Mechanical Systems and Signal Processing 100 (2018) 630–646

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

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

A tacholess order tracking methodology based on a probabilistic

approach to incorporate angular acceleration information into
the maxima tracking process
S. Schmidt a,⇑, P.S. Heyns a, J.P. de Villiers b,c
a
Centre for Asset Integrity Management, Department of Mechanical and Aeronautical Engineering, University of Pretoria, Pretoria, South Africa
b
Department of Electrical, Electronic and Computer Engineering, University of Pretoria, Pretoria, South Africa
c
Defence, Peace, Safety and Security (DPSS), Council for Scientific and Industrial Research (CSIR), PO Box 395, Pretoria 0001, South Africa

a r t i c l e i n f o a b s t r a c t

Article history: Rotational speed or phase measurements are important in the rotating machine condition
Received 16 January 2017 monitoring field. Rotating machines often operate in the presence of fluctuating rotational
Received in revised form 21 July 2017 speeds and its components generate signals which are periodical in the angle domain.
Accepted 30 July 2017
Hence, it is essential to perform order tracking when condition monitoring is performed
under varying speed conditions. Computed order tracking is performed with a measured
tachometer signal and it requires additional hardware to be installed on the machine,
Keywords:
which may not be feasible nor practical. This article presents a tacholess order tracking
Tacholess order tracking
Maxima tracking
method which is capable of accurately estimating the phase of a shaft of interest in the
Vold-Kalman filter presence of large angular accelerations and noise. An improved maxima tracking procedure
Rotational speed estimation is used with an angular-displacement Vold-Kalman filter and the Hilbert transform to esti-
Probabilistic mate the instantaneous phase of the shaft under consideration. The estimated instanta-
Linear regression neous phase is used to resample the vibration signal from the time to the angle domain.
The proposed tacholess order tracking technique is critically investigated on three numer-
ical and three experimental rotational speed profiles. The minimum and maximum phase
error obtained for the experimental data was 0.037697 and 0.05022 radians respectively,
which highlights the potential of the technique.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction

Rotating machines, such as gearboxes, operate under fluctuating operating conditions due to inherent fluctuating char-
acteristics, such as fluctuating wind speeds for wind turbines, that are present during its operation. The resulting vibration
signal, measured on a rotating machine, is best represented in the angle domain, as opposed to the time domain, since the
information within the signal such as localised gear faults and gear mesh interactions are generated at specific angular posi-
tions on the shaft [1].
The vibration signal is transformed from the time to the angle domain using order tracking techniques, where the instan-
taneous phase of the shaft of interest is used to resample the vibration signal into equal angular increments. In computed
order tracking, a tachometer signal is measured simultaneously with the vibration signal by using a stationary probe and

⇑ Corresponding author.
E-mail address: [email protected] (S. Schmidt).

http://dx.doi.org/10.1016/j.ymssp.2017.07.053
0888-3270/Ó 2017 Elsevier Ltd. All rights reserved.
 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646 631

Nomenclature

b envelope of signal of interest

A weight matrix
Q design matrix
Df C frequency resolution of spectrogram [Hz]
Df c feasible bandwidth enforced by the constraint [Hz]
Dt sampling period [s]
DtC time resolution of spectrogram [s]
C short-time Fourier transform
j cost function
ki Lagrangian multiplier of constraint i
L Lagrangian function
m a sample from a Gaussian noise model
/a analytical phase of the shaft [rad]
/r estimated phase of the shaft [rad]
/IF phase of instantaneous frequency [rad]
qi weight parameter of constraint i
r standard deviation
t window function
e non-homogeneity term
f residual component of the data equation
a weighting factor
Bw relative bandwidth
cs fundamental component w.r.t shaft orders
e phase error [rad]
eR relative phase error
fb bandwidth [Hz]
fc centre of the constraint [Hz]
fF a polynomial model of the IF [Hz]
fs sampling frequency [Hz]
f max the IF estimated from maxima tracking [Hz]
f IF the actual, unknown IF [Hz]
fr rotational speed of shaft [Hz]
H the Hilbert transform
j imaginary unit
Lw window length [number of samples]
Lo window overlap [number of samples]
Lz zero padding [number of samples]
Nm number of data points used for linear regression
Np order of the polynomial
Nx number of samples in x
p probability density function
si slack variable of constraint i
t time [s]
wi polynomial weight coefficient i
x the vibration signal

a reference, such as an incremental shaft encoder or a shaft key, on the shaft of interest. This tachometer signal contains the
information pertaining to the instantaneous phase of the shaft, which is used to perform order tracking on the vibration sig-
nal. The order tracking process has been successfully applied for gear fault detection [2,3], misalignment fault diagnosis [4]
and bearing fault diagnosis [5].
Single pulse per revolution (PPR) tachometer signals, obtained from a shaft key and a proximity probe, provide insuffi-
cient resolution to capture large angular accelerations. Incremental shaft encoders, such as zebra tape and digital shaft enco-
ders, provide much finer resolutions to measure the large angular accelerations in the system [6]. However, incremental
shaft encoders require large sampling frequencies to accurately estimate the zero crossing times [7,8], the procurement
and installation of additional expensive instrumentation and equipment and the storage of additional data [9,10]. The shaft
encoders are also in some cases impractical or even impossible to install and are not durable in the harsh operating
conditions [11,12]. The vibration transducer and the optical probe are located on different components and therefore the
632 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646

transmission paths between the measurement equipment and the source of the vibration are different [13]. This transmis-
sion path is frequency dependent and leads to phase distortion when performing computed order tracking in varying speed
conditions [13,14]. Tacholess order tracking approaches, which use the acceleration signal to obtain the instantaneous phase
information, are therefore actively investigated by the engineering community to circumvent the aforementioned limita-
tions and problems when performing computed order tracking [4,9,10,12,15].
Bonnardot et al. [8] proposed using the acceleration signal to perform tacholess order tracking in slightly varying rota-
tional speed applications using narrow-band demodulation on one of the gear mesh frequencies. Combet and Gelman
[11] improved the procedure proposed by Bonnardot et al. [8], by automating the harmonic selection process. Heyns
et al. [16] used a clustering procedure to track an instantaneous frequency component in the spectrogram of the vibration
signal. The IF estimate is low-pass filtered, since it contains noise due to the limited resolution provided by the spectrogram.
The filtered instantaneous frequency is used to resample the signal for fault detection.
Urbanek et al. [10] and Zhao et al. [12] proposed tacholess order tracking procedures in large varying speed conditions.
Urbanek et al. [10] proposed performing maxima tracking on the spectrogram of the vibration signal. The rough estimate of
the instantaneous frequency, obtained from the maxima tracking process, is used to order track the signal and filter out all
components except a single instantaneous frequency. This mono-component signal is reverse resampled to the time domain.
Zhao et al. [12] used the Chirplet transform to extract an instantaneous frequency component and then the angular-velocity
Vold-Kalman filter is applied to the signal to obtain a mono-component signal. The Hilbert transform is used on the mono-
component signal by Urbanek et al. [10] and Zhao et al. [12] to extract the instantaneous phase of the signal over time. This
instantaneous phase provides a mapping from the time to the angle domain, which is used to order track the vibration signal.
Qi et al. [4] used the ensemble empirical mode decomposition algorithm to obtain the mono-component intrinsic mode
functions of the signal, which are subsequently used to obtain the phase information of the shaft of interest. He et al. [17]
proposed using the discrete spectrum correction technique to order track the vibration signal obtained from a wind turbine.
Leclere et al. [9] proposed a multi-order approach for instantaneous angular speed estimation. The spectrogram is applied to
the vibration signal, afterwhich a whitening procedure is applied to attenuate structural resonances. The probability density
function (pdf) within an expected operating frequency range is estimated at each time step using multi order information,
whereafter the pdf is smoothed to calculate the instantaneous angular speed.
In this study, a tacholess order tracking method is proposed which is robust to large varying speed conditions and in
the presence of noise. The estimated rotational speed is investigated as well, because this information is invaluable for
condition-based maintenance decision-making purposes. The benefits of incorporating acceleration information into the
process is highlighted, which can possibly be used to develop multi-order approaches. The tacholess order tracking tech-
nique is developed in Section 2 and validated in Section 3, whereafter conclusions are drawn and recommendations are
made in Section 4.

