Mechanical Systems and Signal Processing 170 (2022) 108827

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

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

Nonlinear responses of a dual-rotor system with rub-impact fault

subject to interval uncertain parameters
Chao Fu a, Weidong Zhu b, Zhaoli Zheng c, d, Chuanzong Sun e, Yongfeng Yang a,
Kuan Lu a, *
a
Institute of Vibration Engineering, Northwestern Polytechnical University, Xi’an 710072, China
b
Department of Mechanical Engineering, University of Maryland, Baltimore County, 1000 Hilltop Circle, Baltimore, MD 21250, USA
c
Key Laboratory of Thermo-Fluid Science and Engineering, Xi’an Jiaotong University, Xi’an 710049, China
d
Science and Technology on Thermal Energy and Power Laboratory, Wuhan 430205, China
e
Wind Energy Institute, Shengyang University of Technology, Shengyang 110023, China

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

Keywords: This paper aims to study the nonlinear steady-state response of a dual-rotor system with rub-
Dual-rotor system impact fault subject to unknown-but-bounded (UBB) uncertainties. Mathematical modelling of
Rub-impact fault the non-linear dynamical system is carried out based on the Lagrangian formulation. The
Nonlinear dynamic response
nonlinear dynamic response of the rubbing dual-rotor system without uncertainty is solved by
Interval uncertain parameter
Non-intrusive surrogate
using the multi-dimensional harmonic balance method coupled with the alternating frequency/
Polar angle interpolation time technique. The arc-length continuation is used to track the solution branches. To predict the
response range subject to uncertainty, a non-intrusive surrogate model in conjunction with the
polar angle interpolation (PAI) with efficiency enhancement is developed to track the propaga­
tions of parametric variabilities. The PAI is dedicated to dealing with collocations where the
responses have multiple solutions. Effects of UBB variables in the physical model and fault-related
parameters are investigated comprehensively. Different features of the variabilities in the steady-
state responses are found under the typical uncertain degrees. The interval scanning method is
used to validate the computation accuracy of the whole procedure. Moreover, the working
mechanism of the PAI method is demonstrated via examples in detail. The results obtained in
simulations can provide useful guidance for the nonlinear dynamic investigations and rub-impact
fault diagnosis of dual-rotor systems under the UBB uncertainties. The proposed non-intrusive
uncertainty quantification framework based on the PAI will also be beneficial to other
nonlinear vibration problems where multiple solutions are involved.

1. Introduction

The dual-rotor configurations have been popular in many industrial applications such as gas turbines and aero-engines due to many
merits over the traditional single rotating systems. An important feature that should be noted is the multi-frequency unbalance
excitation on the system, which is one of the main differences from single rotor systems and it causes the vibration behaviours much
more complex. It will become even more complicated when classical faults are present, such as crack, bearing defect and contact

* Corresponding author.
E-mail addresses: [email protected] (C. Fu), [email protected] (W. Zhu), [email protected] (K. Lu).

https://doi.org/10.1016/j.ymssp.2022.108827
Received 23 September 2021; Received in revised form 28 December 2021; Accepted 7 January 2022
Available online 20 January 2022
0888-3270/© 2022 Elsevier Ltd. All rights reserved.
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

Nomenclature

a0 , ak , bk (k = 1, 2, …) Fourier coefficient vectors for steady state responses

Ascale amplification factor
c0 , ck , dk (k = 1, 2, …) Fourier coefficient vectors for nonlinear forces
C damping matrix
Ci (i = 1, 2, 3) damping of bearings
ds arc length along the response curve
e1 , e2 eccentricities of lower and higher pressure rotors
E, V, D kinetic, potential and dissipation energy of the dual-rotor
fx , fy reaction forces in inter-shaft bearing
f ul (t), f uh (t) unbalance forces of lower and higher pressure rotors
Fn , Fτ normal and tangential forces induced by rub-impact
Fx , Fy projected forces induced by rub-impact
G gyroscopic matrix
H contributing matrix of linear part
̃
H unknown Fourier coefficient vector of rub-impact forces
He projected vector of external forces
Jpi , Jdi (i = l, h) polar and transverse moments of inertia of lower and higher pressure rotors
ri (i = l, h) vibration amplitudes of lower and higher pressure rotors
J Jacobian matrix
Kc contact stiffness between higher-pressure rotor and casing
Ki (i = 1, 2, 3) stiffness of bearings
Khl stiffness of inter-shaft bearing
K stiffness matrix
Li (i = 1, …, 5) geometries of the dual-rotor system
mi (i = l, h) mass of lower and higher pressure rotors
M mass matrix
Nh order of frequency components in Fourier expansion
Q(t) excitation vector
Tj1 ,j2 ,⋯,jn multi-dimensional Chebyshev polynomial basis
U(t) displacement vector
xi , yi (i = l, h) lateral displacements of lower and higher pressure rotors
xbi , ybi (i = 1, 2, 3) displacements of bearings

Greek Symbols
θxi , θyi (i = l, h) rotational displacements of lower and higher pressure rotors
μ friction coefficient between the higher-pressure rotor and casing
δ clearance between the higher-pressure rotor and casing
ω1 , ω2 rotating speeds of lower and higher pressure rotors
υ control parameter in the arc-length continuation
ε predefined iteration tolerance
φj1 ,j2 ,⋯,jn Chebyshev coefficient
α polar angle sample vector
ρ polar sample vector
γ standard interval variable vector
λ rotating speed ratio between lower and higher pressure rotors
γ
̂ sample vector of uncertainties
̃
φ
̃ reorganized surrogate coefficient vector
̂
Υ 2̂
n ×̂
n
matrix of simplex term values at parameter samples
ρ (i+1)
̂ polar amplitude sample vector at parameter samples in (i + 1)th iterative step
2̂n ×1
r reference solutions from the interval scanning method
∼
η error ratio
Z2 two-dimensional integers excluding both being zero

Abbreviations
AFT alternating frequency/time
DFT discrete Fourier transform

2
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

HP higher pressure
HBM harmonic balance method
IDFT inverse discrete Fourier transform
ISM interval scanning method
LP lower pressure
MHBM multi-dimensional harmonic balance method
NAEs nonlinear algebraic equations
NISM non-intrusive surrogate model
PAI polar angle interpolation
UBB unknown-but-bounded
UQ uncertainty quantification

between different components [1–4]. The non-symmetric coaxial rotors in conditions of co– or counter-rotating were considered and
relevant dynamics were presented in [5]. An extended transfer matrix method was developed by Gupta et al. [6] to study the unbalance
response of a dual-rotor system and the natural characteristics were determined experimentally, including the critical speeds and mode
shapes. Wang et al. [7] carried out modelling of a dual-rotor blade casing finite element model based on the actual structure of an aero-
engine and predicted the vibration responses of the casing when blade tip and casing is rubbing. The experimental validation of dual-
rotor system models supported on ball bearing were carried out and the beat vibration response was observed when the two rotors
were at approximate speed [8]. These contributions provided the feasible modelling methods of dual-rotor systems and showed various
dynamical characteristics, which can be of significance to the industry.
As a common fault in rotor systems, the rub-impact between rotor and casing will bring into the system significant effects on the
dynamics and even cause critical failures [9–11]. Experimental observations on the nonlinear dynamics were performed and the
multiple and fractional harmonic components in the responses were found [12]. The modulation characteristics induced by high-
frequency drive speed on a rotor under continuous contact were studied recently [13]. Patel and Darpe [14] investigated the joint
effects of rub-impact and crack faults. Ma et al. [15,16] carried out systematical and comprehensive research on the influences of
different rubbing forms and proposed the model for rub event between blade and casing. The arc-length continuation coupled with the
harmonic balance method (HBM) was introduced into the rotor/stator contact problem in [17]. Xiang et al. [18] established a new
model for an asymmetric rotor-bearing system and the interactions between the nonlinear oil-film force and the rub-impact force were
considered. The mechanisms of sudden unbalance and the rub-impact event caused by blade off in a rotor system were studied
theoretically and experimentally by Wang et al. [19]. From the literature, it is evident that rub-impact between rotor and casing can
cause the systems to behave nonlinearly and very complicated dynamical characteristics are also induced. Sun et al. [20] investigated
the steady-state dynamic responses of a dual-rotor subject to rubbing and the effects of different physical parameters were intensively
discussed. Zhou and Chen [21] constructed a model of a dual rotor-ball bearing-stator coupling system simplified from an industrial
aero-engine. From the above-mentioned works, insights into the rub-impact fault and related nonlinear dynamics of rotor systems can
be obtained, which benefits the engineers for design and maintenance.
As stated previously, the degree of complexity of a dual-rotor system is much higher than that of single rotors in terms of structural
configurations and dynamical characteristics. The inherent uncertainties will be ubiquitous in such complicated mechanical systems
due to a variety of reasons, such as material wear, imprecision or error in manufacturing and property dispersions [22–28]. The forms
of parametric uncertainties and sources can be even more diverse in practical engineering rotating systems, as explained in [29–31].
Currently, stochastic modelling for rotor systems with uncertainties has been widely adopted and relevant dynamics are intensively
studied [32–34]. The polynomial chaos expansion (PCE), which is very popular in stochastic vibrations, has been employed to
investigate the linear and non-linear random vibrations of rotors with or without faults [35–37]. In these studies, the effects of various
random uncertainties in the stiffness of supports, damping and eccentricity on the unbalance dynamic responses and natural char­
acteristics were discussed. An important and unavoidable premise for the application of stochastic methods in the field of uncertainty
analysis is that the probability distribution models of variables are already justified. Otherwise, subjective hypotheses should be made.
To release the hash requirements of the stochastic approaches, researchers have proposed the interval methods, which only utilize the
upper and lower limits of uncertainties and it will be more adaptable for the engineering problems with limited prior information
[38–40]. Wu et al. [41,42] proposed the Chebyshev inclusion function and its improved surrogate method for uncertain structures and
wide applications to different research areas can be found. Recently, Fu et al. [43] employed the interval method to study the dynamics
of a linear dual rotor system under different uncertainties. Zhang et al. [44] used the gPCE to obtain the stochastic nonlinear responses
of a rub-impact rotor system based on probabilistic models. Yang et al. [45] employed a likelihood method to estimate the reliability of
a rotor/stator contact system. Zhang et al. [46] used the Edgeworth series technique to study the effects of random uncertainties on
reliability. To be noted, Didier et al. [47] proposed a combination of the multi-dimensional HBM (MHBM) and the PCE for the sto­
chastic response computation of a rotor excited by the support. Further, they improved the theory by considering the frequencies of a
mechanical system with irregular nonlinearities as stochastic, which allows the prediction of the stochastic nonlinear frequency re­
sponses in the presence of uncertainties [48]. The major difference between the works of Didier et al. [47,48] and the present study on
propagating uncertainty in nonlinear dynamical systems, besides the representation way of uncertainties, is the use of intrusive/non-
intrusive approaches. The intrusive methods usually require a rewriting of the solution process by incorporating uncertainties in every
step, which is not easy for non-experts. This also means different nonlinear vibration problems need to be solved based on their own
derivations. The most important superiority of intrusive methods is computational efficiency, which is higher than that of non-

3
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

intrusive methods. However, non-intrusive methods have obvious advantages in dealing with different problems as they treat
deterministic systems as a black box. This is very useful for engineering problems within various context. What’s more, any established
solvers can be conveniently fitted into the non-intrusive frameworks without modifications. To the best of the authors’ knowledge,
little research has been done to study the uncertain dynamics of a dual-rotor system with rub-impact, where the uncertainties are
inevitable in either the model or the rub-impact fault. More importantly, the multiple solutions for a single control parameter in the
nonlinear vibrations will pose great challenges for the uncertainty propagation methods, which yet are not well addressed [49].
Present methods either use complex intrusive derivations or are associated with a particular nonlinear tool, which is not convenient to
be generalised into other problems.
The main contributions made in this work can be summarised into two points. The first novelty is the proposed polar angle
interpolation (PAI) method that is dedicated to the response prediction of nonlinear forced responses with uncertainty. It is a pure post-
processing technique and can be used in other nonlinear systems, contrary to existing methods in the literature. Moreover, it frees the
user to choose any effective non-intrusive UQ tools. The second contribution lies in the intensive study of the uncertain and nonlinear
dynamic responses of a dual-rotor system involving rub-impact fault. In industry, it is very hard to obtain the friction and contact
stiffness in a dual-rotor system. And the sensitivities of the forced response curves illustrated by the results are of primary importance
to relevant engineers.
The remainder of the content is as follows. The mathematical modelling for the rub-impact dual-rotor using the Lagrangian method
is elaborated in Section 2. In Section 3, the nonlinear approaches for solving the nonlinear system without uncertainty are formulated.
The transformation of the responses based on the PAI method is described in Section 4 to allow accurate prediction for response ranges
with multiple solutions. Section 5 demonstrates the theory of the surrogate modelling method for overall uncertain response esti­
mation. The numerical results and relevant discussions are provided in Section 6. The last section presents the main conclusions.

