Journal of Sound and Vibration 418 (2018) 184–199

Contents lists available at ScienceDirect

Journal of Sound and Vibration

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

Estimation of single plane unbalance parameters of a

rotor-bearing system using Kalman filtering based force
estimation technique
Akash Shrivastava, A.R. Mohanty *
Department of Mechanical Engineering, Indian Institute of Technology, Kharagpur 721302, India

article info abstract

Article history: This paper proposes a model-based method to estimate single plane unbalance parame-
Received 13 August 2017 ters (amplitude and phase angle) in a rotor using Kalman filter and recursive least square
Revised 10 November 2017 based input force estimation technique. Kalman filter based input force estimation tech-
Accepted 11 November 2017
nique requires state-space model and response measurements. A modified system equiva-
Available online XXX
lent reduction expansion process (SEREP) technique is employed to obtain a reduced-order
model of the rotor system so that limited response measurements can be used. The method
Keywords: is demonstrated using numerical simulations on a rotor-disk-bearing system. Results are
Kalman filter presented for different measurement sets including displacement, velocity, and rotational
Model-based fault diagnosis
response. Effects of measurement noise level, filter parameters (process noise covariance and
System equivalent reduction expansion
process
forgetting factor), and modeling error are also presented and it is observed that the unbalance
Unbalance identification parameter estimation is robust with respect to measurement noise.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction

Unbalance is the most common fault and source of vibration in a rotating machinery [1]. Rotating unbalance can be catego-
rized into single and multi-plane based on the unbalance present in single and multi-disks, respectively. Single plane unbalance
occurs in gear wheels, single stage turbine wheels, compressor, pump impellers etc. The present work aims to propose an on-line
estimation technique for single plane unbalance parameters, i.e., amplitude and phase angle.
Unsupervised fault diagnostic techniques of rotor systems mainly rely on signal processing based techniques. These tech-
niques are helpful only for the qualitative understanding of machine’s condition. Furthermore, signal processing based tech-
niques detect faults at a late stage of machinery operations, however for better machinery condition monitoring, incipient fault
detection is necessary. Model-based fault diagnostic techniques utilize all the information obtained from the measured sig-
nal. In comparison to signal processing based techniques, model-based monitoring system is cost effective [2]. Model-based
approaches for rotor system fault diagnosis have been used for the identification of common machinery faults e.g. unbalance
[3,4], misalignment [4], and cracks [5,7] based on load estimation. Full system states are required for equivalent load estimation,
which can be calculated by an observer [5] or modal expansion [4]. It is observed that the modal expansion has a major influ-
ence on the identification results [3]. In model-based techniques, use of the reduced-order technique can limit the use of modal
expansion or full state estimation.
Signal processing based techniques identify faults by their corresponding characteristic frequencies in the measured vibra-
tion signal [6]. However, this technique proves ineffective when different faults show similar characteristics. For example, Pen-

* Corresponding author.
E-mail address: amohanty@mech.iitkgp.ernet.in (A.R. Mohanty).

https://doi.org/10.1016/j.jsv.2017.11.020
0022-460X/© 2017 Elsevier Ltd. All rights reserved.
 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199 185

nacchi and Vania [7] have analyzed real gas turbine vibration data and used a model-based approach to detect actual fault where
shaft bow caused by crack induced 1× vibration which also occurs in the case of unbalance or misalignment. It was concluded
that the model-based approach for fault diagnostic could confirm the occurrence of an actual fault.
A major disadvantage of conventional techniques of balancing is the requirement of a number of runs which is always not
convenient for large machines. Therefore, researchers from the field of machinery vibrations are looking for a balancing tech-
nique that requires less number of runs. The quantification of the amount of unbalance and misalignment present in rotating
machines has been an active area of research. Lees and Friswell [8] presented a technique for estimating bearing parameters and
state of unbalance using a numerical model of the rotor and measured vibration. They verified their technique with a numerical
example of a single plane unbalance for a two bearing system. It was found that a good numerical model for the rotor is required
for the estimation of unbalance. Later, this technique was verified on a small experimental rig [9]. Sinha et al. [10] proposed a
model-based technique for unbalance (amplitude and phase) and misalignment estimation of a rotating machine from a single
run-down/run-up. The method was verified on a small experimental rig, and it was observed that with modeling errors phase
estimation is robust as compared to unbalance amplitude. Tiwari and Chakravarthy [11,12] presented an identification tech-
nique using impulse response measurement and least-square method for the identification of residual unbalances and bearing
dynamic parameters. Frequency domain force and vibration response are used in the identification algorithm.
The occurrence of faults changes the system’s dynamic behavior which is detected via model-based fault diagnostic approach
[13]. Residual generation is one such technique where equivalent forces corresponding to different faults are calculated using
residual vibrations. Then, using a fitting technique such as least-square the phase and residual of a particular fault type can
be estimated either in time-domain [4] or in frequency-domain [14,15]. Usually, fault is identified when a change in vibration
vector exceeds a threshold. In model-based fault diagnostic approach, it is assumed that this change is caused by the fault which
is to be characterized.
Meanwhile, some deterministic-stochastic techniques have been developed in structural dynamics for unknown input force
estimation [16–24]. These techniques consider modeling and measurement errors and provide a better estimate of the inputs.
One such technique has been applied to an inverse heat conduction problem, by developing an on-line estimation technique
based on Kalman filter and a recursive least square estimator [16,18]. The methodology of input estimation based on Kalman
filtering was later implemented on input force estimation of structural systems. Ma et al. [19] have applied this technique for
impact load estimation of single and multi-degree-of-freedom system. Later, this technique was applied to a beam, modeled as
a single-degree-of-freedom system [20] and using finite element method [21]. The response at the tip of the cantilever beam
was measured experimentally to estimate input forces. The calculated normalized root mean square error between true and
estimated forces verifies the effectiveness of the proposed technique. This technique requires only displacement and/or velocity
as the measured response. Eigenvalue realization technique with Markov parameters is used by Liu et al. [22] to obtain system
matrices and then to model a cantilever plate in state-space form. In Ref. [23], this technique was applied for force estimation of
rotating machines where numerical simulations were presented with a rigid rotor model, transverse and radial displacements.
In Ref. [24], Kalman filter based input estimation technique was integrated with fuzzy logic inference system. To the best of
authors’ knowledge, this technique has not been employed in rotating systems for the purpose of fault diagnosis.
Optimal state estimation technique has been an active area of research in the field of fault identification under model-based
approach. Kalman filter and its variants for linear and nonlinear systems are applied to rotordynamics for various purposes e.g.
crack detection [25], parameter identification [26], fault identification [27], etc.
The major challenge of the previous model-based approaches is the estimation of full system response prior to force identi-
fication because of the limited response measurements. Furthermore, measurement noise and modeling error also affect force
identification. Kalman filter-based force estimation technique is capable of dealing with measurement noise and modeling error.
In this paper, Kalman filter and recursive least square based force identification technique is used for unbalance parameter esti-
mation in a rotor-disk-bearing system. Here, single plane unbalance is considered, and thus the location of unbalance force is
known. Finite element modeling is used for mathematical modeling of the system. The reduced-order model using modified
system equivalent reduction expansion process (SEREP) technique is obtained for selected master degrees-of-freedom (DOFs)
and number of modes. The measured displacement/rotational responses are chosen as master states. Effect of different measure-
ment sets (including displacement, velocity, and rotations) are shown and discussed. Results are presented for different shaft
speeds and measurement noise levels. Errors of estimated parameters are calculated for different values of filter parameters.

