Mechanical Systems and Signal Processing 85 (2017) 415–428

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

New procedure for gear fault detection and diagnosis using

instantaneous angular speed
Bing Li, Xining Zhang n, Jili Wu
State Key Laboratory for Manufacturing System Engineering, Xi’an Jiaotong University, 710049 Xi’an, PR China

a r t i c l e i n f o abstract

Article history: Besides the extreme complexity of gear dynamics, the fault diagnosis results in terms of
Received 14 March 2016 vibration signal are sometimes easily misled and even distorted by the interference of
Received in revised form transmission channel or other components like bearings, bars. Recently, the research field
27 June 2016
of Instantaneous Angular Speed (IAS) has attracted significant attentions due to its own
Accepted 22 August 2016
advantages over conventional vibration analysis. On the basis of IAS signal's advantages,
Available online 30 August 2016
this paper presents a new feature extraction method by combining the Empirical Mode
Keywords: Decomposition (EMD) and Autocorrelation Local Cepstrum (ALC) for fault diagnosis of
Multistage gearbox fault diagnosis sophisticated multistage gearbox. Firstly, as a pre-processing step, signal reconstruction is
Instantaneous angular speed
employed to address the oversampled issue caused by the high resolution of the angular
Empirical mode decomposition
sensor and the test speed. Then the adaptive EMD is used to acquire a number of Intrinsic
Autocorrelation local cepstrum
Mode Functions (IMFs). Nevertheless, not all the IMFs are needed for the further analysis
since different IMFs have different sensitivities to fault. Hence, the cosine similarity metric
is introduced to select the most sensitive IMF. Even though, the sensitive IMF is still in-
sufficient for the gear fault diagnosis due to the weakness of the fault component related
to the gear fault. Therefore, as the final step, ALC is used for the purpose of signal de-
noising and feature extraction. The effectiveness and robustness of the new approach has
been validated experimentally on the basis of two gear test rigs with gears under different
working conditions. Diagnosis results show that the new approach is capable of effectively
handling the gear fault diagnosis i.e., the highlighted quefrency and its rahmonics cor-
responding to the rotary period and its multiple are displayed clearly in the cepstrum
record of the proposed method.
& 2016 Elsevier Ltd. All rights reserved.

1. Introduction

Vibration signals collected from the accelerometer usually carry a great deal of information to reflect machine health
conditions. This makes vibration-based fault detection research much more prevalent during past few years. However, in
practice, vibration-based signal analysis still has insurmountable limitations like structure noises and dominant component
interferences for fault diagnosis of the complex gearbox systems. Specifically, vibration signals acquired from a multistage
gearbox are easily susceptible to the interference of the transmission channel since the accelerometer is always mounted on
the outer surface of the gearbox case near the main bearing housing [1,2]. Due to the factors mentioned above, useful
information is inevitably corrupted, which makes it difficult to diagnose gearboxes from such vibration signals [3]. Thus, the

n
Corresponding author.

http://dx.doi.org/10.1016/j.ymssp.2016.08.036
0888-3270/& 2016 Elsevier Ltd. All rights reserved.
416 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428

