Mechanical Systems and Signal Processing 211 (2024) 111187

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

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

Weighted multi-order Viterbi algorithm (WMOVA): Instantaneous

angular speed estimation under harsh conditions
Jinoh Yoo a, Jongmin Park a, Taehyung Kim a, Jong Moon Ha b, *, Byeng
D. Youn a, c, d, *
a
Department of Mechanical Engineering, Seoul National University, Seoul 08826, Republic of Korea
b
Non-destructive Metrology Group, Korea Research Institute of Standards and Science (KRISS), Daejeon 34113, Republic of Korea
c
Institute of Advanced Machines and Design, Seoul National University, Seoul 08826, Republic of Korea
d
OnePredict Inc., Seoul 06160, Republic of Korea

A R T I C L E I N F O A B S T R A C T

Keywords: Instantaneous angular speed (IAS) information is essential for vibration-based fault diagnosis of
Instantaneous speed estimation rotating machinery under non-stationary operating conditions. Possessing the advantages of
Non-stationary operation being cost-effective and less intrusive, tacholess IAS estimation methods that use vibration signals
Time-frequency representation
have been researched and implemented in recent years. However, such methods often suffer from
Harmonic weighting
Rényi entropy
poor performance when subjected to the harsh operating conditions found in real-world settings,
Vibration signal such as large speed variations, extreme noise, and poorly excited harmonics of vibration signals.
To address these challenges, this paper proposes a weighted multi-order Viterbi algorithm
(WMOVA) method for tacholess IAS estimation under harsh conditions. The novel harmonic
weighting of WMOVA enables selective extraction of the correct IAS information, while exploiting
multiple harmonics to construct an accurate ridge in the time–frequency representation (TFR).
The TFR ridge is then tracked by using a modified Viterbi algorithm. The benefits of the proposed
method are demonstrated in this research, first by applying the new approach to simulated vi­
bration signals and then using the proposed approach with data in two case studies. The first case
study examines public data from the 2014 international conference on Condition Monitoring of
Machinery in Non-Stationary Operations (CMMNO); the second study uses data measured from a
gearbox testbed. Comprehensive comparative studies show that the proposed method out­
performs the conventional IAS estimation approach in real-world applications.

1. Introduction

Rotating machinery is a critical component in numerous mechanical systems across various industrial fields [1]. Since the rotating
machinery generally operates in harsh working environments, many vibration-based fault diagnosis techniques have been explored to
prevent unexpected catastrophic failures leading to huge downtime loss [2]. Conventional vibration-based fault diagnosis techniques
are based on frequency-domain analysis, such as detecting the fault characteristic frequency in the frequency spectrum using the
Fourier transform [3]. However, the application of the frequency-domain analysis method is challenging in the actual industrial field

* Corresponding authors at: Non-destructive Metrology Group, Korea Research Institute of Standards and Science (KRISS), Daejeon 34113, Re­
public of Korea (J.M. Ha), Department of Mechanical Engineering, Seoul National University, Seoul 08826, Republic of Korea (B.D. Youn).
E-mail addresses: jmha@kriss.re.kr (J. Moon Ha), bdyoun@snu.ac.kr (B.D. Youn).

https://doi.org/10.1016/j.ymssp.2024.111187
Received 25 February 2023; Received in revised form 23 January 2024; Accepted 25 January 2024
Available online 10 February 2024
0888-3270/© 2024 Elsevier Ltd. All rights reserved.
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Nomenclature

x the vibration signal

n sample number
t time [s]
f frequency [Hz]
fs sampling frequency [Hz]
T sampling period [s]
X short-time Fourier transform
S spectrogram
υ window function
ω instantaneous angular speed of the shaft [Hz]
Ω the actual instantaneous angular speed of the shaft [Hz]
i harmonic order
ok kinematic order with regard to shaft orders
Hi,k ith harmonic of a component with kinematic order ok
fi,k instantaneous frequency of harmonic Hi,k [Hz]
Ai,k amplitude modulation of harmonic Hi,k
Rα Rényi entropy with order α
nci,k complexity of the harmonic Hi,k

because of the spectral smearing problem encountered under non-stationary operating conditions, where most industrial machinery
primarily operates [4]. To overcome this challenge, instantaneous angular speed (IAS) information has become an essential component
of accurate fault diagnosis [5]. IAS information enables the acquisition of equally spaced vibration signals in the angular domain,
which, in turn, facilitates the use of conventional fault diagnosis methods in non-stationary conditions [5]. Previously, IAS information
was measured with encoders or tachometers [6]. However, utilizing a tachometer leads to additional cost and installation problems
[7], which is not preferable in the actual field [8]. As a result, a number of studies (summarized below) have been performed to develop
a tacholess IAS estimation method to extract IAS information directly from vibration signals [7].
The technique based on phase demodulation is the most widely employed IAS estimation method [9]. In this method, a speed-
related harmonic component is isolated by applying the band-pass filter. Then, the IAS is calculated using the phase information of
a speed-related harmonic acquired from an analytic signal based on Hilbert transform [10]. Randall et al. [11,12] applied the Teager
Kaiser Energy Operator (TKEO), which is frequently used for amplitude and phase demodulation in speech analysis. This approach
avoids the errors and excessive time required for causal processing in the time domain by implementing frequency-domain non-causal
processing. Coats et al. [13] performed iterative phase demodulation in multiple stages. These researchers first used the lower har­
monics to increase the allowable speed variation and then used the higher harmonics to obtain accuracy. Nevertheless, phase
demodulation-based methods have a critical disadvantage in that they can only handle limited speed variations [14]. If the speed
variation exceeds 33 %, the overlapping of multiple harmonics in the frequency domain hinders the extraction of mono-component
vibration signals by using a band-pass filter. To mitigate the disadvantage, the above methods divided the vibration signal into
overlapping segments and applied a short-time windowed band-pass filter [11,15]. Still, it poses limitations under large speed vari­
ation profiles, which continues to limit the utility of phase demodulation-based IAS estimation methods in real-world applications.
To resolve this issue, methods based on the time–frequency representation (TFR) can be employed to obtain the IAS information by
tracking one high-amplitude harmonic ridge in the TFR. Urbanek et al. [16] tracked the local maxima in the spectrogram, assuming the
frequency values with the highest amplitude on the spectrogram correspond to instantaneous frequency (IF) at each time. Yet, the
method reveals the inherent challenges associated with time and frequency resolution in TFR. Consequently, a number of studies have
been proposed to increase the resolution of the TFR by implementing various time–frequency analysis. Peng et al. [17] introduced
polynomial chirplet transform (PCT) by deploying the polynomial function instead of the linear chirp kernel, resulting in an enhanced
energy concentration in TFR to estimate instantaneous frequency. Gryllias et al. [18] utilized complex shifted Morlet wavelets for
speed estimation, allowing for selecting the optimal wavelet center frequency and bandwidth. Furthermore, various other TFR-based
methods were explored to improve the ridge-tracking in the TFR. Iatsenko et al. [19] proposed a path optimization method that
considers all possible peak sequences over the whole profile for more accurate IF estimation. Djurovic et al. [20] employed the Viterbi
algorithm to address the cross-terms in Wigner distribution. Schmidt et al. [21] developed the maxima tracking by reflecting the
acceleration information in a probabilistic approach. However, the aforementioned TFR-based IAS estimation methods rely on a single
harmonic, vulnerable to performance degradation due to harsh operating conditions. Poor performance arises when the harmonic
component used for the ridge tracking is disturbed by background noise or confused by ridges that are unrelated to speed [7].
To this end, recent TFR-based approaches employ multiple harmonics. Zimroz et al. [22] detected the IAS results from each fre­
quency band that correspond to multi-harmonic orders and averaged them to overcome the problem of relatively low SNR. Still,
incorrect IAS results that are detected from the higher harmonics with decreased energy induce error in the averaging of the final IAS
estimation result. To manage this issue, Leclére et al. [23] divided the TFR by the frequency bands and multiplied them to obtain the
merged IAS ridge in the probability density function (PDF) map; this method shows the PDF of IAS at each time step. This approach,