2. Mathematical modelling of a dual-rotor system with rub-impact fault

In a dual-rotor system, there are two rotors connected by the inter-shaft bearing, namely the higher pressure (HP) rotor and the
lower pressure (LP) rotor. Fig. 1 shows the schematic diagram of a typical dual-rotor model, which has wide applications in aero-
engines [2,20]. The HP and LP rotors are also called the outer and inner rotors concerning their position, as shown in Fig. 1. The
system is supported on the ends of the two rotors and a linear elastic model is assumed for all the supports. Two thin and rigid discs are
mounted on the massless elastic shafts with O1 and O2 being the geometrical centres. From the left end to the right end of the dual-rotor
system, the geometries of each featured component are marked by Li , i = 1, …, 5. Generally, the HP and LP rotors are rotating at two
frequencies ω1 and ω2 , which are incommensurable. In this paper, the two rotors are co-rotating [50]. Contact often occurs between
the HP rotor and the casing because of the small clearance. The following of this section will establish the mathematical model of the
system with rub-impact excluding the uncertainty.
We first present the contact model between the HP rotor and casing when the rub-impact begins. The motion relationships are
demonstrated in Fig. 2. There are two force components, i.e., the normal impact force Fn and the tangential frictional force Fτ . There is
no contact when the radial deflection of the HP rotor is smaller than the clearance, which will be denoted by δ in the following context.
In such cases, no forces will be added to the dual-rotor system, and it vibrates linearly, i.e., Fn = Fτ = 0. Therefore, the rubbing forces
are non-smooth and depend on the vibrational amplitudes. Assuming the casing can deform marginally and elastically when impacted

Casing
Disc 2
Disc 1

y
Inner rotor Inter-shaft bearing
x

O O1 O2
z

Bearing #1 Bearing #3
Bearing #2
Outer rotor

L1 L2 L3 L4 L5

Fig. 1. Physical model of a dual-rotor system.

4
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

Fig. 2. Contact model between the outer rotor and casing.

by the HP rotor and the friction between them can be described by the Coulomb friction theory, the forces induced by the contact
process can be expressed as [11,15]
⎧ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅
⎨ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Fn (xh , yh ) = Kc ( x2h + y2h − δ)
, if x2h + y2h − δ⩾0 (1)
⎩ F (x , y ) = μF
τ h h n

where xh and yh denote the lateral displacements of the HP rotor, Kc is the impact stiffness noting the normal reaction and μ represents
the frictional coefficient related to the tangential force.
Project the forces Fn and Fτ on the fixed coordinate frame, then the force components on the HP rotor can be expressed as
⎧ δ
⎪
⎪ Fx = Kc (1 − √̅̅̅̅̅̅̅̅̅̅̅̅̅̅ )(xh − μyh )
⎪
⎪
⎪
⎨ xh + y2h
2
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅
, if x2h + y2h − δ⩾0 (2)
⎪
⎪ δ
⎪ Fy = Kc (1 − √̅̅̅̅̅̅̅̅̅̅̅̅̅̅
⎪ )(μ x + y )
⎪
⎩
h h
x2h + y2h

We use the Lagrangian formulation to derive the equations of motion of the dual-rotor system. First, the energy analyses of the system
should be conducted [20]. The two rotors are assumed rigid, and two lateral and two rotational displacements are used to describe each
rotor. Based on the above assumptions, the kinetic energy of the rotors is given by
1 1
(3)
2 2
Ei = mi (ẋ2i + ẏ2i ) + [Jdi (θ̇xi + θ̇yi ) + Jpi ω2i ] − Jpi ωi θ̇yi θxi , i = l, h
2 2

where mi , Jdi and Jpi are the mass, transverse moment of inertia and polar moment of inertia, respectively. θxi and θyi are rotation angles
of the two rotors. A dot over the symbol means derivation with respect to time. Subscripts l and h denote quantities for the LP and HP
rotors, respectively.
The potential energy in the bearings is expressed as
1
Vi = Ki (x2bi + y2bi ), i = 1, 2, 3 (4)
2

where (xbi , ybi ), i = 1, 2, 3 defines the positions of the bearings and Ki , i = 1, 2, 3 represents the stiffness. Their positions can be
calculated by the following formula based on geometrical relationships
{
xb1 = xl − L1 θyl
(5)
yb1 = yl + L1 θxl
{
xb2 = xh − L3 θyh
(6)
yb2 = yh + L3 θxh

5
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

{
xb3 = xl + (L2 + L3 + L4 + L5 )θyl
(7)
yb3 = yl − (L2 + L3 + L4 + L5 )θxl

Subsequently, Rayleigh’s dissipation energy can be written as

1
Di = Ci (ẋ2bi + ẏ2bi ), i = 1, 2, 3 (8)
2

where Ci , i = 1, 2, 3 denotes the bearing damping. The connecting structure of the two rotors is treated as a linear elastic spring and
the supporting forces are calculated by
{
fx = Khl [xl + (L2 + L3 + L4 )θyl − xh − L4 θyh ]
(9)
fy = Khl [yl − (L2 + L3 + L4 )θxl − yh + L4 θxh ]

where Khl is the stiffness of the inter-shaft bearing.

Assuming the initial phases of mass unbalances equal to zero, the unbalance forces in the system can be expressed as
⎧
⎨ f (t) =[m e ω2 cos(ω t), m e ω2 sin(ω t), 0, 0]T
(10)
ul l 1 1 1 l 1 1 1
⎩ f uh (t) =[mh e2 ω2 cos(ω2 t), mh e2 ω2 sin(ω2 t), 0, 0]T
2 2

where e1 and e2 are the eccentricities of the LP and HP rotors.

Denote the system displacement vector as U = [xl , yl , θxl , θyl , xh , yh , θxh , θyh ]T , the second type Lagrangian equation can be
applied as
⎛ ⎞
d ⎜ ∂E ⎟ ∂E ∂D ∂V
⎝ ⎠− + + = Qj , j = 1, …, 8 (11)
dt ∂U̇ j ∂Uj ∂U̇ j ∂Uj

where E = El + Eh , D = D1 + D2 + D3 ,V = ml gyl +mh gyh +V1 +V2 +V3 and Qj represents the generalized force for each degree of
freedom.
The reaction forces and moments induced by the inter-shaft bearing with respect to the displacement vector can be written as
[ ]T
̂ fx ̂
f int (t) = fx , fy , − fy L, L, − fx , − fy , fy L4 , − fx L4 = Kint U(t) (12)

where
⎡ ⎤
⎢ 1 0 0 ̂
L − 1 0 0 − L4 ⎥
⎢ ⎥
⎢ 0 1 ̂
− L 0 0 − 1 L4 0 ⎥
⎢ ⎥
⎢ ⎥
⎢
⎢ 0 ̂
− L ̂2
L 0 0 ̂
L ̂ 4
− LL 0 ⎥⎥
⎢ ⎥
⎢ ̂
L 0 0 ̂2
L ̂
− L 0 0 LL4 ⎥
− ̂
Kint = Khl ⎢
⎢
⎥
⎥ (13)
⎢ − 1 0 0 ̂
− L 1 0 0 L4 ⎥
⎢ ⎥
⎢ 0 − 1 ̂
L 0 0 1 − L4 0 ⎥
⎢ ⎥
⎢ ̂ 4 ⎥
⎢ 0 L4 − LL 0 0 − L4 L24 0 ⎥
⎢ ⎥
⎣ ̂ ⎦
− L4 0 0 − LL4 L4 0 0 L24

where Kint is the stiffness matrix induced by the inter-shaft bearing and ̂
L = L2 + L3 + L4 .
After calculation and re-organization, the following matrices-based equation of motion can be obtained:

MÜ(t) + (C+ω1 G1 + ω2 G2 )U̇(t)+(K + Kint )U(t) = Qs + Qu (t) + Qnl (U, t) (14)

where M, C and K represent the mass, viscous damping and stiffness matrices of the dual-rotor system, respectively, G1 and G2 are the
gyroscopic matrices of the two rotors, and Qs, Qu and Qnl are the static gravitational forces, unbalance excitation vector and the
nonlinear rub-impact forces, respectively. The detailed expressions of them are provided in Appendix A.
The dynamical system shown in Eq. (14) is nonlinear and multiple-frequency excitations are contained. Although the system used
here has only eight degrees of freedom, it captures the main nonlinear dynamics of the dual-rotor and a better illustration of the effects
of uncertain parameters can be obtained. Moreover, the framework proposed in the upcoming sections can be generalised to a system
with more degrees of freedom. For a large system, the required computation cost will increase since non-intrusive UQ methods often
involve sampling of the uncertainties, which implies calls of the deterministic system.

