11806 IEEE SENSORS JOURNAL, VOL. 23, NO.

11, 1 JUNE 2023

Periodic Detection Mode Decomposition and Its

Application in Bearing Fault Diagnosis
Chaoyong Ma , Zhiqiang Yang , Yonggang Xu , Aijun Hu , and Kun Zhang

Abstract—In the complex background noise, it is difficult

to extract the periodic pulse of the rolling bearing fault signal.
In this article, a new modal decomposition method supported
by the singular value ratio (SVR) spectrum is proposed to find
the optimal period of bearing fault data, which can be named
periodic detection mode decomposition (PDMD). To detect
the intervals of periods and eliminate the interference of
useless periods, the initial measurement intervals in this
method are divided according to the theoretical fault periods
of different fault types of bearings. In each interval, the SVR
spectrum is used to detect the appropriate period and sup-
press the influence of noise on the recognition process. This
period is used to construct the optimal Ramanujan subspace
(RS). Finally, harmonic spectral kurtosis (HSK) is used to identify the extracted period as false information, interference,
or fault. Simulation and experimental data verify the effectiveness of the proposed method. This method can effectively
extract and screen periodic pulses and can successfully identify the outer and inner ring faults of bearings.
Index Terms— Periodic component (PC), periodic detection mode decomposition (PDMD), Ramanujan subspace (RS),
rolling bearing, singular value ratio (SVR) spectrum.

I. I NTRODUCTION noise [4], [5]. Therefore, the accurate extraction of periodic

OLLING bearings play a load-bearing role and are pulses in fault signals is the focus and difficulty of current
R widely used in transportation, aviation, metallurgy, elec-
tric power, machinery, and other fields. Rolling bearings often
research.
In recent years, various methods of signal periodicity
work in severe environments and are prone to failure. If the measurement and period decomposition have been proposed
bearing failure cannot be detected in time, it will lead to one after another. Periodic subspace theory is a hot topic
damage to the bearing and equipment, resulting in major in the fields of molecular biology, music, and speech pitch
safety accidents [1]. Equipment condition monitoring based detection [6], [7], [8]. Periodicity transform (PT) [9] gen-
on vibration signal processing is a popular method for bearing erates periodic components (PCs) by projecting the signal
fault diagnosis, which can detect bearing damage in time into different periodic subspaces. However, the period defined
and avoid serious losses [2], [3]. When a rolling bearing by this method is confusing, and the results are affected by
is damaged, the fault signal will show the characteristics the order of extraction. The exact periodic subspace decom-
of a periodic shock, which is usually submerged in strong position (EPSD) [10] determines the energy of the projec-
tion on each orthogonal subspace by maximum likelihood
Manuscript received 20 December 2022; revised 8 March 2023; estimation and thus quickly determines the principal PCs of
accepted 5 April 2023. Date of publication 12 April 2023; date of current the signal. Although the confusion problem of the period
version 31 May 2023. This work was supported by the National Natural
Science Foundation of China under Grant 51775005. The associate edi- is solved, there is an overlapping problem between periodic
tor coordinating the review of this article and approving it for publication subspaces. Subsequently, in view of the overlapping problem
was Prof. Weihua Li. (Corresponding author: Yonggang Xu.) of traditional periodic subspaces, Vaidyanathan [11], [12]
Chaoyong Ma, Zhiqiang Yang, and Kun Zhang are with the Depart-
ment of Materials and Manufacturing, Beijing University of Tech- introduced the Ramanujan subspace (RS) by the Ramanujan
nology, Beijing 100124, China (e-mail: [email protected]; sum sequence [13] and proposed the Ramanujan PT (RPT).
[email protected]; [email protected]). The correlation properties of RSs are discussed in detail,
Yonggang Xu is with the Beijing Engineering Research Center of
Precision Measurement Technology and Instruments, Beijing University and it is proved that any signal can be decomposed into the
of Technology, Beijing 100124, China (e-mail: [email protected]). sum of orthogonal projections on finite RSs. However, this
Aijun Hu is with the Department of Mechanical Engineering, North method needs to satisfy that the length of the signal is a
China Electric Power University, Baoding 071003, China (e-mail:
[email protected]). multiple of the subperiod. Only some PCs can be identified
Digital Object Identifier 10.1109/JSEN.2023.3265377 and extracted. Therefore, the orthogonal RS pursuit (RSP)
1558-1748 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on June 12,2023 at 12:36:30 UTC from IEEE Xplore. Restrictions apply.
MA et al.: PDMD AND ITS APPLICATION IN BEARING FAULT DIAGNOSIS 11807