2
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

called the multi-order probabilistic approach (MOPA), has been found to be one of the most accurate IAS estimation techniques [7,24]
proposed to date. These multi-harmonic methods utilize as many harmonics as possible to improve accuracy and robustness, given that
all harmonics are well excited [7]. However, due to low energy, some poorly excited harmonics may fade in extreme noise under harsh
conditions, leading to significant errors [7,22]. Thus, using multiple harmonics without considering harmonic weighting worsens the
IAS estimation when weak or absent harmonics exist. To overcome this challenge, Peeters et al. [25] attempted to advance MOPA by
assigning different weights to each harmonic order based on importance. However, the method proposed in [25] requires a challenging
prerequisite in that the importance of each harmonic order must be pre-obtained from representative machinery that is equipped with
a tachometer.
Other research studies have combined approaches based on TFR and phase demodulation to capitalize on the advantages of these
methods. Urbanek et al. [9] proposed a two-step procedure where the TFR-based method is first used to get a rough estimation in
situations with large speed variations. The rough IAS is then used for angular resampling to minimize the speed variation. Next, the
method based on phase demodulation with band-pass filtering is used to arrive at a finer estimation. Based on this two-step procedure,
recent studies have sought to utilize multiple harmonics during phase demodulation in the second step. Peng et al. [24] compensated
for the incorrect estimations, where impulse-like errors occur due to severe amplitude modulation, by replacing these segments with
results from other harmonics. Peeters et al. [15] proposed SNR-based time-varying harmonic weighting, while also incorporating the
concept of multiple harmonics into the phase demodulation approach. Still, there is a serious limitation in that the accuracy of this two-
step estimation approach is highly dependent on the TFR-based method in the first step, when dealing with profiles with large speed
variations. Since the band-pass filtering in the second step is implemented based on the rough IAS information acquired from the TFR-
based methods, inaccurate TFR-based estimation impedes extracting the speed-related harmonics completely. Thus, it is crucial to
obtain an accurate IAS by using TFR-based methods in the first step.
Despite the extensive research on various IAS estimation methods, as previously discussed, there remains a need for an accurate
TFR-based IAS estimation method under harsh operating conditions characterized by large and rapid speed variation, excessive noise,
and operating regime changes that alter the harmonics’ SNR over time [7,15]. To solve the abovementioned challenges, this paper
proposes a TFR-based IAS estimation method utilizing multi-harmonic information with novel harmonic weighting, requiring no prior
knowledge of the relative importance of each harmonic component. For this purpose, the proposed method employs an energy-based
harmonic weighting approach, which assigns large weights to harmonics with high instantaneous energy that potentially contain more
accurate IAS information while mitigating the risk of intermittent harmonics’ fading. In addition, this paper proposes an innovative
strategy considering the signal complexity to cope with practical challenges that act as disturbance factors, such as overlapping,
crossing, and extreme noise of signals [7,15,21]. These interfering components can induce an undesired energy overestimation while
also increasing the harmonics’ complexity, causing unexpected harmonic weights when considering only the harmonics’ energy. Our
proposed harmonic weighting utilizes the time-varying energy normalized by the complexity of harmonics to solve these difficulties.
Therefore, harmonic weighting enables focus on high-energy harmonics and suppresses faded, overlapped, or crossed harmonics with
high complexity. Consequently, the proposed method can estimate accurate IAS under harsh non-stationary conditions, including large
speed variation, extremely low SNR, and the presence of speed-unrelated components. We call this new method the weighted multi-
order Viterbi algorithm (WMOVA). First, novel energy-based harmonic weights are formed by calculating the harmonics’ instanta­
neous energy and assessing their complexity through Rényi entropy. Next, a global probability density function (PDF) map is con­
structed using the weighted product of PDFs of IAS acquired from each harmonic order. Finally, the Viterbi algorithm is applied to
track the IAS ridge in the PDF map.
Therefore, the primary contribution of this study lies in developing a novel TFR-based harmonic weighting approach, integrating
the concept of energy and complexity. This novel harmonic weighting strategy enables our method to fully utilize the more accurate
and reliable IAS-relevant information inherent in vibration harmonics, thereby enhancing the IAS estimation accuracy even in the
presence of significant disturbance components and substantial speed variation. These challenges are inherent in real-world appli­
cations, such as wind turbines and rotating equipment in heavy industry. The rest of this paper is structured as follows. Section 2
reviews the background knowledge, including methods to measure complexity and the Viterbi algorithm. Section 3 explains the
proposed method in detail. Section 4 demonstrates the effectiveness of the proposed method using a simulation model. Section 5
verifies the proposed method by applying it to two experimental case studies. Lastly, Section 6 summarizes the conclusions of the work.

2. Background knowledge

In this section, we describe the background knowledge that is required for a better understanding of the proposed method. First, we
introduce the basic concept of TFR-based IAS estimation methods. Then, subsequently, the theories of measuring complexity using
Rényi entropy and the Viterbi algorithm are explained. In particular, the rationale for selecting third-order Rényi entropy as the
complexity measure will be described.

2.1. Basic concept of TFR-based methods

Time-frequency analysis allows us to decide which frequencies are dominant at each time in a non-stationary signal. The TFR is the
energy distribution on the time–frequency plane, as described by the time–frequency analysis. Short-time Fourier transform (STFT) is
the most widely used TFR analysis method because of its simplicity and functionality [26]. The continuous STFT X(t, f) can be written
as

3
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

∫ ∞
X(t, f ) = x(τ)υ(τ − t)e− j2π f τ
dτ (1)
− ∞

√̅̅̅̅̅̅̅
where x(t) is the vibration signal, υ(t) denotes the window function such as hanning window, and j = − 1. The spectrogram, the
corresponding TFR, is merely the squared magnitude of STFT as:
⃒∫ ∞ ⃒2
⃒ ⃒
S(t, f ) = |X(t, f )|2 = ⃒⃒ x(τ)υ(τ − t)e− j2πf τ dτ⃒⃒ (2)
− ∞

In TFR, multi-harmonics in the vibration make numerous ridges or local maxima of instantaneous frequencies (IF) related to the IAS.
The connection between the IAS of the shaft and the IF of the harmonics is expressed as follows [21]:
fi,k = i ⋅ ok ⋅ Ω = Hi,k Ω (3)

where Ω is the actual IAS, ok implies the kinematic order with respect to the shaft, such as a gear ratio or a meshing order (e.g., in a
drivetrain), and Hi,k implies the ith harmonic order of the frequency component ok Ω. TFR-based methods calculate IAS by using Eq. (3)
after tracking the harmonic ridges.

2.2. Measuring complexity using Rényi entropy in TFR

Signal complexity is the most important factor when analyzing multi-component non-stationary signals. Complexity or information
content in the TFR is associated with the number of components and the inverse of energy concentration, quantifying the uncertainty
or randomness of the signal [27–29]. Namely, signals characterized by a multitude of components and low concentration must have a
high complexity value since they have relatively more information and high uncertainty [27,28]. Among feasible methods, complexity
measure using Rényi entropy has been widely utilized because of its useful properties, such as component counting, amplitude
sensitivity, and noise sensitivity [27,28,30–33]. These properties are particularly suitable for the TFR of the vibration signal, having
multiple components with different amplitude values and background noise. Rényi entropy-based complexity measure exploits the
analogy between TFR and the probability theory [27,28]. Therefore, TFRs, such as a spectrogram, are viewed as 2-D probability
density functions by using the following equation [34]:
S(t, f )
Cs (t, f ) = ∫ ∞ ∫ ∞ (4)
− ∞ − ∞
S(t, f )dtdf

where the denominator normalizes the TFR with the signal energy, Cs denotes the 2-D PDF using the whole frequency range, and S(t, f)
denotes the spectrogram. In turn, Rényi entropy employing the 2-D PDF takes the form of [32,34]
∫∞ ∫∞
1
Rα (Cs ) = log2 Csα (t, f )dtdf (5)
1− α − ∞ − ∞

where Rα is the Rényi entropy of order α > 0 (α→1 produce Shannon entropy). The component counting property of Rényi entropy-
based complexity assumes the absence of interfering cross-terms in TFR [27,28]. However, actual TFRs do not satisfy this condition,
including the spectrogram and Wigner-Ville distribution [28,35]. This issue can be addressed by utilizing the odd α ≥ 3 that decays the
information of cross-terms asymptotically to zero [28]. Therefore, the third-order Rényi entropy (R3 ) is adopted throughout this paper.
Let us denote the ideal TFR for the mono-component signal x1 (t) without noise as Cx1 (t,f). Then, the two-component signal TFR, which
is composed of x1 (t) and its shifted signal x2 (t) = x(t − Δt), is denoted as Cx1 +x2 (t, f). Then, the counting property of Rényi entropy
describes the relationship between the information quantities of two signals as follows [28]:
R3 (Cx1 +x2 (t, f )) = R3 (Cx1 (t, f )) + 1 (6)

In other words, the difference between these signals is exactly one bit. Thus, by generalizing Eq. (6), the complexity of the signal is
formulated as [31]

nc = 2R3 (Cx1 +x2 +⋯+xnc )− R3 (Cx1 )

(7)

Eq. (7) performs well when components have a similar form, and the Rényi entropy of a mono-component signal (R3 (Cx1 )) is available
[34].