6
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

3. Nonlinear solution procedures for the dynamical system

3.1. Deterministic analysis with MHBM-AFT

The nonlinear problem of the dual-rotor system involving rub-impact can be solved efficiently by the MHBM [51], which is the
extended version of the conventional HBM. Due to the two incommensurable frequencies in unbalance excitations, the dynamical
response of the dual-rotor system is non-periodic, and its frequency components contain possible combinations of ω1 and ω2 as
k1 ω1 + k2 ω2 (15)

with ki = − Nh , − Nh + 1, …, − 1, 0, 1, …, Nh − 1, Nh for i = 1, 2. Nh denotes the order of the two fundamental frequencies

retained. The implementation of the MHBM coupled with the alternating frequency/time (MHBM-AFT) technique will be used.
The steady-state dynamic response of the system can be approximately expressed using a generalized Fourier series as
Nh
∑ Nh
∑
U(t) = (ak1 k2 cos( k1 ω1 + k2 ω2 )t + bk1 k2 sin( k1 ω1 + k2 ω2 )t) (16)
k1 =− Nh k2 =− Nh

where ak1 k2 and bk1 k2 are the Fourier coefficients in the decomposition for cosine and sine terms. In condensed form, it is denoted by
∑
U(t) = a0 + (ak cos( k ⋅ ω)t + bk sin( k ⋅ ω)t) (17)
k∈Z2

where operator ( k ⋅ ω) represents the inner product, k= [k1 , k2 ] denotes the harmonic number set in different frequency directions
and ω = [ω1 , ω2 ] stands for the two fundamental frequencies in the solutions. However, only positive combinations of the two fre­
quencies are kept in the nonlinear forces and solutions [52], i.e., ( k ⋅ ω) = k1 ω1 + k2 ω2 > 0.
For convenience, Eq. (14) is rewritten as
̃ U̇(t)+KU(t)
MÜ(t)+C ̃ ̃
= Q(t) (18)

where K̃ = K + Kint , C
̃ = C +ω1 G1 +ω2 G2 and Q(t) ̃ = Ql (t) + Qnl (t), in which Ql (t) = Qs + Qu (t) represents the unbalance excitations
and the gravitational forces, which are constant quantities or can be expressed by the first-order harmonic components. The Qnl (t)
corresponds to the rub-impact forces, which can also be written in the condensed form of the generalized Fourier series
∑
Qnl (t) =c0 + (ck cos( k ⋅ ω)t + dk sin( k ⋅ ω)t) (19)
k∈Z2

where ck and dk are the unknown Fourier coefficients for the nonlinear forces. By submitting Eqs. (17) and (19) into Eq. (18), a set of
nonlinear algebraic equations (NAEs) can be obtained
̃
g(X) = HX + H(X) − He = 0 (20)

where X regroups the unknown Fourier coefficients for the nonlinear dynamical response

X = [a0 , a1 , b2 , ⋯, aNc , bNc ]T (21)

with Nc being the total number of harmonic terms in Eq. (20), which is a variant depending on the values of ω and the order of the
fundamental frequencies Nh . The notations H, H(X)
̃ and He are the linear diagonal stiffness matrix, the Fourier coefficients of the rub-
impact forces Qnl (t) and the projected vector of the external forces Ql (t), respectively. The detailed expressions of these matrices and
vectors are given in Appendix B.
The NAEs shown in Eq (20). should be solved by the Newton-Raphson iteration method to find a fixed solution
[ ]− 1
X(i+1) = X(i) − J(X(i) ) g(X(i) ) (22)

where X(i) and X(i+1) represent two consecutive values of vector X at the ith and (i + 1)th steps. J(X(i) ) is the Jacobian matrix of g(X)
calculated at the ith iteration step. The termination criterion for the iteration can be defined by
⃦ (i+1) ⃦
⃦X − X(i) ⃦∞ ⩽ε (23)

in which ‖ ⋅ ‖ denotes the Euclidean norm and ε is a predefined tolerance. The detailed calculation process of the Jacobian matrix J via
the AFT technique can be found in Appendix C.

3.2. Path following by using the predictor–corrector method

It is difficult to achieve convergence or find the right path for the dynamic responses with large curvature and turning points. The

7
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

continuation algorithm, also called the predictor–corrector method, is widely used to ensure the proper path following of the nonlinear
response curves. This type of method usually involves two main steps. The first step is to produce an estimated point for the next
solution based on the predictor and the second is to correct the estimated one to approach the true solution iteratively by using the
corrector. Many dedicated algorithms can be found for the path following. The arc-length continuation [53] will be used here to track
the multiple solutions for the rub-impact dual-rotor system. An additional equation is added to the previous system to meet the arc-
length conditions
g(X, υ) = 0
g(X) = ̃ (24)
( )2 ( )2
‖dX‖ dυ
+ =1 (25)
ds ds

where dX = Xj+1 − Xj and dυ = υj+1 − υj . Subscript j denotes the last convergent solution point and j + 1 represents the estimated next
point. An initial guess of the start values can be set as the vibration amplitudes of the underlying linear system. υ is a control parameter
and it is usually set as the rotating frequency of the rotors. ds is the continuation parameter denoting the length of the curvilinear step,
which is used to predict the next point along the response curve. This procedure searches the possible solutions on the hypersphere
space with the incremental curve length.
The value of the incremental distance should be adjusted automatically according to the increments in the response curves. It will
be reduced if abrupt changes are found, and the next estimated point will be close. In the case of no significant changes being detected,
the increment of arc length should be increased to promote computation efficiency [51]. In the arc-length continuation, both the
rotational speed and the corresponding amplitude are variables, and they will be determined according to the predefined iteration
rules. Therefore, the dynamic response curves under different parameter sets will have different speed samples and the data length will
be changing in the same speed interval.

4. Transformation of the responses with multi-solutions via the PAI

As discussed in the introduction, the dynamical responses of nonlinear systems will have multiple solutions near the turning points.
The nonlinear nature and the multi-solution property pose challenges for uncertainty analysis, especially for non-intrusive uncertainty
quantification (UQ) methodologies. It is feasible to predict the uncertain vibrations of nonlinear systems using time history [40,41],
which is in fact a single-valued form by nature and cannot show the amplitude-frequency characteristics in a certain rotation speed
range. Currently, few non-intrusive UQ approaches have been reported that can automatically deal with multiple solutions without
extra manipulations. Moreover, the continuation method described in Section 3.2 uses variable speed steps in different simulation
conditions, which will produce different lengths of the response vectors. In other words, the sample response vectors with different
parameter sets cannot be further processed under the rotational speed points based on a constant incremental step. New sampling
strategies should be incorporated to ensure that the sample response vectors are on ‘the same condition’. Sinou et al. [37] treated the
rotating speed of a flexible rotor with nonlinearities as a random variable and expanded it using the polynomial chaos expansion to
allow the response prediction at turning points. Sarrouy et al. [54] proposed to sample in the phase space of vibrations and applied it
into a linear vibration system. Panunzio et al. [55] proposed the arc length ratio to deal with uncertain variables in nonlinear problems,
where the turning points should be marked. The sole goal of the above methods is to transform the vibration responses of nonlinear

Fig. 3. Illustration of fixed rotation speed sampling.

8
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

systems into a single-valued mode that can be sampled using a new control parameter (rather than the rotational speed) based on fixed
incremental steps. Thus, all the response sample data will be on ‘the same condition’ that can be further processed by most of the UQ
methods. Here, we propose the coordinate transform from the Cartesian coordinates to the polar coordinates and apply the PAI method
for the nonlinear response range prediction of dual-rotor systems with rub-impact fault.
To explain the PAI method, we give the expression for the global steady-state dynamic responses (the vibration amplitudes) of the
LP and HP rotors as
⎧ (√̅̅̅̅̅̅̅̅̅̅̅̅̅̅ )
⎪
⎪ rl = max x2l + y2l
⎨
(√̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ) (26)
⎪
⎪
⎩ rh = max x2h + y2h

where rl and rh represent the deflections of the LP and HP rotors for a specified rotating speed pair. When the rotating speed is
considered in a range, there will be arrays of the vibration amplitudes ̃ r and a row of speed samples ω
̃ . As mentioned previously, two
critical problems arise for UQ in nonlinear problems with turning points where the arc-length continuation is applied. Firstly, the
multi-solutions at different solution branches for a single rotating speed lead to failure of the UQ methods. To show the situation, Fig. 3
demonstrates the sampling mode for different nonlinear response curves under the fixed rotating speeds. In Fig. 3, the displayed curves
are five samples for the amplitude-frequency response of the dual-rotor system around the second resonance with an uncertain
imbalance in the LP rotor. Secondly, the same rotating speeds are not necessarily chosen in the continuation for the different parameter
samples during the UQ and the size of them will also be a variable depending on the chosen physical parameter set. It makes different
response curves technically incomparable to each other. To overcome the two issues, the response curves are transformed into the
polar sub-coordinates firstly as
⎧
⎪
⎪ ̃r − r0
⎪
⎪ αi = arctan( i )
⎪
⎨ ω̃ i − ω0
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ , i = 1, 2, 3, … (27)
⎪
⎪ ( )2 ( )2̅
⎪
⎪ ρ = ω
̃ − ω + ̃r − r
⎪
⎩ i i 0 i 0

where P0 (ω0 , r0 ) denotes a polar origin added to the coordinate system and R(̃ωi , ̃ri ) is a data point on the response curve, αi and ρi are
the polar angle and polar radius, respectively. The transform of the polar angles is implemented as the four-quadrant inverse tangent
function, which can produce angles from − π to π. Zero will be assigned to the polar angle when ω ̃ i = ω0 . The polar origins can be
multiple and the only principle in selecting them is to ensure that only one single vibration amplitude of the original response curve can
be mapped to every polar angle. Fig. 4 presents the mechanism of the polar transform for the same situation as shown in Fig. 3. By
adding three polar origins whose ordinates are equal to the clearance value between the rotor and casing, the response curves in the
range considered are divided into three segments. In Fig. 4, a single polar angle starting from the polar origins will only correspond to
one amplitude on a single response curve, i.e., they are in a single-valued mode. As can be seen from Fig. 4 that the value of the abscissa
is much higher than the ordinate, causing the transformed polar points to be clustered in narrow ranges (i.e., all the polar angles are
close to zero) and further leading to inaccuracy in the UQ. Therefore, an amplification factor is introduced to scale the ordinates and

P1 P3
P2

y=

Fig. 4. Illustration of fixed polar angle sampling.

9
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

make the transformed polar points more scattered. It can be viewed as an informal normalization. Then, the transformation from the
Cartesian coordinates to the polar one is re-written as
⎧
⎪
⎪ A (̃r − r0 )
⎪
⎪ αi = arctan( scale i
̃ )
⎪
⎨ ω̃ i − ω0
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ , i = 1, 2, 3, … (28)
⎪
⎪ ( )2 [ ]2̅
⎪
⎪ ρ
̃ = ω
̃ − ω + A (̃r − r )
⎪
⎩ i i 0 scale i 0

where Ascale is a scaling factor to the original ordinates, ̃

αi and ̃ρi represent the scaled polar angle and polar radius after amplification. At
this point, the first issue raised previously due to the possible multi-solutions in the nonlinear steady-state response curves is solved.
Next, we need to address the inconsistency in steps of the rotating speed (the control parameter in the arc-length continuation) for
the simulations at different samples. Actually, there is no such approach available to make the speed samples unified. We can sample in
every sub-coordinate using the polar angle as a new free parameter alternatively with fixed incremental step

αend − αstart
α(j) = αstart + j , j = 0, 1, …, N1 (29)
N1

where α(j) is the jth polar angle sample, N1 +1 denotes the number of the polar angle samples. The αstart represents the polar angle of the
first point in the current polar system and αend corresponds to the polar angle of the last point. Their values usually take − π, − π/2, π

Cart2pol Interp1(Spline)

Pol2cart Interval UQ
Fig. 5. The technical route of the polar angle interpolation method.