method was proposed by Deng and Han [14]. RSP iteratively

constructs RSs, and the main PCs are selected by the energy
of each component. Meanwhile, in order to improve the
recognition of hidden periods, the conjugate subspace pursuit
with periodicity and energy (CSPPE) method based on dual
strategy measurement was proposed by Deng and Han [15].
This method has the characteristics of time-varying periodicity
and high computational efficiency. In view of the lack of
self-adaptability of RSP and CSPPE methods, an adaptive
periodic mode decomposition (APMD) method based on RSs
was proposed by Cheng et al. [16]. The APMD determines the
main periods through energy and adaptively decomposes the
corresponding main PCs, which is applied to bearing feature
extraction. However, the core of its periodicity measurement
is based on the autocorrelation function, and the accuracy is
greatly disturbed by noise. In addition, the wrong periodicity
measurement may lead to a large number of false components,
and it is limited to select periodic pulses by the maximum
correlation with the original signal.
In this article, a periodic detection mode decomposition
(PDMD) assisted by the singular value ratio (SVR) spec-
trum [17], [18] to find the optimal period of bearing fault
data is proposed. First, based on the theoretical fault periods
of different fault types of bearings, the initial measurement
intervals are divided to detect the periods and eliminate the
interference of useless periods. Second, the SVR spectrum
is used to detect the appropriate period in each interval and
construct the optimal RS. The influence of noise on the
recognition process is greatly suppressed. Finally, harmonic
spectral kurtosis (HSK) [19] is applied to identify the extracted
period as false information, interference, or fault. Simulation
and experimental results show that this method can effectively
extract and filter periodic pulse components to realize fault
diagnosis. The measurement accuracy of decimal step size Fig. 1. Flowchart of the OPMD method.
can be achieved in the aspect of periodic measurement. The
rest of this article is organized as follows. In Section II, the
PDMD algorithm is described in detail. The simulation signal 3) Calculate the orthogonal projection on the RS according
is applied to verify the superiority of PDMD in Section III. to the optimal period as well as extract and reconstruct
The PDMD is applied to bearing fault diagnosis in Section IV. the suspected fault component.
The conclusion is arranged in Section V. 4) The HSK is used to identify the faulty component and
II. P ERIODIC D ETECTION M ODE D ECOMPOSITION obtain the component containing fault information.
When the bearing of rotating machinery fails, the vibration
signal has quasi-periodic characteristics. But it is usually
affected by complex environments, and the periodicity of A. Optimal Period Screening and Modal Extraction
the time domain is not obvious. In this article, the periodic Based on SVR Spectrum and RS
information that may be the fault in the intervals is detected RS theory has an excellent ability to extract PCs. For
by the SVR spectrum, and the RSs are constructed to extract periodic signal x (t) = x (t + T ), whether x (t) is a sinu-
and reconstruct the suspected fault components. The proposed soidal signal or a harmonic signal, RS can well reveal the
PDMD can not only accurately extract the faulty PCs, but composition of PCs in the signal. RS theory constructs dif-
also reduce the false components and suppress the interference ferent orthogonal projection matrices through different peri-
of noise. The flowchart of the proposed PDMD algorithm is ods to extract the corresponding PCs. Therefore, it is very
shown in Fig. 1. important to accurately measure the period of the signal.
1) Collect the bearing vibration acceleration signal and set In this article, the periodic information of possible faults in
the initial period detection intervals according to the the divided intervals is detected by the SVR spectrum, and
fault characteristic frequency of the bearing inner ring, the RSs are constructed to reconstruct the suspected fault
outer ring, cage, and ball. components.
2) SVR spectrum is used to detect the periods in each 1) The periodic information of the signal is detected by the
interval and extract its optimal period. SVR spectrum.
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on June 12,2023 at 12:36:30 UTC from IEEE Xplore. Restrictions apply.
11808 IEEE SENSORS JOURNAL, VOL. 23, NO. 11, 1 JUNE 2023

For signal x ∈ R N of any length, the process of periodic The Ramanujan sum sequence is defined as follows:
mode extraction is as follows. Tk
First, the initial period detection intervals
X
cTk (n) = e j2πln/Tk (5)
[round( f s /( f k + 3)), round( f s /( f k − 3))] is set
l=1
according to the fault characteristic frequency f k (l, Tk ) = 1
(k = 1, 2, 3, 4) of the bearing inner ring, outer ring,
cage, and ball to reduce the interference of irrelevant where (l, Tk ) = 1 means l and Tk is prime to each other.
areas. Where round (·) is a four-in-five function, f s is RS STk is defined as the column space of an integer
the sampling frequency. period matrix BTk
Second, the signal x is continuously truncated according 
cTk (0) cTk (Tk − 1) · · · cTk (1)

to the length L of each row, which is defined as the  cTk (1) cTk (0) ··· cTk (2) 
decomposition matrix A ∈ R M×L  cTk (2) cTk (1) cTk (3) 
 