signal source selection becomes somewhat of a challenge in obtaining particular gear conditions accurately.
Recently, the research field of Instantaneous Angular Speed (IAS) has attracted significant attention from the research
community, which is attributable to its own advantages over conventional vibration analysis [1,4,5]. The idea behind using
IAS signal as a fault detection tool is that any fault in the rotating machine has a direct impact on the angular speed of the
rotary shaft. It usually contains a considerable amount of information on fault components reflecting health and usage
status of rotating machinery [6]. Arising from these benefits, different advanced IAS-based techniques have been developed
so far in some specific occasions such as in: rotors [4], diesel engines [7,8], bearings [9], induction motors [4,6] and so on.
Even so, because of complexity and non-stationary, little literature has applied the IAS signal for gear fault diagnosis and
hence it remains a poorly understood area. Besides, due to the high sample frequency in collecting the raw signal, a great
deal of data can be generated, which will bring a heavy burden for further analysis and computation. On closer observation,
it is noticeable that the peaks above a certain threshold are dramatically reduced, which means the IAS signal is over-
sampled using the original sample frequency.
Fortunately, many pragmatic identification approaches such as time–frequency representation, empirical mode de-
composition (EMD), wavelet transform, spectral kurtosis have been successfully established to enhance the signal feature
and simultaneously provide potent proofs for maintenance decision-making [10–14]. Among these diagnostic technologies,
the EMD technique and ensemble empirical mode decomposition (EEMD) have been widely used to analyze vibration
signals in the field of fault diagnosis, particularly for signals in non-stationary cases and signals with a larger size [11,15].
EEMD, developed by Wu and Huang [16], is able to eliminate the problem of mode mixing effectively by means of adding
finite white noise to the raw signal, but it also increases the computing time since many tries and interactions have to be
taken into account. In addition, both EMD and EEMD try to decompose a signal into different intrinsic mode functions
(IMFs), or mono-component functions. Nevertheless, not all IMFs are sensitively and closely related to the faults. The se-
lection of IMFs are usually relied on the user's experience by visual inspections. To figure out this issue, Ricci et al. [11]
introduced the merit index to assist the IMF selection and Hilbert transform. Satisfactory results were achieved when the
automatic IMF selection was used to deal with the vibration signals from spiral bevel gears. However, the specific parameter
k used in that paper relied on the mechanical system and the default threshold should be also specified initially. In the
literature [15], Lei et al. proposed another procedure to divide all the IMFs into two clusters, and both the simulation and
experiment were engaged to demonstrate the effectiveness of the IMF selection. However, for this method, the normal
signal must be considered.
In order to obtain the diagnosis information more sufficiently, recently two improved cepstrums have been developed
successively [17,18]. Both Local Cepstrum (LC) analysis and its improved form Autocorrelation Local Cepstrum (ALC) focus on
achieving the maximum utility of the local (partial) spectrum information. By searching for the optimal frequency band of
interest, local cepstrum is capable of effectively addressing the crucial issue of interference reduction of aperiodic com-
ponents [18]. However, the optimal frequency band selection is usually time consuming and quite difficult in the LC cal-
culation. Meanwhile, the noise reduction performance of LC is also limited since it does not concern too much about the
signal de-nosing issue. As a simple yet effective method, ALC has been successfully developed by combining LC with au-
tocorrelation and successfully utilized to deal with the gearbox diagnosis using vibration signals [17]. On the consideration
of IAS spectrum signatures, it is worth noting that both LC and ALC are very convenient and suitable for IAS signal processing
with intensive frequency distributions and they are capable of emphasizing periodic fault components with the background
noise and other components minimized.
After adequately considering the issues mentioned above, in this paper, different issues will be fixed using different
methods. First of all, signal reconstruction is delivered as a pre-processing method to deal with the oversample obstruct.
Then, a new feature extraction and fault diagnosis method in terms of EMD and ALC for the sophisticated multistage
gearbox is established. Meanwhile, the cosine similarity measure [19,20] that is most widely reported for similarity mea-
surement will be introduced for the sensitive IMF selection.
The main structure of this article is briefly organized as follows. Section 2 provides the principle and measurement of the
conventional IAS. In Section 3, some relevant theories that support the proposed method, including signal reconstruction,
EMD, ALC, are reported successively. Then study cases including two types of gearboxes are investigated in Section 4.
Besides, the fusion method is applied to these two gearboxes under different health statuses and some results and dis-
cussions are established thereafter in this section. Finally, some important conclusions and future works of this article are
carefully drawn.

2. Measuring principles of the instantaneous angular speed

In this section, the IAS measuring principles are mainly described. The reason why signals extraction based on speed
fluctuations has the ability to deal with the fault diagnosis is primarily put forward. Then, the measuring model is briefly
introduced.

2.1. Foundations of angular signal characteristic analysis

IAS signals acquired from some complex machineries such as a multistage gearbox, can be treated as a combination of
 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428 417

periodic signal and structure borne noise. The frequency aliasing phenomenon is inevitably shown in the spectrum due to
various noises produced by random fluctuations of load, electric supply, filtering and digitalization [4,21]. Besides, a gear
with defects will theoretically generate an uneven distribution of tangential force between the meshing teeth, which results
in periodical perturbation of speed along with its rotation. Every time when a tooth passes through the disturbance, a speed
fluctuation happens, which in return enables the shaft rotation speed to contain dynamic responses of fault components.
That is to say, the IAS signal is actually non-stationary with the characteristics of frequency modulation (FM) and varies with
speed and loading conditions. More importantly, signals derived from angular domain with strong anti-disturbance ability
and great robustness, are not susceptible to the interference of the transmission channel and other components like
bearings, bars. In particular, the IAS signal has the high noise suppression capability and great performance since this
algorithm is theoretically developed on the basis of frequency modulation, and it is usually measured at the shaft of the
machine where speed fluctuates are directly affected by the faults. In a word, signal extraction on the basis of speed
fluctuations is capable of dealing with fault detection and diagnosis due to the abundant information that reflects the health
status of a machinery component.

2.2. Modeling IAS measurement

According to the prior research [22], measurement performance of IAS signals is based on two basic principles: counting
the number of pulses in a given time duration and measuring the elapsed time for a single cycle. The conventional IAS
definition is relied on the following equation:

2π 60fh
ω= ∙
Nf Ni (1)

where fh represents the clock frequency or sample frequency; Ni is the number of data points between two rising edges of
the encoder signal; Nf is the resolution of the encoder. Note that here ω is the angular velocity in the unit of rpm.
Another factor that directly affects the measurement result is the sample frequency. The minimal sample frequency of
the encoder's square signal has to be set to a large value because of the resolution of the encoder and the rotating speed. On
the basis of the sampling theorem and computing method of the minimum sample frequency in literature [4], the sample
frequency can be derived in advance according to the following equation to minimize the computational complexity and
errors.

fsmin =4×⎡⎣ Nf f0 + Nhf0 + Nf P ⎤⎦

( ) (2)

where fsmin , f0 , P and Nh respectively denote the minimal sample frequency, the rotation frequency, the estimated shaft
speed variation and the multiples of the rotational speed in the frequency band of interest.