10
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

and 0. The complete set of samples of the polar angles will remain unchanged for different simulations.
For the sampled polar angle vector α, the corresponding polar radius vector ρ may not be readily available. In such a case, the
interpolation technique is employed to extract the sample values. The cubic spline interpolation method can be used with satisfactory
performance. It generates a cubic polynomial on consecutive polar points as

W(α) = w3 α3 + w2 α2 + w1 α + w0 (30)

where w3 , w2 , w1 and w0 are the four coefficients to be determined. There are many mature methods for this derivation, and it will not
be detailed here. Once the cubic polynomial is constructed, the sample values of the radii can be obtained as ρ = W(α). The above
operation solves the second issue that the samples are not on ‘the same condition’.
To this point, the samples of both the polar angles and the radius can be processed by the UQ algorithms, which will be described in
the next section. The UQ algorithm predicts the interval range of polar radius on every unified polar angle step. To present the obtained
interval results in the Cartesian coordinates, the response curves should be transformed back from the polar frames as a reverse of Eq.
(28). It can be expressed as
⎧
⎪
⎨ω
⎪ ̃i = ̃ρi cos̃
αi
ρ
̃ α
siñ i
, i = 1, 2, 3, … (31)
⎪ i
⎩ ̃ri = A
⎪
scale

The afore-mentioned three steps pave the road for the following uncertainty analysis of the nonlinear dynamical responses, which is
now the same as UQ in linear vibration problems. For better reader comprehension, the technical route of the PAI method is shown in
Fig. 5.

5. Surrogate modelling for the propagation of non-random uncertainty

5.1. Representation of non-random uncertainty in the dual-rotor system with contact

According to the classification of different types of uncertainties, the non-random or UBB descriptions are more suitable for dual-
rotor systems with limited prior distribution information of variables. For the system studied in this paper, we consider the variabilities
in the two main sources, i.e., the physical parameters of the model and key parameters related to the rub-impact fault. More specif­
ically, the uncertainties in the imbalances, bearing damping, speed ratio, contact stiffness, friction coefficient and clearance are
investigated. Without sufficient statistical information, they are modelled mathematically as the non-random interval quantities using
two bounds. For example, the uncertain imbalances of the two rotors induced by wear or material degradations can be described as

eIj = [ecj − βe,j ecj , ecj + βe,j ecj ], j = 1, 2 (32)

where the subscripts indicate the two rotors. The eIj , ecj and βe,j are the interval imbalance, mid-value and variability coefficient,
respectively. Let Δej = βe,j ecj , j = 1, 2, it further leads to

eIj = ecj + Δej [− 1, 1], j = 1, 2 (33)

Then the variables [ej ] in the intervals of imbalances can be transformed as

[ej ] − ecj
γj = , [ej ] ∈ eIj , γj ∈ [− 1, 1], j = 1, 2 (34)
Δej

The reorganization in Eq. (34) transfers the variables in arbitrary intervals to new ones constrained in the standard interval, which
will benefit the surrogate modelling in the next section. Reverse transform can be applied from the new variables to the actual pa­
rameters for physical mapping.
Similar expressions can be obtained for the interval damping, contact stiffness, speed ratio, friction coefficient and clearance using
interval models. Correspondingly, a new variable is introduced for each of the interval uncertainty and a vector containing them will
be generated:

γ = [γ1 , γ2 , ⋯, γ i , ⋯] ∈ [− 1, 1]n , i = 1, 2, ⋯, n (35)

where n denotes the number of interval uncertainties. We then can deal with the new set of uncertain variables instead of the practical
uncertainties in a standard way.

5.2. Non-intrusive uncertainty propagation via the surrogate method

The previous sections have addressed the difficulties in applying uncertainty propagation analysis. In this subsection, the interval
UQ based on non-intrusive computation will be elaborated which allows to track the variabilities of the nonlinear dynamic response of
the rub-impact dual-rotor system with UBB quantities. We use the steady-state amplitudes as the objective quantities in the UQ, but it

11
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

should be noted that the interval polar amplitude should be first calculated and then transformed back to the interval amplitudes and
rotational speeds. For the interval UQ, many algorithms are available with their respective strengths and limitations. The Chebyshev
inclusion function proposed by Wu et al. [41] has many merits in applications to the engineering uncertain problems, such as
transparency in calculation, non-intrusiveness and excellent control on the overestimations. However, it will be overwhelmed by the
underlying computational burden in multi-dimensional uncertain systems due to the tensorial sampling times. An improved collo­
cation scheme was further developed to alleviate the simulation cost with fewer samples [56], which will be incorporated here to
estimate the interval nonlinear dynamic response. Compared with conventional methods such as the perturbation technique and
Taylor series method, the prerequisites of small variations of uncertainties and derivative information can be ignored.
When uncertain variables γI are considered, the NAEs expressed in Eq. (20) should be re-written in interval representation
̃ I , γI ) − He (γI ) = 0
g(XI , γI ) = H(γI )X(γI ) + H(X (36)

where γI = [γ, γ] denotes the interval vector, γ and γ are the lower and upper limits. Then, the Newton-Raphson iteration procedure
evolves to

X(i+1) (γI ) = X(i) (γI ) − [J(X(i) (γI ), γI )]− 1 g(X(i) (γI ), γI ) (37)
It is found that the interval system expressed in Eq. (37) cannot be directly iterated due to the potential enormous overestimations,
which will cause the obtained interval results meaningless. For clarity in the following deduction, the interval polar amplitudes under
unified polar angle samples will be handled and the interval derivations based on Section 3 and 4 will be omitted. To obtain the interval
solutions, the non-intrusive surrogate model (NISM) for the uncertain dynamic response (i.e., the interval polar amplitude) is firstly
expressed by the q-order Chebyshev approximation as
q
∑ ∑ q
1
ρ(i+1) (γI ) ≈ ⋯ φ (ρ(i+1) )Tj1 ,j2 ,⋯,jn (γI ) (38)
j1 =0 jn =0
2ν j1 ,j2 ,⋯,jn

with ν being the number of 0 in the index set j1 , j2 , ⋯, jn and φj1 ,j2 ,⋯,jn (ρ(i+1) ) denotes the unknown Chebyshev coefficient, which is a
function of the polar amplitude ρ(i+1) obtained in the (i + 1)th iteration step. Tj1 ,j2 ,⋯,jn (γI ) is the n-dimensional Chebyshev polynomial
series

Tj1 ,j2 ,⋯,jn (γI ) = Tj1 (γI1 )Tj2 (γI2 )⋯Tjn (γ In )

(39)
= cos(j1 ϕ1 )cos(j2 ϕ2 )⋯cos(jn ϕn )

where
ϕi = arccosγ i ∈ [0, π ], γi ∈ [− 1, 1], i = 1, 2, …, n (40)
Generally, the Gauss-Chebyshev quadrature is used to solve the Chebyshev coefficients in Eq. (38) numerically as
( )n ∫ 1 ∫ 1 (i+1) I
2 ρ (γ )Tj1 ,j2 ,⋯,jn (γI )
φj1 ,j2 ,⋯,jn = ⋯ √̅̅̅̅̅̅̅̅̅̅̅̅̅ √̅̅̅̅̅̅̅̅̅̅̅̅̅ dγ1 ⋯dγn
π − 1 − 1 1 − γ21 ⋯ 1 − γ2n
(41)
( )n ∑
q+1 q+1
∑
2
≈ ⋯ ρ(i+1) (γj1 , …, γjn )Tj1 ,j2 ,⋯,jn (γj1 , …, γjn )
q + 1 j1 =1 jn =1

The computation cost is high, however, because of the tensorial sampling in the Chebyshev interpolation. It should be noted that
the Chebyshev basis can also be expressed in polynomial form
{
T0 (γ) = 1, T1 (γ) = γ
(42)
Tj+1 (γ) = 2γTj (γ) − Tj− 1 (γ)

As the desired NISM order is q, redundant terms of the same order and terms with orders higher than q in Eq. (38) can be deleted
without reducing the accuracy. Then, the refined surrogate is given by
∑
ρ(i+1) (γI ) ≈ φ̃ j1 ,j2 ,⋯,jn (ρ(i+1) )Tj1 ,j2 ,⋯,jn (γI )
0⩽j1 +j1 +⋯+jn ⩽q
∑ (43)
= ̃
̃ j1 ,j2 ,⋯,jn (ρ(i+1) )γ j11 γj22 ⋯γjnn
φ
0⩽j1 +j1 +⋯+jn ⩽q

where φ̃ j1 ,j2 ,⋯,jn (ρ(i+1) ) represents the new coefficient to be determined. The Chebyshev polynomial basis is replaced by the simplex
̃
polynomial basis equivalently, which changes the coefficients, but the accuracy will remain the same. The number of terms on the
right-hand side of Eq. (43) is

12
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

(caption on next page)

13
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

Fig. 6. The flowchart of the uncertainty analysis framework.

(n + q)!
n=
̂ (44)
n!q!

Therefore, it will only need ̂

n samples to determine the surrogate model for the uncertain responses theoretically. However, it is
suggested that robust estimations will be produced when two times of the required least number of parameter samples in surrogate
modelling are used [57]. The 2 ̂
n samples for the uncertain parameters can be extracted randomly from the candidate pool, which is the
tensorial form grid of samples in every uncertain dimension. For each dimension, the Chebyshev roots are used as the interpolation
points
2m − 1
̂γ m = cos π , m = 1, 2, ⋯, q + 1 (45)
2(q + 1)
For the extracted samples, we label them sequentially as
̂
γ k = {̂
γ k,1 , ̂γ k,2 , …, ̂γ k,n }, k = 1, 2, …, 2̂
n (46)

Then, a vector of the sampled polar amplitude at the (i + 1)th iteration step ̂ can be calculated using the deterministic polar
(i+1)
ρ
n×1
2̂
amplitude at the ith step based on Eq. (22) by assigning the uncertain parameters to the samples. At a specific sample for uncertainties,
it is expressed as

X(i+1) (γj1 , …, γ jn ) = X(i) (γj1 , …, γjn )

(47)
− [J(X(i) (γ j1 , …, γjn ), (γj1 , …, γ jn ))]− 1 g(X(i) (γ j1 , …, γjn ), (γj1 , …, γ jn ))
{ }
ρ (i+1) (γj1 , …, γjn ) = Cart2pol X(i+1) (γj1 , …, γjn )
̂ (48)

For convenience, we write Eq. (43) as

̃
̃ (ρ(i+1) )Υ T
ρ(i+1) (γI ) = φ (49)

with
̃ ̃
̃ = {φ
φ ̃
̃1 , φ ̃
̃ 2 , …, φ
̃̂n } (50)

and
{
Υ = {Υ 1 , Υ 2 , …, Υ̂n }
(51)
Υ k = γ j11 γj22 ⋯γjnn , k = 1, 2, …, ̂
n

Similarly, a matrix of the simplex basis values Υ

̂
2̂
n ×̂
n
can be obtained after a traverse calculation of the parameter samples. In this
way, the unknown coefficients in Eq. (43) can be evaluated by the following regression model:

̃ ̂TΥ
̃ = inv( Υ
φ ̂T̂
̂ )Υ ρ
(i)
(52)
Once the unknown coefficients are solved, the surrogate is established. It is a polynomial function of standard interval variables,
and the max/min values can be easily obtained. In other words, the interval polar amplitude under all the unified polar angles will be

Table 1
Model parameters of the dual-rotor system.
Description Value Description Value

Length of shaft, L1 0.25 m Bearing stiffness, K1 7 × 106 N/m

Length of shaft, L2 0.07 m Bearing stiffness, K2 8 × 106 N/m
Length of shaft, L3 0.15 m Bearing stiffness, K3 7 × 106 N/m
Length of shaft, L4 0.10 m Bearing stiffness, Khl 7 × 108 N/m
Length of shaft, L5 0.05 m Contact stiffness, Kc 2.2 × 107 N/m
Clearance value, δ 2.8 × 10− 4
m Bearing damping, C 9.95 N ⋅ s/m
Mass of disc, ml 16.25 kg Imbalance of disc, e1 4.5 × 10− 5
m
Polar inertia, Jpl 0.186 kg ⋅ m2 Diameter inertia, Jdl 0.093 kg ⋅ m2
Mass of disc, mh 4.30 kg Imbalance of disc, e2 2.0 × 10− 4
m
Polar inertia, Jph 0.0215 kg ⋅ m2 Diameter inertia, Jdh 0.0108 kg ⋅ m2

14
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

solved. A transform expressed in Eq. (31) should lead to the interval estimations of the vibrational amplitudes of the faulted dual-rotor
in the Cartesian coordinates. Although the data points of the outputs in the Cartesian coordinates are not distributed on unified
rotational speed steps, the overall varying range of the interval response has been successfully estimated. For the reader compre­
hension, the computation flowchart of the established uncertainty analysis framework coupled with the PAI is given in Fig. 6.
The computation performance of the non-intrusive uncertainty method in multi-dimensional uncertainty problems is enhanced by
the regression strategy compared with the classic Chebyshev interval method since the total samples of uncertainties have been
reduced. Compared with previous frameworks with HBM/MHBM, whether the cost required in this method depends on several factors.
The first point is that the intrusive methods are usually non-sampling methods, which will be more efficient but are more difficult to
derive, such as [48]. For non-intrusive methods for propagating uncertainty in nonlinear vibrations, the order may be another factor
that influences the efficiency while maintaining sufficient accuracy. The sparse grid technique [55] can be used for the efficiency
enhancement. Therefore, the computational performance of the surrogate model in terms of accuracy and efficiency is satisfactory
within the current context but a quantitative comparison of CPU times with probabilistic methods coupled with HBM/MHBM may
require additional validations.