···
BTk =  .. .. .. .. .
   
x [1] x [2] · · · x [L]  . . . . 
 x [L + 1] x [L + 2] · · · x [2L]   cTk (Tk − 2) cTk (Tk − 3) · · · cTk (Tk − 1) 
 
A= 
 : : : :  cTk (Tk − 1) cTk (Tk − 2) · · · cTk (0)
x [(M −1) L +1] x[(M −1)L +2)] · · · x [M L]
(6)
(1)
The orthogonal projection matrix PTk ∈ R Tk ×Tk for
where M = floor (N /L), L belongs to the detection
extracting PCs is expressed as
intervals, and floor (·) is the downward rounding func-
tion. However, since the actual fault period (data point) BTk
PTk = . (7)
of the bearing is usually not an integer, the construction Tk
of matrix A will produce the accumulation of rounding 3) Calculate the projection x Tk ∈ R N on RS STk .
errors, resulting in the failure of period detection. There- The projection matrix PTk is extended periodically
fore, to avoid the accumulation of rounding errors, [18]  
improved the construction of decomposition matrix A, PTk · · · PTk
which is defined as follows: Z Tk =  ... ..
.
.. 
.  (8)

 
x [1] x [2] ··· x [L] PTk ··· PTk
 x [l1 + 1] x [l1 + 2] · · · x [l1 + L] 
Aτ =  
where Z Tk ∈ R (M×T k )×(M×T k ) and M = floor (N /Tk ).
 :   :  :  :
 
Let signal x ′ ∈ R M×Tk be expressed as follows:

x lm−1 + 1 x lm−1 + 2 · · · x lm−1 + L
(2) x ′ = [x1 , x2 , . . . , x M×Tk ]T . (9)

 L = round(l)
 Then the projection of the signal x ′ is calculated as
lk = round(kl) (3) follows:
m = floor(N /l)


x T′ k = Z Tk x ′ . (10)
where l is a positive real number, and l is increased by
0.1 in each detection interval to search. The PC of any signal x ∈ R N is expressed as
Then singular value decomposition is performed on  ′ 
x Tk
the decomposition matrix Aτ . There must be orthog- x Tk =
x T∗k
(11)
onal matrices U = (u 1 , u 2 , . . . , u m ) ∈ R m×m and
V = (v1 , v2 , . . . , v L ) ∈ R L×L satisfying the following where x T∗k denotes the first N − M×T k sequences
equation: of x T′ k .
The optimal period in each interval is detected through the
Aτ = U SV T (4)
above steps, and four PCs are extracted. In the SVR spectrum,
where S = (diag (σ1 , σ2 , . . . , σa ) , 0) or its transpose, matrix A is constructed by continuously truncating the signal
diag (·) is a diagonal matrix, a = min (m, L), and the as each row of the matrix. Construct a sinusoidal signal with
singular value σ1 ≥ σ2 ≥ · · · ≥ σa ≥ 0. a period of 13.6 sampling intervals, and the green line in
Finally, the principal SVR is calculated using σ1 /σ2 . Fig. 2 is the starting point of each row of matrix A. The
A series of principal SVRs are obtained by constantly interval between the purple line and the green line represents
changing l in matrix Aτ , and the SVR spectrum is the starting point error of each row of matrix A (see Fig. 2).
drawn. The length L corresponding to the maximum It can be seen that the starting point error of the previous data
principal SVR in each interval is defined as the optimal segment will continue to accumulate backward, which may
periodic information Tk (k = 1, 2, 3, 4) of the suspected lead to the failure of the result.
fault. In order to avoid the accumulation of starting point error,
2) The RS STk for extracting PCs is constructed according matrix Aτ no longer continuously truncates the data. Matrix
to the optimal periodic information Tk . Aτ redefines the starting point position of each row to ensure

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on June 12,2023 at 12:36:30 UTC from IEEE Xplore. Restrictions apply.
MA et al.: PDMD AND ITS APPLICATION IN BEARING FAULT DIAGNOSIS 11809