2. Mathematical modeling of rotor system

2.1. Full-order model

Finite element model of a rotor system is used in the present work. The shaft element consists of two nodes with 4 DOFs
(two translational and two rotational) each. The bearings are modeled as linear springs with same stiffness as in the two lateral
directions.
The linear equation of motion at a defined rotational speed is described by:

𝐌𝐱̈ (t ) + 𝐂d 𝐱(
̇ t) + 𝐊𝐱(t) = 𝐒F 𝐅(t) (1)

where M is the matrix containing mass and inertial, 𝐂d is the damping matrix and K is the stiffness matrices. F(t) is the force
vector, 𝐒F is a binary matrix with entries 1 correspond to DOFs where forces are acting and 0 for remaining DOFs, 𝐱(t ) is a vector
186 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199

of nodal displacement and the dot over the vector represents differentiation with respect to time. Gyroscopic effect is neglected
in the present work.
Element matrices for shaft and bearing can be found in Ref. [4]. The matrix 𝐂d includes internal and bearing damping only,
because gyroscopic effect is neglected in the present study. It is a fact that the identification of damping properties of mechanical
systems is more challenging than mass and stiffness characteristics. Experimental measurement is a common practice for more
accurate prediction of the extent and nature of structural damping. However, for study purposes, researchers use some approxi-
mate models for damping, e.g. proportional damping. Rayleigh damping form [28] is considered in the present case to determine
the viscous part of the damping matrix 𝐂d , which is expressed by the linear combination of mass and stiffness matrices,

𝐂dv = 𝛼R 𝐌 + 𝛽R 𝐊
where
[ ] [ ]{ }
𝛼R 𝜔 𝜔 𝜔2 −𝜔1 𝜁1
= 2 21 22
𝛽R 𝜔2
− 𝜔1 −1∕𝜔2 1∕𝜔1 𝜁2
herein 𝜔1 and 𝜔2 represents first and second critical speeds in rad/s, 𝜁1 and 𝜁2 are corresponding modal damping ratios, respec-
tively.
The shaft is discretized into finite Euler-Bernoulli beam elements with 8 DOFs (4 displacement and 4 rotations), a typical
beam element with nodal coordinates is shown in Fig. 1. Transverse displacements in x and y directions are presented by Vx and
Vy , respectively. Rotational DOFs about x and y axes are presented by 𝜃 x and 𝜃 y , respectively. The nodal DOFs are presented by
‘qi ,’ where i is 1–4 for first node and 5–8 for second node.

2.2. Reduced-order model

The accuracy of finite element methods depends on the fineness of discretization. However, computational effort increases
with fineness. Full-order model cannot be used because it requires full system response and due to cost-effectiveness in real
life, vibrations are measured only at limited locations on the rotating machinery. To circumvent this problem, model reduction
techniques have been proposed by many researchers. Friswell et al. [29] presented SEREP technique with right eigenvectors for
model reduction of the rotor-shaft system considering gyroscopic effect. Das and Dutt [30] proposed modified SEREP technique
using right and left eigenvectors for model reduction of rotor system and then applied the same for vibration control [31]. In the
present work, modified SEREP technique is employed for model reduction.
On-line input estimation techniques require state-space representation of the governing equation of system. Therefore, Eq.
(1) must be transformed into a state-space form.
The continuous time state-space form can be written as:
̇ t) = 𝐀c 𝐗(t) + 𝐁c 𝐅(t)
𝐗( (2)

𝐘(t ) = 𝐆c 𝐗(t ) + 𝐉c 𝐅(t ) (3)

[ ]
0 𝐈
where 𝐗(t ) = {𝐱1 𝐱2 · · · 𝐱n 𝐱̇ 1 𝐱̇ 2 · · · 𝐱̇ n } is the state vector, 𝐀c = is the system matrix,
−𝐌−1 𝐊 −𝐌−1 𝐂d
[ ]
0
𝐁c = is input influence matrix, 𝐒F is the force selection matrix.
𝐌 − 1 𝐒F
The output (or measurement) vector 𝐘(t ) in general, is the combinations of displacement, velocity and acceleration response.
̈ t) + 𝐒v 𝐗(
Output vector can be represented as 𝐘(t ) = 𝐒a 𝐗( ̇ t) + 𝐒d 𝐗(t) where 𝐒a , 𝐒v and 𝐒d are the selection matrices for accel-
eration, velocity and displacement response, respectively. The matrix 𝐆c = [𝐒d − 𝐒a 𝐌−1 𝐊 𝐒v − 𝐒a 𝐌−1 𝐂d ] is the state output

Fig. 1. Finite shaft element and coordinates.

