fault signals, respectively. Antoni further proposed the SE
1/2/SES
1/2 infogram to characterize the impact characteristics and cyclostationarity. The SE
1/2/SES
1/2 infogram was defined by Equation (3):
2.1. Teager Energy Operator
Teager energy operator (TEO) can enhance transient impact components, which is suitable for detecting impact characteristics in signal, and has good effect in fault impact feature extraction. For signal
, the Teager energy operator
can be defined as Equation (4):
where
and
are the first- and second-order derivatives of signal
x relative to time
t. The general expression of the amplitude modulation frequency modulation signal is shown in Equation (5):
where
a(t) is the instantaneous amplitude of
and
is the instantaneous phase of
. The instantaneous frequency of the signal
is
. The effect of the energy operator on
is shown in Equation (6):
Because the change of the carrier signal is much faster than the modulated signal, the instantaneous amplitude and the instantaneous frequency of the modulated signal are relatively slow relative to the high-frequency carrier, and can be approximately equal to the constant. Therefore, if
,
, then the Formula (6) can be approximately expressed as Equation (7).
Therefore, the envelope signal (instantaneous amplitude)
obtained by demodulation and separation of signal s can be expressed as Equation (9):
The signal energy in transmission is defined as the square of the amplitude of the signal. If the impact amplitude is small, the impact component may be drowned by other components. The Teager energy operator is the product of squared instantaneous amplitude and instantaneous frequency. Due to the high vibration frequency of transient impact, the Teager energy operator can effectively enhance transient impact components and significantly suppress noise.
In order to verify the superiority of the Teager energy operator spectrum (TEOS) compared with the envelope spectrum (ES) and the square envelope spectrum (SES), the following simulation signals are defined as Equation (10).
where
ξ is the damping ratio,
is the amplitude,
(
),
is the natural frequency,
= 25 Hz is the characteristic frequency and
is the Gauss white noise. The parameters of the simulation signal are shown in
Table 1.
The waveforms of the synthetic signal with fault impulses and noise are shown in
Figure 1,
Figure 1a shows a periodic signal, which is used to simulate the fault impulses.
Figure 1b shows a synthetic signal with fault impulses and noise (–12 dB). The envelope spectrum (ES), the square envelope spectrum (SES) and the Teager energy operator spectrum (TEOS) of the simulation signal are analyzed, respectively. The results are shown in
Figure 2. It can be seen from
Figure 2 that the amplitude of TEOS is higher than the amplitude of ES and the amplitude of SES at the fault characteristic frequency spectrum line. However, in the rest of the spectrum, the result is just the opposite. That is to say, the signal-to-noise ratio (SNR) of the Teager energy operator spectrum is much higher than that of the square envelope spectrum and the envelope spectrum.
2.2. Shannon Entropy
In 1948, Shannon put forward the concept of “Shannon entropy” and solved the problem of quantitative measurement of information. There is a direct relationship between the amount of information in a signal and its uncertainty, so from this point of view, we can think that the measure of the amount of information is equal to how much uncertainty there is. Shannon entropy reflects the degree of disordering (ordering) of a system, the more orderly a system is, the lower the entropy, and vice versa. In view of this, Shannon entropy can reflect the order of vibration signals. Given a random sequence
, the calculation method of Shannon entropy is shown in Equation (11):
where
is the value of Shannon entropy,
is the probability mass associated with the value
and
is an arbitrary positive constant that dictates the units.
The Shannon entropy index can reflect the regularity of the vibration signal, and its value has a certain relation with the SNR, so it can be used as an indicator to reflect the intensity of the impact of the fault. To verify this conclusion, the simulation signal model in Equation (10) is added with different degrees of noise to simulate the numerical changes of Shannon under different SNR. It can be seen from
Figure 3 that the numerical regularity of Shannon entropy decreases with the increase of SNR.
2.3. Teager Energy Entropy
In view of the superiority of the Teager energy operator relative to the square envelope spectrum, we propose a Teager Energy Entropy (TEE) index. Inspired by infogram, we define the Teager energy entropy in the time domain (TEEt) and the Teager energy entropy in the frequency domain (TEEf) respectively as Equations (12) and (13).
In order to optimize the proportion of time domain and frequency domain components, an adaptive weighting method is proposed. Firstly, the regularization of TEEt and TEEf is carried out using Equations (14) and (15).
In these,
m = 1, 2, …,
M (
n = 1, 2, …,
N) represents the
M(
N)-th signal component. Then the mean value (
) and standard deviation (
) of the two regularization sequences are calculated separately using Equations (16)–(19).
Then the weight coefficients of the regularized sequence are calculated as Equations (20) and (21).