2. Proposed tacholess order tracking methodology

The proposed tacholess order tracking method, illustrated in Fig. 1, is very similar to the methods proposed by Urbanek
et al. [10] and Zhao et al. [12]. In the proposed tacholess order tracking method, the maxima tracking algorithm is improved
by incorporating acceleration information into the maxima tracking process which renders it more robust in the presence of
angular acceleration and noise components. The angular-displacement Vold-Kalman filter is investigated as opposed to the
angular-velocity Vold-Kalman filter used by [12].
The rotational speed profile, f r versus time, and the acceleration signal, abbreviated by acc., are presented in (a) of Fig. 1.
The spectrogram of the vibration signal in (a), is presented in (b). An instantaneous frequency (IF) component is tracked with
the proposed maxima tracking procedure to identify the relationship between the IF and time. The tracked IF component is
superimposed on the spectrogram in (c).
The tracked instantaneous frequency in (c) of Fig. 1 is used as the centre frequency for an angular-displacement Vold-
Kalman filter (VKF), discussed in Section 2.3. The VKF is a bandpass filter with a time-varying centre frequency. The spectro-
gram of the filtered vibration signal is shown in (d), where it can be seen that it is a mono-component signal only containing
the tracked IF. The phase of this mono-component signal is calculated from the Hilbert transform and shown in (e). Note that
the phase of the IF needs to be scaled to reflect the phase of the shaft of interest. The phase of the shaft of interest, presented
in (f), is used to order track the vibration signal. The ordinate (or phase of the shaft) in Figure (f) is sampled with equal angle
lengths and then new time sampling points are determined from the instantaneous phase profile. Each step of this process is
discussed in more detail in the subsequent subsections.

2.1. Time-frequency distribution

The time and frequency domain information in the vibration signal are investigated simultaneously in a time-frequency
distribution so that the IF of interest can be estimated at each time step. Urbanek et al. [10] used the spectrogram for maxima
tracking and Zhao et al. [12] used the spectrogram to estimate the chirp rate so that the Chirplet transform could be used to
estimate the component of interest. Even though the spectrogram renders a limited time–frequency resolution, it is used for
maxima tracking in this article and motivated by:
 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646 633

(a)
3

fr [Hz]
2.5
2
0 5 10 15 20
Acc. [m/s2 ]
2
0
-2
0 5 10 15 20
Time [s]

×104 (e) (f)

7 Instantaneous phase
Phase of shaft [rad]

250
Phase of IF [rad]

6 Equal angle samples

200 New time samples
5
4 150
3
100
2
1 50
0 0
5 10 15 5 10 15
Time [s] Time [s]
Fig. 1. Key steps in the proposed tacholess order tracking method: (a) The acceleration signal and the rotational speed profile (f r ) over time; (b) the
spectrogram of the vibration signal; (c) the maxima tracked IF superimposed on the spectrogram in (b); (d) the spectrogram of the Vold-Kalman filtered
signal using the maxima tracked IF in (c); (e) the phase of the IF; (f) the resampling process using the phase of the shaft.

 It does not contain any interference or cross-terms such as the Wigner-Ville distribution [18].
 Fewer hyperparameters need to be optimised than for the Chirplet transform, which requires the chirp rate to be deter-
mined, and the continuous wavelet transform, which requires the scales as well as the wavelet basis function to be
determined.
 It is fairly computationally efficient compared to the Chirplet transform.

The short-time Fourier transform (STFT), performed on a continuous signal xðtÞ is in the form of,

Z 1
Cðt; f Þ ¼ xðsÞtðs  tÞ expð2pfjsÞds ð1Þ
1

pffiffiffiffiffiffiffi
where a window function such as a Hamming window, is denoted by tðtÞ and j ¼ 1. The spectrogram is related to the
STFT in Eq. (1) with jCðt; f Þj2 . Eq. (1) is easily implemented for discrete signals, since the fast Fourier transform is repeatedly
applied on each windowed vibration signal segment to construct the discrete STFT and to obtain the discrete spectrogram.
The discrete spectrogram is described by the integer variables n and l in this paper to obtain jC½n; lj2 , where t ¼ nDt C seconds
and f ¼ lDf C hertz. The time between horizontally adjacent windows in the spectrogram is denoted by Dt C and the frequency
between vertically adjacent windows is denoted by Df C and they describe the time and frequency resolution of the spectro-
gram, respectively.
634 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646

A discrete signal evaluated at time increment n is denoted by x½n, where a continuous signal or a function is denoted by
xðtÞ in this article. The relationship between the discrete time increment n and the continuous time t is,
t ¼ nDt ð2Þ
where Dt is the sampling period. Please note that all discrete indices, such as n in x½n is in the range of ½0; N x  1, if the num-
ber of indices in x is N x for example.

2.2. Proposed maxima tracking procedure

Maxima tracking is performed to find, isolate and track an IF, denoted by f IF ,

f IF ðtÞ ¼ k  cs  f r ðtÞ ð3Þ

in a time-frequency distribution. The IF is related to the rotating frequency or speed of the shaft, denoted by f r ðtÞ, with the
time-invariant factor k  cs . This indicates that it is the kth harmonic of a component with a frequency of cs  f r ðtÞ, such as a
gear mesh frequency or another shaft in the system. This IF presents itself as a ridge or a local maximum in the discrete time-
frequency distribution, where the IF is estimated by solving a constrained optimisation problem in the standard form of,

minimise  jC½n; lj2

l
ð4Þ
2 2
subject to ðl  Df C  f c ½nÞ 6 Df c

where f c denotes the centre of the constraint, which is the predicted value of the IF at time increment n; Df c denotes the fea-
sible bandwidth enforced by the constraint; and l  Df C is the current frequency that is investigated in the spectrogram in
Hertz. If maxima tracking is performed as an unconstrained optimisation problem, different components will be tracked
which results in erroneous phase estimation and order tracking results. The solution of Eq. (4) at time increment n, denoted
by f max ½n, is assumed to be related to the actual IF by,
f IF ðnDtC Þ  f max ½n ð5Þ
where Dt C is the time difference between horizontally adjacent window centres in the spectrogram. Note that the noise and
the fixed time-frequency resolution can result in some differences between the estimated IF f max ½n and the actual IF f IF ,
which motivate the approximation used in Eq. (5). Urbanek et al. [10] centred the constraint in Eq. (4) about the maxima
tracked frequency estimated at the previous time increment,
f c ½n ¼ f max ½n  1 ð6Þ
with the frequency gradient i.e. acceleration information not taken into account. The feasible region (represented by Df c )
must be increased to accommodate the errors made by Eq. (6) when the rotating machine operates under large angular
accelerations. However, if a large Df c is used in the maxima tracking process, it is more susceptible to tracking incorrect fre-
quency components such as other harmonics and background noise which are present in the large feasible region.
Higher order frequency information needs to be incorporated into the maxima tracking process to make it more robust to
track the IF in the presence of large accelerations. The Taylor series expansion of the IF in the signal about the previous time
step,
2
d 1 d
f IF ðtÞ ¼ f IF ðt  DtC Þ þ Dt C f ðt  Dt C Þ þ Dt 2C 2 f IF ðt  DtC Þ þ    ð7Þ
dt IF 2 dt