 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199 187

matrix and 𝐉c is the direct transmission matrix.

Right eigenvector, 𝚽R is obtained by solving

̇ t) = 𝐀c 𝐗(t).
𝐗( (4)

And left eigenvector, 𝚽L is obtained by solving

̇ t) = 𝐀𝐓 𝐗(t)
𝐗( (5)
c

The state vector can be expressed in terms of left and right eigenvector as,

𝐗red = 𝚽R 𝜂 (6)

𝜂 = 𝚽𝐓L 𝐗 (7)

where 𝜂 is the modal vector. By selecting the dominant modes of vibration we can reduce the size of left and right eigenvector
matrices.

𝐗 = 𝚽R 𝜂 (8)

𝜂 = 𝚽𝐓L 𝐗 (9)

Now in the second stage of reduction process, the measurable active states are selected out of the total states. Transverse
displacements are chosen because it is easy to measure them as compared to rotational vibration. However, with the use of
non-contact rotational laser vibrometer, rotational vibration can also be measured. By dividing states into two groups namely
‘active’ and ‘non-active’, Eq. (6) can be decomposed as
{ } [ ]
𝐗active 𝚽active
red = R 𝜂 (10)
𝐗deleted
red
𝚽deleted
R

where number of active and non-active states are ‘a’ and ‘d’, respectively. Substituting Eq. (8) into Eq. (10) we get,

𝐗active
red
= 𝐓𝐗red (11)

where 𝐓 = 𝚽active
R
𝚽𝐓
L
is a transformation matrix. The number of active DOFs (a) and selected number of modes nm may not be
same, therefore, 𝚽active
R
cannot be inverted conventionally.
Inverting Eq. (11) we get,

𝐗red = 𝐓rev 𝐗active

red
(12)

The reverse transformation matrix 𝐓rev between full and reduced state can be expressed as,
{
𝚽R 𝚽𝐓 [𝚽𝐓 ]+ 𝚽active , for a ≥ nm
𝐓rev = L L R
𝚽R 𝚽TR [𝚽TL ]+ 𝚽active
L
, for a < nm

For detailed derivation please refer [30]. Finally the reduced equation of motion can be written as,

𝐗̇ red = 𝐀red 𝐗red + 𝐁red F(t ) (13)

𝐘(t ) = 𝐆red 𝐗red + 𝐉red F(t ) (14)

where,

𝐀red = 𝐓𝐀c 𝐓rev (15)

𝐁red = 𝐓𝐁c (16)

𝐆red = 𝐆c 𝐓rev (17)

188 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199

3. Kalman filter based input force estimation

In the present work, Kalman filter based input estimation technique is used for model-based fault estimation of a rotating
machinery. This technique has been successfully applied for various kinds of inverse problems such as inverse heat conduction
[16,18] and input force estimation of beam [20]. These techniques require measurement of responses at all the locations, which
is not feasible for a real world problem, where the size of the structures is very large. In the proposed method, the reduced
model based on the theory described in the previous section is employed for the unknown unbalance force estimation.
Under a stochastic environment, the discrete time linear dynamic system can be represented as,
𝐱(k + 1) = 𝐀𝐱(k) + 𝐁𝐅(k) + 𝐰(k) (18)

𝐲(k) = 𝐆𝐱(k) + 𝐉𝐅(k) + 𝐯(k) (19)

where k is the time index, 𝐱(k) = 𝐗(kΔt ), 𝐅(k) = 𝐅(kΔt ) and assuming the zero order hold for the inter-sample behavior of the
input, 𝐀 = e𝐀red Δt , 𝐁 = [𝐀 − 𝐈]𝐀red
−1 𝐁 , 𝐆 = 𝐆
red red and J = 0 because only displacement/velocity are used in the measurement
vector thus direct transmission matrix is zero. While discretizing the system matrices, we choose a sampling interval such that
it is able to capture the dynamics of the systems. In general, the sampling frequency should be higher than twice the largest
𝜋
eigenfrequency i.e. Δt ≤ .
max(𝜔i )
𝐰(k) and 𝐯(k) are the system and measurement noise vectors and contains uncorrelated white Gaussian random sequences with
variance Q and R, respectively.
The algorithm can be divided into two steps, first Kalman gain and innovation covariance are calculated by using the simple
Kalman filter without any input, and then these values are used in the recursive least-square algorithm. For the sake of brevity
only final expressions are given here, the derivations can be found in Refs. [16,32].
State prediction(without input), 𝐱[k∕k − 1] = 𝐀 𝐱[k − 1∕k − 1] (20)

State prediction covariance, 𝐏[k∕k − 1] = 𝐀 𝐏[k − 1∕k − 1] 𝐀𝐓 + 𝐐 (21)

Innovation covariance, 𝐒[k] = 𝐆𝐏[k∕k − 1]𝐆𝐓 + 𝐑 (22)

Kalman gain, 𝐊a [k] = 𝐏[k∕k − 1]𝐆𝐓 𝐒−1 [k] (23)

Updated state covariance, 𝐏[k∕k] = [𝐈 − 𝐊a [k]𝐆]𝐏[k∕k − 1] (24)