6. Numerical results and discussions

Numerical results under various uncertain conditions as well as accuracy validation will be presented in this section based on the
methodologies described previously. The deterministic vibration responses and the performance of the polar transformation procedure
will also be demonstrated. Based on the fact that the lower-order combinations of the two fundamental frequencies usually have the
most significant amplitudes [52], the number of harmonic components in the MHBM is further truncated and selected as |k1 | +
|k2 |⩽Nh . It is proved to be sufficient for computational accuracy when Nh = 5 by extra validations. The physical parameters of the dual-
rotor system and the clearance value between the rotor and casing are given in Table 1. During the contact process, the friction co­
efficient between the rigid surfaces μ equals 0.25. The rotating speed ratio between the HP and the LP rotors is set as λ = 1.2. It is
interesting to analysis the critical speeds of the dual-rotor system with an inter-shaft bearing. Based on the method described in [50],
the Campbell diagram is calculated by solving the eigenproblem with the damping matrix included, which is shown in Fig. 7. In the
figure, two dashed lines correspond to the LP and HP rotors. The intersection points (magenta markers, only forward modes are
excited) of the eigenmode lines and the synchronous excitation lines indicate that the first order critical speeds of the HP and LP rotors
are 886 rad/s and 1065 rad/s, respectively. To be noted, the first order modes are indeed influenced by the gyroscopic effects in this
model, but they are very small. The second order critical speeds of the two rotors are both over 2500 rad/s. The results are validated by
extra calculations of the frequency response curves of the linear dual-rotor system, part of which will be illustrated next.
First, the deterministic nonlinear responses of the system without considering uncertainty are provided in Fig. 8 for both the LP and
HP rotors. It is observed from the amplitude-frequency diagrams that the steady-state dynamic characteristics of the two rotors have
the same pattern, but the amplitudes of the HP are smaller. In the considered speed band, there are two resonance peaks and the second
has a higher amplitude. Linear vibration mechanism is noticed without the rub-impact fault when the clearance is set to a very large
value (δ = 1 × 104 m, for example). When the clearance is set to δ = 2.8 × 10− 4 m, two rubbing regions are present for both rotors and
the amplitudes in these speed intervals are decreased due to contacts. In other speed ranges, the amplitude of the system has not
reached the limit of the clearance and it will be identical to that of the linear case without rubbing. Moreover, a hysteresis feature of the
vibration peaks occurs, which is similar to the hardening effects in typical nonlinear systems. Consequently, multi-solution regions can
be found, which are predicted by the arc-length path-following technique. The above analyses provide some initial and fundamental

Fig. 7. Campbell diagram of the linear dual-rotor system.

15
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

(a) (b)

Fig. 8. Steady-state responses of the dual-rotor system: (a) HP rotor and (b) LP rotor.

Table 2
Parameter sets for numerical cases.
Case e1 e2 C3 λ Kc δ μ

1 10% – – – – – –
2 – 10% – – – – –
3 – – 10% – – – –
4 – – – 1% – – –
5 – – – – 10% – –
6 – – – – – 5% –
7 – – – – – – 5%
8 10% – 10% – 10% – –
9 5% 5% – – – – 3%

Fig. 9. Comparison of the interpolated response and the original response.

16
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

understandings of the steady-state dynamics of the rub-impact dual-rotor system.

In the following of this section, different cases with a variety of uncertain parameters will be considered and the effects of the
bounded uncertainties will be compared and discussed. For simplicity and conciseness, all the interval results will be provided
regarding the nonlinear responses of the HP rotor. To begin with, the calculation performance and some insights of the PAI-based
sampling strategy are illustrated in detail. The prediction accuracy of the uncertainty analysis method is validated by the interval
scanning method (ISM). Then, the effects of the single uncertain parameters are studied. The joint effects of multiple interval un­
certainties on the nonlinear dynamic responses of the dual-rotor system will be investigated in the last two numerical cases. Table 2
summarises the corresponding uncertain parameters and their respective uncertainty coefficients for each simulation case, which can
provide navigation of the numerical results and benefit the reader’s comprehension. The relevant case numbers are marked in the
captions of figures as well.

6.1. Verifications of the polar interpolation and the UQ scheme

In this subsection, we give the results of the polar interpolation and the interval range predicted by the non-intrusive surrogate
model in the polar coordinate. Further, the accuracy of the response estimations in the Cartesian coordinate is validated by the ISM,
which is crude and time-consuming. To show the effectiveness of the PAI method, the deterministic dynamic response of the HP rotor

(a)

(b)

Fig. 10. Response range estimation in the polar coordinate for the multiple solution regions (Solid lines—response samples of the surrogate; dashed
lines—interval bounds from the surrogate): (a) Interval range around the first peak and (b) Interval range around the second peak.

17
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

and its interpolated response is plotted in Fig. 9. The interpolated response is generated based on the procedures described in Section 4
using the spline interpolation and the amplification factor for the amplitudes is 500. Five polar origins are used to implement the
transformation whose abscissas are 0, 887, 967, 1068 and 1500 rad/s. The ordinate of these origins is fixed and set to the magnitude of
the clearance as 2.8 × 10− 4 m. As stated in Section 4, they can be chosen as other points as long as there is only one response value for
each polar angle. The sampling step for all the polar angles is π/2000 rad, which will generate 8000 polar angle samples in total. From
Fig. 9, we can see that the two response curves are identical to each other in both the linear vibration speed bands and the multi-
solution regions where rub-impact occurs. It proves that the transformation between the two coordinate systems will not change
the responses while the polar angle sampling will bring much convenience for the uncertainty analysis. At this stage, we investigate the
response prediction of the surrogate in the multi-solution speed bands since the linear parts of the vibration are not challenging. When
10% uncertainty in the imbalance of the LP rotor is considered, the sample responses (solid lines) and the bounds estimated by the
surrogate (dashed lines) near the two peaks are shown in Fig. 10. As we can see from Fig. 10, the bounds enclose all the sample re­
sponses and most importantly, they are all in a single-valued mode although multi-solutions may exist for a single rotating speed. In
such a way, it is similar to the UQ in linear systems. The values of polar amplitude in Fig. 10 have been scaled for robust polar
transformations and should be scaled down for the practical physical interpretations. One can adjust the scaling factors to make
different polar coordinates continuous at connection points. To demonstrate the direct comparison of the response bounds, Fig. 11
gives the response ranges (dashed lines) predicted by the surrogate and the responses (solid lines) from the ISM in which 21 evenly
distributed samples are used. From Fig. 11, we can notice that the bounds of the proposed method envelop all the sample responses
generated by the ISM and there are no obvious overestimations. However, it could be impossible to provide quantitative error analysis
for the results in the Cartesian system since the response data is not on ‘the same condition’. Instead, the response residual or error ratio
in the polar coordinates is defined by taking the results of the ISM as accurate ones. It is calculated as
⃒ ⃒
⃒ ⃒
⃒r − r⃒
η= ∼
× 100% (53)
r
∼

where r is the polar bound generated by the ISM and r is the response limit obtained from the surrogate model. The evolution of
∼
residual rates with the sampling sequence is given in Fig. 12 for the LB and UB. The results shown in Fig. 12 demonstrate that the results
from the two approaches are in good agreement with the largest error being about 0.11%. In other sampling points, the errors are very
small for both two response bounds. For a deeper view, the peaks in the error diagram are found to be the connection points of different
polar systems (Sequence numbers 1000, 3000, 5000 and 7000) and the nonlinear response peaks (Sequence numbers around 2320 and
6390). The first type of error is caused by numerical interpolations at the start or end of a polar system. The second corresponds to the
nonlinear vibration peaks whose vicinity is of high curvature, i.e., the response is very sensitive. To be noted, they can be further
reduced by improving the interpolation model in Eq. (30) and increasing the interpolation number, i.e., a step length less than π/2000
rad. Nevertheless, it is sufficiently small here.
The above analyses verify the accuracy of the PAI procedure and the surrogate for the uncertainty propagation. The dilemma faced
in the UQ algorithms of multi-solutions for a fixed rotating speed is successfully solved. Case studies with different uncertain pa­
rameters as well as comparisons and discussions will be presented in the next subsections.

Fig. 11. Results validation in the Cartesian coordinate (Solid lines—response samples of the ISM; dashed lines—interval bounds from
the surrogate).

18
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

Fig. 12. Error diagram of the interval bounds obtained from the surrogate model.

6.2. Steady-state responses of the system under uncertainty

In this subsection, the uncertain vibration responses of the rubbing dual-rotor system under several individual physical parameters
as well as multiple uncertainties simultaneously will be investigated. More specifically, the first four cases are associated with un­
certain model properties, including imbalances on the two discs, bearing damping and the rotational speed ratio. Then, Case 5 ~ 7
correspond to the uncertainties related to the rub-impact fault, i.e., the clearance, friction coefficient and contact stiffness. At last, two
cases will be studied where three uncertain variables are considered. In the results, the deterministic response curve is also given to
provide a straightforward illustration of the effects of the bounded uncertainties.

6.2.1. Effects of uncertain model parameters

Factors such as material wear and rust can cause variations in the balancing condition and the imbalance on rotors should be
considered uncertain. For this purpose, the imbalances on the two discs are modelled as interval quantities and their varying per­
centage of the uncertain portion is 10%. In single-dimensional uncertain problems, all the other physical parameters are constant
quantities with their deterministic values except for the considered uncertain variable. The effects of imbalance uncertainties on the
nonlinear vibration responses of the system are demonstrated in Fig. 13 and Fig. 14, which correspond to Case 1 and 2. The uncertain
imbalances will influence the unbalance excitation vector Qu in Eq. (14). It is found that uncertainty in the imbalance on the HP rotor
mainly affects the first vibration peak while the second one remains almost unchanged. For the uncertainty in the imbalance on the LP

Fig. 13. Vibration response with 10% uncertainty in the imbalance of the HP rotor (Case 1) (Red area—response range estimated by the surrogate;
blue solid line—deterministic response).