3. Basic concepts of the involving signal processing approaches

IAS-based signal has been proved to be effective and sensitive to handle the fault diagnosis issues in some specific
research owing to its distinct benefits. However, the efficiency of IAS-based signal essentially relies on several factors like
proper estimation algorithm, sensor resolution and irregularity, interpolation algorithms, frequency aliasing, quantization
effect and background noise. Therefore, much research has to be concerned to overcome these difficulties.

3.1. IAS signal pre-processing

Due to the high resolution of the angular sensor, i.e., optical encoder, more detailed information or diagnosis results can
be achieved. Meanwhile, a high sample frequency is always needed and hence a large amount of data are generated in-
evitably, which of course can result in time consuming issue and high computational complexity simultaneously. Sometimes
the large data length can even dramatically affect the signal processing result.
After calculating the IAS signal, by spectral observing, we can clearly find the oversampled phenomenon in the spectrum
where above a certain threshold, the dominant frequency peaks are totally replaced by background noises. This over-
sampled phenomenon is mainly relied on the reason that the sample frequency for the original encoder signal is too larger
for the IAS signal and hence it may be inexpedient to use the whole spectrum.
In order to improve the computation efficiency, as the pre-processing method, an effective strategy of signal re-
construction is put forward by discarding the useless information exceeding the specific threshold. Note that, the frequency
resolution remains the same, which ensures that the low frequency diagnosis information can be completely preserved for
the further investigations.
418 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428

3.2. EMD algorithm and effective IMF selection

3.2.1. EMD algorithm

Typically, EMD method is well-developed and widely applied to the analysis of non-linear and non-stationary signals in
fault diagnosis of the rotating machinery field. Under the simple assumption, any signal can be decomposed into a series of
almost mono-components called IMFs. Being independent of the others, each IMF must satisfy the following conditions [15].

1) In all data set, the number of extremes and the number of zero-crossings must either be equal or only a difference of one
is allowed.
2) At any point, the mean value of the envelope defined by the local maxima and the envelope defined by the local minima
must be equal to zero.

A review of the EMD algorithm is demonstrated briefly in this subsection based on literature [15]. Any raw signal x(t ) can
be decomposed by the EMD algorithm as follows:

1) After extracting all the local extremes (including maxima and minima) of x(t ), separately interpolate all the local maxima
and the minima by cubic spline lines to form upper and lower envelopes.
2) Calculate the mean value of upper and lower envelope value and define the mean signal asm1(t ), the difference between
x(t ) and m1(t ) is considered as the first component h1(t ), i.e., h1( t )=x( t ) − m1( t ).
3) Replace x(t ) by h1(t ), figure out whether h1(t ) meets the IMF conditions and if not, repeat steps (1) and (2) until h1(t )
becomes an IMF. Thus, c1( t ) = h1(t ) is designated as the first IMF.
4) Separate c1(t ) from x(t ) to get the residual signal r1(t ) which will subsequently substitutes the original signal for the further
IMF derivation.
5) Derivation will continue until the stoppage criterion of the signal's decomposition is fulfilled when the final residual
signal rn(t ) becomes a monotonic function from which no more IMFs can be extracted.
6) At last, the original signal x(t ) is decomposed into a series of empirical modes and a residue rn(t ).

Based on the EMD algorithm, an original signal x(t ) can be considered as a linear superposition of IMFs with a residue
rn(t ) :
n
x( t ) = ∑ ci(t ) + rn( t )
i=1 (3)

Theoretically, the frequency bandwidths of the empirical modes ci( t ),i = 1,2, … , n change from high to low. The first few
IMFs describe the high frequency phenomena while the last ones are related to the low frequency components of the signals
[23]. Sometimes, the IMFs covering very low frequency components are discarded arbitrarily. For more details on the theory
of EMD algorithm and mathematical discussion, the reader is suggested to [15].

3.2.2. Effective IMF selection using cosine similarity measures

Accordingly, the frequency components contained in each IMF usually change with the variation of signal x(t ). Some of
them are closely related to the fault while the others may be less effective. In general, when the IMFs obtained by EMD are
used for fault diagnosis, the diagnosis results are usually quite different with different IMFs. Nowadays, the choice of IMFs to
be analyzed is still depended on the user's experience through visual inspections [11]. Unfortunately, the experiment-based
choice sometimes misleads the user's determination. Thus, methods for sensitive IMFs selection are still important to
distinguish the fault-related information from the useless IMFs [24].
The basic idea behind the IMFs selection is to select some sensitive fault-related IMFs instead of the raw signal for the
aim of fault detection and diagnosis. To expect the selection process automatic and effective, a conventional measure named
cosine measure metric that has been widely used in the measure of vector similarity is employed in this paper. The simi-
larity measure method is originally defined as the inner product of two vectors X =[x1, x2, …,xN ] and Y =[y1, y2 ,… ,yN ] divided by
the product of their lengths. The original similarity metric between the two vectors is formally established as [19]:
N
∑k = 1 xk yk
Simcos ( X , Y ) = cos (α )=
N N
∑k = 1 xk2 × ∑k = 1 yk2 (4)

where N is the number of the elements in the vector.