Innovation, 𝐙[k] = 𝐲[k] − 𝐆𝐱[k∕k − 1] (25)

Updated state estimation, 𝐱[k∕k] = 𝐱[k∕k − 1] + 𝐊a 𝐙[k] (26)

where 𝐱[k∕k − 1] is the predicted state, 𝐏[k∕k − 1] is state prediction covariance matrix, 𝐒[k] is the innovation covariance matrix,
and 𝐊a [k] is Kalman gain matrix. The filter initializes with an initial state and error covariance, which are unknown quantities,
in practice, arbitrary initial states are assumed with a large value of error covariance. However, as the time advances, the error
reduces with new measurements.
The bias free Kalman filter provides initial inputs (𝐒[k], 𝐊a [k] and 𝐙[k]) to the least square algorithm. The equations for
recursive least-square algorithm are
𝐁s [k] = 𝐆[𝐀𝐌s [k − 1] + 𝐈]𝐁 (27)

𝐌s [k] = [𝐈 − 𝐊a [k]𝐂][𝐀𝐌s [k − 1] + 𝐈] (28)

𝐊b [k] = 𝜆−1 𝐏b [k − 1]𝐁𝐓

s
[k] [𝐁s [k]𝜆−1 𝐏b [k − 1]𝐁𝐓
s
[k] + 𝐒[k]]−1 (29)

𝐏b [k] = [𝐈 − 𝐊b [k]𝐁s [k]]𝜆−1 𝐏b [k − 1] (30)

̂
𝐅[k] = ̂
𝐅[k − 1] + 𝐊b [k][𝐙[k] − 𝐁s [k]̂
𝐅[k − 1]] (31)

where 𝐁s [k] and 𝐌s [k] are the sensitivity matrices, ̂

𝐅[k] is the estimated input force vector, and 𝐏b [k] is the error covariance of
the estimated input vector. For time-varying case, 𝐊b [k] is prevented from shrinking to zero by forgetting factor 𝜆, and algorithm
preserves its updating ability.
A large value for initial error variance matrices P and Pb are assumed which provide some initial errors in estimated input.
However, it can be observed from the expressions of gains and sensitivity matrices that they can be calculated prior to the
analysis. Thus, the initial estimation errors can be avoided.
 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199 189

4. Scheme of the fault diagnostic technique

In this section, complete procedure for unbalance identification is described, the corresponding flow diagram is presented
in Fig. (2). First, a full order model of the rotor system is formed using mass, stiffness, and damping matrices obtained from
finite element model. The full order model is converted into state-space form, which is expressed in Eq. (2) and Eq. (3). The full
order model can be used directly for input estimation, but it require response measurements at all DOFs. Since, measurements
at all node positions are not possible, reduced-order model is needed. In this paper, modified SEREP technique is used, which is
suitable for rotating system. Kalman filter and recursive least square based input estimation requires state-space model of the
system and responses to identify input to the system. By this way, unbalance forces acting on the rotor disk can be identified
along with its angular location. Least-square technique is used in the present work for the estimation of unbalance parameters
using estimated unbalance force and unbalance force model.

5. Modeling of unbalance system

In model-based fault diagnostic technique, unbalance fault is represented by a mathematical model which describes relations
between unbalance force and unbalance parameters (magnitude and phase of unbalance). Schematic diagram of an unbalanced
disk is shown in Fig. (3).
By considering only the transverse vibrations, the unbalance forces at disk location in lateral directions can be expressed by
following equation,

𝐅xu = me𝜔2 cos(𝜔t + 𝜙) (32)

Fig. 2. Scheme of the fault diagnostic technique.

Fig. 3. Disk unbalance.

190 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199

𝐅yu = me𝜔2 sin(𝜔t + 𝜙) (33)

Here, the static equilibrium is considered as a reference for the axis system, therefore, force due to gravity acting in vertical
direction is excluded. Shaft bow is also neglected in the present work.

6. Least square fitting

A least squares algorithm is used to achieve the best fit between the estimated forces, ̂
𝐅 and the model forces 𝐅(β, t). The
objective function is expressed as,
∑
min‖𝐅(𝛃, t) − ̂
𝐅(t)‖22 = min (𝐅(𝛃, t(i)) − ̂
𝐅(i))2 (34)
𝛃 𝛃
i

For the case of unbalance the fault model force is expressed by,

𝐅x (𝛃, t) = βx1 𝜔2 cos(𝜔t + βx2 ) (35)

𝐅y (𝛃, t) = βy1 𝜔2 sin(𝜔t + βy2 ) (36)

where 𝐅x and 𝐅y are the unbalance forces in X and y directions, respectively.

7. Numerical example

Finite element model of a rotor-bearing system (Fig. 4) is shown in Fig. 5. This model is selected for numerical simulation
and will be used for experimental verification also. The rotor consists of a shaft of 15.875 mm diameter, supported on three
bearings and carries a disk. The detail of the geometrical parameters is shown in Fig. 4. Density, Young’s modulus and Poisson’s
ratio of shaft material are 7800 kg/m3 , 0.29 and 200 GPa, respectively. The stiffness and damping coefficients of the bearings
are 2.57 × 107 N/m and 250 Ns/m [37]. The helical coupling between the electric motor and the shaft is modeled as a hollow
shaft, its mass is 124.6 g. Shaft is divided into 14 Euler-Bernoulli beam elements with 4 DOFs at each node (two translation
and two rotations). The use of Timoshenko beam elements is recommended for finite element modeling of the beam/shaft.
However, in many instances, Euler-Bernoulli elements provide satisfactory results e.g. for the shafts with low slenderness ratio
(diameter/length). The slenderness ratio of the shaft is 0.029, therefore, Euler-Bernoulli beam elements are selected for finite
element modeling. The shaft is divided into finite elements such that bearings and disk locations coincides with nodes. The
flexible coupling is considered as a hollow shaft is modeled by considering it as a combination of two shaft element connected
by a frictionless joint (for more details we refer the reader to [4,33]).