19
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

Fig. 14. Vibration response with 10% uncertainty in the imbalance of the lower pressure rotor (Case 2) (Red area—response range estimated by the
surrogate; blue solid line—deterministic response).

Fig. 15. Vibration response with 10% uncertainty in the damping of bearing #3 (Case 3) (Red area—response range estimated by the surrogate;
blue solid line—deterministic response).

rotor, it is the other way around that only the second vibration peak is significantly deviated and the first one is less influenced. The
above features are in agreement with the results of the parametrical investigations in Ref. [2]. There are also shifts of resonance
rotational frequencies due to uncertainties in imbalances, which cannot be observed in linear rotor systems. In detail, the positions of
peaks in the upper bounds are shifted to the right compared to the deterministic ones while those of the lower bounds are to the left.
With the same uncertainty variation range, the deviations of the dynamic response due to the uncertainty in imbalance on the LP rotor
are more significant than that in the imbalance on the HP rotor, which indicates the higher sensitivity of the system to the former.
Uncertainty in damping can be introduced in varying working conditions. For example, the properties of a squeeze film damper are
influenced by the temperature. The variability of the nonlinear response is given in Fig. 15 (Case 3) when there is 10% uncertainty in
the viscous damping of bearing #3. Uncertainties in bearing damping cause dispersion in the damping matrix C in Eq. (14). It can be
noticed that uncertainty in damping only affects the responses in the adjacent areas of the vibration peaks, which is similar to linear
cases. The difference is that the magnitude of vibration is increased or decreased due to uncertainty in damping towards the principal
evolution direction rather than up and down vertically at a fixed frequency. The second resonance peak is more deviated under
damping uncertainty than the first resonance peak. The slight shifts of the resonance peaks demonstrate the small variations of the
critical speeds of the dual-rotor system.

20
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

Fig. 16. Vibration response with 1% uncertainty in the speed ratio (Case 4) (Red area—response range estimated by the surrogate; blue solid
line—deterministic response).

The rotation speed ratio is a critical parameter in dual-rotor systems and has great influences on vibration behaviours. The pre­
dicted nonlinear responses of the system considering 1% interval uncertainty in the speed ratio are presented in Fig. 16, which is
denoted as Case 4. The speed ratio is associated with the rotating speeds of the HP and LP rotors, thus its variability has effects on the
gyroscopic matrices G1, G2 and unbalance excitation vector Qu in Eq. (14). We can find that significant deviations in the second vi­
bration resonance occur under only 1% fluctuation in the speed ratio and minor variations are observed around the first resonance. It
shows the high sensitivity of the system response to the variations of rotation speed ratio. The influence mechanism is quite different
from that of the uncertainty in imbalance on the LP rotor although they both mainly affect the second resonance. The largest vibration
magnitude almost remains the same compared to the deterministic one while it has oscillated in the case of uncertain imbalance. A
high-amplitude speed band is noticed where the amplitude stays large and almost the same, which is similar to the phenomenon
observed in the case of uncertain support stiffness in single rotor systems [58]. The hysteresis region has been widened due to the
uncertainty in speed ratio. When the rotational ratio is changed, the ratio of the two resonance frequencies varies as well. As the first
resonant frequency of the HP rotor remains almost unchanged, the second resonant frequency will change to satisfy the varying ratio.

6.2.2. Effects of uncertain fault parameters

In this subsection, uncertainties in the three key parameters related to the rub-impact fault are considered to reveal their effects on
the nonlinear dynamic responses. All fault-related uncertain parameters will introduce uncertainties of the nonlinear force vector Qnl.
The result of the predicted response range of the HP rotor is plotted in Fig. 17(a) when the contact stiffness between the HP and casing
rotors has 10% uncertainty, which is represented by Case 5 in Table 2. From Fig. 17, we can see that the deterministic response is not
affected in rotation speed bands with linear vibrations. Deviations begin when the rub-impact fault happens. The width of variation
ranges increases along with the rotating speed until the largest vibration amplitudes. The second vibration peak is more influenced
than the first one. After the vibration peaks, the influences of interval uncertainty in the contact stiffness become weak gradually and
disappear when the vibration amplitude is smaller than the clearance. The above features are reasonable since the contact stiffness will
only take effect when the rub and impact actually occur. Compared to the high-amplitude speed band observed in Case 4, the difference
in Case 5 is that the curvature of the high-amplitude band is in accordance with the principle direction of the nonlinear response curve,
which can be further evidenced in the zoomed views of the two peaks provided in Fig. 17(b) and (c). Moreover, it is found that the
deterministic amplitudes at the principal curvature direction will approach the upper bounds very closely but not penetrating.
Considering 5% uncertainty in the clearance, Fig. 18 shows the corresponding interval response of the HP rotor (Case 6). Not
surprisingly, it influences only the response when rub-impact happens just like the uncertainty in contact stiffness. However, un­
certainty in clearance has greater impacts on the two peaks throughout the whole rubbing process from the start to the end. At the
rotation speeds when the rub-impact fault occurs or disappears, the deviated response ranges appear or stop immediately at the
deterministic response curves. A decrease of the effects is noticed around the first resonance after the vibration peak until the rub-
impact event disappears. Similar high-amplitude speed bands are noticed as in Case 5. Moreover, a shifting mechanism of the reso­
nances contrary to Case 1–3 is observed that the peak in the UB is shifted to the left and that of the LB to the right. It can be described as
the widening of the hysteresis range of the dual-rotor system. The featured phenomenon indicates critical variations in the system
characteristics induced by fluctuations in the clearance.
The third rub-impact related uncertainty considered in this subsection is the bounded friction coefficient, which is noted by Case 7.
Fig. 19 presents the response variability under 5% fluctuations of the deterministic value. As can be noticed in Fig. 19, only weak
deviations are visible, and the influenced speed range is restrained very closely around the two vibration peaks. It indicates the low

21
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

(a)

(b) (c)

Fig. 17. Vibration response with 10% uncertainty in the contact stiffness (Case 5) (Red area—response range estimated by the surrogate; blue solid
line—deterministic response): (a) response in the full rotating speed range, (b) zoom view of the first peak and (c) zoom view of the second peak.

sensitivity of the nonlinear response to the small variations of the friction coefficient during the rub-impact event.
The results presented in Fig. 13-Fig. 19 capture the characteristics of uncertainty in nonlinear responses of the dual-rotor system
with different interval parameters. It is observed that the nonlinearity will be induced when rub-impact happens while other parts of
the curve are linear vibrations in essence. In the literature, works about propagating uncertainty in nonlinear dynamical systems
considered several types of nonlinearities [37,48,55], such as contact with stick–slip motion and cubic stiffness. Compared with the
results in these publications, the nonlinear responses reported in this work focus specifically on the rub-impact dual-rotor system
within an engineering background, where the curvature of dynamic responses is medium. For the nonlinear systems with more
complex frequency responses, which have extremely high curvatures, multiple turning points and more than three solutions for a single
frequency, the effectiveness of the proposed method may decrease.

6.2.3. Effects of multi-bounded uncertainties

In this subsection, two cases with three interval uncertain parameters are investigated (Case 8 and 9). The studies of single un­
certainty can show the effects of individual parameters and indicate their sensitivities. It is more likely that multiple uncertainties exist
simultaneously in an engineering dual-rotor system. The first set of parameters is the imbalance on the HP rotor, the contact stiffness
and the viscous damping in bearing #3 with 10% uncertainties in all of them. Response of the system under such circumstances is
shown in Fig. 20, which is noted as Case 8. The first resonance as well as the two peaks are significantly affected by the joint effects of
the three uncertainties, as demonstrated in Fig. 20. Wide hysteresis ranges can be seen around the two peaks. Naturally, there are no
obvious deviations after the second resonance according to the effects of the individual parameters. Complex shifting of peaks in the LB

22
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

Fig. 18. Vibration response with 5% uncertainty in the clearance (Case 6) (Red area—response range estimated by the surrogate; blue solid
line—deterministic response).

Fig. 19. Vibration response with 5% uncertainty in the friction coefficient (Case 7) (Red area—response range estimated by the surrogate; blue solid
line—deterministic response).

and UB of the response is observed. The result shows the variability of the dynamics of the dual-rotor subject to multiple interval
variables.
Case 9 considers 3% uncertainty in the clearance and 5% uncertainties in imbalances on both the HP and LP rotors. Fig. 21 shows
the nonlinear response range under such conditions. Deviations of the deterministic response in the whole speed range are observed in
Fig. 21 although only small uncertain coefficients are included for each uncertainty. The characteristics of individual uncertain pa­
rameters are all integrated into the system response. As the nature of uncertainty, the actual dynamic response of the dual-rotor may
take any possible value in the estimated range. To further reveal the performance of the proposed method, the sample responses and
the predicted bounds are illustrated in Fig. 22. It can be seen from Fig. 22 that the bounds (dashed lines) enclose perfectly all the sample
response curves. Intersections of different sample response curves are common as shown in Fig. 22, which are caused by the nonlinear
nature. The excellent wrapping of ranges proves the robustness of the uncertainty propagation model even in those points where
irregular bounds are developed. It should be noted that it will be very time-consuming to conduct a numerical simulation with too
many samples due to the nonlinear nature of the deterministic problem. Therefore, it can be computationally prohibitive for the
traditional sampling approaches such as the ISM to calculate the uncertain response ranges under multiple interval variables
rigorously.

23
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

Fig. 20. Vibration response with 10% uncertainties in the contact stiffness, damping in bearing #3 and the imbalance of the HP rotor (Case 8).

Fig. 21. Vibration response with 3% in the clearance and 5% uncertainties in the imbalances of disc 1 and disc 2 (Case 9).

7. Conclusions

The uncertain steady-state dynamical behaviours of a typical dual-rotor system with rub-impact fault and interval parameters in
both the physical model and fault event are investigated based on the non-linear analysis and non-intrusive uncertainty propagation.
The proposed polar angle interpolation is particularly dedicated to the nonlinear frequency response estimation in multi-solution
regions. This transformation has satisfactory accuracy and paves the road for non-intrusive uncertainty analysis. The non-intrusive
surrogate model based on the Chebyshev polynomials has proved to be effective in predicting the interval steady-state vibration
amplitudes of the system. Its efficiency and accuracy have been validated by the conventional large volume sampling-based method.
Parameters like friction and contact stiffness are extremely difficult to quantify in real applications, so the sensitivity of the nonlinear
forced responses to them is of great importance to the industry. It is useful to investigate the effects of different uncertain parameters
because they may have different influence patterns with respect to speed range and deviation amplitude. Moreover, they may
accumulate to produce overall impacts. In the numerical simulations, various uncertain cases are included and discussed in detail.
Based on the results, some conclusions are summarized as follows:

• The uncertainties in imbalances of the higher and lower pressure rotors affect the nonlinear vibrations around the first and second
resonances respectively.

24
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

Fig. 22. Enclosure of the response bounds in the Cartesian coordinate (Case 9) (Solid lines—discrete response samples; dashed lines—interval
bounds from the surrogate).