can be used to predict the value of the IF at the current time step, denoted by t ¼ nDt C , if all of the derivatives of f IF are con-
tinuous. This prediction is used to centre the constraint,
f c ½n ¼ f IF ðnDtC Þ ð8Þ
used in Eq. (4). Eq. (6) can be obtained from Eqs. (7) and (8) by neglecting the higher order information i.e. no acceleration is
present and by assuming Eq. (5) is valid. The problem with Eq. (7) is that the gradients of the IF are unknown and the esti-
mated value of the IF at the previous time increment f max ½n  1 is known instead of the actual value f IF ðt  DtC Þ. The spec-
trogram contains discrete time steps and therefore the gradients in Eq. (7) can only be estimated by using finite difference
schemes on the previous IF estimates i.e. f max ½n  1; f max ½n  2, etc. The estimate of the IF contains noise due to ridge smear-
ing and background noise in the spectrogram, which makes finite difference schemes poor approximations to the gradients
in Eq. (7).
A probabilistic approach is proposed in this article to infer the actual IF so that higher order information can be incorpo-
rated into the maxima tracking process. It is assumed that the true IF, f IF , is related to its representation in the spectrogram or
its estimate from the spectrogram, f max , with,
f max ½n ¼ f IF ðnDt C Þ þ m ð9Þ
 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646 635

where m is a sample from a zero mean Gaussian with variance r2 , which represents the noise that is attributed to ridge
smearing and other noise components in the spectrogram. It is assumed that the noise in the spectrogram is Gaussian dis-
tributed and therefore the estimated IF (or maxima tracked frequency) and the true IF is related with the Gaussian proba-
bility density function p given by,
!
  1 ðf ½n  f IF ðnDt C ÞÞ
2
p f max ½njf IF ðnDt C Þ; r2 ¼ pffiffiffiffiffiffiffiffiffiffiffiffi exp  max ð10Þ
2r2 p 2r 2

where the true IF, denoted by f IF , remains unknown. A model of the true IF, denoted by f F and represented by a N p -order
polynomial in the form of,

f F ðtÞ ¼ w0 þ w1 t 1 . . . wNp tNp ð11Þ

is used instead of the unknown IF f IF . Table 1 summarises the difference between the different notations for the IF. The poly-
nomial weights w ¼ ½w0 ; w1 ; . . . wNp T are estimated by performing maximum likelihood,

Y
n1
 
^ ¼ argmax
w p f max ½ijf F ðiDtC Þ; r2 ð12Þ
w i¼nN m

over the previous N m time steps by assuming that the data are identically and independently distributed. The resulting
weights [19],
1
^ ¼ ðQ T Q Þ Q T f max
w ð13Þ
of the maximum likelihood solution are computed from the N m  ðN p þ 1Þ design matrix Q and the previous N m estimates of
the IF, denoted by a N m  1 vector f max . The design matrix of the N p -order polynomial, in the form of,
2 3
1 Dt C  ðn  1Þ ... ðDtC  ðn  1ÞÞNp
6 7
61 Dt C  ðn  2Þ . . . ðDtC  ðn  2ÞÞNp 7
6 7
Q ¼6. .. .. .. 7 ð14Þ
6. 7
4. . . . 5
1 Dt C  ðn  Nm Þ . . . ðDtC  ðn  N m ÞÞNp
and the vector of the previous N m estimates of the IF,
2 3
f max ½n  1
6 f ½n  2 7
6 max 7
f max ¼ 6
6 ..
7
7 ð15Þ
4 . 5
f max ½n  Nm 
are used in Eq. (13) to obtain the weights used in Eq. (11). The polynomial function in Eq. (11) satisfies the Taylor series
expansion of the IF in Eq. (7), which means that if a first order polynomial function is used, then the assumption is made
that the angular acceleration is constant.
It is strongly suggested to use a first order polynomial to ensure that only a small number of training points (N m ) are
required to avoid overfitting, solving Eq. (13) repeatedly remains computationally efficient and that errors in the extrapola-
tion process are minimised. The number of training points are N m ¼ 5 in this article, which makes the first order approxi-
mation sensible for DtC , which was in the range of ½0:028; 0:05 seconds for all spectrograms in this article. By using Eq.
(6), the IF estimate at the first time increment in the maxima tracking process needs to be provided by the user. If the con-
straint in Eq. (4) is centred by the prediction made by Eq. (11) from a first order polynomial, then the initial gradient needs to
be provided as well. To make this process easier, it is assumed that the initial gradient is zero, which essentially results in Eq.
(6) to be used for the initial steps. This assumption is reasonable if other frequency components are not in close proximity to
the IF under consideration, otherwise a rough estimate of the gradient can be provided as input if the aforementioned
assumption leads to incorrect results. In this study the first three estimates of the IF are obtained by using Eq. (4) with
Eq. (6) and thereafter Eq. (11) can be used to estimate f c . The initial estimate of the IF is obtained by inspecting the spectro-
gram and cs  k, in Eq. (3) and is estimated by roughly knowing the operating range of the machine or using a simple device
such as a hand-held tachometer.

Table 1
The differences between f max ; f IF and f F .

f IF ðtÞ Is the true, continuous IF that need to be estimated by the process

f max ½n Is the IF that is inferred at discrete time steps from the maxima tracking process
f F ðtÞ Is a model that captures the local gradient characteristics of the estimated IF
636 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646

Since it is a constrained minimisation problem, the Lagrangian in the form of,

   
2 2 2 2
Ln ðl; k; sÞ ¼ jC½n; lj2 þ k1 ðlDf C  f max ½nÞ  Df c1 þ s21 þ k2 ðlDf C  f F ðnDt C ÞÞ  Df c2 þ s22 ð16Þ

is minimised to find the frequency f ¼ lDf C that minimises the cost function jC½n; lj2 and adheres to the constraints. The
Lagrangian multiplier and slack variable associated with the ith constraint are denoted by ki and si , respectively. Eq. (16) can
be minimised by solving a set of linear or quadratic programming sub-problems, by solving a penalised unconstrained cost
function etc. [20]. The penalised unconstrained cost function in the form of,
h i h i
jðl; q; nÞ ¼ jC½n; lj2 þ q1 max 0; ðlDf C  f max ½n  1Þ2  Df 2c1 þ q2 max 0; ðlDf C  f F ðnDtC ÞÞ2  Df 2c2 ð17Þ

is minimised using a brute force approach, since the limited number of feasible solutions at each time step, due to the dis-
crete spectrogram, makes it easy to obtain a minimum. The bandwidth describing the feasible region enforced by the two
constraints are denoted by Df c1 and Df c2 , respectively. The parameters q1 and q2 must be chosen so that the cost function
in Eq. (17) is dominated by the constraint terms if the constraints are violated. A procedure is proposed in Section 2.5 to
choose the values of q1 and q2 . In this paper the feasible bandwidths (Df c1 and Df c2 ) in Eq. (17) have the same value. The
components of the vector q ¼ ½q1 ; q2 , in Eq. (17), have the following values,
(
109 if n ¼ 0; 1; 2
q1 ¼ ð18Þ
0 otherwise

and,
(
109 if n > 2
q2 ¼ ð19Þ
0 otherwise

where only one of the constraints are active at a time due to the choice of q. The estimate of the instantaneous frequency is
obtained from,
f max ½n ¼ Df C  argmin jðl; q; nÞ ð20Þ
l

where jðl; q; nÞ is given by Eq. (17). Note that the solution of Eq. (17) is also the solution of Eq. (4) if both constraints are
incorporated into Eq. (4) similarly to Eq. (17). The maxima tracking process is illustrated in Fig. 2. The method implemented
by Urbanek et al. [10] essentially uses q1 > 0, while q2 is always zero in Eq. (17). In this article the proposed maxima tracking
algorithm is compared to the method by Urbanek et al. [10] to highlight the benefits of including acceleration information
into the maxima tracking process.