It can be easily seen that the cosine value reflects the magnitude of similarity between two vectors. The cosine of the
angle between two vectors is used to measure how “similar” they are, which in turn, is a measure of the similarity of these
IMFs.
Among the most extensive applications including text classification, speaker recognition, information retrieval [19,20],
the cosine similarity usually concerns about the vectors with non-negative elements, such as weights of features. In this
way, the magnitude usually lies in [0,1], where in the sense the larger the cosine value is, the higher the similar degree of
 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428 419

two vectors would be and vice versa. However, when the vectors in Eq. (4) are substituted by the IMFs, the negative
similarity values will be generated since both positive and negative values are allowed for the elements in IMF signals.
Therefore, the magnitude range of the similarity would be extended into [ 1, 1]. Intuitively, for calculating the similarity
between the original and the IMF signals, when the cosine similarity value is equivalent to 1, it just implies that the maps
between the two are identical; the similarity value will be equal to 1 when the similarity calculation happens between the
original and the – IMF signals. It can be reported as a negative correlation between the two signals. Thus, for easy un-
derstanding, the cosine similarity metric should be modified as:
N
| ∑k = 1 xk yk |
Simcos ( X , Y ) = cos (α )=
N N
∑k = 1 xk2 × ∑k = 1 yk2 (5)

Finally, the modified cosine similarity of their redirection maps can be developed to determine the similarity degree
between the two signals X and Y .

3.3. Related theories of autocorrelation local cepstrum analysis

Generally, when a fault occurs in a gear, families of sideband components may show up around the meshing frequency
and its harmonics. In addition, at a constant speed, these sidebands with equal spaces usually distribute widely in the
spectrum when the faults are severe. Traditional cepstrum analysis is quite suitable to handle this issue since it has the
ability in simplifying the equally spaced sideband feature and extracting the periodical components in the spectrum.
However, the cepstrum would be not preferable for the case of early gear failure, where sidebands to be recognized are
easily overwhelmed by the background components. Meanwhile, poor diagnosis results can be achieved when this tech-
nique is employed to explore the spectrum with concentrated sideband distribution, because the periodic features would be
further weakened by the average effect of Fourier transform.
In order to overcome these shortcomings, the autocorrelation local cepstrum (ALC) is recently established. ALC analysis
mainly focuses on achieving the maximum utility of the local (partial) spectrum information and eliminating the limitation
of highly localized distribution of periodic fault features. The flowchart of this method is shown in Fig. 1 and the basic
algorithm and procedure of this method can be briefly summarized as follows.

1) Obtain the analytical signal x( t ) using the corresponding equipment or calculation.

2) Employ autocorrelation technique in time domain to identify the overwhelming periodic components primarily. Here the
autocorrelation signal is defined as Rx( τ ).
3) Apply the Fourier transform to achieve the spectrum of Rx( τ ) and shift the zero-frequency component to the center of the
spectrum. Here the shifted signal is named as Sx′( ω).
4) Employ autocorrelation technique to Sx′( ω) to gain the signal Rx′( ω), which is able to enhance the convergence of periodic
fault features near the zero frequency.
5) Perform symmetrical filter to the signal Rx′( f ) in the spectrum center with an optimal window length of interest and
thereafter use inverse shift to swap the left and right half frequency components.
6) Use inverse Fourier transform of the logarithm of the inverse shifted signal to achieve the final ALC result.

Accordingly, three major obstacles are successfully fixed by using the above procedure: a) The interference can be sig-
nificantly reduced using the related de-noising process. b) The optimal frequency band of interest with periodic fault
features is effectively concentrated around the zero frequency, which is much convenient for further exploration and ex-
traction. c) The optimal frequency band can be easily selected using an appropriate window function around the middle of
the shifted spectrum Sx′( ω).

3.4. Principle of the improved EMD method with ALC

A new fusion technology by combining EMD with ALC is put forward in this paper to obtain much evident features for
condition recognition and fault diagnosis. The flowchart of the proposed method is apparently illustrated in Fig. 2.
After the squared signal collected from an optical encoder, firstly, the discrete IAS signal is extracted from the normalized
signal using the IAS measurement. As a pre-processing step, signal reconstruction is primarily delivered herein to reduce the
oversampled effect and improve calculation efficiency at the same time. Then EMD is employed to deal with the re-
constructed signal for the IMF extraction subsequently. The most distinctive IMF containing representative features of gear
fault is chosen on the basis of cosine similarity metric for further analysis thereafter. When ALC method with an appropriate
window length is put forward to process the synthesized signal, a de-noised cepstrum sequence is acquired and utilized for
the purpose of gear fault detection and diagnosis. Next, the IAS signal will be greatly enhanced by minimizing the noise and
other components. Finally, the faults can be diagnosed thereby.
420 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428

Fig. 1. The flowchart of autocorrelation local cepstrum analysis.