2.3. Viterbi algorithm

The Viterbi algorithm is widely used to track a ridge of TFR for vibration-based IAS estimation. The Viterbi algorithm consists of
two penalty functions (i.e., q and g), formulated as follows [20,36]:

4
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

{ nf nf − 1
}
∑ ∑
̂l(nt ) = arg min q(S(nt , l(nt ))) + g(l(nt ), l(nt + 1)) (8)
l(nt )∈L
nt =ns nt =ns

[ ]
where nt is the time step, L is a set of all possible instantaneous frequency (IF) paths l(nt ) within the time interval nt ∈ ns , nf , ̂l(nt ) is
the optimal IF path, and S(nt , l(nt )) is the corresponding spectrogram value. The penalty function q is designed to trace the largest-
amplitude ridge in the TFR, and the penalty function g prevents an abrupt change of IF. In detail, q assigns a lower penalty for
having a higher amplitude at the time step by defining the following relationship:
( ( ))
q(S(nt , l(nt ))) = q S nt , lj = j − 1 (9)

where lj is the frequency bin that has the jth biggest magnitude in the sorted magnitude sequence at the time step nt , given by
⃒ ( )⃒
|S(nt , l1 ) |⩾|S(nt , l2 ) |⩾⋯⩾⃒S nt , lj ⃒⩾⋯⩾|S(nt , lF ) |, j ∈ [1, F] (10)

where F is the number of frequency bins in the spectrogram. The path function g determines the penalty by measuring the path length
between two adjacent time steps as [20,36]:
{
0, for |a − b|⩽Δ
g(a, b) = (11)
η(|a − b| − Δ ), for |a − b| > Δ

where Δ is the threshold for the expected maximum variation of IF, and η is the weight coefficient. Despite the promising performance
of the Viterbi algorithm, it suffers from a huge computational cost.
The background knowledge explained above is employed in the proposed method, which will be discussed thoroughly in the
following section.

3. Proposed method

During harsh non-stationary operations of rotating machinery, large speed variations, high levels of time-varying amplitude
modulation, poorly excited low-energy harmonics, and extreme noise are often prevalent. These harsh conditions induce difficulties
such as crossing, overlapping, and fading of the harmonics [7]. Hence, harmonic weighting in the TFR is designed to consider har­
monics differently according to the time-varying energy of the harmonics. Consequently, a more accurate IAS ridge is constructed in
the global PDF map by reflecting the multiple harmonics and their weights. Then, the modified Viterbi algorithm is utilized to extract
the IAS ridge. In this context, our proposed approach is named the weighted multi-order Viterbi algorithm (WMOVA). In this section,

Fig. 1. Flowchart of the proposed method.

5
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

we explain the proposed WMOVA method in detail. Fig. 1 shows the overall procedure of the proposed method.

3.1. Probability density function of the IAS derived from the instantaneous spectrum

In Section 3.1, Eqs. (12)–(14) are exploited from the MOPA for construction of the PDF [23]. Before building PDFs of IAS, the
spectrogram is whitened to reduce the constant harmonics that arise from structural resonances [23,37].
S(t, f )
W(t, f ) = (12)
〈S(t, f )|S(t, f ) < Sn (f )〉t

where W(t, f) is a whitened spectrogram, Sn (f) is a magnitude value exceeded by n% of S(t, f) for the whole time at the frequency f, and
〈⋯〉t is the mean operator over time. In this paper, we use 90% for all cases.
The PDFs of IAS are constructed by use of the instantaneous spectrum, as it contains the information from multiple harmonics [23].
With the given prior knowledge of the minimum and maximum angular speed Ωmin , Ωmax , the PDF at time t of the harmonic Hi,k is
( )
constructed by using the whitened instantaneous spectrum with the frequency range of Hi,k Ωmin , Hi,k Ωmax , as follows [23]:
⎧ ( ⃒ )
⎪ 1 ( )
⎨ p Ωt ⃒Hi,k = W t, Hi,k ω for Ωmin < ω < Ωmax
ξi,k (13)
⎪
⎩ ( ⃒⃒ )
p Ωt Hi,k = 0 for ω < Ωmin | ω > Ωmax

where Ωt denotes the actual IAS Ω at time t, W(t, f) is the whitened spectrogram, and ξi,k is a normalization factor.
∫ Ωmax ( )
ξi,k = W t, Hi,k ω dω (14)
Ωmin

Eq. (13) is available when the maximum frequency range Hi,k Ωmax is below the upper limit frequency fmax , which is the Nyquist
frequency fNy in most cases. If the maximum range Hi,k Ωmax goes beyond the upper limit, the harmonic Hi,k is not considered for speed
estimation.

3.2. Energy-based harmonic weighting

The PDFs provided by Eq. (13) provide information about the IAS from each harmonic in a probabilistic manner. However, if some
harmonics fade due to low energy, PDFs extracted from them will deliver incorrect information for IAS estimation. In this case, uniform
weighting for each PDF induces a significant error. Therefore, novel harmonic weighting is proposed to focus on the PDFs derived from
high-energy harmonics. To reflect the time-varying energy of the harmonic Hi,k , the square root of the instantaneous energy in the
( )
corresponding frequency range Hi,k Ωmin , Hi,k Ωmax is calculated as [38]:
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∫ ̅
Hi,k Ωmax
Mi,k (t) = W(t, f ) df (15)
Hi,k Ωmin

However, when a large speed variation exists, an overlap of adjacent multiple harmonics occurs in the frequency range. Moreover,
speed-unrelated vibrations from other sources and high noises also occur under harsh conditions. In these situations, Eq. (15) over­
estimates the energy and gives high weights to overlapped or crossed harmonics; this has an effect opposite to our intention. Therefore,
the complexity of the harmonic Hi,k should be considered to suppress the negative effect of overlapping, crossing, and noise. The
overlapping and crossing contribute to an increase in the number of components in the interested frequency range, while the extreme
noise reduces the energy concentration level of the speed-related harmonic. These factors can be characterized by the complexity or
uncertainty increases in the corresponding frequency range [28,30,39], which causes inaccurate IAS estimation results. To address this
issue, we propose a novel harmonic weighting strategy that utilizes the instantaneous energy measure normalized by the complexity of
harmonics. This weighting strategy will emphasize harmonics that exhibit both high energy and low complexity. For example, a
harmonic component with high energy but also with high complexity at the same time will be assigned a low weight value due to the
increased likelihood of its energy overestimation and high uncertainty. For that reason, altering Eqs. (4) and (5) from the whole
frequency range to the range of the harmonic Hi,k yields:

W(t, f ) ( )
Ci,k (t, f ) = ∫ Hi,k Ωmax ∫ ∞ Hi,k Ωmin ⩽f ⩽Hi,k Ωmax (16)
Hi,k Ωmin − ∞
W(t, f )dtdf

∫ Hi,k Ωmax ∫ ∞
( ) 1
Rα Ci,k = log α
Ci,k (t, f ) dtdf (17)
1− α 2 Hi,k Ωmin − ∞

where Ci,k and Rα (Ci,k ) indicate the 2-D PDF of the TFR and Rényi entropy regarding the harmonic Hi,k . Still, to properly assess the
complexity, Eq. (7) should be revised. In most practical cases, getting the Rényi entropy of the ideal mono-component vibration signal

6
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

(R3 (Cx1 )) is impossible. Therefore, the mono-component Rényi entropy is replaced with the minimum Rényi entropy from among the
harmonics that have the same fundamental kinematic order with regard to the shaft.
Rα (Ci,k )− min(Rα (Ci,k ))
nci,k = 2 ∀i (18)

where nci,k is the complexity of the harmonic Hi,k . Rather than choosing the smallest entropy from the available harmonics, the
minimum that has the same kinematic order is chosen to guarantee that the form between harmonic components is as similar as
possible. Finally, the harmonic weight can be defined by dividing Eq. (15) by Eq. (18) to consider both the energy and the complexity of
harmonic Hi,k :

Mi,k (t)
ci,k (t) = (19)
nci,k

where ci,k is the time-varying harmonic weight for harmonic Hi,k , and Mi,k (t) is the instantaneous energy of the harmonic Hi,k calculated
by Eq. (15). The normalized version of ci,k is

ci,k (t)
wi,k (t) = ∑ ∑ (20)
k i ci,k (t)

The weighted product using Eq. (13) and (20) allows the focus to be on PDFs from harmonics with high energy and low complexity
suitable to IAS estimation while utilizing as many harmonics as possible.
Ni ∏
∏ Nk
( ⃒ )w (t)
p(Ωt )∝ p Ωt ⃒Hi,k i,k (21)
i=1 k=1

where p(Ωt ) is the global PDF that indicates the probabilistic IAS Ω at time t, Nk is the number of considered kinematic orders, and Ni is
the number of considered harmonics orders. Then, the global PDF map is built by combining the global PDF p(Ωt ) across the entire
range of the time step.