2.3. The Vold-Kalman filter

The Vold-Kalman filter (VKF) is a bandpass filter with a varying centre frequency and bandwidth. If the centre frequency
of the constraint in Eq. (4) is set equal to the estimated IF, denoted by f max , and the bandwidth is set sufficiently fine, then a
mono-component signal, denoted by xIF can be extracted from a multicomponent vibration signal x. Note that a discrete sig-
nal is investigated in this section and therefore the whole signal is denoted by x i.e. a one-dimensional vector over all time
steps. The signal at increment n is denoted by x½n. The Vold-Kalman filter allows slight errors in the maxima tracking process
and this is the reason why a Vold-Kalman filter is used, instead of integrating the estimated maxima tracked frequency,
obtained from Eq. (20).
Two generations of the VKF exist, namely the angular-velocity (or first generation) and the angular-displacement (or sec-
ond generation) VKF. The angular-velocity VKF aims to estimate the mono-component signal, while the angular-
displacement VKF aims to estimate the envelope of the mono-component signal with the supplied frequency content
[21]. The signal containing the IF that needs to be extracted, xIF ½n is related to the vibration signal x½n in the data equation
of the VKF,

Spectrogram properties (|Γ[n, l]|2 for each l, ΔtΓ , ΔfΓ ) and ρ, Δfc1 , Δfc2

l
fmax [n − 1] Find l which minimises Eq. (17) × fmax [n]

ΔfΓ
fF (nΔtΓ )

Fig. 2. The maxima tracking process is illustrated in this figure at time increment n and this is performed for each time increment n. Note that for example if
q1 in Eq. (17) is zero, then f max ½n  1 do not need to be used in the process.
 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646 637

x½n ¼ xIF ½n þ f½n; ð21Þ

where f½n represents the other sinusoidal components and the background noise at time t ¼ n=f s . The structural equation for
the angular-velocity VKF [21],

xIF ½n  2 cosð2pf IF ½nf s ÞxIF ½n þ 1 þ xIF ½n þ 2 ¼ e½n

1
ð22Þ

and of the angular-displacement VKF,

b½n  b½n þ 1 ¼ e½n ð23Þ

describes the characteristic of interest (either the signal component or its envelope) and e½n is known as the non-
homogeneity term at time increment n which represents the error in the left-side of the structural equation. Eq. (23) is
for a one-pole angular-displacement filter, where the structural equation for other poles are found in the paper by Tuma
[21]. The envelope b½n is related to the signal of interest by,
!
2pj X n
xIF ½n ¼ b½n exp f ½i ð24Þ
f s i¼0 IF

pffiffiffiffiffiffiffi
where f s is the sampling frequency of the signal xIF ; j ¼ 1 and the instantaneous frequency f IF ½i is obtained from Eq. (5)
and by using the maxima tracking information. There are three sets of unknowns in Eq. (21)–(23) namely e; f and either b for
the angular-displacement VKF or xIF for the angular-velocity VKF. The dimensions of the aforementioned vectors are N x  1,
where N x is the number of samples in the signal x. This results in an underdetermined system of equations, which is solved
by finding the characteristic of interest (xIF or b) that simultaneously minimises e and f. This is achieved by solving the multi-
objective optimisation problem with the weighted sum method in the form of [21],

j ¼ a2 eT e þ fT f ð25Þ

where a is the weighting factor of the Vold-Kalman filter and it describes the relative importance of minimising eT e instead of
fT f. The weighting factor is given by [21],
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffi ffi
21
a¼ ð26Þ
2ð1  cos ð2pf b =f s ÞÞ

for a one-pole angular-displacement VKF filter, where f b is the bandwidth of the filter. A unique weighting factor, found in
Eq. (26), can be applied at each time step, which provides more flexibility to the VKF,

j ¼ eT AT Ae þ fT f ð27Þ

where the N x  N x weight matrix A is a diagonal matrix containing the local weight factors and N x is the number of samples
in the signal x. The component at time increment n, denoted by A½n; n, describes the importance of minimising eT e instead of
fT f at the time increment. The choice of weighting factor at time increment n, denoted by A½n; n, determines the rise time as
well as the bandwidth of the filter and is very important to the success of the VKF [22]. The angular-displacement VKF allows
multiple components to be tracked, while the angular-velocity VKF is constrained to a single component [21]. Due to the
nature of the data and structural equations, the optimal solution of the cost function, where the cost function is given by
Eq. (27), can be obtained in closed form if only a single component is tracked [21].
Pan and Lin [22] found that the angular-displacement VKF is able to estimate the component of interest accurately in the
presence of crossing orders. It obtains a better rise time and bandwidth combination than the angular-velocity VKF as well.
The angular-displacement VKF does not have any frequency nor slew rate limitations [23]. In the authors’ experience, it is
found that the one-pole angular-displacement filter is able to extract the component of interest using a small relative band-
width, given by,

f b ½n
Bw ¼ ð28Þ
f IF ½n

without any numerical problems. A one-pole angular-displacement VKF used with a small relative bandwidth makes the
phase error independent of the harmonic that is filtered i.e. k in Eq. (3), which reduces the number of hyperparameters that
need to be optimised. Hence a one-pole angular-displacement VKF is used in this study to obtain a mono-component signal
from the multicomponent vibration signal. If Bw is large, it means that relatively large maxima tracking errors are accommo-
dated in the process at the risk of other frequency components being passed through the passband as well. The values of Bw
are provided in Section 3 for each investigation.
638 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646

2.4. Instantaneous phase estimation and resampling

The phase estimation process is presented in this section for continuous signals, however it is easily extended to discrete
signals obtained from the maxima tracking and the Vold-Kalman filtering processes. The mono-component signal obtained
from the VKF in the previous section, denoted by xIF , is transformed with the Hilbert transform,
Z 1
1 1
HðxIF ðtÞÞ ¼ xIF ðsÞ ds ð29Þ
p 1 ts

and used to obtain the instantaneous phase of this signal,

 
HðxIF ðtÞÞ
/IF ðtÞ ¼ arctan ð30Þ
xIF ðtÞ

which is constrained to ½p; p radians due to the arctangent function. This phase, obtained from Eq. (30), is unwrapped to
obtain the cumulative phase of the estimated IF component over time. The tracked IF component is related to the rotating
shaft of interest with Eq. (3) if Eq. (5) is valid. The aforementioned relationship is used to obtain the instantaneous phase of
the shaft,

1
/r ðtÞ ¼ unwrapð/IF ðtÞÞ ð31Þ
kcs

at time step t. The instantaneous phase of the shaft over time, obtained from Eq. (31), is used to resample the signal into
equal angular increments as shown by (f) in Fig. 1.