4. Experimental verification

Due to its complexity, fault detection of gearbox requires desperately proper signal processing techniques to extract the
hidden information in the signal [21]. The purpose of this section is to show how the fusion technique works for the fault
diagnosis of the gearboxes. Specifically, two experimental gearbox rigs with different gear statuses were concerned. Note
that many tests and comparisons have been developed and analyzed during the investigation process, nevertheless, on the
consideration of the paper length, only one test from each test rig has been used to demonstrate the high performance and
robustness of the new algorithm in the subsections.

4.1. Case 1: fault diagnosis on the helical gearbox

4.1.1. Experiment setup

Fig. 3 illustrates the structure chart of the helical gearbox test rig together with the internal structure of the gearbox. The
whole test rig mainly consisted of a drive AC motor with a nominal power at 4 kW, a DC motor, an incremental optical
encoder, an excitation regulator and two gearboxes with the same parameters, i.e., the right gearbox was used as a reduction
gearbox, while the other was used for accelerating speed. The output of the accelerating gearbox was connected by the DC
motor, which was separately excited and acted as a generator. An electric stove with a maximum output power of 1500 W
was connected to the generator so that the electricity generated by the DC generator can be consumed timely. To measure
the IAS, the rotary encoder installed with a resolution of 1500 pulses per revolution (ppr) was flexibly attached to the output
rotor of the reduction gearbox using a specific holder, while the encoder's holder was fixed at the rig base through a
magnetic base.
Early failure was introduced artificially to the gear Z4 in reduction gearbox by making three adjacent teeth crests thin and
sharp (see Fig. 4(a)). In this experiment, the actual power of the stove was set at 625 W by the excitation regulator. The
actual input speed of the reduction gearbox was approximately equal to 1770 rpm and the corresponding frequency
characteristic of fault gear Z4 was around 2.326 Hz. A PCI-9846H high speed data acquisition card was utilized here to collect
 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428 421

Fig. 2. The relevant flow chart of the fusion method.

the square pulse signal from the incremental encoder. Each file of the encoder signal consisted of 4 s of data with a sampling
rate of 192 kHz.

4.1.2. Results and discussions

The IAS signals for the undamaged and damaged gears are calculated and intensively illustrated in Fig. 5. By comparing
the two time serials, it is observed that as a reference condition without fault, the time record appears unremarkable with
little periodic modulation characteristics, while the time recorder in Fig. 5(b) shows a series of impulse responses with

Fig. 3. The experimental rig with helical gearboxes.

422 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428

Fig. 4. Different fault gears (a) case1: helical gear with three adjacent teeth crests thin and sharp (b) case 2: spur gear with a broken tooth.

Fig. 5. Original IAS waveforms and spectrums with enlarged version ranging from 100 Hz to 350 Hz. (a) and (c) are waveform and spectrum of undamaged
gear; (b) and (d) are waveform and spectrum of damaged gear.

smeared amplitude modulation periodically spaced at approximately one revolution of the output shaft. Meanwhile, the
oversampled phenomenon can be noticed in the two spectrums where frequency peaks above 500 Hz reduce dramatically.
Thus it is reasonable to employ the signal reconstruction strategy to handle this problem. On closer inspection on the larger
version, we can find that there are more prominent sideband features in the damaged case than those in the undamaged
one. Even though, the fault diagnosis is still not much straightforward in this phase.
Thereafter, the EMD is further applied to the IAS signals to gain the corresponding IMFs, while IMFs with very low
frequency components are not concerned arbitrarily. Then, the similarity between each IMF and the original IAS is evaluated
by the cosine similarity metric for the purpose of capturing the most sensitive IMF. The values obtained by Eq. (5) for the
gears under two different conditions are listed in Table 1 and Table 2 respectively, where the highest values of similarities
are selected to diagnose gear faults while the others are discarded as irrelevant ones.
According to the cosine similarities, IMF 2 for the undamaged gear and IMF 1 for the damaged gear are selected as the
most sensitive IMFs respectively as shown in Fig. 6. As can be seen, differences are highlighted in both amplitude and wave
shape for the two IMFs. Note that, the impulse components have been reserved in Fig. 6(b). Nevertheless, the fault char-
acteristics are still not highlighted and effective enough to distinguish which gear suffers from failure because of the

Table 1
Results of cosine similarities based on the undamaged helical gear.

IMF 1 2 3 4 5 6 7 8

Simcos 0.5398 0.7943 0.2165 0.0512 0.0529 0.0352 0.0846 0.0134

 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428 423

Table 2
Results of cosine similarities based on the damaged helical gear.