Antoni [20] defined the Wold–Cramer decomposition spec-

trum of nonstationary signals
Z +∞
y(t) = e j2π f t P(t, f )d X ( f ) (12)
−∞

where P(t, f ) is the time–frequency complex envelope of

y(t).
Cyclostationary random process is a special nonstationary
Fig. 2. Rounding error caused by different decomposition matrices. random process, and its statistical characteristics show periodic
and stable changes
X
g (t, s) = g (t + T, s) = gk (s)e j2πkt/T . (13)
k

Similarly, g0 (t, 1t) is described as a periodic pulse capture

function in the nt local region (−1t, 1t) of g (t, s)
X
g0 (t, 1t) = gn (nt)e j2πkt/T . (14)
n
Fig. 3. SVR spectrum of (a) matrix A and (b) matrix Aτ .
In order to measure the local energy intensity of the
time–frequency complex envelope, the second-order harmonic
instantaneous moment is defined as follows:
that the starting point error of each row is less than half
the sampling interval, as shown in the interval between the E{|Pn (t, f )d X ( f ) |2 }
S2H Y (t, f ) = = |Pn (t, f )|2 · S2X .
purple line and the red line in Fig. 2. Matrix Aτ considers df
the influence of the noninteger period, which can be accurate (15)
to decimal sampling interval, and improves the accuracy and
adaptability of the SVR spectrum. These two matrices are Its spectral moments are
used for periodic detection of the above sinusoidal signals,
S2H Y ( f ) = E{S2H Y (t, f )} = E{|Pn (t, f )|2 } · S2X . (16)
and the SVR spectrum is shown in Fig. 3. When matrix
A processes the nonintegral period signal, the period is not HSK is defined as the energy-normalized fourth-order har-
accurately detected due to the influence of error accumulation monic spectral cumulant
[see Fig. 3(a)]. The matrix Aτ is detected at a 0.1 sampling
interval, and an obvious peak appears at a period of 13.6 [see C4H Y ( f ) S4H Y ( f )
HSKY ( f ) = = 2 ; f ̸= 0 (17)
Fig. 3(b)]. The construction of matrix Aτ can accurately detect S2Y ( f )
2 S2Y ( f )
the period, which is more suitable for detecting the periodic
where C4H Y ( f ) = S4H Y ( f ) denotes the normalized
information of vibration signals. For vibration signals, if the
fourth-order harmonic spectral moment of y(t).
search length l of matrix Aτ is equal to the signal period T ,
The false component mainly contains irrelevant information
each row of data of matrix Aτ is similar. In the singular value
such as noise. The difference between the mean value and
sequence of matrix Aτ , the first singular value σ1 is more
peak value of the instantaneous moment calculated by HSK
prominent, and the subsequent singular values are smaller.
is small, and its HSK value is low. The fault information is
In the SVR spectrum, the data length corresponding to the
the dominant component in the signal. When the frequency of
maximum SVR is the main period of the signal.
HSK extraction is related to the fault information, the HSK
value will be very high. Therefore, screening fault components
B. Component Screening Method Based on HSK according to HSK value has good antiinterference ability.
The four extracted PCs may have false components, so it
is necessary to eliminate false components. Although the III. S IMULATION S IGNAL A NALYSIS
fault signal is accompanied by noise and other interfer- In this article, the SVR spectrum is introduced for periodic
ence, the energy of the fault information is mainly concen- measurement, and according to the characteristics of bearing
trated in the resonant frequency band. If there is a certain fault, the measurement intervals are divided to reduce the
type of fault, the PC corresponding to this period should interference of irrelevant regions. In addition, only one prin-
be mainly periodic fault impact information. Its envelope cipal PC is extracted from each interval to further reduce the
spectrum contains obvious fault characteristic frequencies and generation of false components, and the HSK index is used
harmonics. If the fault does not exist, the PC should be to eliminate false components. This section uses simulation
mainly the noise in the signal. In order to quantify the fault signals to verify the superiority of this method.
information in the envelope spectrum, we propose HSK by The simulation signal s(t) is constructed to simulate the
extracting the instantaneous moment of the envelope spectrum inner ring and outer ring composite fault signal under noise
frequency [19]. interference. The signal is composed of an inner ring fault

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on June 12,2023 at 12:36:30 UTC from IEEE Xplore. Restrictions apply.
11810 IEEE SENSORS JOURNAL, VOL. 23, NO. 11, 1 JUNE 2023

Fig. 4. Time domain and spectrum of the simulation signal. Fig. 6. (a) Components extracted by PDMD. (b) HSK of components.

Fig. 7. Kurtosis of components.