2.5. Guidelines for choosing parameters

The adjustable parameters in the proposed tacholess order tracking process are the

 properties of the spectrogram namely, the window length (Lw ), window overlap (Lo ), window type and zero padding (Lz );
where Lw ; Lo and Lz are given in the number of samples that is used.
 initial value of the estimated instantaneous frequency f max and its relationship with the rotational speed of the shaft cs  k.
 properties of the maxima tracking process such as N m and N p .
 allowable bandwidth of the constraints of the maxima tracking process Df c and q.
 relative bandwidth, type and order of the Vold-Kalman filter.

Some guidelines are proposed in this section to select the unknown parameters. Please note that the maxima tracking
properties such as N m and N p are discussed and motivated in Section 2.2 while the Vold-Kalman properties are thoroughly
discussed and motivated in Section 2.3 and are therefore not repeated in this section. The aforementioned parameters are
summarised in Section 3.2.3 for each investigated numerical and experimental case.
The time and frequency resolution of the spectrogram is related to the window length, window overlap and zero padding
by Dt C / Lw  Lo and Df C / 1=ðLw þ Lz Þ, respectively. The appropriate window length is problem dependent and should be
chosen to have sufficient time resolution to estimate varying rotational speeds without a significant loss in frequency res-
olution. In this paper, typical window lengths of 0.1–0.5 s are investigated. Large window overlaps are suggested (typically
50–90% of Lw ) and zero padding of Lz ¼ 2  ð2LLw  Lw Þ where LLw ¼ ceilðlog2 ðLw ÞÞ to ensure that relatively good time and fre-
quency resolutions are obtained. The ceiling function is denoted by ceilðÞ. Rectangular windows are used in this paper similar
to [9], but other windows can be investigated if desired.
The initial value of f max can be chosen by inspecting the spectrogram and looking for dominant meshing components,
while cs will be the number of teeth on the gear that is connected to the shaft under consideration. The constant k can be
estimated by approximately knowing the speed of the shaft under consideration. Hand-held tachometers can for example
be used to obtain a rough estimate of the operating range of the shaft under consideration, which can subsequently be used
to estimate k.
The feasible bandwidth Df c indicates the allowable error when centring the constraint f c and the appropriate character-
istics are dependent on the harmonic that is tracked as well as the local frequency and noise characteristics in the spectro-
gram. Hence, in many cases the appropriate Df c is chosen on a trail and error basis. Due to the fact that a first order
polynomial is used, some errors can be present in the prediction due to unpredictable changes in operating conditions
and therefore using Df c ¼ 0 is unrealistic. It is suggested to start with Df c in the range of 5  20Hz for the datasets that
are investigated.
The parameter q ensures that the constraint violation dominate the cost function in Eq. (17). Assume that if constraint i,
governed by qi , is violated by ðq  1ÞDf c where q > 1, then the objective function jðl; q; nÞ > 0 for all integers l and n. This
condition will guarantee that f max , obtained from the optimisation process, will be within the range ½f c  qDf c ; f c þ qDf c  for
 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646 639

 
all time increments. The solution to this problem is qi > jCmax j2 = ðq2  1ÞDf c where jCmax j2 is the absolute maximum value
2

of the spectrogram and q can for example be 1:0001 to simulate a 0:01% constraint violation.

3. Validation

Three numerical and three experimental datasets are used to validate the proposed tacholess order tracking method. The
performance of the proposed maxima tracking algorithm is critically compared to the performance of the maxima tracking
method used by [10], where the latter is used as a benchmark. After the maxima tracking process is completed, the same
process (i.e. the Vold-Kalman filter, Hilbert transform, etc.) is followed for both maxima tracking algorithms. The results from
the benchmark maxima tracking algorithm is indicated by (a) and the proposed method is indicated by (b) in all subsequent
figures.
The vibration signal is order tracked using the estimated instantaneous phase and therefore the performance of the tac-
holess order tracking method (or the maxima tracking algorithm) is quantified using the estimated phase error. The phase
error, in this study, is computed from,
eðtÞ ¼ j/r ðtÞ  /a ðtÞj ð32Þ
where the cumulative analytical and estimated phase of the shaft, in radians, are denoted by /a and /r respectively. The
maximum relative phase error is computed from,
maxðeðtÞÞ
eR ¼ ð33Þ
maxð/a ðtÞÞ
where maxð/a ðtÞÞ denotes the maximum value of the analytical instantaneous phase for the time interval and eðtÞ is calcu-
lated from Eq. (32). The numerical data are considered first in this section, whereafter the experimental data are presented.
Note that the parameters that are used in the tacholess order tracking process are summarised for the numerical and exper-
imental data in Section 3.2.3.

3.1. Numerical data

Two rotational speed profiles are investigated on the numerical data. The first rotational speed profile of the shaft of inter-
est in Hertz,
f r;1 ðtÞ ¼ 200t þ 50 ð34Þ
has a constant angular acceleration of 400p [rad/s2] over 20 s. The second rotational speed profile in Hertz,
1 1:25
f r;2 ðtÞ ¼ cosð0:1ptÞ þ ð35Þ
5 p
has a varying angular acceleration. Note that the rotational speed in Eq. (35) is very small, but a gear mesh frequency of a
simulated gearbox, given by f r;2  40 is introduced and tracked on the spectrogram of the simulated signal, which has a sig-
nificantly larger frequency gradient.
The first simulated vibration signal is in the form of,
 Z t 
x1 ðtÞ ¼ sin 2p f r;1 ðsÞds ð36Þ
0

and contains a frequency component given by Eq. (34). Note that the signal is a mono-component signal and therefore the
Hilbert transform can directly be applied. However, this signal is investigated to evaluate the performance of the proposed
method on a mono-component dataset and then an additional component is incorporated in the second simulated vibration
signal.
The second simulated vibration signal is in the form of,
 Z t   
x2 ðtÞ ¼ sin 2p f r;1 ðsÞds þ sin 2p103 t ð37Þ
0

where it contains the rotational speed given in Eq. (34) and an additional constant frequency component. This rotational
speed invariant component simulates noise generated by resonance bands or by environmental effects. The effect of adding
spurious components to the signal is investigated by comparing the results of the maxima tracking algorithms on the signal
given by Eqs. (36) and (37). The third and last simulated vibration signal is in the form of,
3
X  Z t   Z t 

x3 ðtÞ ¼ 1:5 sin 2:0ps f r;2 ðsÞds þ sin 80:0ps f r;2 ðsÞds þ 0:5 sin ð2p½1:65t þ 2:0t Þ þ 0:5 sinð80:0ptÞ
s¼1 0 0