Fig. 4. Schematic diagram of rotor system.

Fig. 5. Finite element model of the rotor system.

 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199 191

At 0 rev/min, the first two natural frequencies of the rotor system shown in Fig. 5 are (−6.86-685.236i) and
(−6.86 + 685.236i), the real parts are due to damping. The natural frequencies occur in pair because inertia and stiffness prop-
erties in the x and y directions are identical [34].
Here, the comparison of full and reduced order model are presented, first two modes are included in transformation. In Fig. 6,
magnitude of complex frequency response function calculated at node 9 for full and reduced order model is shown.
For the rotor system shown in Fig. 5, a total of 120 multi-type responses are possible, including 30 displacements and velocity
responses and 30 rotational displacements and velocities. But practically it is not possible to measure all of the responses.
Bearing pedestals are the most common locations for the installation of the sensors e.g. accelerometers. However, with laser
vibrometer, velocities at any location of the rotor can be measured. Optimal placement of sensors has been an active area of
research in the field of fault diagnostic [35]. However, study of optimal sensor placement is not the scope of the present work.
Here, we have selected five different measurement sets as shown in Fig. 7, named M1-M5. As it is shown, displacement
sensors are represented with square and rotational with circle. At selected locations, displacements are measured at both the
transverse directions (x and y) and rotations are measured at xz and yz planes. Because of the symmetry in horizontal and vertical
planes, sensors are shown only in one plane (yz plane). These five sets are selected to analyze the effects of number, type, and
distance between sensors.

Fig. 6. Comparison of magnitude of frequency response at node 9 for full and reduced order model.

Fig. 7. Five sensor placement configuration (a) M1 (b) M2 (c) M3 (d) M4 (e) M5.
192 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199

Fig. 8. Exact and estimated unbalance force for shaft speed 30 Hz and for measurement sets (a) M1, M2, M3 (b) M4, M5.

8. Results and discussions

Steady-state responses are generated from full order model of the rotor system. Displacement/velocity/rotational responses
are obtained using Eq. (18) and Eq. (19). The responses are measured with the sampling frequency of 10000 Hz. To simulate
real measurements, we have added random noise which can be expressed as, noise = 𝛾𝜎 rk , where 𝛾 is the noise level, 𝜎 is the
standard deviation of the exact response, and rk is the normal random sequences. The measurement noise, which is Gaussian
in nature, is expressed by some percentage of the standard deviation of the exact signal [36]. This choice is made to maintain
a constant signal-to-noise ratio of the responses for all cases. The measurement noise covariance matrix needs to be set prior
to the filtering process, which can be chosen based on some percentage of the peak measurement response. As we know, the
magnitude of response increases with the rotor speed for a particular unbalance, therefore, the noise level is maintained for
every rotor speed. For the case when the noise level is same for every speed, the noise will be less for higher rotor speeds as
compared to the lower speeds, which may lead to the wrong conclusion. The initial state is set to zero. Initial error covariance
P(0|0) and initial error covariance of estimated force Pb (0) are set to 1010 , since these values are usually unknown [19]. The
measurement error covariance R can be calculated as (𝛾𝜎 )2 . Reduced-order model obtained from modified SEREP technique is
used for force estimation.
Results are presented for shaft speeds from 10 Hz to 50 Hz with 5 Hz spacing. The unbalance mass and its distance from
center is kept constant which are 5 g and 0.07 m, respectively, thus the amount of unbalance is 3.5 × 10−4 kg-m. Simulations
are performed for different noise levels which are 1%, 5% and 10%.
For comparison between all five measurement sets, exact and estimated forces are shown in Fig. 8. It can be easily seen that,
as compared to other sets the force estimation is less accurate when measurement set 4 is used. In fourth measurement set,
difference between response levels at both the nodes is very large, this could be the possible reason for poor performance of
estimation.
Errors in the estimated amount of unbalance is defined by,
Exact amount of unbalance − Estimated amount of unbalance
Error(%) = × 100 (37)
Exact amount of unbalance
Error in estimated phase is expressed by |Exact phase − Estimated phase|. These errors are calculated for different shaft
speeds and measurement noise levels as shown in Fig. 9 for different measurement sets. The modeling errors/process noise
are not considered and hence a small value (10−20 × [I]n×n , n is the number of active states) is assigned to the process noise
covariance Q. It can be observed that the estimation is good when displacement response at bearing 2 and bearing 3 are included
in the measurement vector (measurement set 1). The performance of estimator is increased when rotational responses are
included in the measurement vector (measurement set 2). Increase in estimation accuracy is due to the increase in measured
DOFs. However, the location of sensor is important which can be verified by comparing results for M1, M3, M4 and M5.
Errors in both amplitude and phase are almost same for all three levels of measurement noise i.e. 1%, 5% and 10%. This is due
to the fitting of estimated forces using least-square technique. In Table 1, errors of estimated amount of unbalance and phase
angle are presented for different measurement sets and measurement noise levels.
The unbalance force is estimated prior to unbalance parameter estimation, and the estimated unbalance force (Eq. (31)) is
directly related to the biased innovations 𝐙[k]. As we know that the unbalance force and response 𝐲[k] is proportional to the
square of the shaft speed, which causes increase in the values of innovations with shaft speed. The filter parameters such as
process noise covariance Q and forgetting factor 𝜆 are set to a constant value, therefore, the error is changing with the speed.
 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199 193

Fig. 9. Errors in estimated amount of unbalance (a (phase, 0), b (phase, 20), c (phase, 40)) and phase (d (phase, 0), e (phase, 20), f (phase, 40)) for different shaft speeds and
measurement noise level.
194 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199

Table 1
Estimated unbalance parameters for different measurement sets and measurement noise levels. (Rotor speed 40 Hz).