3.3. Tracking ridges using the Viterbi algorithm

The Viterbi algorithm was chosen to track ridges in the global PDF map because of its robustness in noisy conditions [20]. However,
( )
because the Viterbi algorithm has high computational complexity O F2 T , the large number of bins F required for accurate estimation
leads to an enormous computational cost. To reduce this computational cost, only a few IAS candidate bins with high likelihood values
are selected at each time step in the global PDF map. In this paper, the 3 median absolute deviation (MAD) threshold was used to decide
candidate bins utilized for the modified Viterbi algorithm in each time step.
Unlike the original VA, which uses Eq. (8) and is applied to the TFR, modifications must be made for use in the global PDF map. The
Viterbi algorithm for the PDF map can be redefined as:
{ n nf − 1
}
∑ f ∑
̂l ω (nt ) = arg min q(p(Ωt = lω (nt )) ) + g(lω (nt ), lω (nt + 1)) (22)
lω (nt )∈Lω
nt =ns nt =ns

[ ]
where Lω is a set of all possible IAS path functions lω (nt ) within the time interval nt ∈ ns , nf , and ̂l ω (nt ) is the estimated IAS path.
Instead of Eq. (9) the penalty function q assigns a lower penalty for having a higher likelihood value at the time step by the following
relationship:
( ( ))
q(p(Ωt = lω (nt )) ) = q p Ωt = lωj = j− 1 (23)

where lωj represents the IAS candidate bins in the global PDF, having the jth biggest likelihood in the sorted likelihood sequence at the
time step nt , given by
⃒ ( )⃒ ⃒ ( )⃒ ⃒ ( )⃒ ⃒ ( )⃒
⃒p Ωt = lω ⃒⩾⃒p Ωt = lω ⃒⩾⋯⩾⃒⃒p Ωt = lω ⃒⃒⩾⋯⩾⃒p Ωt = lω ⃒, j ∈ Fω (24)
1 2 j Fω

where Fω is the number of IAS candidate bins in the global PDF, p(Ωt ). The path penalty function g remains the same as that shown in
Eq. (11). In this paper, we decide the threshold parameter Δ in the path penalty function using the prior information of maximum
acceleration as follows:
/
Δ = amax Δt Δpdf (25)

where Δt is the time step of the spectrogram, and Δpdf is the bin width of the PDF.

7
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

4. Simulation model

A simulated non-stationary vibration signal was generated to explain the proposed method effectively and to highlight its ad­
vantages. Initially, the simulated data is generated to emulate the vibration signals under harsh conditions in real-world applications.
Then, the proposed method is applied to the simulated data, followed by the quantitative analyses. Finally, parameter studies are also
conducted to show the robustness of the proposed method.

4.1. Simulated data

The simulated vibration data has 200, 000 samples with sampling frequency fs = 10, 000Hz. The speed profile is designed to
demonstrate a range of speed variations and accelerations, which progressively increase over time, as shown in Fig. 2. The poles in the
speed profile are labeled from A to J, where corresponding shaft speeds and variation values are shown in Table 1. Using this shaft
speed profile, the overall form of the simulated non-stationary vibration signal is generated [15,21], which consists of four harmonics
of shaft frequency and gear mesh frequency, namely Ni = 4 and Nk = 2. In addition, speed-unrelated components that induce crossings
and extreme white Gaussian noise, ν with a − 10dB SNR, are added to the simulated signal. Thus, the vibration signal under harsh
conditions was simulated as [15,21]:
( ) ( )
∑Ni ∑Nk ∑
N− 1 ∑
2 ∑
N− 1
x(n) = Ai,k (n)cos 2πT fi,k (n) + Aun
s cos 2 π T fs
un
(n) + ν(n) (26)
i=1 k=1 n=0 s=1 n=0

f1,k=2 = ok=2 f1,k=1 = 20f1,k=1 (27)

where n = 0, 1, ⋯, N − 1 is the sample number, k = 1, 2 is the shaft and GMF component, ok=2 is the kinematic order of GMF with
regard to the shaft, T is the sampling period, fi,k is the instantaneous frequency of harmonic Hi,k , Ai,k is the time-varying amplitude
modulation, and Aun un
s , fs are the amplitude and instantaneous frequency of the speed-unrelated signal with order s, respectively. The
first term in Eq. (26) denotes the vibration signal made by the shaft and GMF components. A significant amplitude modulation (AM)
function of cos8 is applied to these harmonics to consider the vibration attenuation, which causes the effect of harmonic fading [24].
Further, to reflect the general characteristics of the vibration signal, a smaller amplitude is assigned to the higher harmonic order,
which causes the higher harmonics to fade in extreme background noise. Additionally, initial phases are assigned differently to shaft
components and GMF components.
Thereby, the amplitude modulation functions are modeled as follows [24]:

Ai,k (n) = mi,k cos8 (0.25πn/fs + ϕk ) + 22− i

(28)

where ϕk is the initial phase, mi,k is the magnitude of the AM function, and fs is the sampling frequency. In the simulation model, ϕ1 =
2.75 and ϕ2 = 1.25 are used for the initial phases and mi,1 = 24− i and mi,2 = 5 • 22− i are used for the magnitude values.
In practical applications, vibration signals typically contain unwanted speed-unrelated components, which disturb the rotating
machinery condition monitoring [21,40]. In the context of IAS estimation, these speed-irrelevant noise components interfere with
accurate estimation by crossing with speed-related harmonics [21]. Hence, to emulate the unwanted components of the vibration
signal under harsh conditions, the second term in Eq. (26) contains two speed-unrelated components; the speed profiles of these signals
are expressed as [21,40]:
( / )
2 80 un
f un
s=1 (n) = 3cos 0.6π n fs + + 82.5, f un
s=2 = f (29)
3 3 s=1

Fig. 2. Shaft speed profile of the simulation model.

8
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Table 1
IAS values and speed variation percentages at each point in Fig. 2.
A B C D E F G H I J

IAS Ω [Hz] 19.4 21.5 17.5 23.5 15.5 25.5 13.5 27.5 11.6 29.4
Variation [%] 3 7.5 12.5 17.5 22.5 27.5 32.5 37.5 42 47

The amplitude values of speed-unrelated signals are selected to be Aun un

1 = A2 = 3.5. In Fig. 3(a), the simulated vibration signal is
presented in the time domain. The spectrogram of the simulated vibration signal is presented in Fig. 3(b). Fig. 4 illustrates the enlarged
spectrogram to show the lower frequency range clearly, where shaft harmonics exist. In Fig. 3(b) and Fig. 4, the regions of the shaft and
GMF harmonics with low amplitude modulation values are faded, as a result of the extreme background noise. Significantly, the poorly
excited 3rd and 4th harmonics of the shaft and GMF are almost completely faded due to the low energy. Further, the 3rd and 4th shaft
harmonics and 4th GMF harmonics are crossed with speed-unrelated components.

4.2. Assessment of the proposed method on simulated data

The proposed method is demonstrated using the simulated signal, while the whitening process is omitted due to the absence of
constant harmonic noise. Suppose we have the prior knowledge of the minimum and maximum shaft speed, Ωmin = 10Hz and Ωmax =
32Hz, respectively. The proposed method first divides the spectrogram into frequency bands corresponding to each harmonic by using
Eq. (3), as shown in Fig. 5. The instantaneous frequency components for each harmonic Hi,k can be found in the corresponding
frequency-ranged spectrogram, from which PDFs can be obtained for IAS estimation through the use of Eqs. (13) and (14).
However, the presence of harmonics crossing, overlapping, and fading in the instantaneous spectrums can lead to an inaccurate
prediction of the IAS. To demonstrate these challenges, while presenting the detailed advantages of the proposed method, this paper
analyzes two representative examples t = 14.7s and t = 12.375s, as indicated by the dotted lines in Fig. 5. First, Fig. 6 shows the
instantaneous spectrums at t = 14.7 corresponding to point H in Fig. 1, where the actual IF fi,k (14.7s) = Hi,k Ωt=14.7 = 27.5Hi,k Hz is
plotted in red-dashed lines. Instantaneous spectrums of the 2nd, 3rd, and 4th shaft harmonics in Fig. 6(a), and of the 3rd to 4th GMF
harmonics in Fig. 6(b) have low magnitude at the actual IF due to their low energies. PDFs from these harmonic components deliver
incorrect IAS information, leading to inaccurate estimations. To address this issue, the harmonic energy contained in each instanta­
neous spectrum can be considered by Eq. (15). However, overlapping caused by large speed variation, and crossing from the unrelated
vibration component can cause energy overestimation, regardless of the importance of the information. Overlapping and crossing
phenomena are evident in the 2nd to 4th GMF instantaneous spectra in Fig. 6(b) and in the instantaneous spectra of the 2nd to 4th shaft
harmonic in Fig. 6(a). In addition, extreme noise also contributes to energy overestimation, especially in higher harmonics that contain
wide frequency bands. Therefore, the complexity should be considered simultaneously to weaken the negative effects of overlapping,
crossing, and extreme noise. The harmonic weighting proposed in Eq. (19) is defined by dividing the square root of the energy by the
complexity calculated by applying Eqs. (16)–(18) to the frequency-ranged spectrograms in Fig. 5. Table 2 summarizes the instanta­
neous energy, complexity, weight, and normalized weight of each harmonic at t = 14.7. It can be seen that low harmonic weights are
assigned to the shaft harmonics that have a low instantaneous energy value Mi,1 , compared to the GMF harmonics. Among the GMF
harmonics, the 1st GMF harmonic contains the most valuable information with the highest SNR, as shown in Fig. 6(b). However,
Table 2 shows that higher GMF harmonics have larger energy due to the overlapping, crossing, and extreme noise components. In this

Fig. 3. (a) The simulated vibration signal in the time domain, (b) spectrogram of the simulated signal.