ð38Þ
640 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646

and contains a shaft component given by Eq. (35) and its three harmonics, a gear mesh frequency (40  f r;2 Hz) and its three
harmonics, a linear component and a constant frequency component. The two latter frequency components are independent
of the rotation speed and complicates the maxima tracking task.
The ratio of the Df c to spectrogram frequency resolution, is selected as 16 for the first two simulated signals and 2.73 for
the last signal for the maxima tracking process. Note that it should be significantly larger for the first two simulated signals
due to the large frequency gradient. The VKF relative bandwidth is selected to be 0.1 for the first two signals and 104 for the
last signal. Gaussian noise with a standard deviation of 1.0 is added to the first two signals and Gaussian noise with a stan-
dard deviation of 2.5 is added to the last simulated signal.
The spectrogram of the signal given by Eq. (36) is computed and the maxima tracking algorithms are performed on the
spectrogram, separately. The result of the benchmark and the proposed maxima tracking procedures are superimposed on
the spectrogram in Fig. 3i and indicated by (a) and (b), respectively. The angular-displacement VKF is applied to the signal,
whereafter the Hilbert transform is applied to calculate the phase of the component. The phase error is calculated with Eq.
(32) and given in Fig. 3ii. Both the maxima tracking approaches resulted in very similar phase errors, with the proposed max-
ima tracking method performing slightly better. Note that the phase error resulted from a constant frequency error which
results from the limited time-frequency resolution of the spectrogram and the large frequency gradient. The constant fre-
quency errors can be reduced by increasing the relative bandwidth of the VKF. However care must be taken in multicompo-
nent signals where there is a risk of tracking another frequency component. Even though there is a linear phase error, the
final phase error is significantly smaller than the final instantaneous phase of 251324.899 radians which results in a max-
imum relative phase error, calculated by Eq. (33), of 0.3849% and 0.3724% for the benchmark and the proposed method,
respectively.
The same process is followed for the second simulated vibration signal given in Eq. (37). The results from the two maxima
tracking methods are superimposed on the spectrogram of the vibration signal in Fig. 4i. The signal contains two frequency
components with equal amplitudes, where the one frequency component is unrelated to the rotational speed of the shaft of
interest. The two components can only be distinguished from one another in the spectrogram by evaluating the frequency
gradient. The results in Fig. 4i indicate that the benchmark method follows the signal component with a gradient of zero,
which is expected. In contrast, the proposed method tracks the correct frequency component. The phase errors are compared
in Fig. 4ii, which proves that the proposed method performs significantly better and it is significantly more robust on a signal
with spurious components. The performance of the proposed maxima tracking algorithm does not change significantly
between signal 1 and signal 2 (i.e. the addition of a spurious component), however the performance of the benchmark
method is sensitive to the addition of another component.
The last simulated signal, given in Eq. (38), is investigated similarly to the previous methods. The resulting maxima track-
ing results are superimposed in Fig. 5i on the spectrogram and result in the phase errors given in Fig. 5ii. A similar result is
observed in Fig. 5 as in Fig. 4. The benchmark maxima tracking algorithm is incapable of distinguishing between the constant
frequency components and the IF that needs to be tracked in Fig. 5i. The proposed maxima tracking algorithm tracks the
correct IF for the whole duration of the signal. This is reflected in the phase errors compared in Fig. 5ii.
The performance of the two maxima tracking approaches is summarised in Table 2. The roughly estimated signal-to-noise
ratio in decibels (db),
0 2 1
PN1 
B i¼0 j C IF ½ij
2
C
¼ 10 log 10@ 2 A ð39Þ
PN1 
SNRdb
i¼0 j C noise ½ij 2

105
104 (a)
103 (b)
Phase error [rad]

102
101
100
10−1
10−2
10−3
10−4
10−5
0 5 10 15
Time [s]
(i) (ii)

Fig. 3. The result of the proposed tacholess order tracking method on the first simulated signal, given by Eq. (36). The maxima tracking results are
superimposed on the spectrogram of the signal in figure i with the resulting phase errors, calculated from Eq. (32), shown in figure ii. The benchmark
maxima tracking algorithm is indicated as (a) in the figures and the proposed maxima tracking algorithm is indicated by (b).
 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646 641

105
(a)
104 (b)

Phase error [rad]

103
102
101
100
10−1
10−2

0 5 10 15
Time [s]
(i) (ii)

Fig. 4. The result of the proposed method on the second simulated signal, given by Eq. (37). The maxima tracking results are superimposed on the
spectrogram of the signal in figure i with the resulting phase errors, calculated from Eq. (32), shown in figure ii.

(a)
100 (b)

Phase error [rad]

10−1

10−2

10−3

10−4

0 5 10 15 20
Time [s]
(i) (ii)

Fig. 5. The result of the proposed method on the third simulated signal, given by Eq. (38). The maxima tracking results are superimposed on the
spectrogram of the signal in figure i with the resulting phase errors, calculated from Eq. (32), shown in figure ii.

Table 2
Summary of the maxima tracking results from the numerical signals. All the phase units are in radians, except the maximum relative phase error.

Signal 1 Signal 2 Signal 3

Maximum phase completed 251324.899 251324.899 314.1593
Signal-to-noise ratio (db) 52.79 22.71 20.35
Benchmark method [10]
RMS phase error 576.3152 54636.0666 1.4576
Maximum phase error 967.3627 141189.2577 5.3663
Maximum relative phase error (%) 0.38491 56.1780 1.70815
Proposed method:
RMS phase error 540.9771 459.872 0.039169
Maximum phase error 935.9197 829.7571 0.063935
Maximum relative phase error (%) 0.372394 0.330153 0.020351

is presented in the table as well, where jCIF ½ij2 is the magnitude of the spectrogram of the IF at time step i and jCnoise ½ij2
presents the amplitude of the noise around the IF in the spectrogram at time step i. The local signal-to-noise ratio is pre-
sented to give the approximate conditions in which the maxima tracking was performed in. The proposed maxima tracking
algorithm, in contrast to the benchmark algorithm, is able to track the correct frequency component in the presence of other
spurious components. The spurious noise components emulate resonance bands which are excited due to the presence of
damage and other noise components during data acquisition, which is present when performing condition monitoring on
rotating machines. The quality of the order tracked vibration signal depends on the accuracy of the estimated instantaneous
phase and is investigated on the experimental data in the next section.

3.2. Experimental data

Three aspects of the proposed tacholess order tracking method is evaluated in this section: The ability of the proposed
tacholess order tracking method,
642 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646

1. to estimate the instantaneous phase correctly;

2. to order track the vibration signal;
3. to estimate the rotational speed of the shaft of interest.

3.2.1. Experimental setup

The experimental setup in Fig. 6 was designed by Stander and Heyns [24] and was refurbished to perform gear fatigue
tests under fluctuating conditions. The electrical motor supplies rotational energy to the system and the alternator, which
is connected to a resistor bank, dissipates the energy from the system by applying a counteracting load to the motor. A per-
sonal computer is used to control the rotational speed and the load in the system. Three gearboxes are present in this setup,
with the gearbox that is monitored with accelerometers labelled as the monitored gearbox. The input shaft of the monitored
gearbox contains a zebra tape shaft encoder, with 88 PPR, and an optical probe. The output of the monitored gearbox con-
tains a proximity probe that is triggered by the shaft key and produces 1 PPR.
The axial vibration signal, obtained from the tri-axial accelerometer that is on the bearing housing of the monitored gear-
box, is used to extract the rotational information from. The axial vibration signal is used because the helical gears cause
strong axial excitations which result in prominent meshing components in its vibration spectrum. The tachometer signals
are both sampled at 51.2 kHz and the acceleration signal is sampled at a rate of 25.6 kHz with an OROS OR35 data acquisition
device.
Three rotational speed profiles, shown in Fig. 7, were applied to the system during which vibration measurements and
tachometer measurements were made. The unique rotational speed profiles contained relatively large rotational speed fluc-
tuations to validate the proposed methodology. The tachometer signal that is associated with the 88 PPR zebra tape shaft
encoder, geometrically compensated with the technique proposed by Diamond et al. [6], is used to compute the rotational
speeds in Fig. 7 as well as the analytical phase of the experimental data.