Measurement set (%) noise level Actual unbalance parameters Estimated unbalance parameters (%) error in (deg) Absolute
amount of unbalance difference in phase
Amount of Phase Amount of Phase
unbalance (kg-m) (deg) unbalance (kg-m) (deg)

1 1 3.5 × 10−4 40 3.209 × 10−4 38.42 8.31 1.58

5 3.5 × 10−4 40 3.209 × 10−4 38.45 8.30 1.54
10 3.5 × 10−4 40 3.207 × 10−4 38.35 8.35 1.65

2 1 3.5 × 10−4 40 3.324 × 10−4 39.27 5.02 0.73

5 3.5 × 10−4 40 3.324 × 10−4 39.25 5.02 0.75
10 3.5 × 10−4 40 3.327 × 10−4 39.31 4.92 0.69

3 1 3.5 × 10−4 40 3.133 × 10−4 37.83 10.48 2.17

5 3.5 × 10−4 40 3.133 × 10−4 37.83 10.47 2.17
10 3.5 × 10−4 40 3.130 × 10−4 37.89 10.55 2.11

4 1 3.5 × 10−4 40 2.958 × 10−4 35.66 15.48 4.34

5 3.5 × 10−4 40 2.958 × 10−4 35.63 15.48 4.37
10 3.5 × 10−4 40 2.961 × 10−4 35.60 15.39 4.4

5 1 3.5 × 10−4 40 3.026 × 10−4 36.45 13.53 3.55

5 3.5 × 10−4 40 3.027 × 10−4 36.42 13.49 3.58
10 3.5 × 10−4 40 3.026 × 10−4 36.43 13.52 3.57

However, the results are provided for the different values of filter parameters in the following subsections to show its effect on
filter performance.
Variation of error estimates for measurement set 2 is different as compared with other sets. This variation is due to underes-
timation and overestimation of unbalance forces before and after a certain speed (≈30 Hz in the present case). In Fig. 10, actual
and estimated unbalance forces are shown for 10 Hz and 40 Hz shaft speed. For low shaft speed, the amplitude of transverse
rotation is less as compared with high shaft speed, this can be a possible reason for such variation. The overall performance of
unbalance parameter estimation is good when measurement set 2 is used.

8.1. Effect of type and location of sensors

In this section, effect of rotational response is investigated while keeping the same number of measured DOFs (6 in the
present case) unlike the previous case. As shown in Fig. 11, four different measurement sets are used, set M6 contains only trans-
lational responses and other sets contain one rotational and two translational responses. Following parameters are used in the
simulation: shaft speed = 40 Hz, measurement noise level = 10%, eccentricity = 3.5 × 10−4 kg-m, and phase angle = 40
deg. Estimated unbalance fault parameters are tabulated in Table 2.
Using measurement set M6, error in estimated amount of unbalance decreased to 5.637% from 8.35% in M1, but error in
the phase estimate increased which was previously observed for the case of M4. In measurement sets M7, M8 and M9 rota-
tional responses are measured at bearing 1, 2 and 3, respectively. The estimated unbalance parameters are more accurate when

Fig. 10. Actual and estimated unbalance force obtained using measurement set 2 for shaft speed (a) 10 Hz and (b) 40 Hz.
 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199 195

Fig. 11. Different sensor placement configuration (a) M6 (b) M7 (c) M8 (d) M9.

Table 2
Estimated unbalance parameters for different measurement sets. (Rotor speed 40 Hz).

Measurement set Actual unbalance parameters Estimated unbalance parameters (%) error in (deg) Absolute
amount of unbalance difference in phase
Amount of unbalance Phase Amount of unbalance Phase
(kg-m) (deg) (kg-m) (deg)

6 3.5 × 10−4 40 3.302 × 10−4 35.998 5.637 4.001

7 3.5 × 10−4 40 3.239 × 10−4 35.600 7.438 4.399
8 3.5 × 10−4 40 3.339 × 10−4 36.207 4.583 3.792
9 3.5 × 10−4 40 3.354 × 10−4 36.753 4.171 3.246

rotational responses are measured at node 6 and 12.

8.2. Effect of process noise covariance

The process noise covariance is one of the important parameters in Kalman filter and recursive least square based input esti-
mation. The effect of process noise covariance on filter performance and consequently on the estimation of unbalance parame-
ters are presented in this section. Here, for a fixed measurement noise level (10%), errors of estimated unbalance parameters are
calculated for different shaft speeds and assumed process noise covariance. Results are presented for five different measurement
sets.

Fig. 12. Errors in estimated amount of unbalance (a) and phase (b) for different shaft speeds and process noise covariance. (Measurement noise level 10%).
196 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199

Fig. 13. Details of actual and estimated unbalance force with respect to different values of process noise covariance for measurement set 4. (LS-least square fitting).

It is easily observed from the error plots Fig. 12 that the unbalance parameter estimation is not affected with the location
of the unbalance mass, i.e. insensitive to position of hot-spots. During response generation, the process noise is not included.
Then, the unbalance parameters are estimated for three different values of process noise covariance i.e. 10−20 , 10−19 , and 10−18
Error in estimated amount of unbalance is almost same for all the values process noise variances, with maximum error around
20% for fifth measurement set. However, phase error increases with the magnitude of process noise variance, maximum error is
around 10◦ for fourth measurement set and 3◦ for second measurement set.
The variation on estimated force with respect to different levels of Q is shown in Fig. 13 for fourth measurement set. Increase
in time-delay can be observed with increase in process noise covariance.