The Teager energy entropy, combined with the time domain and the frequency domain, is finally expressed as Equation (22):
2.4. Teager Energy Entropy Ratio of Wavelet Packet
Lei et al. introduced wavelet packet transform (WPT) into the kurtogram algorithm to replace the traditional decomposition method (FIR, STFT) in kurtogram. Meanwhile, Stępień et al. confirmed that the entropy-based WPT is the most suitable approach that is not limited to one level of decomposition, and it allows us to find an optimal decomposition tree [
17]. Inspired by this idea, the algorithm is further improved by the TEE index in this paper. The paving of the kurtogram and the TEEgram are shown in
Figure 4.
In many cases, the components that reflect the characteristics of fault impact are likely to be drowned out by noise. This can easily cause misdiagnosis during the confirmation of the resonance frequency band. To illustrate these problems, different degrees of noise are added to the simulated signals in Equation (10).
Figure 5a shows the simulated fault signal with SNR = −12 dB,
Figure 5b shows the simulated fault signal with SNR = −16 dB,
Figure 5c shows the TEEgram of (a), while
Figure 5d shows the TEEgram of (b). When the signal-to-noise ratio (SNR) is −12 dB, the TEEgram can accurately locate the resonance frequency band (
Table 1; the natural frequency is about 1500 Hz), but when the SNR is reduced to −16 dB, the resonance frequency band is inaccurate.
In view of the defect that kurtogram is too sensitive to strong background noise and singular points, Wang et al. proposed an improved method named SKRgram [
18]. Subsequently, Miao et al. used the Gini coefficient as an index to further improve the SKRgram [
19]. The essence of the SKRgram is to compare the change of spectral kurtosis before and after failure. Based on this, the potential bands of fault impact hidden in noise and other constant interference components can be found. Inspired by this idea, we further modify the TEE index as Equation (23):
2.5. Process of the Proposed Method
The bearing vibration signals are collected under healthy state and fault state conditions (
Figure 6a,b). According to the bearing parameters, the characteristic frequency of bearing fault is calculated based on the theoretical equations (Equation (24)).
where
,
,
and
respectively represent inner ring fault, outer ring fault, rolling element fault and cage fault.
is the ball diameter,
is the pitch diameter,
is the contact angle,
is the speed of rotation,
is the number of the balls.
Wavelet packet transform (WPT) is applied, in turn, to vibration signals in healthy state and fault state. The time domain Teager energy entropy (
) and the frequency domain Teager energy entropy (
) are calculated for each sub-band of the signal under the healthy condition.
and
are weighted to obtain Teager energy entropy (
) in the healthy state. Similarly, the energy entropy of Teager (
) under the fault condition is obtained. Teager energy entropy diagram (TEEgram) of wavelet sub-band of health status and fault state signals is constructed respectively (
Figure 6c,d).
According to the Teager energy entropy of the signal in the healthy state (
) and the Teager energy entropy of the signal in the fault state (
), the Teager energy entropy rate (
) is calculated. Teager energy entropy ratio diagram (TEERgram) of wavelet sub-band is constructed as shown in
Figure 6e. It can be seen from
Figure 6e that the location of the relative change of TEE value is the fault resonance frequency band. The wavelet subband with the largest TEER value is selected as the optimal band. The center frequency and bandwidth are extracted to set the parameters of the bandpass filter, and the filter is used to filter the selected wavelet sub-bands. Then the envelope spectrum of the filtered signal is analyzed (
Figure 6f). The fault type is analyzed by comparing the characteristic frequency of the spectrum line with the theoretical characteristic frequency calculated in step 1.
3. Simulation Analysis
The effectiveness of the proposed algorithm is verified by the mathematical model of the inner ring fault simulation signal in Equation (25) [
20].
where
is the ith shock; T is the period of shock;
is the rotating frequency;
is the amplitude modulation with
as a cycle;
is an exponential decay pulse;
is the attenuation coefficient of vibration caused by damping; and
,
represent the phase. The other parameters of the simulation signal are shown in
Table 2.
Figure 7a shows the waveform of the inner fault impulses;
Figure 7b shows the noise;
Figure 7c shows the synthetic signal with fault impulses and noise (−16 dB).
Figure 7d shows the spectrum of
Figure 7c. It can be seen from
Figure 7d that the fault feature information cannot be found, and there is a suspected resonance frequency band near 2500 Hz.
The simulation signal is respectively analyzed by kurtogram, kurtosis of the wavelet packet transform (KWPT) and the proposed method. The results of kurtogram analysis are shown in
Figure 8. The results of KWPT analysis are shown in
Figure 9. The two methods are inaccurate in locating the resonant frequency band, and are significantly affected by noise. In the envelope spectrum of their sub-bands, the characteristic frequencies of faults can be found.
Figure 10 is the method proposed in this paper. By calculating the Teager energy operator ratio (TEER) of the signal before and after the fault, the resonant frequency band that changes before and after the fault is obtained, that is, the resonance frequency band caused by the fault.