IMF 1 2 3 4 5 6 7 8

Simcos 0.7511 0.6079 0.2299 0.056 0.0536 0.0318 0.0815 0.0439

Fig. 6. Sensitive IMFs for (a) undamaged (b) damaged helical gears.

weakness of the fault component related to the gear fault.

Finally, the ALC method is delivered to further eliminate the strong residual noise and achieve the diagnosis purpose,
where each sensitive IMF component is successively processed by ALC. Note that determined by trial and error, the window
length was set to 400 Hz for ALC calculation in this case. In addition, in order to demonstrate the preferable improvement of
the proposed method, different cepstrum analysis results under undamaged and damaged conditions have been calculated
and intensively shown in Figs. 7 and 8 respectively.
When the new fusion method applies to the raw signals under undamaged and damaged conditions, some anticipated
results are achieved finally. Obviously, by means of ALC with sensitive IMF, the cepstrum in Fig. 8(d) is almost totally
dominated by quefrency and its rahmonics components that coincide to the rotary period of gear Z4 while there is no sign of
periodic impulse as shown in Fig. 7(d) for the undamaged case. This indicates that the proposed method is quite effective to
address the fault diagnosis issue of helical gearbox on the basis of IAS signal. Moreover, relying on the comparison results
between each cepstrum, a brief summary can be conducted as follows. a) After taking into the consideration of the partial
information, the magnitude has been significantly improved, especially for the damaged case. b) The effectiveness of the
proposed method can be definitely illustrated since the period of shock responses corresponding to the fault characteristics
apparently shows up in Fig. 8(d) while the periodical feature in Fig. 7(d) is remarkably reduced. c) Although the ALC is
effective in dealing with the gear fault, one can see that the first, third and fourth order are still quite smeared in Fig. 8
(c) compared with Fig. 8(d), which means without considering the sensitive IMF, the ALC is incapable of handling the task of
fault feature extraction on gears with structure noises and dominant component interferences.
In fact, thanks to the advantages of the IAS signal, the gear's fault features can be more visible. Some key steps like the
signal reconstructions with sensitive IMFs and feature extraction using ALC technology are necessary for the complex gear
diagnosis.

4.2. Case 2: fault diagnosis on the spur gearbox

4.2.1. Experiment setup

Another two-stage gearbox test rig with its internal structure is shown in Fig. 9. This rig was roughly composed by a two-
424 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428

Fig. 7. Different cepstrum analysis results under undamaged condition (a) cepstrum of original IAS signal (b) cepstrum of the sensitive IMF (c) ALC of
original IAS signal (d) ALC of the sensitive IMF.

Fig. 8. Different cepstrum analysis results under damaged condition (a) cepstrum of original IAS signal (b) cepstrum of the sensitive IMF (c) ALC of original
IAS signal (d) ALC of the sensitive IMF.

stage gearbox, an induction motor, a magnetic powder brake, two controllers, i.e., one was for speed control while the other
was for load control, and two encoders with the same resolution of 1024 ppr fixedly mounted on the input and out shafts
respectively. The induction motor was for power supplying while the magnetic powder brake was for load supplying.
For this current experiment, the speed of the input shaft was about 600 rpm and the load of the output shaft was set to
10 N m. An artificial fault was seeded to Z1 with one tooth completely removed (see Fig. 4(b)). Besides, the test gear mounted
on the input shaft could be easily replaced from normal to fault. The same data acquisition card was also employed here to
 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428 425

Fig. 9. The experimental rig with spur gearbox and its internal structure.

collect the encoder signals from the two encoders. In this experiment, the encoder on the input shaft was used to validate
the new procedure. Other parameters, such as sample frequency and data length were set just the same as the previous case.

4.2.2. Results and discussions

First of all, the waveforms and spectrums of IAS signals collected from undamaged and damaged gears are shown in
Fig. 10. Besides, each spectrum also includes a larger version around the fundamental meshing frequency. The amplitude
modulation with a period of 3.5 Hz related to the frequency of the middle shaft can be distantly observed from the time
recorder of the normal gear, which may be caused by frequent replacements during different experiments or some un-
avoidable manufacturing and installing errors. Fortunately, this specific amplitude modulation is greatly repressed or re-
placed by the severe fault signature corresponding to the input shaft. In addition, both of the spectrums provide additional
evidences to exhibit the sidebands around meshing frequency. However, the fault characteristics at this step are ambiguous
and the working status is still pending.
In the next step, EMD is applied to each of the IAS signals and corresponding IMFs are obtained. The similarity between
each IMF and the original IAS signal is gained by means of cosine similarity metric. Note that only the first eight IMFs are
considered and evaluated since meshing frequency only presents in the higher order IMFs. Tables 3 and 4 show the

Fig. 10. Original IAS waveforms and spectrums with enlarged version ranging from 300 Hz to 450 Hz. (a) and (c) are waveform and spectrum of un-
damaged spur gear; (b) and (d) are waveform and spectrum of damaged spur gear.
426 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428

Table 3
Results of cosine similarities based on the undamaged spur gear.