9
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 4. Enlarged spectrogram of the simulated vibration signal.

study, the issue of energy overestimation is addressed by incorporating the Rényi entropy-based complexity nci,k , which reflects the
number of components and extreme noise. Since the overlapping, crossing, and extreme noise, which increase the complexity or
randomness, induce energy overestimation, the proposed harmonic weighting normalizes the square root of the instantaneous energy
by the complexity value. This normalization aims to focus on harmonics with high energy and low complexity, which have more
accurate and reliable information. In Table 2, the high complexity values are assigned to the higher GMF harmonics, having over­
lapping and crossing components, thus successfully alleviating the energy overestimation. As a result, the harmonic weighting focuses
primarily on the 1st GMF, followed by the 2nd GMF, while keeping the weights of higher GMF harmonics low by considering both the
instantaneous energy and the complexity. Fig. 7 presents the global PDF of IAS p(Ωt=14.7 ) at t = 14.7 with and without considering the
harmonic weights. The global PDF using all harmonics equally without the weights in Fig. 7(a) shows the highest likelihood value at
the wrong IAS, caused by the speed-unrelated peaks in the 3rd shaft harmonic, and in the 4th GMF harmonic in Fig. 6. On the other
hand, the global PDF that incorporates the harmonic weights shows a prominent likelihood value at the actual IAS.
Fig. 8 presents the instantaneous spectrums at t = 12.375 with actual IF fi,k (12.375s) = Hi,k Ωt=12.375 = 15.1Hi,k Hz , where shaft
harmonics have bigger AM values than GMF harmonics. This runs contrary to the example t = 14.7 shown in Fig. 6. The instantaneous
spectra of the 3rd and 4th shaft harmonics and GMF harmonics have low magnitude at the actual IF, with incorrect peaks that arise
from the crossing of speed-unrelated components and overlapping of adjacent harmonics. Thus, these instantaneous spectrums should
result in inaccurate PDFs, which will in turn lead to inaccurate IAS estimation. In Table 3, the harmonic weights primarily focus more
on each of the 1st and 2nd shaft harmonics than on the corresponding GMF harmonics, while preventing the use of higher harmonics.
This result confirms that the proposed harmonic weighting can deal with the time-varying energy difference. Consequently, the global
PDF with harmonic weighting shown in Fig. 9(b) shows a better likelihood value at the actual IAS than the PDF without harmonic
weighting shown in Fig. 9(a).
Fig. 10 shows the proposed harmonic weights and the amplitude modulation values for each harmonic throughout the entire time.
The harmonic weights coincide well with the designed time-varying amplitude modulation pattern for the simulation, proving that the
proposed weighting can select proper PDFs by reflecting the instantaneous energy and complexity under harsh non-stationary oper­
ating conditions. Fig. 11 compares the global PDF maps with and without harmonic weighting. Fig. 11(a), made without weighting,
shows a blurry instantaneous angular speed ridge. Fig. 11(a) also presents secondary maxima ridges caused by overlapping starting at
10 Hz, and further caused by crossing with speed-unrelated components at around 20 Hz and 27 Hz, which hinders the accuracy of the
IAS estimation. The global PDF map made with harmonic weighting using Eq. (21) in Fig. 11(b) shows a more distinct IAS ridge and
suppresses the other ridges. The IAS ridge shown in Fig. 11(b) is then tracked by applying the Viterbi algorithm, using Eq. (22).
To emphasize the estimation accuracy of the proposed method (WMOVA), comparison studies are presented with single-step IAS
prediction methods, including single harmonic phase demodulation (SHD) [10], frequency domain energy operator (FDEO) [11,12],
and multi-order probabilistic approach (MOPA) [23] methods. Furthermore, we also present an ablation study by removing the
harmonic weighting from the proposed method, this is referred to as MOVA. The input parameters and prior knowledge, for each
method are summarized in Table 4 [7]. The parameters used for MOVA are identical to the parameters used in WMOVA, thus, these
details are omitted in Table 4. The Hanning window is employed to calculate the STFT for the TFR-based methods.
Fig. 12 displays the estimated IAS profiles of each method. Both the SHD and the FDEO, which are the phase demodulation-based
methods, show errors at around t = 6.5s and after t = 13.5s. As the SHD and the FDEO only utilize the single 1st shaft harmonic, they
provide inaccurate estimation when the 1st shaft harmonic fades due to amplitude modulation, as shown in Fig. 4. The fact that phase
demodulation-based methods failed to estimate after t = 13.5s confirms that these methods are only effective for small speed varia­
tions, 33 % at most. MOPA shows better performance than phase demodulation-based methods because it considers multiple har­
monics at once. However, MOPA has a weakness in large acceleration regions due to its assumption of low acceleration, which is used
for the smoothing method named a priori of continuity in MOPA; this approach is not suitable under harsh conditions [23]. Moreover,
the absence of weighting in MOPA and MOVA induces the confusing secondary ridge that is shown in Fig. 11(a), resulting in errors, as
shown in the zoomed-in depiction in Fig. 12. The proposed WMOVA outperforms the conventional methods by accurately tracking the
actual IF, even in the presence of large speed variation, speed-unrelated components, and extreme noise. Fig. 13 shows each method’s

10
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 5. Divided spectrogram with frequency ranges corresponding to each harmonic.

IAS estimation percentage error at each time to clearly illustrate where the local errors occur. In Fig. 13, the tremendous errors of SHD
and FDEO confirm that the phase-demodulation based methods are not appropriate for large speed variation profiles. MOPA shows a
large error at t = 17.4s, indicating the disadvantage in a large acceleration region. The large errors of MOPA and MOVA compared to
the proposed WMOVA at t = 19.2s, which is the same region as the zoomed-in depiction in Fig. 12, verify the advantage of the
proposed harmonic weighting in the presence of overlapping and crossing. Fig. 14 summarizes the mean and median absolute errors of
each method. The proposed WMOVA shows the minimum values for both mean and median absolute error. In Fig. 15, the order
spectrums of the simulated vibration data are presented by performing order tracking using the estimated IASs of each method. The IAS
estimation method causing smeared order spectrum can be interpreted as having inaccurate IAS results [41]. Among the IAS esti­
mation methods, the order spectrum obtained by the WMOVA is the least smeared and has the highest peak value. Therefore, all the
quantitative analyses show that the proposed method has the best accuracy for the simulated vibration data.

4.3. Simulation model parameter study

A parameter study is conducted to assess the robustness of the proposed method by varying the parameters employed within the
simulation model. Mainly, the SNR value and the initial phase of GMF harmonics’ AM functions ϕ2 are examined. Fig. 16(a) compares

11
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 6. Actual IF value (dotted red line) and instantaneous spectrum (black) at t = 14.7 (a) of shaft harmonics (Hi,1 ), (b) of GMF harmonics (Hi,2 ).
(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Table 2
Square root of instantaneous energy (Mi,k ), complexity (nci,k ), harmonic weight (ci,k ), and normalized harmonic weight (wi,k ) of each harmonic (Hi,k )
at t = 14.7.
H1,1 H2,1 H3,1 H4,1 H1,2 H2,2 H3,2 H4,2
( )
Mi,k × 104 0.9076 1.0509 1.4505 1.5008 4.1931 5.5495 5.6600 6.4715
nci,k 1 1.3243 5.2919 6.4761 1 2.6095 16.466 21.115
( )
ci,k × 104 0.9076 0.7936 0.2741 0.2317 4.1931 2.1267 0.3437 0.3065
wi,k 0.0989 0.0865 0.0299 0.0253 0.4569 0.2317 0.0375 0.0334

Fig. 7. The global PDF of IAS (in black) and the actual IAS (in dotted red line) at t = 14.7 (a) without harmonic weights, (b) with harmonic weights.
(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

the mean absolute error of each method with various SNR values in the simulation model. Due to the limitation of phase demodulation-
based methods for large speed variation, SHD and FDEO show far more inaccurate results compared to the TFR-based methods
regardless of the SNR values. Consequently, Fig. 16(b) confines the comparison exclusively to the TFR-based methods. Among these
methods, the proposed WMOVA consistently demonstrates the highest accuracy across all the SNR values. Fig. 17(a) performs the
parameter study for the initial phase value ϕ2 of GMF harmonics’ AM functions. In Fig. 17(a), similar to Fig. 16(a), the SHD and FDEO
show inaccurate results regardless of the initial phase value. Fig. 17(b) compares the MAE only with the TFR-based methods, where the

12
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 8. Actual IF value (dotted red line) and instantaneous spectrum (black) at t = 12.375 (a) of shaft harmonics (Hi,1 ), (b) of GMF harmonics (Hi,2 ).
(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Table 3
Square root of instantaneous energy (Mi,k ), complexity (nci,k ), harmonic weight (ci,k ), and normalized harmonic weight (wi,k ) of each harmonic (Hi,k )
at t = 12.375.
H1,1 H2,1 H3,1 H4,1 H1,2 H2,2 H3,2 H4,2
( )
Mi,k × 10 4 3.3520 2.3050 2.7077 2.1255 2.7555 4.0479 5.3207 6.3261
nci,k 1 1.3243 5.2919 6.4761 1 2.6095 16.466 21.115
( )
ci,k × 104 3.3520 1.7406 0.5117 0.3282 2.7555 1.5512 0.3231 0.2996
wi,k 0.3086 0.1603 0.0471 0.0302 0.2537 0.1428 0.0297 0.0276

Fig. 9. The global PDF of IAS (black) and actual IAS (dotted red line) at t = 12.375 (a) without harmonic weights, (b) with harmonic weights. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

WMOVA also shows the best accuracy across all the initial phase values. Therefore, the results from both parameter studies show the
robustness of the proposed method.
In the subsequent section, the actual vibration data from rotating machinery under non-stationary operations are exploited to
further highlight the proposed method.