Fig. 5. SVR spectrum.

signal s A (t), an outer ring fault signal s B (t), and random noise
g(t)
50
 q
−2π ζ f nit
X
 s A (t) = 1 − ζ 2 t)u(t − kTi )c(t)



 4e sin(2π f ni


 k=1
 c(t) = (1 − cos(2π f t))3/2 × cos(2π f t)

r r Fig. 8. Time domain and spectrum. (a) Original signal.
30 q (b) Component A. (c) Component B.
5e−2π ζ fno sin(2π f no 1 − ζ 2 t)u(t − kTo )
 X t
s B (t) =





k=1



s (t) = s A (t) + s B (t) + g (t)

and B are much larger than that of components C and D.
(18) In addition, due to noise and other factors, RS has waveform
distortion. Therefore, kurtosis may not be sensitive here for
where the natural frequency f ni = 500 Hz, f no = 800 Hz, con- identifying fault components (see Fig. 7). The components
version frequency fr = 16 Hz, sampling frequency 8192 Hz, A and B filtered by HSK and the original signal are shown
damping ratio ζ = 0.1, inner ring fault frequency f i = 100 Hz, in Fig. 8. The spectrum of components A and B has signif-
that is, period Ti is 82. Refer to the proportional relationship icant resonance frequency band characteristics. The envelope
of various types of fault frequencies of NSK-6307 bearing, spectrum contains prominent fault frequency and harmonic
and set the outer ring fault frequency f o = 62 Hz, that is, the information (see Fig. 9). The correlation coefficient is used to
period To is 132. Gaussian white noise g(t) = SNR(−6 dB) is evaluate the similarity of resonance frequency bands between
added. The time-domain waveform and spectrum of the mixed components A and B and the original signal components.
signal s(t) are shown in Fig. 4. Since the inner ring signal has amplitude modulation, only the
Considering the data length and calculation amount, the signal-to-noise ratio (SNR) of the outer ring signal component
cage interval here is set to [1000, 1100]. The PDMD method is calculated here to quantify the extraction effect of PDMD.
proposed in this article is used to extract signal components. Fig. 10 shows that PDMD has the ability to accurately extract
The SVR spectrum of the measurement intervals is shown periodic pulses.
in Fig. 5. In the SVR spectrum, the main period of each APMD method is used to decompose the signal. Fig. 11
interval is marked by A–D, respectively. The intervals of A–D shows the periodic measurement of the original signal. First
correspond to the periodic detection intervals of the bearing of all, it is not very prominent during the fault period. Second,
inner ring, outer ring, rolling element, and cage fault, respec- the APMD method will preferentially extract components with
tively. The maximum SVRs of the inner ring and outer ring a period of 11 as the first PC. Due to the small energy ratio of
intervals are significantly larger than that of other intervals, the first PC, it is difficult to meet the stopping decomposition
and the corresponding periods are the theoretical fault periods. condition. As a result, a large number of false components
Fig. 6 shows the four components extracted by PDMD and are decomposed, with a number of 81. At the same time,
the corresponding HSK. The HSK values of components A APMD uses the maximum correlation with the original signal

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on June 12,2023 at 12:36:30 UTC from IEEE Xplore. Restrictions apply.
MA et al.: PDMD AND ITS APPLICATION IN BEARING FAULT DIAGNOSIS 11811

Fig. 9. Envelope spectrum. (a) Component A. (b) Component B.

Fig. 13. Waveform, spectrum, and envelope spectrum of PC15.

Fig. 10. (a) SNR of component B and the original signal. (b) Correlation
coefficient between components A and B, and the original signal spec-
trum.

Fig. 14. Fault experiment system.

Fig. 11. Periodic measurement of APMD.

Fig. 15. Time domain and spectrum of the signal.

Fig. 12. Correlation coefficient between components and the original

signal.