IMF 1 2 3 4 5 6 7 8

Simcos 0.6448 0.3438 0.4623 0.2021 0.0924 0.1696 0.1137 0.2501

Table 4
Results of cosine similarities based on the damaged spur gear.

IMF 1 2 3 4 5 6 7 8

Simcos 0.6867 0.2358 0.187 0.1183 0.1351 0.1729 0.2541 0.3919

(a)
0.5
Amplitude[rad/s]

-0.5

0 0.5 1 1.5 2 2.5 3 3.5 4

Time[s]

(b)
0.5
Amplitude[rad/s]

-0.5

0 0.5 1 1.5 2 2.5 3 3.5 4

Time[s]

Fig. 11. Sensitive IMFs for (a) undamaged (b) damaged spur gears.

similarity results of spur gears under undamaged and damaged statuses respectively.
Slightly different from the previous case, IMF 1 with the largest similarity value is selected as the most sensitive IMF for
each condition. The waveforms of the two sensitive IMFs are illustrated in Fig. 11.
By comparing Fig. 11 with Fig. 10, it is found that the amplitude modulation is enhanced for the undamaged gear while
some impulses can be noticed corresponding to fault characteristics for the damaged gear. Nevertheless, the fault period is
still insufficient for the gear fault diagnosis due to the plenty of background noises.
According to the proposed procedure, the ALC is employed to each of the sensitive IMFs thereafter. Here a window length
of 500 Hz was selected, which was also relied on trial and error. By means of different cepstrum calculating methods, Figs. 12
and 13 intensively represent different cepstrum analysis results under different gear conditions.
Similarly to the previous signal in Fig. 7, little remarkable period characteristics can be observed in the four cepstrums
under the undamaged condition, where non periodical transient represents clearly to reveal the abnormal condition even
though the ambiguous characteristic related to the middle gear shaft has been effectively eliminated. Meanwhile, it is worth
noting that not only the traditional cepstrum and IMF-based cepstrum are ineffective to handle the diagnosis issue, but also
the ALC itself fails to extract the fault signature when it is directly used to process the original signal. No meaningful
information concerned to gear fault period is acquired in these three cepstrums in Fig. 13(a), (b) and (c). Fortunately, the ALC
on the basis of sensitive IMF has highlighted the fault characteristics. After applying the proposed method, unevenly spaced
peaks in Fig. 13(c) tend to become dominant components in Fig. 13(d). Finally, the fault of gear Z1 with a broken tooth can be
traced clearly by the highlighted quefrency and its rahmonics. In this way, this experiment once again verifies that the
proposed method has good robustness and adaptability for fault detection and diagnosis.
 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428 427

Fig. 12. Different cepstrum analysis results of spur gear under undamaged condition (a) cepstrum of original IAS signal (b) cepstrum of the sensitive IMF
(c) ALC of original IAS signal (d) ALC of the sensitive IMF.

Fig. 13. Different cepstrum analysis results of spur gear under damaged condition (a) cepstrum of original IAS signal (b) cepstrum of the sensitive IMF
(c) ALC of original IAS signal (d) ALC of the sensitive IMF.

5. Conclusions

In this article, a fusion technology involving EMD, sensitive IMF selection, ALC has been presented to deal with the IAS
signals from different multistage gearboxes. On the basis of experimental investigations for undamaged and damaged gears,
the innovations of this study mainly lie in:

1) As the core analysis source, the IAS signal employed in this paper is capable of highlighting the whole performance for the
428 B. Li et al. / Mechanical Systems and Signal Processing 85 (2017) 415–428

multistage gearbox diagnosis. Nevertheless, the sample frequency for the raw encoder signal is definitely oversampled
and hence it is unnecessary for fault diagnosis using the whole spectrum. Signal reconstruction method has been involved
and proved to be much helpful to overcome this issue.
2) The modified cosine similarity metric is introduced to automatically select the sensitive IMFs since not all IMFs extracted
by EMD are sensitive and closely to the fault and many IMFs are useless or even mislead the fault detection.
3) Two experiment cases with different multistage gearboxes have been implemented successfully to demonstrate the ef-
fectiveness and robustness of the proposed method. Meanwhile, in each case, gears under different conditions are also
concerned.
4) On the basis of the selected IMF, ALC is effectively used for the purpose of signal de-noising and feature extraction. Taking
the consideration of local spectrum information, the new approach is proved to be capable of enhancing fault features
where the gear with fault can be traced clearly by the highlighted quefrency and its rahmonics in each case.

In particular, possible future research topics would mainly concentrate on employing some new methods to enhance
fault identification in minor fault and other different types of failure cases. It can also be expected that this approach could
be introduced in real-world machinery health condition monitoring applications, where a strong background noise always
exists.

Acknowledgments

The article is published with the support from the project (51275379) and (51421004) of the National Natural Science
Foundation of China, which is highly appreciated by the authors. Also the authors would thank anonymous reviewers for
their precious time and valuable suggestions.