13
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 10. The proposed harmonic weights (red) compared to the actual amplitude modulation values (black) for each harmonic. (For interpretation
of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 11. The global PDF map of IAS made (a) without weighting and (b) with harmonic weighting.

14
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Table 4
Summary of the input parameters of each method for the simulation model.
Parameter # Method name

SHD FDEO MOPA WMOVA

1 finit = 20Hz finit = 20Hz Ωmin = 10Hz Ωmin = 10Hz

2 Bw = 12Hz Bw = 12Hz Ωmax = 32Hz Ωmax = 32Hz
3 Nw = 10, 000 Nw = 10, 000 γ = 5Hz/s amax = 18Hz/s
{ } { }
4 NOverlap = 3, 000 NOverlap = 3, 000 H1,k = {1, 20} H1,k = {1, 20}
5 Nw = 5, 000 Nw = 5, 000
6 NFFT = 20, 000 NFFT = 20, 000
7 NOverlap = 4, 750 NOverlap = 4, 750
8 Δpdf = 0.01 Δpdf = 0.01
9 Kw = 20 η =2

with finit is the IF at the initial time index, Bw is the bandwidth of the band-pass filter, Nw is the sample size of the window, NFFT is the sample size of the
FFT, NOverlap represents the overlapped samples between windows, γ is the expected or mean acceleration of the IAS, amax is the maximum acceleration
of the IAS, Ωmin , Ωmax are the minimum and maximum speed, Δpdf is the bin width of the PDF of IAS, η is the weight coefficient for Viterbi algorithm,
{ }
Kw is the number of windows used for smoothing, and H1,k is the list of considered fundamental kinematic orders with regard to the shaft order.

Fig. 12. Estimated instantaneous angular speed profiles of each method for the simulation model.

5. Case studies

This paper presents two case studies to validate the proposed method. First, the proposed approach is applied to the well-known
wind turbine gearbox dataset, taken from the 2014 International Conference on Condition Monitoring of Machinery in Non-Stationary
Operations (CMMNO), to show the performance of the proposed method under non-harsh conditions. Second, the method is applied to
vibration data collected from a gearbox testbed to demonstrate the effectiveness of the proposed method under extremely harsh
conditions. Similar to Section 4, SHD, FDEO, and MOPA, which are single-step IAS estimation methods, are utilized for comparison.
Furthermore, a parameter study is conducted to evaluate the impact of the window size parameter Nw , a key parameter for the TFR-
based methods.

5.1. CMMNO wind turbine gearbox data set

The CMMNO dataset, which is widely used to demonstrate the IAS estimation method, contains the vibration signal of a wind
turbine gearbox measured by an accelerometer located on the gearbox housing [23]. The duration of the signal is approximately 550s,
and the sampling rate is fs = 5kHz. Estimating the IAS of the high-speed shaft was the primary goal of the test.
The proposed WMOVA and MOPA use the first ten harmonics of every six vibration sources in the gear train, and their fundamental
kinematic orders are listed in Table 5 [7,15]. The input parameters used for every method are arranged in Table 6. The parameters for
CMMNO data are determined based on reference [7]. The raw spectrogram and the whitened spectrogram using a Hanning window are

15
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 13. The IAS estimation percentage error of each method at each time for the simulation model.

Fig. 14. Mean and median absolute errors of the results of each method for the simulation model.

Fig. 15. Order spectrum of the simulated data zoomed in the shaft order using encoder and the estimated IASs of each method.

given in Fig. 18. The whitening process removes the constant harmonic noise that arises from resonances around 1300Hz in Fig. 18(a).
Then, the proposed WMOVA is applied to the whitened spectrogram in Fig. 18(b) to obtain the global PDF map, as shown in Fig. 19.
The IAS ridge in the PDF map is tracked by the modified Viterbi algorithm to estimate the IAS profile.
Fig. 20 displays the IAS estimation results obtained from each method. The zoomed view of the estimation results in Fig. 20 shows
the large local errors of SHD and FDEO. Since SHD and FDEO only utilize a single 2nd GMF harmonic, they are vulnerable to the

16
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 16. Mean absolute error with various SNR values in the simulation model (a) for all comparison methods and (b) for TFR-based methods.

Fig. 17. Mean absolute error with various initial phase value ϕ2 used in the AM function of the simulation model’s GMF harmonics (a) for all
comparison methods and (b) for TFR-based methods.

Table 5
Fundamental kinematic orders with regard to a
high-speed shaft in the gear train.
Gear pair Order value

1 1
2/3, 1/2 1.0255
4/5 5.3167
6/7 29
8/9 15.225
10/11 6.6196

impulse-like noises caused by the low-amplitude modulation values of the harmonic they used [24]. The smoothing process on these
impulse-like errors introduces the overfitting issue and sequentially makes a large local error [24]. WMOVA and MOPA overcome this
issue by considering multiple harmonics simultaneously. The profiles of WMOVA and MOPA appear to be the same as the encoder
profile. Fig. 21 shows the IAS estimation percentage error at each time, where both the MOPA and WMOVA show very small error
values. To clearly see the performances of each method, mean and median absolute errors are calculated in Fig. 22. As expected, SHD
and FDEO show larger errors than MOPA and WMOVA. MOPA performs best, while WMOVA achieves almost identical results. Given
that the harmonics in the CMMNO data show high SNR values, the harmonic weighting does not have a significant effect. Furthermore,
the smoothing method used in MOPA works well due to the small acceleration of the data. Still, the difference between WMOVA and

17
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Table 6
Summary of the input parameters of each method for the CMMNO data set.
Parameter # Method name

SHD FDEO MOPA WMOVA

1 finit = 53Hz finit = 53Hz Ωmin = 5Hz Ωmin = 15Hz

2 Bw = 8Hz Bw = 8Hz Ωmax = 35Hz Ωmax = 35Hz
3 Nw = 50, 000 Nw = 50, 000 γ = 0.4Hz/s amax = 0.91Hz/s
{ } { }
4 NOverlap = 15, 000 NOverlap = 15, 000 H1,k = Table 3 H1,k = Table 3
5 Nw = 5000 Nw = 5000
6 NFFT = 10, 000 NFFT = 10, 000
7 NOverlap = 4750 NOverlap = 4750
8 Δpdf = 0.01 Δpdf = 0.01
9 Kw = 20 η =2

Fig. 18. Spectrogram of the CMMNO vibration signal (a) raw, (b) whitened.

Fig. 19. The global PDF map of CMMNO data made by the proposed method.

18
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 20. Estimated instantaneous angular speed profiles of each method on the CMMNO data.

Fig. 21. The IAS estimation percentage error of each method at each time for the CMMNO data.

Fig. 22. Mean and median absolute errors for each method on the CMMNO data.

MOPA can be considered negligible, given the improved accuracy compared to SHD and FDEO. Fig. 23 shows the order spectrum of the
CMMNO data by performing order tracking using the estimated IASs of each method. The order spectrums obtained by the MOPA and
WMOVA closely resemble the order spectrum obtained by the encoder. Thus, the quantitative analyses of the CMMNO case study prove
that WMOVA works properly under well-organized conditions.

19
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 23. Order spectrum of the CMMNO data using encoder and the estimated IASs of each method.

Fig. 24. Overview of the gearbox testbed.

5.2. Gearbox testbed data set

To acquire vibration data under harsh operating conditions, an experiment using a gearbox testbed at the Korea Research Institute
of Standards and Science (KRISS) was performed. The vibration sensor was attached to the planetary gearbox housing, as shown in
Fig. 24. The sampling frequency is 25,600 Hz, and the length of the signal is about 60 s. The sampling frequency of the encoder signal is
25,600 Hz, and the smoothing window length of 0.25s is applied. The goal is to estimate the IAS of the high-speed shaft (HSS) linked to
the planetary gearbox. The maximum speed variation of the IAS profile is more than 55 %, and the maximum acceleration value is

Fig. 25. Raw spectrogram of the KRISS vibration signal.

20
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 26. Whitened spectrogram of the KRISS vibration signal.