3.2.2. Maxima tracking and phase estimation results

The performance of the maxima tracking algorithms is evaluated on the experimental data in this section, where a similar
process to the numerical investigation is followed. The Df c is chosen to be equal to 6.4 times the frequency resolution of the
spectrogram and the relative bandwidth of the VKF is 104 on all investigated cases. The spectrogram of the vibration signal
associated with the first rotational speed profile (see (a) in Fig. 7) is superimposed with the results of the two maxima track-
ing approaches in Fig. 8i, where the benchmark and the proposed method are indicated by (a) and (b), respectively. The gear
mesh frequencies of the monitored gearbox are tracked in Fig. 8i, however it is immersed in noise induced by the resonance
bands and other frequency components. This adversely affects the performance of the benchmark maxima tracking algo-
rithm in the regions with large frequency gradients as seen in Fig. 8i. In contrast, the proposed maxima tracking algorithm,
is able to track the correct frequency component and resulted in a significantly better phase estimate as shown in Fig. 8ii. The
large phase errors associated with the benchmark method is attributed to the maxima tracking algorithm drifting off from
the true estimate.
The spectrogram of the vibration signal, associated with the second rotational speed profile (see (b) in Fig. 7), is shown in
Fig. 9i. The results from the maxima tracking processes are superimposed in Fig. 9i, where it is seen that both algorithms are
able to track the component of interest. The resulting phase errors of the two maxima tracking approaches used in the pro-
posed tacholess order tracking method, presented in Fig. 9ii, confirms both maxima tracking algorithms performed well.
Even though the benchmark maxima tracking algorithm resulted in the maximum phase error, on average it performs
slightly better.
The spectrogram of the last experimental vibration signal, associated with profile (c) in Fig. 7, is superimposed with the
maxima tracking results of the two algorithms illustrated in Fig. 10i. The resulting phase errors are shown in Fig. 10ii, where
it can be seen that the proposed maxima tracking method performs significantly better than the benchmark maxima track-
ing algorithm.
The results of the three experimental profiles are summarised in Table 3. The signal-to-noise ratio is calculated with Eq.
(39) and is presented in Table 3. The relative phase error of the proposed tacholess order tracking method, calculated from
Eq. (33) for the different experimental signals, is very small. The proposed maxima tracking algorithm results in significant
improvements in profile 1 and profile 3, and performs almost equally well with the benchmark maxima tracking algorithm
on profile 2.

Fig. 6. Experimental setup.

 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646 643

(a) (b) (c)

18 15 15

Rotational speed [rad/s]

14 14
16
13 13
14
12 12

12 11 11
0 5 10 15 0 10 20 30 0 10 20 30
Time [s] Time [s] Time [s]

Fig. 7. Experimental rotational speed profiles with Profile 1 indicated by (a), Profile 2 indicated by (b) and Profile 3 indicated by (c) in this figure.

102
101
100

Phase error [rad]

10−1
10−2
10−3
10−4
10−5
(a)
10−6 (b)
10−7
5 10 15
Time [s]
(i) (ii)

Fig. 8. The spectrogram of the vibration signal from the first experimental rotational speed profile is superimposed with the maxima tracking results in
figure i. The resulting phase estimation errors of the two results are compared in figure ii. Note that the benchmark maxima tracking method’s results are
indicated by (a) and the results from the proposed method are indicated by (b).

0.1
Phase error [rad]

0.0316

0.01

0.0032
(a)
(b)
0.001
10 20 30
Time [s]
(i) (ii)

Fig. 9. The spectrogram and the maxima tracked results are presented in figure i with the resulting phase errors for the benchmark and the proposed
maxima tracking algorithms in figure ii using the proposed tacholess order tracking method. This is performed for the second experimental signal shown as
(b) in Fig. 7.

3.2.3. Tacholess order tracking parameters

The parameters used in the tacholess order tracking process for the numerical and the experimental datasets are pre-
sented in Table 4 with the caption describing the parameters.

3.2.4. Order tracking results

The instantaneous phase is used in the previous sections to quantify the performance of the proposed tacholess order
tracking method with respect to using the tachometer signal obtained from the zebra tape shaft encoder and the optical
probe. The instantaneous phase is used to resample the vibration signal with a constant angular frequency as illustrated
by (f) in Fig. 1. In this section the spectrum of the order tracked signals, using the proposed tacholess order tracking method
644 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646

10−1

10−2

Phase error [rad]

10−3

10−4
(a)
(b)
−5
10
5 10 15 20 25 30
Time [s]
(i) (ii)

Fig. 10. The spectrogram and the maxima tracked results are presented in figure i with the resulting phase errors for the benchmark and the proposed
maxima tracking algorithms in figure ii using the proposed tacholess order tracking method. This is performed for the third experimental signal shown as
(c) in Fig. 7.

Table 3
Summary of the maxima tracking results from the experimental rotational speed profiles. All the phase units are in radians, except the maximum relative phase
error which is in %.

Profile 1 Profile 2 Profile 3

Maximum phase completed 295.2322 439.753 439.7533
Signal-to-noise ratio (db) 19.35 20.49 18.04
Benchmark method by [10]
RMS phase error 8.3951 0.030019 0.053025
Maximum phase error 16.8094 0.048112 0.069044
Maximum relative phase error (%) 5.6936 0.010941 0.015701
Proposed method
RMS phase error 0.016662 0.033053 0.037182
Maximum phase error 0.037697 0.04343 0.050224
Maximum relative phase error (%) 0.012768 0.009876 0.011421

Table 4
The table contains the parameters that were used in the numerical (Num.) and experimental (Exp.) analyses. The sampling frequency of the vibration signal f s ;
the window length, window overlap, zero padding and window type that are used for the spectrogram; The number of data points Nm and the order of the
polynomial N p used in the maxima tracking process; cs and k which are used to calculate the instantaneous phase; Df c describes the feasible region of the
maxima tracking process; q is used in the maxima tracking optimisation process; Bw is used in the Vold-Kalman filter.

Unit Num. 1 Num. 2 Num. 3 Exp. 1 Exp. 2 Exp. 3

fs Hz 10,000 10,000 1500 25,000 25,000 25,000

Window length Samples 4000 4000 750 3500 5000 5000
Window overlap Samples 3600 3600 675 2800 4000 4000
Zero padding Samples 4192 4192 1298 4692 11384 11384
Window type – Rect. Rect. Rect. Rect. Rect. Rect.
Nm – 5 5 5 5 5 5
Np – 1 1 1 1 1 1
cs – 1 1 40 37 37 37
k – 1 1 3 7 5 5
Df c Hz 20 20 2 20 10 10
q – 109 109 109 109 109 109
Bw – 0.1 0.1 0.0001 0.0001 0.0001 0.0001

and the tachometer signal from the zebra tape shaft encoder are compared. The linear amplitude spectrum of the two
approaches are compared in Fig. 11 for the first experimental profile (See (a) in Fig. 7).
The resulting spectrum using the tacholess and the tacho-based approaches, in Fig. 11, are very similar. A more detailed
view of the lower frequency spectrum is shown in Fig. 12 for the same order tracked signal as used in Fig. 11. The frequency
components of the shafts and the gear mesh frequencies are easily identified in the spectrum with clear and distinct peaks.
The results in Fig. 11 and in Fig. 12, attest to the ability of the proposed method to order track the measured vibration signal
without using a tachometer.
 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646 645

0.04
Zebra tape shaft encoder
Magnitude [m/s2 ] Tacholess
0.03

0.02

0.01

0
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150
Frequency [Order]

Fig. 11. The linear spectrum of the order tracked vibration signal using the two order tracking approaches.

10−2
Magnitude [m/s2 ]

10−4

10−6
Zebra tape shaft encoder
Tacholess
10−8
0 1 2 3 4 5 6 7 8 9 10
Frequency [Order]

Fig. 12. The 0–10 order spectrum of the order tracked vibration signal using the two order tracking approaches.

(a) (b)
Rotational speed [rad/s]
Rotational speed [rad/s]

18 16.4

16.3
16
16.2
14 Zebra tape shaft encoder
fmax /(cs × k) 16.1
dφr (t)/dt
12 16
0 5 10 15 3 3.2 3.4
Time [s] Time [s]

Fig. 13. The rotational speed estimates of three approaches on the first experimental speed profile. The full view is shown in (a), while a zoomed view is
shown in (b).