8.3. Effect of forgetting factor (𝜆)

The ‘forgetting’ or ‘fading’ factor 𝜆, which is a weighting factor in the second step of the filter and whose value lie between 0
and 1, is used in the input estimation algorithm to compromise between fast adaptive capability and loss of estimate accuracy
[21]. In most of the literature discussed in the introduction section, a constant value of 𝜆 is used for input estimation. However,
some improvement in the selection of 𝜆 is also reported by the researchers, e.g., in Ref. [17], an adaptive 𝜆 is proposed based on
biased innovation, and in Ref. [24], a fuzzy weighting function is used instead of 𝜆. For different values of 𝜆 errors of estimated
amount of unbalance and phase are calculated which is presented in Fig. 14. The variation on estimated force with respect to
different levels of 𝜆 is shown in Fig. 15 for first measurement set. The shaft speed is 50 Hz and measurement noise level is 10%. It
is evident from Fig. 15 that with increase in the value of 𝜆, algorithm is less sensitive to disturbances and a time-delay between
exact and estimated force is increased.

Fig. 14. Errors in estimated amount of unbalance (a) and phase (b) for different shaft speeds and forgetting factor (𝜆). (Measurement noise level 10%).
 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199 197

Fig. 15. Details of actual and estimated unbalance force with respect to different values of process noise covariance for measurement set 4.

Fig. 16. Actual and estimated unbalance force for different levels of modeling error using measurement set 2.

8.4. Effect of modeling error

Different modeling errors are considered here to test the sensitivity. A mathematical model of a real system cannot be perfect.
Incorrect values of physical parameters are a most probable cause of this discrepancy. Here, to incorporate modeling error, the
stiffness matrix K is changed to K + ΔK, where ΔK is some percentage of K. The analysis is performed for 5% and 10% change in
stiffness matrix K.
In estimation of unbalance forces, model with erroneous stiffness matrix is used. In Fig. 16, estimated and actual unbalance
force is represented for different levels of modeling error in stiffness matrix. Under-estimation and over-estimation of unbalance
forces are obvious for lower and higher values of stiffness matrix, respectively.

9. Conclusions

In this paper, we present a novel technique to estimate rotor unbalance parameters (amplitude and phase angle) of a rotor-
disk-bearing system. The method is based on unbalance force estimation using Kalman filter and recursive least square based
input estimation technique. The proposed technique uses reduced-order model of the system and limited response measure-
ments (including displacements and rotations). A numerical simulation study has been carried out for a rotor-disk-coupling-
bearing system with different measurement sets. Based on the results presented in this paper, the following conclusions can be
drawn.

• Unlike model-based approaches available in the literature, estimation of full system state (e.g using modal expansion) prior
to force estimation is not required.
• Since a reduced-order model is used for force estimation, the analysis is fast and suitable for online estimation of unbalance
fault parameters.
198 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199

• Different types of responses are generated from full-order model and contaminated with white Gaussian noise of different
levels (1%, 5% and 10%) to simulate measurements from a real system.
• Effects of multi-type sensors are presented and it was observed that the accuracy of estimation is increased when rotational
responses are included in measurement vector along with transverse displacements.
• It has been observed that the estimation is robust to the measurement noise.
• The effect of different values of Q has been shown. A slight variation in estimated parameters is found for different value of
Q.
• Effect of 𝜆 which accounts for fast adaptive capability and estimation accuracy is shown. It has been observed that with
the increase in 𝜆 time-delay between actual and estimated unbalance force increases, it affects the estimation of unbalance
phase.
• Because finite element modeling is used which uses physical parameters, effect of modeling error is also presented. Unbal-
ance forces are estimated with errors in stiffness matrix, it mainly affects the amplitude of unbalance.

The present work is a part of an ongoing research on model-based techniques for unbalance identification and is restricted
to the estimation of single-plane unbalance only. However, authors believe that the present approach can be easily extended
for multi-plane unbalance identification. A general approach for static and dynamic balancing with experimental verification is
remains as a future work.