Table 7
Summary of the input parameters of each method for the KRISS data set.
Parameter # Method name

SHD FDEO MOPA WMOVA

1 finit = 780Hz finit = 780Hz Ωmin = 5Hz Ωmin = 5Hz

2 Bw = 312Hz Bw = 312Hz Ωmax = 30Hz Ωmax = 30Hz
3 Nw = 12, 800 Nw = 12, 800 γ = 4.4Hz/s amax = 18Hz/s
{ } { }
4 NOverlap = 2, 560 NOverlap = 2, 560 H1,k = {1, 23.37} H1,k = {1, 23.37}
5 Nw = 12, 800 Nw = 12, 800
6 NFFT = 51, 200 NFFT = 51, 200
7 NOverlap = 11, 520 NOverlap = 11, 520
8 Δpdf = 0.01 Δpdf = 0.01
9 Kw = 20 η =2

about 18 Hz/s, leading the adjacent harmonics to overlap. Fig. 25 shows the raw Hanning-windowed spectrogram of the KRISS vi­
bration signal. The constant harmonic noises from resonances are present in the entire frequency range. In particular, high-energy
constant noises at around 77 Hz and 200 Hz make the harmonics fade in the low-frequency band. The whitened spectrogram in
Fig. 26 makes these harmonics visible. Still, higher harmonics with low energy are faded or even absent at specific time regions, such as
around 19 s, 34 s, and 52 s.
For estimation, SHD and FDEO utilize the strong 2nd GMF harmonic, starting at around 780 Hz. WMOVA and MOPA use the first

Fig. 27. The global PDF map of KRISS data made by the proposed method.

21
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 28. Estimated instantaneous angular speed profiles of each method for the KRISS data.

ten harmonics of HSS and the planetary gearbox GMF. Table 7 summarizes the input parameters for each method. The primary dif­
ferences for the utilized input parameters are the harmonic orders related to the system’s kinematic orders and the size of the window
length Nw . A shorter window length of 0.5s is adopted to handle the rapid speed variation. The global PDF map constructed by the
proposed method is shown in Fig. 27.
Fig. 28 displays the estimated IAS profiles obtained by each method. As expected, SHD and FDEO failed to estimate the IAS due to
the large speed variation. MOPA performs much better than SHD and FDEO by using multiple harmonics in the TFR. However, MOPA
shows two disadvantages. First, MOPA fails to estimate the IAS in the large acceleration regions as a result of the convolution
smoothing method. Second, MOPA makes critical errors in regions where low-energy harmonics fade because of the absence of
weighting. Fig. 29 shows the IAS estimation percentage error of each method. In Fig. 29, SHD and FDEO have large errors over the
entire time, indicating the limitation of the phase demodulation-based method. MOPA exhibits large local errors in specific time bins,
which belong to the regions where harmonic fading appears in Fig. 26 or the large acceleration occurs. The result of MOPA confirms
the disadvantages of weight absence and smoothing. The proposed WMOVA approach follows the encoder profile best in Fig. 28 and
exhibits the smallest percentage errors in Fig. 29 by using energy-based harmonic weighting. Fig. 30 calculates the mean and median
absolute errors to easily compare the accuracy, where the WMOVA shows the best result. The order spectrums of the KRISS data using
the estimated IASs of each method are shown in Fig. 31. Again, the order spectrum acquired by the WMOVA shows the best perfor­
mance, having the least smeared spectrum among the IAS estimation methods. Consequently, the best results are obtained in the

Fig. 29. The IAS estimation percentage error of each method at each time for the KRISS data.

22
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

Fig. 30. Mean and median absolute errors for each method for the KRISS data.

Fig. 31. Order spectrum of the KRISS data zoomed in the shaft order using encoder and the estimated IASs of each method.

Fig. 32. Mean absolute error of the MOPA and WMOVA approach with various window lengths on the KRISS data.

quantitative analyses by the newly proposed method, which confirms the accuracy of WMOVA under harsh conditions.
It is important to note that determining the optimal window length autonomously for each case is challenging without true IAS
measurements from an encoder system. Therefore, we also conducted a parameter study on window length Nw for TFR-based methods
with the KRISS data. Fig. 32 displays the MAE values of MOPA and WMOVA on the KRISS data. For both methods, it is evident that
shorter window length performs better when dealing with rapid speed variation profiles. Further, the results demonstrate the better
accuracy and robustness of the proposed method across all the available ranges of window length.

6. Conclusion

In this paper, we proposed a weighted multi-order Viterbi algorithm (WMOVA) method to estimate the instantaneous angular
speed (IAS) under harsh non-stationary conditions. Under harsh conditions, many difficulties, such as overlapping, fading, crossing of
harmonics, and extreme noise, coexist in the vibration signal, hindering the accurate IAS estimation. Complex rotating systems in real

23
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

engineering applications frequently encounter such harsh conditions, complicating the implementation of precise IAS estimation
methods. To cope with these difficulties that increase the complexity or uncertainty, a novel harmonic weighting approach is proposed.
The proposed method prioritizes harmonics with high energy and low complexity for an accurate IAS estimation. Specifically, the
square root of the instantaneous energy of harmonics is employed as the numerator, with the harmonics’ complexity serving as the
denominator. Consequently, the novel energy-based harmonic weighting enables extracting accurate and reliable IAS-related infor­
mation utilizing multiple harmonics under harsh operating conditions. The modified Viterbi algorithm is applied to track the IAS ridge
without smoothing, which can cope with large acceleration and noisy conditions. The proposed method was first deeply analyzed using
a simulated vibration signal that represents extremely harsh conditions. Two experimental case studies were then conducted with
quantitative analyses to demonstrate the proposed WMOVA method. Additionally, parameter studies were performed to verify the
robustness. The results show that WMOVA outperformed conventional IAS estimation methods. The proposed method could produce
accurate and robust IAS estimation results, even under harsh operating conditions. Such accurate and reliable IAS estimation results
have the potential for broader application in vibration-based non-stationary fault diagnosis of rotating machinery across real-world
industrial locations.
The principal contribution of this study lies in introducing the energy-based harmonic weighting methodology. The innovative
aspects of this harmonic weighting method are summarized as follows:

1) A time-varying energy-based harmonic weighting strategy is employed to effectively utilize the information from multiple har­
monics. This approach facilitates extracting accurate IAS information from high-energy harmonics while mitigating the risk of
intermittent harmonics’ fading.
2) A complexity-based novel harmonic weighting strategy is proposed to address the challenges associated with undesired energy
overestimation caused by disturbance factors, such as overlapping, crossing, and extreme noise under harsh conditions. This
strategy employs complexity, as measured by Rényi entropy, to account for the inherent uncertainty in the vibration signals that
could hinder accurate IAS estimation.

However, the proposed method has some noteworthy limitations. First, similar to MOPA, one of the most accurate IAS estimation
methods, it necessitates prior knowledge of the system’s minimum and maximum speed values. The proposed method cannot accu­
rately estimate the IAS if the prior knowledge is unavailable. Second, the high computational cost of the Viterbi algorithm remains an
obstacle to its application in real industrial sites. Therefore, future studies must be undertaken to resolve both limitations associated
with the necessity of the system’s prior knowledge and the high computational cost. Future improvements can be achieved by
investigating the approaches that enable automated estimation of the system’s prior knowledge, including minimum and maximum
speed or the kinematic orders, thus facilitating the IAS estimation. The research on ridge tracking or optimization techniques with
noise robustness and low cost is recommended as well to enhance versatility. Although the case studies exhibit the superior perfor­
mance of the proposed method, there still remains room for further development and validation regarding fault diagnosis performance
assisted by the estimated IAS information. In case 2, despite the notable reduction in IAS estimation error compared to conventional
methods in such harsh operating scenarios, the localized errors due to the inherent resolution limitations of the TFR result in a more
smeared order spectrum than the encoder-based order spectrum. Since the smeared effect acts as an obstacle for non-stationary fault
diagnosis, the future research direction to overcome the resolution limitations is also necessary for the tacholess order tracking. Po­
tential solutions may encompass the implementation of the two-step IAS estimation approach or the time–frequency post-processing
approaches, such as the reassignment method and the synchrosqueezing transform, which improve the readability by enhancing the
energy concentration of the TFR. Lastly, the proposed method needs to be verified on real industrial data to demonstrate its advantages
in practical applications.

CRediT authorship contribution statement

Jinoh Yoo: Writing – original draft, Visualization, Validation, Software, Methodology, Investigation, Formal analysis, Concep­
tualization. Jongmin Park: Writing – review & editing, Software, Resources. Taehyung Kim: Writing – review & editing, Software.
Jong Moon Ha: Conceptualization, Data curation, Investigation, Resources, Supervision, Writing – review & editing. Byeng D. Youn:
Data curation, Funding acquisition, Project administration, Resources, Supervision, Writing – review & editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability

The authors do not have permission to share data.

Acknowledgements

This research was supported by the International Research & Development Program of the National Research Foundation of Korea

24
J. Yoo et al. Mechanical Systems and Signal Processing 211 (2024) 111187

(NRF) funded by the Ministry of Science and ICT (No. 2022K1A4A7A0409632911) and National Research Foundation of Korea (NRF)
funded by the Ministry of Science and ICT (NRF-2021M2E6A1084687).