3.2.5. Rotational speed estimation

The last aspect that is briefly investigated in this study is the ability of the proposed method to estimate the rotational
speed of the shaft of interest. The normalised maxima tracked frequency, f max =ðcs  kÞ, and the derivative of the Hilbert trans-
form phase is superimposed with the rotational speed calculated from the tachometer signal associated with the zebra tape
shaft encoder in Fig. 13. It is evident from the results that the derivative of the instantaneous phase of the shaft, obtained
from the Hilbert transform, and the maxima tracked frequency performs well in estimating the rotational speed of the shaft.
The RMS error, obtained from the estimated rotational speed by f max =ðcs  kÞ is 0.0423 rad/s. The RMS error of the rotational
speed obtained from the Hilbert transform is 0.0425 rad/s, which is slightly higher than the f max =ðcs  kÞ estimate due to the
end point effects of the Hilbert transform. The estimated rotational speed does not contain diagnostic information similar to
the zebra tape shaft encoder signal, however it provides invaluable information as well. This information can assist with
prognostics and the maintenance task in general, since the number of cycles completed by the different components in cer-
tain operating regimes can be estimated.
646 S. Schmidt et al. / Mechanical Systems and Signal Processing 100 (2018) 630–646

4. Conclusions and recommendations

The proposed tacholess order tracking method is validated on three numerical and three experimental datasets. The
instantaneous phase, which is used to resample the vibration signal from the time to the angle domain, is accurately esti-
mated with the proposed method on the six datasets. The results also indicate that it is more robust to noise and other fre-
quency components than the benchmark maxima tracking algorithm. The low phase estimation errors result in very good
order tracking results, which is the main objective when applying this technique. It is also possible to accurately estimate
the mean rotational speed with the proposed tacholess order tracking method, which can for example be used to support
maintenance decisions and can be used in the remaining useful life estimation process of the rotating machine components.
It is however suggested that the technique needs to be employed on datasets, obtained from industrial environments so
that it can fully be validated. It is recommended that techniques need to be explored which can help to estimate a good ini-
tial value of the IF and the k  cs term so that the tacholess order tracking process can automatically be performed. The char-
acteristics of the angular-displacement Vold-Kalman filter can be investigated further as well to ensure that the performance
is optimal for a broad range of applications. Incorporating acceleration information into the process can also be investigated
with multi-order techniques in future investigations, because it can lead to even more robust tacholess order tracking
techniques.

Acknowledgements

The authors gratefully acknowledge the support that was received from the Eskom Power Plant Engineering Institute
(EPPEI) in the execution of the research.

References

[1] J. Antoni, F. Bonnardot, A. Raad, M. El Badaoui, Cyclostationary modelling of rotating machine vibration signals, Mech. Syst. Sign. Process. 18 (6) (2004)
1285–1314.
[2] T. Heyns, P.S. Heyns, J.P. De Villiers, Combining synchronous averaging with a Gaussian mixture model novelty detection scheme for vibration-based
condition monitoring of a gearbox, Mech. Syst. Sign. Process. 32 (2012) 200–215.
[3] C.J. Stander, P.S. Heyns, W. Schoombie, Using vibration monitoring for local fault detection on gears operating under fluctuating load conditions, Mech.
Syst. Sign. Process. 16 (6) (2002) 1005–1024.
[4] X. Qi, Z. Yuan, X. Han, Diagnosis of misalignment faults by tacholess order tracking analysis and RBF networks, Neurocomputing 169 (2015) 439–448.
[5] D. Abboud, J. Antoni, S. Sieg-Zieba, M. Eltabach, Envelope analysis of rotating machine vibrations in variable speed conditions: a comprehensive
treatment, Mech. Syst. Sign. Process. 84 (2017) 200–226.
[6] D.H. Diamond, P.S. Heyns, A.J. Oberholster, Online shaft encoder geometry compensation for arbitrary shaft speed profiles using Bayesian regression,
Mech. Syst. Sign. Process. 81 (2016) 402–418.
[7] K. Fyfe, E. Munck, Analysis of computed order tracking, Mech. Syst. Sign. Process. 11 (2) (1997) 187–205.
[8] 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. Sign. Process. 19 (4) (2005) 766–785.
[9] Q. Leclère, H. André, J. Antoni, A multi-order probabilistic approach for instantaneous angular speed tracking debriefing of the CMMNO’14 diagnosis
contest, Mech. Syst. Sign. Process. 81 (2016) 375–386.
[10] J. Urbanek, T. Barszcz, J. Antoni, A two-step procedure for estimation of instantaneous rotational speed with large fluctuations, Mech. Syst. Sign.
Process. 38 (1) (2013) 96–102.
[11] F. Combet, L. Gelman, An automated methodology for performing time synchronous averaging of a gearbox signal without speed sensor, Mech. Syst.
Sign. Process. 21 (6) (2007) 2590–2606.
[12] M. Zhao, J. Lin, X. Wang, Y. Lei, J. Cao, A tacho-less order tracking technique for large speed variations, Mech. Syst. Sign. Process. 40 (1) (2013) 76–90.
[13] J. Lin, M. Zhao, A review and strategy for the diagnosis of speed-varying machinery, in: 2014 IEEE Conference on Prognostics and Health Management
(PHM), IEEE, 2014, pp. 1–9.
[14] C.J. Stander, P.S. Heyns, Transmission path phase compensation for gear monitoring under fluctuating load conditions, Mech. Syst. Sign. Process. 20 (7)
(2006) 1511–1522.
[15] K.C. Gryllias, I.A. Antoniadis, Estimation of the instantaneous rotation speed using complex shifted Morlet wavelets, Mech. Syst. Sign. Process. 38 (1)
(2013) 78–95.
[16] T. Heyns, P.S. Heyns, R. Zimroz, Combining discrepancy analysis with sensorless signal resampling for condition monitoring of rotating machines under
fluctuating operations, in: Ninth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies, vol. 2 (2), 2012,
pp. 52–58.
[17] G. He, K. Ding, W. Li, X. Jiao, A novel order tracking method for wind turbine planetary gearbox vibration analysis based on discrete spectrum
correction technique, Renew. Energy 87 (2016) 364–375.
[18] Z.K. Peng, F.L. Chu, Application of the wavelet transform in machine condition monitoring and fault diagnostics: a review with bibliography, Mech.
Syst. Sign. Process. 18 (2) (2004) 199–221.
[19] C.M. Bishop, Pattern Recognition and Machine Learning, Springer, 2009.
[20] J.S. Arora, Introduction to Optimum Design, Elsevier/Academic Press, 2011.
[21] J. Tuma, Setting the passband width in the Vold-Kalman order tracking filter, in: Twelfth International Congress on Sound and Vibration, Lisbon, 2005,
pp. 1–8.
[22] M.-C. Pan, Y.-F. Lin, Further exploration of Vold-Kalman-filtering order tracking with shaft-speed information – I: Theoretical part, numerical
implementation and parameter investigations, Mech. Syst. Sign. Process. 20 (5) (2006) 1134–1154.
[23] J.R. Blough, A survey of DSP methods for rotating machinery analysis, what is needed, what is available, J. Sound Vib. 262 (2003) 707–720.
[24] C.J. Stander, P.S. Heyns, Instantaneous angular speed monitoring of gearboxes under non-cyclic stationary load conditions, Mech. Syst. Sign. Process. 19
(4) (2005) 817–835.