• The speed ratio is of primary importance for engineers when designing a dual-rotor system. Uncertainty in this parameter has great
impacts on the steady-state vibrations, i.e., a small uncertainty will lead to more impact than other uncertainties.
• Fluctuations in damping and imbalance can lead to proportional deviations along the principal curvature direction of the response,
which indicates variations in the natural characteristics.
• Fault-related uncertainties have effects only in the rub-impact regions. The order of influence from high to low is the uncertain
clearance, contact stiffness and friction coefficient. However, the variability pattern of dynamic responses is different for the three
fault uncertainties.
• In some cases, the frequency shifts and high-amplitude speed bands in the nonlinear response appear, in which the upper bound
closely wraps the deterministic response curve.
• Propagation of multi-source uncertainties brings global deviations to the vibration response and critical changes may happen in the
dynamics of dual-rotor systems.

It should be noted that the results are obtained on a particular rotor–stator system. However, the proposed PAI as well as the UQ
framework are both non-intrusive approaches, which can be generalised to other engineering rotor set-ups without any difficulty. The
analysis methods developed in the current study are generic and can be adapted to a wide range of nonlinear vibration problems under
uncertainty, especially those with multi-solution regions. The dynamic characteristics obtained provide references for the design and
dynamic investigations of dual-rotor systems subject to typical faults and parametric variabilities. In the future, research focusing on
the effects of uncertainties on the bifurcations and stability of solutions will be carried out.

Declaration of Competing Interest

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

Acknowledgements

This work was supported by the Fundamental Research Funds for the Central Universities (No. G2021KY0601), the National
Natural Science Foundation of China (No. 11972295, 12072263, 11802235), Doctoral Research Foundation from the Department of
Science & Technology of Liaoning Province (No. 2019BS182) and Natural Science Foundation of Shaanxi Province (No. 2020JQ-129).

Appendix. A

The matrices and vectors in the equation of motion denoted by Eq. (14) can be expressed as
[ ] [ ] [ ] [ ]
C1 0 ̃ 0 0 K1 0
C= G1 = G1 0 G2 = ̃ K= (A1)
0 C2 0 0 0 G2 0 K2

M= diag(ml , ml , Jdl , Jdl , mh , mh , Jdh , Jdh ) (A2)

25
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

⎡ ⎤
⎢ C1 + C3 0 0 C3 ̃
L − C1 L1 ⎥
⎢ 0 C1 + C3 ̃
C1 L1 − C3 L 0 ⎥
⎢ ⎥
C1 = ⎢
⎢ 2 ⎥
⎥ (A3)
⎢ 0 ̃ 2 ̃
C1 L1 − C3 L C1 L1 + C3 L 0 ⎥
⎣ 2
⎦
C3 ̃
L − C1 L1 0 0 ̃
C1 L21 + C3 L

⎡ ⎤
C2 0 0 − C2 L3
⎢ ⎥
⎢ 0 C2 C2 L3 0 ⎥
⎢ ⎥
C2 = ⎢ ⎥ (A4)
⎢ 0 C2 L 3 C2 L23 0 ⎥
⎣ ⎦
− C2 L3 0 0 C2 L23

⎡ ⎤
0 0 0 0
⎢ 0 ⎥
̃1 = ⎢ 0
G
0 0 ⎥ (A5)
⎣0 0 0 Jpl ⎦
0 0 − Jpl 0
⎡ ⎤
0 0 0 0
⎢0 0 0 0 ⎥
̃ ⎢
G2 = ⎣ ⎥ (A6)
0 0 0 Jph ⎦
0 0 − Jph 0
⎡ ⎤
⎢ K1 + K3 0 0 K3 ̃
L − K1 L1 ⎥
⎢ 0 K1 + K3 K1 L1 − K3 ̃
L 0 ⎥
⎢ ⎥
K1 = ⎢
⎢ 2 ⎥
⎥ (A7)
⎢ 0 K1 L1 − K3 ̃
L 2 ̃
K1 L + K3 L
1 0 ⎥
⎣ 2
⎦
K3 ̃
L − K1 L1 0 0 ̃
K1 L21 + K3 L

⎡ ⎤
K2 0 0 − K2 L3
⎢ ⎥
⎢ 0 K2 K2 L3 0 ⎥
⎢ ⎥
K2 = ⎢ ⎥ (A8)
⎢ 0 K2 L3 K2 L23 0 ⎥
⎣ ⎦
− K2 L3 0 0 K2 L23

Qs = [0, − ml g, 0, 0, 0, − mh g, 0, 0]T (A9)

(A10)
T
Qu = [ml e1 ω21 cos(ω1 t), ml e1 ω21 sin(ω1 t), 0, 0, mh e2 ω22 cos(ω2 t), mh e2 ω22 sin(ω2 t), 0, 0]

Qnl = [0, 0, 0, 0, Fx , Fy , 0, 0]T (A11)

where
̃
L = L2 + L3 + L4 + L5 (A12)

Appendix. B

Expressions of quantities in Eq. (20) are given by

He = [h0 , h1 , 0, …, h2 , …, 0]T (B1)

̃
H(X) = [c0 , c1 , d1 , ⋯, cNc , dNc ]T (B2)
⎡ ⎤
̃
K 0 0 ⋯ 0
⎢0 Λ1 0 ⋯ 0 ⎥
⎢ ⎥
H=⎢
⎢0 0 Λ2 ⋯ 0 ⎥ ⎥ (B3)
⎣⋮ ⋮ ⋮ ⋱ ⋮ ⎦
0 0 0 ⋯ Λ Nc
where

26
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

h0 = [0, − ml g, 0, 0, 0, − mh g, 0, 0]T (B4)

(B5)
T
h1 = [ml e1 ω21 , 0, 0, 0, 0, 0, 0, 0, 0, ml e1 ω21 , 0, 0, 0, 0, 0, 0]

(B6)
T
h2 = [0, 0, 0, 0, mh e2 ω22 , 0, 0, 0, 0, 0, 0, 0, 0, mh e2 ω22 , 0, 0]
[ ]
̃ − ( k ⋅ ω)2 M
K ̃
( k ⋅ ω)C
Λk = , k = 1, 2, …, Nc (B7)
̃
− ( k ⋅ ω)C ̃ − ( k ⋅ ω)2 M
K

Appendix. C

The Jacobian matrix J is computed as

∂g(X)
J(X) = = H+̃
J(X) (C1)
∂X

with
⎡ ∂c ∂c0 ∂c0 ∂c0 ∂c0 ⎤
0
⋯
⎢ ∂a0 ∂a1 ∂b1 ∂aNc ∂bNc ⎥
⎢ ⎥
⎢
⎢ ∂c1 ∂c1 ∂c1 ∂c1 ∂c1 ⎥
⎥
⎢ ⋯ ⎥
⎢ ∂a0 ∂a1 ∂b1 ∂aNc ∂bNc ⎥
⎢ ⎥
⎢ ⎥
̃ ⎢ ∂d1 ∂d1 ∂d1 ∂d1 ∂d1 ⎥
∂H(X) ⎢ ⋯ ⎥
̃
J(X) = =⎢
⎢ ∂a0 ∂a1 ∂b1 ∂aNc ∂bNc ⎥
⎥ (C2)
∂X ⎢ ⎥
⎢ ⋮ ⋮ ⋮ ⋱ ⋮ ⋮ ⎥
⎢ ⎥
⎢ ∂cN ∂cNc ∂cNc ∂cNc ∂cNc ⎥
⎢ c
⋯ ⎥
⎢ ∂a ∂a1 ∂b1 ∂aNc ∂bNc ⎥
⎢ 0 ⎥
⎢ ⎥
⎣ ∂d ∂dNc ∂dNc ∂dNc ∂d ⎦
Nc Nc
⋯
∂a0 ∂a1 ∂b1 ∂aNc ∂bNc
The partial derivatives in Eq. (C2) cannot be directly solved because the force vector Qnl (t) related to the rub-impact is piecewise
nonlinear to the displacement vector U(t) and the explicit relationships for terms ci , di and ai , bi are unavailable in the frequency
domain. To completely determine the Jacobian matrix, the AFT technique [59,60] can be employed to transform the nonlinear forces
and the displacements from the frequency domain to the time domain, in which their explicit expressions are already defined. The
solution process can be summarised into three main steps. First, convert the current displacements from the frequency domain back to
the time domain via the inverse discrete Fourier transform (IDFT). Then, calculate the corresponding nonlinear forces by using Eq. (2).
Last, transfer the calculated nonlinear forces from the time domain to the frequency domain based on the discrete Fourier transform
(DFT). The above alternating between the two domains enables the computation of the partial derivatives in Eq. (C2).
To implement the AFT technique, the dynamical response of the dual-rotor system at the (n1 ,n2 )th discrete time is calculated based
on the IDFT for the defined harmonic components
Nh
∑ Nh (
∑ )
2k1 π 2k2 π 2k1 π 2k2 π
U(n1 , n2 ) = ak1 k2 cos( n1 + n2 ) + bk1 k2 sin( n1 + n2 ) (C3)
k1 =− Nh k2 =− Nh
N N N N

where n1 , n2 = 0, 1, 2, …, N − 1 with N being the number of the discrete times. Then, we can easily get the (n1 ,n2 )th nonlinear rub-
impact forces Qnl (n1 , n2 ) in the time domain based on Eq. (2).
Next, the representations of the rub-impact nonlinear forces in the frequency domain are obtained using the DFT as
( )
κ ∑ N− 1 ∑N− 1
2k1 π 2k2 π
̃ck1 k2 = Q (n1 , n2 )cos n 1 + n2 (C4)
N 2 n1 =0 n2 =0 nl N N

( )
κ ∑
N− 1 ∑
N− 1
2k1 π 2k2 π
̃
dk1 k2 = 2 Qnl (n1 , n2 )sin n1 + n2 (C5)
N n1 =0 n2 =0 N N

in which κ is an integer and

27
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

{
1, if n1 = n2 = 0;
κ= (C6)
2, otherwise.

To this point, the partial derivatives in Eq. (C2) can be solved by

̃ ∂H
∂H ̃ ∂Ξ
= (C7)
∂X ∂Ξ ∂X
[ ]T
where Ξ = …, ̃ck1 k2 , ̃
dk1 k2 , … with ki = − Nh , − Nh + 1, …, − 1, 0, 1, …, Nh − 1, Nh for i = 1, 2.
Thus, the Jacobian matrix is obtained via the ATF technique and the iterative algorithm in Eq. (22) can be carried out. The criterion
of non-negative combinations of the frequencies will apply throughout the investigation, although the formulation here is written in
complete form.