References

[1] J. Park, Y. Kim, K. Na, B.D. Youn, Y. Chen, M.J. Zuo, Y.-C. Bae, An image-based feature extraction method for fault diagnosis of variable-speed rotating
machinery, Mech. Syst. Sig. Process. 167 (2022) 108524.
[2] H. Kim, C.H. Park, C. Suh, M. Chae, H. Yoon, B.D. Youn, MPARN: multi-scale path attention residual network for fault diagnosis of rotating machines, J. Comput.
Des. Eng. 10 (2023) 860–872.
[3] M. Vishwakarma, R. Purohit, V. Harshlata, P. Rajput, Vibration analysis & condition monitoring for rotating machines: a review, Mater. Today:. Proc. 4 (2017)
2659–2664.
[4] R.B. Randall, J. Antoni, Rolling element bearing diagnostics—a tutorial, Mech. Syst. Sig. Process. 25 (2011) 485–520.
[5] D. Rémond, J. Antoni, R.B. Randall, Editorial for the special issue on Instantaneous Angular Speed (IAS) processing and angular applications, Mech. Syst. Sig.
Process. 44 (2014) 1–4.
[6] J.M. Ha, B.D. Youn, H. Oh, B. Han, Y. Jung, J. Park, Autocorrelation-based time synchronous averaging for condition monitoring of planetary gearboxes in wind
turbines, Mech. Syst. Sig. Process. 70–71 (2016) 161–175.
[7] C. Peeters, Q. Leclère, J. Antoni, P. Lindahl, J. Donnal, S. Leeb, J. Helsen, Review and comparison of tacholess instantaneous speed estimation methods on
experimental vibration data, Mech. Syst. Sig. Process. 129 (2019) 407–436.
[8] J.M. Ha, B.D. Youn, Fault diagnosis of a planetary gearbox by D norm-based time synchronous averaging (DTSA) with roughly estimated phase information
under an encoder-less operating condition, J. Sound Vib. 520 (2022).
[9] J. Urbanek, T. Barszcz, J. Antoni, A two-step procedure for estimation of instantaneous rotational speed with large fluctuations, Mech. Syst. Sig. Process. 38
(2013) 96–102.
[10] F. Bonnardot, M. El Badaoui, R.B. Randall, J. Danière, F. Guillet, Use of the acceleration signal of a gearbox in order to perform angular resampling (with limited
speed fluctuation), Mech. Syst. Sig. Process. 19 (2005) 766–785.
[11] R. Randall, W. Smith, Use of the Teager Kaiser Energy Operator to estimate machine speed, PHM Society European Conference (2016).
[12] R.B. Randall, W.A. Smith, Uses and mis-uses of energy operators for machine diagnostics, Mech. Syst. Sig. Process. 133 (2019).
[13] M.D. Coats, N. Sawalhi, R. Randall, Extraction of tacho information from a vibration signal for improved synchronous averaging, Proc. Acoust. (2009) 1–8.
[14] M.D. Coats, R. Randall, Order-Tracking with and without a tacho signal for gear fault diagnostics, Proc. Acoust. (2012) 1–6.
[15] C. Peeters, J. Antoni, Q. Leclère, T. Verstraeten, J. Helsen, Multi-harmonic phase demodulation method for instantaneous angular speed estimation using
harmonic weighting, Mech. Syst. Sig. Process. 167 (2022).
[16] J. Urbanek, T. Barszcz, N. Sawalhi, R. Randall, Comparison of amplitude-based and phase-based methods for speed tracking in application to wind turbines,
Metrol. Meas. Syst. 18 (2011) 295–304.
[17] Z.K. Peng, G. Meng, F.L. Chu, Z.Q. Lang, W.M. Zhang, Y. Yang, Polynomial Chirplet transform with application to instantaneous frequency estimation, IEEE
Trans. Instrum. Meas. 60 (2011) 3222–3229.
[18] K.C. Gryllias, I.A. Antoniadis, Estimation of the instantaneous rotation speed using complex shifted Morlet wavelets, Mech. Syst. Sig. Process. 38 (2013) 78–95.
[19] D. Iatsenko, P.V. McClintock, A. Stefanovska, Extraction of instantaneous frequencies from ridges in time–frequency representations of signals, Signal Process.
125 (2016) 290–303.
[20] I. Djurović, L. Stanković, An algorithm for the Wigner distribution based instantaneous frequency estimation in a high noise environment, Signal Process. 84
(2004) 631–643.
[21] S. Schmidt, P.S. Heyns, J.P. de Villiers, A tacholess order tracking methodology based on a probabilistic approach to incorporate angular acceleration
information into the maxima tracking process, Mech. Syst. Sig. Process. 100 (2018) 630–646.
[22] R. Zimroz, J. Urbanek, T. Barszcz, W. Bartelmus, F. Millioz, N. Martin, Measurement of instantaneous shaft speed by advanced vibration signal processing –
application to wind turbine gearbox, Metrol. Meas. Syst. 18 (2011).
[23] Q. Leclère, H. André, J. Antoni, A multi-order probabilistic approach for Instantaneous Angular Speed tracking debriefing of the CMMNO14‫ ׳‬diagnosis contest,
Mech. Syst. Sig. Process. 81 (2016) 375–386.
[24] D. Peng, W.A. Smith, R.B. Randall, Z. Peng, C.K. Mechefske, Speed estimation in planetary gearboxes: a method for reducing impulsive noise, Mech. Syst. Sig.
Process. 159 (2021).
[25] C. Peeters, Q. Leclere, J. Antoni, J. Helsen, Instantaneous angular speed estimation for multi-stage wind turbine gearboxes, Instantaneous Angular Speed
Estimation for Multi-Stage Wind Turbine Gearboxes (2017) 50–57.
[26] L. Cohen, Time-Frequency Analysis, Prentice Hall, New Jersey, 1995.
[27] P. Flandrin, R.G. Baraniuk, O. Michel, Time-frequency complexity and information, in: Proceedings of ICASSP’94. IEEE International Conference on Acoustics,
Speech and Signal Processing, IEEE, 1994, pp. III/329-III/332 vol. 323.
[28] R.G. Baraniuk, P. Flandrin, A.J. Janssen, O.J. Michel, Measuring time-frequency information content using the Rényi entropies, IEEE Trans. Inf. Theory 47
(2001) 1391–1409.
[29] T.-H. Sang, W.J. Williams, Renyi information and signal-dependent optimal kernel design, in: 1995 International Conference on Acoustics, Speech, and Signal
Processing, IEEE, 1995, pp. 997-1000.
[30] M.B. Malarvili, V. Sucic, M. Mesbah, B. Boashash, Renyi entropy of quadratic time-frequency distributions: effects of signal’s parameters, in: 2007 9th
International Symposium on Signal Processing and Its Applications, IEEE, 2007, pp. 1–4.
[31] J. Lerga, N. Saulig, V. Mozetic, R. Lerga, Number of EEG signal components estimated using the short-term Rényi entropy, in: 2016 International
Multidisciplinary Conference on Computer and Energy Science (SpliTech), IEEE, 2016, pp. 1–6.
[32] A. Rényi, On measures of entropy and information, Proceedings of the fourth Berkeley symposium on mathematical statistics and probability, Berkeley,
California, USA, 1961.
[33] L. Stanković, A measure of some time–frequency distributions concentration, Signal Process. 81 (2001) 621–631.
[34] V. Sucic, N. Saulig, B. Boashash, Estimating the number of components of a multicomponent nonstationary signal using the short-term time-frequency Rényi
entropy, EURASIP J. Adv. Signal Process. 2011 (2011) 1–11.
[35] W.-K. Lu, Q. Zhang, Deconvolutive short-time Fourier transform spectrogram, IEEE Signal Process Lett. 16 (2009) 576–579.
[36] N.A. Khan, M. Mohammadi, I. Djurović, A modified viterbi algorithm-based if estimation algorithm for adaptive directional time–frequency distributions,
Circuits Systems Signal Process. 38 (2019) 2227–2244.
[37] X. Yu, Z. Feng, D. Zhang, Adaptive high-resolution order spectrum for complex signal analysis of rotating machinery: Principle and applications, Mech. Syst. Sig.
Process. 177 (2022).
[38] S. Loutridis, Instantaneous energy density as a feature for gear fault detection, Mech. Syst. Sig. Process. 20 (2006) 1239–1253.
[39] W.J. Williams, M.L. Brown, A.O. Hero III, Uncertainty, information, and time-frequency distributions, Advanced signal processing algorithms, architectures, and
implementations ii, SPIE (1991) 144–156.
[40] S. Chen, M. Du, Z. Peng, M. Liang, Q. He, W. Zhang, High-accuracy fault feature extraction for rolling bearings under time-varying speed conditions using an
iterative envelope-tracking filter, J. Sound Vib. 448 (2019) 211–229.
[41] D. Peng, Y. Chen, M.J. Zuo, C.K. Mechefske, Assessment and improvement of the accuracy of tacholess instantaneous speed estimation, Mech. Syst. Sig. Process.
202 (2023) 110706.

25