References

[1] A.R. Mohanty, Machinery Condition Monitoring: Principles and Practices, CRC Press, 2014.
[2] T. Ishihara, J. Yoon, Prediction of Dynamic Response of Wind Turbines Using SCADA Data and Updated Aeroelastic Model, European Wind Energy Associa-
tion, Copenhagen, Denmark, 16–19 April 2012, EWEA, 2012. Paper no. PO.25.
[3] R. Markert, R. Platz, M. Seidler, Model based fault identification in rotor systems by least squares fitting, Int. J. Rotating Mach. 7 (2001) 311–321.
[4] A.K. Jalan, A.R. Mohanty, Model based fault diagnosis of a rotor-bearing system for misalignment and unbalance under steady-state condition, J. Sound Vib.
327 (2009) 604–622.
[5] H. Bach, R. Markert, Determination of the fault position in rotors for the example of a transverse crack, Struct. Health Monit. (1998) 325–335.
[6] V.K. Rai, A.R. Mohanty, Bearing fault diagnosis using FFT of intrinsic mode functions in Hilbert-Huang transform, Mech. Syst. Signal Process. 21 (2007)
2607–2615.
[7] P. Pennacchi, A. Vania, Diagnostics of a crack in a load coupling of a gas turbine using the machine model and the analysis of the shaft vibrations, Mech.
Syst. Signal Process. 22 (2008) 1157–1178.
[8] A.W. Lees, M.I. Friswell, The evaluation of rotor imbalance in flexibly mounted machines, J. Sound Vib. 208 (1997) 671–683.
[9] S. Edwards, A.W. Lees, M.I. Friswell, Experimental identification of excitation and support parameters of a flexible rotor-bearings-foundation system from
a single run-down, J. Sound Vib. 232 (2000) 963–992.
[10] J.K. Sinha, A.W. Lees, M.I. Friswell, Estimating unbalance and misalignment of a flexible rotating machine from a single run-down, J. Sound Vib. 272 (2004)
967–989.
[11] R. Tiwari, V. Chakravarthy, Simultaneous identification of residual unbalances and bearing dynamic parameters from impulse responses of rotor-bearing
systems, Mech. Syst. Signal Process. 20 (2006) 1590–1614.
[12] R. Tiwari, V. Chakravarthy, Simultaneous estimation of the residual unbalance and bearing dynamic parameters from the experimental data in a rotor-bear-
ing system, Mech. Mach. Theor. 44 (2009) 792–812.
[13] S. Edwards, A.W. Lees, M.I. Friswell, Fault diagnosis of rotating machinery, Shock Vib. Digest 30 (1998) 4–13.
[14] N. Bachschmid, P. Pennacchi, A. Vania, Identification of multiple faults in rotor systems, J. Sound Vib. 254 (2002) 327–366.
[15] P. Pennacchi, N. Bachschmid, A. Vania, G.A. Zanetta, L. Gregori, Use of modal representation for the supporting structure in model-based fault identification
of large rotating machinery: part 1-theoretical remark, Mech. Syst. Signal Process. 20 (2006) 662–681.
[16] P.C. Tuan, C.C. Ji, L.W. Fong, W.T. Huang, An input estimation approach to on-line two-dimensional inverse heat conduction problems, Numer. Heat Tr.
B-FUND 29 (1996) 345–363.
[17] P.C. Tuan, W.T. Hou, Adaptive robust weighting input estimation method for the 1-D inverse heat conduction problem, Numer. Heat Tr. B-FUND 34 (1998)
439–456.
[18] T.C. Chen, P.C. Tuan, Input estimation method including finite-element scheme for solving inverse heat conduction problems, Numer. Heat Tr. B-FUND 47
(2005) 277–290.
[19] C.K. Ma, P.C. Tuan, D.C. Lin, C.S. Liu, A study of an inverse method for the estimation of impulsive loads, Int. J. Syst. Sci. 29 (1998) 663–672.
[20] C.K. Ma, D.C. Lin, Input forces estimation of a cantilever beam, Inverse Probl. Eng. 8 (2000) 511–528.
[21] C.K. Ma, J.M. Chang, D.C. Lin, Input forces estimation of beam structures by an inverse method, J. Sound Vib. 259 (2003) 387–407.
[22] J.J. Liu, C.K. Ma, I.C. Kung, D.C. Lin, Input force estimation of a cantilever plate by using a system identification technique, Comput. Meth. Appl. Mech. Eng.
190 (2000) 1309–1322.
[23] C.K. Ma, D.C. Lin, J.M. Chang, Estimation of forces generated by a machine mounted upon isolators under operating conditions, J. Franklin Inst. 336 (1999)
875–892.
[24] M.H. Lee, T.C. Chen, Intelligent fuzzy weighted input estimation method for the forces generated by an operating rotating machine, Measurement 44 (2011)
917–926.
[25] S. Seibold, K. Weinert, A time domain method for the localization of cracks in rotors, J. Sound Vib. 195 (1996) 57–73.
[26] B.A. Miller, A.S. Howard, Identifying bearing rotor-dynamic coefficients using an extended Kalman filter, Tribol. Trans. 52 (2009) 671–679.
[27] S. Khanam, J.K. Dutt, N. Tandon, Extracting rolling element bearing faults from noisy vibration signal using Kalman filter, J. Vib. Acoust. 136 (2014)
031008.
[28] H. Ma, Q. Zhao, X. Zhao, Q. Han, B. Wen, Dynamic characteristics analysis of a rotor-stator system under different rubbing forms, Appl. Math. Model. 39
(2015) 2392–2408.
[29] M. Friswell, J.E.T. Penny, S.D. Garvey, Model reduction for structures with damping and gyroscopic effects, in: Proceedings of ISMA-25 Leuven, Belgium,
September, 2000, pp. 1151–1158.
[30] A.S. Das, J.K. Dutt, Reduced model of a rotor-shaft system using modified SEREP, Mech. Res. Commun. 35 (2008) 398–407.
[31] A.S. Das, J.K. Dutt, A reduced rotor model using modified SEREP approach for vibration control of rotors, Mech. Syst. Signal Process. 26 (2012) 167–180.
[32] C. Loh, A. Wu, J.N. Yang, C. Chen, T. Ueng, Input force identification using Kalman filter techniques: application to soil-pile interaction, in: Proceedings of
the SPIE, Society of Photo-optical Instrumentation Engineers, 693223, 2008, pp. 1–9.
[33] E. Kramer, Dynamics of Rotors and Foundations, Springer Science \& Business Media, 2013.
[34] M.I. Friswell, J.E.T. Penny, S.D. Garvey, A.W. Lees, Dynamics of Rotating Machines, Cambridge University Press, 2010.
 A. Shrivastava and A.R. Mohanty / Journal of Sound and Vibration 418 (2018) 184–199 199

[35] S. Fatima, S.G. Dastidar, A.R. Mohanty, V.N.A. Naikan, Technique for optimal placement of transducers for fault detection in rotating machines, Proc. Inst.
Mech. Eng. O J. Risk Reliab. 227 (2013) 119–131.
[36] E. Lourens, C. Papadimitriou, S. Gillijns, E. Reynders, G. De Roeck, G. Lombaert, Joint input-response estimation for structural systems based on
reduced-order models and vibration data from a limited number of sensors, Mech. Syst. Signal Process. 29 (2012) 310–327.
[37] A.K. Jalan, Model Based Fault Diagnosis of a Rotor System for Unbalance, Misalignment and Crack, Ph.D. thesis, IIT Kharagpur, 2010.