to determine the main PCs. Accuracy may be affected by noise,

which increases the complexity of judgment. The correlation
coefficient between each PC and the original signal is shown in
Fig. 12. The 15th PC (PC15) has the greatest correlation, but Fig. 16. SVR spectrum.
it is not prominent. Its time domain, spectrum, and envelope
spectrum are shown in Fig. 13 and do not contain fault
information. Therefore, the periodic recognition and extraction band features in the spectrum contain a lot of noise, and there
ability of APMD is limited. are no significant fault features in the envelope spectrum.
Compared with APMD, the proposed PDMD has a higher The PDMD in this article is used to process the data, and
periodic recognition ability. Second, the division of local the periodic measurement is shown in Fig. 16, where A–D
periodic measurement intervals further reduces the interference correspond to the main period of the periodic detection interval
of unrelated regions. The simulation signal verifies that PDMD of the bearing inner ring, outer ring, rolling element, and cage
can effectively identify and extract fault information. fault, respectively. Considering the data length and calculation
amount, the cage interval here is set to [1500, 1700]. Fig. 17(a)
IV. A PPLICATIONS shows the extracted four suspected fault components. The
In this experiment, the vibration signals of the outer ring HSK of components is calculated as shown in Fig. 17(b),
and inner ring are collected by the fault experiment system as and the HSK of component B is much larger than that of
shown in Fig. 14. The bearing model is 6307, the sampling other components. In addition, the traditional kurtosis index is
frequency is 15 360 Hz, and the motor speed is 1496 r/min. calculated as shown in Fig. 18. However, the kurtosis value
After calculation, the fault characteristic frequency of the outer of component B is not prominent. Fig. 19 shows the time
ring is 76.52 Hz and that of the inner ring is 122.95 Hz. domain, spectrum, and envelope spectrum corresponding to
component B, including obvious outer ring fault characteristic
A. Analysis of Bearing Outer Ring Fault Data frequency and frequency doubling. PDMD can accurately
Fig. 15 shows the collected outer ring vibration signal. Peri- screen the optimal period and effectively extract the periodic
odic pulses cannot be seen in the time domain, the resonance pulse component in the bearing signal.
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on June 12,2023 at 12:36:30 UTC from IEEE Xplore. Restrictions apply.
11812 IEEE SENSORS JOURNAL, VOL. 23, NO. 11, 1 JUNE 2023

Fig. 21. Correlation coefficient between components and the original

signal.

Fig. 17. (a) Components extracted by PDMD. (b) HSK of components.

Fig. 22. Time domain, spectrum, and envelope spectrum of PC41.

Fig. 18. Kurtosis of components.

Fig. 23. Time domain and spectrum of the signal.

Fig. 19. Time domain, spectrum, and envelope spectrum of

component B.

Fig. 20. Periodic measurement of APMD. Fig. 24. SVR spectrum.

The APMD method is used to process data. Fig. 20 shows and the fault characteristic harmonics in the envelope spec-
the periodic measurement of the original signal. The first main trum are not significant enough. The periodic measurement
period is not a fault period. Due to the small energy ratio of of PDMD is shown in Fig. 24, where A–D correspond to
the first PC, APMD produces over-decomposition, resulting the main period of the periodic detection interval of the
in 127 components. The correlation coefficient between each bearing inner ring, outer ring, rolling element, and cage fault,
component and the original signal is shown in Fig. 21. The respectively. Fig. 25(a) shows the extracted four suspected
correlation coefficient of the 41st PC (PC41) is the largest, but fault components. The HSK of components is calculated as
it is not prominent. The waveform, spectrum, and envelope shown in Fig. 25(b), and the HSK of component A is much
spectrum of PC41 are shown in Fig. 22. It does not contain larger than that of other components. In addition, the kurtosis
fault information and cannot effectively diagnose faults. Com- of components is not prominent, and the effect of screening
pared with APMD, the proposed PDMD improves the periodic components here is poor (see Fig. 26). Fig. 27 shows the time
recognition ability. The experimental signals further verify that domain, spectrum, and envelope spectrum corresponding to
PDMD can effectively identify and extract the outer ring fault component A, including obvious inner ring fault information.
information. PDMD can be well applied to inner ring fault diagnosis.
Similarly, the APMD method is used to process the data.
B. Analysis of Bearing Inner Ring Fault Data However, the performance of periodic measurement is poor
The inner ring vibration signal is shown in Fig. 23. The time here, and the accuracy is disturbed by noise (see Fig. 28).
domain and the spectrum contain a lot of noise interference, Due to the small energy ratio of the first PC, APMD produces

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on June 12,2023 at 12:36:30 UTC from IEEE Xplore. Restrictions apply.
MA et al.: PDMD AND ITS APPLICATION IN BEARING FAULT DIAGNOSIS 11813

Fig. 30. Waveform, spectrum, and envelope spectrum of PC14.

shown in Fig. 30. Its frequency does not correspond to the

Fig. 25. (a) Components extracted by PDMD. (b) HSK of components.
fault frequency. Compared with APMD, it is further verified
that PDMD can effectively identify and extract inner ring
fault information and can suppress noise interference and
over-decomposition.