References

[1] C.J. Stander, P.S. Heyns, Instantaneous angular speed monitoring of gearboxes under non-cyclic stationary load conditions, Mech. Syst. Signal Process.
19 (2005) 817–835.
[2] P. McFadden, Detecting fatigue cracks in gears by amplitude and phase demodulation of the meshing vibration, J. Vib. Acoust. 108 (1986) 165–170.
[3] J. Hongkai, H. Zhengjia, D. Chendong, C. Peng, Gearbox fault diagnosis using adaptive redundant Lifting Scheme, Mech. Syst. Signal Process. 20 (2006)
1992–2006.
[4] F. Gu, I. Yesilyurt, Y. Li, G. Harris, A. Ball, An investigation of the effects of measurement noise in the use of instantaneous angular speed for machine
diagnosis, Mech. Syst. Signal Process. 20 (2006) 1444–1460.
[5] F. Combet, R. Zimroz, A new method for the estimation of the instantaneous speed relative fluctuation in a vibration signal based on the short time
scale transform, Mech. Syst. Signal Process. 23 (2009) 1382–1397.
[6] A. Fourati, N. Feki, A. Bourdon, D. Remond, F. Chaari, M. Haddar, Electrical modeling for faults detection based on motor current signal analysis and
angular approach, Lecture Notes in Mechanical Engineering, Springer, 2016, pp. 15–25.
[7] M. Dereszewski, A. Charchalis, S. Polanowski, Analysis of diagnostic utility of instantaneous angular speed fluctuation of diesel engine crankshaft, J.
KONES 18 (2011) 123–128.
[8] Z. Li, X. Yan, C. Yuan, Z. Peng, Intelligent fault diagnosis method for marine diesel engines using instantaneous angular speed, J. Mech. Sci. Technol. 26
(2012) 2413–2423.
[9] W. Moustafa, O. Cousinard, F. Bolaers, K. Sghir, J. Dron, Low speed bearings fault detection and size estimation using instantaneous angular speed, J.
Vib. Control (2014).
[10] Y.G. Lei, J. Lin, Z.J. He, Y.Y. Zi, Application of an improved kurtogram method for fault diagnosis of rolling element bearings, Mech. Syst. Signal Process.
25 (2011) 1738–1749.
[11] R. Ricci, P. Pennacchi, Diagnostics of gear faults based on EMD and automatic selection of intrinsic mode functions, Mech. Syst. Signal Process. 25
(2011) 821–838.
[12] Q. Liu, F. Chen, Z. Zhou, Q. Wei, Fault diagnosis of rolling bearing based on wavelet package transform and ensemble empirical mode decomposition,
Adv. Mech. Eng. 5 (2015). 792584-792584.
[13] J. Antoni, R.B. Randall, The spectral kurtosis: application to the vibratory surveillance and diagnostics of rotating machines, Mech. Syst. Signal Process.
20 (2006) 308–331.
[14] S. Mallat, A Wavelet Tour of Signal Processing, Academic Press, United States, 2009.
[15] N.E. Huang, Z. Shen, S.R. Long, M.L.C. Wu, H.H. Shih, Q.N. Zheng, N.C. Yen, C.C. Tung, H.H. Liu, The empirical mode decomposition and the Hilbert
spectrum for nonlinear and non-stationary time series analysis, Proc. R. Soc. Lond. A – Math. Phys. 454 (1998) 903–995.
[16] Z. Wu, N.E. Huang, Ensemble empirical mode decomposition: a noise-assisted data analysis method, Adv. Adapt. Data Anal. 1 (2009) 1–41.
[17] X. Zhang, B. Li, W. Lei, An improved local cepstrum analysis method, J. Xi'an Jiaotong Univ. 50 (2016) 1–5.
[18] B. Li, X. Zhang, Periodical feature extraction and fault diagnosis for gearbox using local cepstrum technology, ASME 2015 International Mechanical
Engineering Congress and Exposition, American Society of Mechanical Engineers, 2015, V04BT04A059-V004BT004A059.
[19] W. Zong, F. Wu, L.K. Chu, D. Sculli, A discriminative and semantic feature selection method for text categorization, Int. J. Prod. Econ. 165 (2015)
215–222.
[20] J. Ye, Cosine similarity measures for intuitionistic fuzzy sets and their applications, Math. Comput. Model. 53 (2011) 91–97.
[21] S.K. Roy, A.R. Mohanty, C.S. Kumar, Fault detection in a multistage gearbox by time synchronous averaging of the instantaneous angular speed, J. Vib.
Control (2014).
[22] Y. Li, F. Gu, G. Harris, A. Ball, N. Bennett, K. Travis, The measurement of instantaneous angular speed, Mech. Syst. Signal Process. 19 (2005) 786–805.
[23] H. Konstantin-Hansen, H. Herlufsen, Envelope and cepstrum analyses for machinery fault identification, Sound Vib. 44 (2010) 10.
[24] Y.G. Lei, M.J. Zuo, Fault diagnosis of rotating machinery using an improved HHT based on EEMD and sensitive IMFs, Meas. Sci. Technol. 20 (2009)
125701.