References

[1] T. Gao, S. Cao, Paroxysmal impulse vibration phenomena and mechanism of a dual–rotor system with an outer raceway defect of the inter-shaft bearing, Mech.
Syst. Sig. Process. 157 (2021), 107730.
[2] L. Hou, Y. Chen, Y. Fu, H. Chen, Z. Lu, Z. Liu, Application of the HB–AFT method to the primary resonance analysis of a dual-rotor system, Nonlinear Dyn. 88
(2017) 2531–2551.
[3] Y.C. Shan, X.D. Liu, T. He, Q.H. Li, Research on the finite element impact-contact analytical model of dual-rotor system and its diagnosis method, Journal of
Aerospace, Power 20 (2005) 789–794.
[4] J. Ma, H. Zhang, S. Lou, F. Chu, Z. Shi, F. Gu, A.D. Ball, Analytical and experimental investigation of vibration characteristics induced by tribofilm-asperity
interactions in hydrodynamic journal bearings, Mech. Syst. Sig. Process. 150 (2021), 107227.
[5] G. Ferraris, V. Maisonneuve, M. Lalanne, Prediction of the dynamic behavior of non-symmetrical coaxial co- or counter rotating rotors, J. Sound Vib. 195 (1996)
649–666.
[6] K. Gupta, K.D. Gupta, K. Athre, Unbalance response of a dual rotor system: theory and experiment, J. Vib. Acoust. 115 (1993) 427–435.
[7] N. Wang, C. Liu, D. Jiang, K. Behdinan, Casing vibration response prediction of dual-rotor-blade-casing system with blade-casing rubbing, Mech. Syst. Sig.
Process. 118 (2019) 61–77.
[8] G. Luo, H. Zhou, F. Wang, X. Yang, Dynamic response of co-rotating and counter-rotating dual-rotor system supported on ball bearing, Journal of Aerospace,
Power 8 (2012) 1887–1894.
[9] L. Chen, Z. Qin, F. Chu, Dynamic characteristics of rub-impact on rotor system with cylindrical shell, Int. J. Mech. Sci. 133 (2017) 51–64.
[10] A. Muszynska, P. Goldman, Chaotic responses of unbalanced rotor/bearing/stator systems with looseness or rubs, Chaos, Solitons Fractals 5 (9) (1995)
1683–1704.
[11] J. Padovan, F.K. Choy, Nonlinear dynamics of rotor/blade/casing rub interactions, J. Turbomach. 109 (1987) 527–534.
[12] F. Chu, W. Lu, Experimental observation of nonlinear vibrations in a rub-impact rotor system, J. Sound Vib. 283 (2005) 621–643.
[13] A. Alzibdeh, M. Alqaradawi, B. Balachandran, Effects of high frequency drive speed modulation on rotor with continuous stator contact, Int. J. Mech. Sci.
131–132 (2017) 559–571.
[14] T.H. Patel, A.K. Darpe, Vibration response of a cracked rotor in presence of rotor–stator rub, J. Sound Vib. 317 (2008) 841–865.
[15] 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.
[16] H. Ma, C. Shi, Q. Han, B. Wen, Fixed-point rubbing fault characteristic analysis of a rotor system based on contact theory, Mech. Syst. Sig. Process. 38 (2013)
137–153.
[17] G. Von Groll, D.J. Ewins, The harmonic balance method with arc-length continuation in rotor/stator contact problems, J. Sound Vib. 241 (2001) 223–233.
[18] L. Xiang, A. Hu, L. Hou, Y. Xiong, J. Xing, Nonlinear coupled dynamics of an asymmetric double-disc rotor-bearing system under rub-impact and oil-film forces,
Appl. Math. Model. 40 (2016) 4505–4523.
[19] C. Wang, D. Zhang, Y. Ma, Z. Liang, J. Hong, Theoretical and experimental investigation on the sudden unbalance and rub-impact in rotor system caused by
blade off, Mech. Syst. Sig. Process. 76 (2016) 111–135.
[20] C. Sun, Y. Chen, L. Hou, Steady-state response characteristics of a dual-rotor system induced by rub-impact, Nonlinear Dyn. 86 (2016) 91–105.
[21] H. Zhou, G. Chen, Dynamic response analysis of dual rotor-ball bearing-stator coupling system for aero-engine, Journal of Aerospace, Power 24 (2009)
1284–1291.
[22] W. Gao, D.i. Wu, K. Gao, X. Chen, F. Tin-Loi, Structural reliability analysis with imprecise random and interval fields, Appl. Math. Model. 55 (2018) 49–67.
[23] E. Jacquelin, S. Adhikari, J.-J. Sinou, M.I. Friswell, Polynomial chaos expansion in structural dynamics: Accelerating the convergence of the first two statistical
moment sequences, J. Sound Vib. 356 (2015) 144–154.
[24] L. Wang, C. Xiong, X. Wang, M. Xu, Y. Li, A dimension-wise method and its improvement for multidisciplinary interval uncertainty analysis, Appl. Math. Model.
59 (2018) 680–695.
[25] K. Guo, J. Jiang, Z. Li, Diffusion and persistence of rotor/stator synchronous full annular rub response under weak random perturbations, J. Vib. Eng. Technol. 8
(4) (2020) 599–611.
[26] Z. Zheng, Y. Xie, D. Zhang, Numerical investigation on the gravity response of a two-pole generator rotor system with interval uncertainties, Applied Sciences 9
(2019) 3036.
[27] J. Ma, C. Fu, H. Zhang, F. Chu, Z. Shi, F. Gu, A.D. Ball, Modelling non-Gaussian surfaces and misalignment for condition monitoring of journal bearings,
Measurement 174 (2021), 108983.
[28] Z. Lv, Z. Qiu, An iteration method for predicting static response of nonlinear structural systems with non-deterministic parameters, Appl. Math. Model. 68
(2019) 48–65.
[29] C. Fu, X. Ren, Y. Yang, Y. Xia, W. Deng, An interval precise integration method for transient unbalance response analysis of rotor system with uncertainty, Mech.
Syst. Sig. Process. 107 (2018) 137–148.
[30] J.-J. Sinou, L. Nechak, S. Besset, Kriging Metamodeling in Rotordynamics: Application for Predicting Critical Speeds and Vibrations of a Flexible Rotor,
Complexity 2018 (2018) 1–26.
[31] C. Fu, Y. Xu, Y. Yang, K. Lu, F. Gu, A. Ball, Dynamics analysis of a hollow-shaft rotor system with an open crack under model uncertainties, Commun. Nonlinear
Sci. Numer. Simul. 83 (2020), 105102.
[32] E.H. Koroishi, A.A. Cavalini Jr, A.M. de Lima, V. Steffen Jr, Stochastic modeling of flexible rotors, J. Braz. Soc. Mech. Sci. Eng. 34 (2012) 574–583.
[33] T.G. Ritto, R.H. Lopez, R. Sampaio, J.E.S.D. Cursi, Robust optimization of a flexible rotor-bearing system using the Campbell diagram, Eng. Optim. 43 (2011)
77–96.

28
C. Fu et al. Mechanical Systems and Signal Processing 170 (2022) 108827

[34] Y. Ma, Y. Wang, C. Wang, J. Hong, Nonlinear interval analysis of rotor response with joints under uncertainties, Chin. J. Aeronaut. 33 (2020) 205–218.
[35] J. Didier, J.J. Sinou, B. Faverjon, Study of the non-linear dynamic response of a rotor system with faults and uncertainties, J. Sound Vib. 331 (2012) 671–703.
[36] C. Fu, Y. Xu, Y. Yang, K. Lu, F. Gu, A. Ball, Response analysis of an accelerating unbalanced rotating system with both random and interval variables, J. Sound
Vib. 466 (2020), 115047.
[37] J.J. Sinou, J. Didier, B. Faverjon, Stochastic non-linear response of a flexible rotor with local non-linearities, Int. J. Non Linear Mech. 74 (2015) 92–99.
[38] C. Li, B. Chen, H. Peng, Z. Sheng, Sparse regression Chebyshev polynomial interval method for nonlinear dynamic systems under uncertainty, Appl. Math.
Model. 51 (2017) 505–525.
[39] X.Y. Long, C. Jiang, X. Han, J.C. Tang, F.J. Guan, An enhanced subinterval analysis method for uncertain structural problems, Appl. Math. Model. 54 (2017)
580–593.
[40] C. Fu, D. Zhen, Y. Yang, F. Gu, A. Ball, Effects of bounded uncertainties on the dynamic characteristics of an overhung rotor system with rubbing fault, Energies
12 (2019) 4365.
[41] J. Wu, Y. Zhang, L. Chen, L. Zhen, A Chebyshev interval method for nonlinear dynamic systems under uncertainty, Appl. Math. Model. 37 (2013) 4578–4591.
[42] J. Wu, Z. Luo, H. Li, N. Zhang, A new hybrid uncertainty optimization method for structures using orthogonal series expansion, Appl. Math. Model. 45 (2017)
474–490.
[43] C. Fu, G. Feng, J. Ma, K. Lu, Y. Yang, F. Gu, Predicting the dynamic response of dual-rotor system subject to interval parametric uncertainties based on the non-
intrusive metamodel, Mathematics 8 (2020) 736.
[44] Z. Zhang, X. Ma, H. Hua, X. Liang, Nonlinear stochastic dynamics of a rub-impact rotor system with probabilistic uncertainties, Nonlinear Dyn. 102 (2020)
2229–2246.
[45] L. Yang, K. He, Y. Guo, Reliability analysis of a nonlinear rotor/stator contact system in the presence of aleatory and epistemic uncertainty, J. Mech. Sci.
Technol. 32 (2018) 4089–4101.
[46] Y. Zhang, B. Wen, A.Y. Leung, Reliability analysis for rotor rubbing, J. Vib. Acoust. 124 (2002) 58–62.
[47] J. Didier, J.J. Sinou, B. Faverjon, Multi-dimensional harmonic balance with uncertainties applied to rotor dynamics, J. Vib. Acoust. 134 (2012), 061003.
[48] J. Didier, J.J. Sinou, B. Faverjon, Nonlinear vibrations of a mechanical system with non-regular nonlinearities and uncertainties, Commun. Nonlinear Sci.
Numer. Simul. 18 (2013) 3250–3270.
[49] K. Worden, G. Manson, T.M. Lord, M.I. Friswell, Some observations on uncertainty propagation through a simple nonlinear system, J. Sound Vib. 288 (2005)
601–621.
[50] M. Guskov, J.J. Sinou, F. Thouverez, O.S. Naraikin, Experimental and numerical investigations of a dual-shaft test rig with intershaft bearing, Int. J. Rotating
Mach. 2007 (2007), 075762.
[51] M. Guskov, J.J. Sinou, F. Thouverez, Multi-dimensional harmonic balance applied to rotor dynamics, Mech. Res. Commun. 35 (2008) 537–545.
[52] N. Coudeyras, S. Nacivet, J.J. Sinou, Periodic and quasi-periodic solutions for multi-instabilities involved in brake squeal, J. Sound Vib. 328 (2009) 520–540.
[53] M.A. Crisfield, An arc–length method including line search and acceleration, Int. J. Numer. Meth. Eng. 19 (2010) 1269–1289.
[54] E. Sarrouy, E. Pagnacco, E.S. de Cursi, A constant phase approach for the frequency response of stochastic linear oscillators, Mechanics and Industry 17 (2016)
206.
[55] A. Panunzio, L. Salles, C. Schwingshackl, Uncertainty propagation for nonlinear vibrations: A non-intrusive approach, J. Sound Vib. 389 (2017) 309–325.
[56] J. Wu, Z. Luo, N. Zhang, Y. Zhang, A new uncertain analysis method and its application in vehicle dynamics, Mech. Syst. Sig. Process. 50–51 (2015) 659–675.
[57] S.S. Isukapalli, A. Roy, P.G. Georgopoulos, Stochastic response surface methods (SRSMs) for uncertainty propagation: application to environmental and
biological systems, Risk Anal. 18 (2010) 351–363.
[58] C. Fu, X. Ren, Y. Yang, W. Qin, Dynamic response analysis of an overhung rotor with interval uncertainties, Nonlinear Dyn. 89 (2017) 2115–2124.
[59] T.M. Cameron, J.H. Griffin, An alternating frequency/time domain method for calculating the steady–state response of nonlinear dynamic systems, J. Appl.
Mech. 56 (1989) 149–154.
[60] Y.B. Kim, S.T. Noah, Quasi-periodic resposne and stability analysis for a non-linear Jeffcott rotor, J. Sound Vib. 190 (1996) 239–253.

29