V. C ONCLUSION
In this article, a PDMD is proposed, which effectively
improves the ability of period recognition. PDMD realizes
fault diagnosis by identifying and extracting four suspected
Fig. 26. Kurtosis of components. fault components in the signal and effectively eliminating false
components. On the one hand, PDMD can achieve periodic
measurement accuracy of decimal step size. In addition, the
local periodic search intervals are reasonably divided. It not
only avoids the interference of unrelated regions, but also
reduces the generation of false components. On the other hand,
HSK is applied to eliminate false components, which improves
the accuracy of identifying fault information. The simulation
and experimental results show that the method can accurately
search the optimal period and extract the periodic pulse,
Fig. 27. Time domain, spectrum, and envelope spectrum of component which is suitable for the fault diagnosis of rolling bearings.
A. However, when the bearing operates abnormally, the load of
each impact is different, especially the inner ring signal with
amplitude modulation. The component amplitude extracted by
RS theory will appear averaging phenomenon, and the original
true amplitude information will be lost, which needs further
study.

R EFERENCES
Fig. 28. Periodic measurement of APMD. [1] S. Wang, G. Mei, B. Chen, and Y. Cheng, “An improved time-varying
morphological filtering and its application to bearing fault diagnosis,”
IEEE Sensors J., vol. 22, no. 21, pp. 20707–20717, Nov. 2022.
[2] K. Zhang, Y. Xu, Z. Liao, L. Song, and P. Chen, “A novel fast entrogram
and its applications in rolling bearing fault diagnosis,” Mech. Syst. Signal
Process., vol. 154, Jun. 2021, Art. no. 107582.
[3] Z. Liao, X. Song, B. Jia, and P. Chen, “Bearing fault feature enhance-
ment and diagnosis based on statistical filtering and 1.5-dimensional
symmetric difference analytic energy spectrum,” IEEE Sensors J.,
vol. 21, no. 8, pp. 9959–9968, Apr. 2021.
[4] D. Liu, L. Cui, W. Cheng, D. Zhao, and W. Wen, “Rolling bearing fault
Fig. 29. Correlation coefficient between components and the original severity recognition via data mining integrated with convolutional neural
signal. network,” IEEE Sensors J., vol. 22, no. 6, pp. 5768–5777, Mar. 2022.
[5] Y. An, K. Zhang, Q. Liu, Y. Chai, and X. Huang, “Rolling bearing fault
diagnosis method base on periodic sparse attention and LSTM,” IEEE
Sensors J., vol. 22, no. 12, pp. 12044–12053, Jun. 2022.
over-decomposition, resulting in 119 components. The corre- [6] H. Qu, M. Zhang, and L. Pang, “Mental workload classification method
lation coefficient between each component and the original based on EEG cross-session subspace alignment,” Mathematics, vol. 10,
signal is shown in Fig. 29. The 14th PC (PC14) is selected no. 11, p. 1875, May 2022.
[7] X. Zhang and D. Feng, “An efficient MUSIC algorithm enhanced by
through maximum correlation, but it is not prominent. The iteratively estimating signal subspace and its applications in spatial
waveform, spectrum, and envelope spectrum of PC14 are colored noise,” Remote Sens., vol. 14, no. 17, p. 4260, Aug. 2022.

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on June 12,2023 at 12:36:30 UTC from IEEE Xplore. Restrictions apply.
11814 IEEE SENSORS JOURNAL, VOL. 23, NO. 11, 1 JUNE 2023

[8] L. Zhou, K. Ye, J. Qi, and H. Sun, “DOA estimation based on pseudo- Zhiqiang Yang received the B.S. degree in
noise subspace for relocating enhanced nested array,” IEEE Signal mechanical engineering from the Henan Univer-
Process. Lett., vol. 29, pp. 1858–1862, 2022. sity of Technology, Zhengzhou, China, in 2020.
[9] W. A. Sethares and T. W. Staley, “Periodicity transforms,” IEEE Trans. He is currently pursuing the M.S. degree in
Signal Process., vol. 47, no. 11, pp. 2953–2964, Nov. 1999. mechanical engineering with the Beijing Univer-
[10] D. D. Muresan and T. W. Parks, “Orthogonal, exactly periodic sub- sity of Technology, Beijing, China.
space decomposition,” IEEE Trans. Signal Process., vol. 51, no. 9, His current research interests include equip-
pp. 2270–2279, Sep. 2003. ment fault diagnosis technology and signal
[11] P. P. Vaidyanathan, “Ramanujan sums in the context of signal processing.
processing—Part I: Fundamentals,” IEEE Trans. Signal Process., vol. 62,
no. 16, pp. 4145–4157, Aug. 2014.
[12] P. P. Vaidyanathan, “Ramanujan sums in the context of signal
processing—Part II: FIR representations and applications,” IEEE Trans.
Signal Process., vol. 62, no. 16, pp. 4158–4172, Aug. 2014.
[13] S. Ramanujan, “On certain trigonometrical sums and their applications in
the theory of numbers,” Trans. Cambridge Philos. Soc., vol. 22, no. 13,
pp. 259–276, 1918.
[14] S.-W. Deng and J.-Q. Han, “Ramanujan subspace pursuit for signal peri-
odic decomposition,” Mech. Syst. Signal Process., vol. 90, pp. 79–96, Yonggang Xu received the B.S. and Ph.D.
Jun. 2017. degrees in mechanical engineering and automa-
[15] S.-W. Deng and J.-Q. Han, “Signal periodic decomposition with tion from Xi’an Jiaotong University, Xi’an, China,
conjugate subspaces,” IEEE Trans. Signal Process., vol. 64, no. 22, in 1998 and 2003, respectively.
pp. 5981–5992, Nov. 2016. From 2004 to 2006, he worked with the Post-
[16] J. Cheng, Y. Yang, X. Li, and J. Cheng, “Adaptive periodic mode doctoral Research Station, Tianjin University,
decomposition and its application in rolling bearing fault diagnosis,” Tianjin, China. Since 2006, he has been an
Mech. Syst. Signal Process., vol. 161, Dec. 2021, Art. no. 107943. Associate Professor with the College of Mechan-
[17] P. P. Kanjilal and S. Palit, “On multiple pattern extraction using singular ical Engineering and Applied Electronics Tech-
value decomposition,” IEEE Trans. Signal Process., vol. 43, no. 6, nology, Beijing University of Technology, Beijing,
pp. 1536–1540, Jun. 1995. China. His research interests include condition
[18] H. Liu, J. Li, Y. Zhao, and L. Qu, “Improved singular value decomposi- monitoring and fault diagnosis of electromechanical equipment and
tion technique for detecting and extracting periodic impulse component modern signal processing.
in a vibration signal,” Chin. J. Mech. Eng., vol. 17, no. 3, pp. 340–345,
2004.
[19] K. Zhang, P. Chen, M. Yang, L. Song, and Y. Xu, “The harmogram:
A periodic impulses detection method and its application in bearing
fault diagnosis,” Mech. Syst. Signal Process., vol. 165, Feb. 2022,
Art. no. 108374.
[20] J. Antoni, “The spectral kurtosis: A useful tool for characterising
non-stationary signals,” Mech. Syst. Signal Process., vol. 20, no. 2,
pp. 282–307, 2002. Aijun Hu received the B.S. and M.S. degrees
in mechanical engineering and the Ph.D. degree
in thermal engineering from North China Electric
Power University, Beijing, China, in 1993, 1998,
and 2008, respectively.
He is currently a Professor with the Depart-
ment of Mechanical Engineering, North China
Electric Power University. His research interests
include signal process, mechanical fault diagno-
sis, and industry data analysis.

Chaoyong Ma received the B.S. degree from

the Hebei Institute of Technology, Tianjin, China,
in 1993, and the M.S. degree in mechanical
manufacturing and automation and the Ph.D.
degree in monitoring device technology from the
Beijing University of Technology, Beijing, China, Kun Zhang was born in Zhangjiakou, China.
in 1997 and 2013, respectively. He received the B.S. degree in mechanical
From 1997 to 2003, he was a Senior Engi- engineering from the North China University of
neer with the General Institute of Architec- Science and Technology, Qinhuangdao, China,
tural Research, Ministry of Metallurgy of China, in 2015, the M.S. degree in mechanical engi-
Beijing. Since 2003, he has been a Senior Engi- neering from the Beijing University of Tech-
neer with the College of Mechanical and Electrical Engineering, Beijing nology, Beijing, China, in 2019, and the Ph.D.
University of Technology. He has been engaged in the research of degree from Mie University, Tsu, Japan, in 2022.
intelligent monitoring technology for key parts of automobiles and fault He is now a Lecturer with the Beijing University
diagnosis technology of rotating machinery. He presided over a number of Technology. His research interests include
of horizontal projects of school–enterprise cooperation. His research condition monitoring and intelligent fault diagno-
interests include automotive parts performance testing, rotating equip- sis of rotating equipment, such as bearings and gears and digital signal
ment condition testing, and fault diagnosis. processing.

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on June 12,2023 at 12:36:30 UTC from IEEE Xplore. Restrictions apply.