Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

Structural Engineering and Mechanics, Vol. 81, No.

4 (2022) 411-428
https://doi.org/10.12989/sem.2022.81.4.411 411

Model order reduction for Campbell diagram analysis


of shaft-disc-blade system in 3D finite elements
Ty Phuor and GilHo Yoona

School of Mechanical Engineering, Hanyang University, Seoul 04763, Republic of Korea

(Received August 8, 2021, Revised October 20, 2021, Accepted November 7, 2021)

Abstract. This paper presents the Campbell diagram analysis of the rotordynamic system using the full order model (FOM)
and the reduced order model (ROM) techniques to determine the critical speeds, identify the stability and reduce the
computational time. Due to the spin-speed-dependent matrices (e.g., centrifugal stiffening matrix), several model order reduction
(MOR) techniques may be considered, such as the modal superposition (MS) method and the Krylov subspace-based MOR
techniques (e.g., Ritz vector (RV), quasi-static Ritz vector (QSRV), multifrequency quasi-static Ritz vector (MQSRV),
multifrequency/ multi-spin-speed quasi-static Ritz vector (MMQSRV) and the combined Ritz vector & modal superposition
(RV+MS) methods). The proposed MMQSRV method in this study is extended from the MQSRV method by incorporating the
rotational-speed-dependent stiffness matrices into the Krylov subspace during the MOR process. Thus, the objective of this note
is to respond to the question of whether to use the MS method or the Krylov subspace-based MOR technique in establishing the
Campbell diagram of the shaft-disc-blade assembly systems in three-dimensional (3D) finite element analysis (FEA). The
Campbell diagrams produced by the FOM and various MOR methods are presented and discussed thoroughly by computing the
norm of relative errors (ER). It is found that the RV and the MS methods are dominant at low and high rotating speeds,
respectively. More precisely, as the spinning velocity becomes large, the calculated ER produced by the RV method is
significantly increased; in contrast, the ER produced by the MS method is smaller and more consistent. From a computational
point of view, the MORs have substantially reduced the time computing considerably compared to the FOM. Additionally, the
verification of the 3D FE rotordynamic model is also provided and found to be in close agreement with the existing solutions.
Keywords: Campbell diagram; Krylov subspace; Modal Superposition (MS); Model Order Reduction (MOR); Multi-
frequency/Multi-spin-Speed Quasi-Static Ritz Vector (MMQSRV); Ritz Vector (RV); rotordynamic

1. Introduction shaft. It was found that a good agreement was reached; but,
some discrepancies were revealed due to (1) the
The shaft-disc-blade assemblies that are the engines' key oversimplified model of the shaft flexibility and (2)
components and compressors play an essential role in difficulties in determining the stiffness parameters of the
designing the machines (Wang et al. 2019, Wang et al. system. In addition, Huang and Ho (1996) developed an
2015). The great significance of the design is to develop an analytical approach to study the dynamic coupling between
accurate dynamic model and gain deep insight into the shaft torsion and blade bending of a rotating shaft-disc-
vibration characteristics of the blades and rotors systems. blade system using the weighted residual method. This
So, these issues have been attracted an enormous number of approach was found to be useful (1) for physical
researchers and engineers over the past decades (Genta comprehension, (2) for computation of the interaction
2005, Han 2014, Li et al. 2019, Okabe et al. 1991, Wang et forces, and (3) for determining the sensitivity of the
al. 2019, Wang et al. 2015, Yuan et al. 2017). However, system’s frequencies to the variation in some parameters;
most of them have focused only on the rotordynamics even so, some drawbacks were identified in the numerical
systems by assuming the blade and disc to be rigid and root finding. Moreover, by using the energy method in
tuned or on the vibration of the blade and disc by neglecting conjunction with the assumed modes method, She et al.
the effect of the shaft flexibility (Wang et al. 2019). For (2018) established the continuum model of flexible shaft-
example, Crawely et al. (1986) conducted the experimental disc-blade coupling system to investigate the coupling
investigation and developed a simplified model to vibration among shaft-bending, shaft-torsion, disc-
determine the natural frequencies of the flexible blade and transverse, and blade-bending. The authors found that the
rigid disc coupling supported by a flexible cantilevered critical speed of the system could be significantly affected
by the shaft and disc flexibility. Nevertheless, the
weaknesses in their simulations were reported due to the
Corresponding author, Postdoctoral Researcher model simplification (Wang et al. 2019).
E-mail: ty168.utp@gmail.com To overcome the above limitation, the 3D solid FE
a
Professor model is utilized because (1) the detailed geometry can be
E-mail: ghy@hanyang.ac.kr accurately captured, and (2) the flexibility of all
components may be taken into account due to its elasticity
Copyright © 2022 Techno-Press, Ltd.
http://www.techno-press.com/journals/sem&subpage=7 ISSN: 1225-4568 (Print), 1598-6217 (Online)
412 Ty Phuor and GilHo Yoon

Table 1 The comparison of the capabilities of the finite element method to model a rotating shaft-disk system (Wagner et
al. 2010)
Prismatic solid An abrupt change
Tapered
or hollow section in cross-section
Model Rotary Gyroscopic Shear Solid two Hollow
Two
(Authors, year) inertia effect deformation axisymmetric principal axisymmetric Non-
Axisymmetric principal Axisymmetric
axes axisymmetric
axes
Rayleigh shaft element
  ×   × × × × ×
(Nelson and McVaugh 1976)
Conical shaft element based
on Timoshenko beam     ×  ×  × ×
(Genta and Gugliotta 1988)
Modulus corrected elements
         ×
(Vest and Darlow 1990)
Axisymmetric finite element
    ×  ×   ×
(Stephenson and Rouch 1993)
Three-dimensional solid
finite element          
(Nandi and Neogy 2001)

theory (Wang et al. 2019, Wang et al. 2015). More implementation of one zero-frequency and one zero-spin-
precisely, Table 1 provides the comparisons of the speed in the Krylov subspace for the MOR process. Wagner
capabilities of FEM proposed to model a rotating shaft-disk et al. (2010) conducted a survey and review study on the
system. Obviously, the characteristic behaviors of the 3D model order reduction (MOR) methods for the
solid FE model are dominant. Moreover, such elements rotordynamic system and reported no single most efficient
were also widely used in modeling and successfully applied MOR method.
for the blade-disc analysis (Castanier and Pierre 2002, In contrast to the Krylov subspace-based model order
Martel and Sánchez-Álvarez 2018, Petrov et al. 2013, reduction, Khulief and Mohiuddin (1997) and Wagner et al.
Sternchüss 2009) and the complex rotor systems (Chatelet (2010) provided two different methods of the modal
et al. 2005, Lazarus et al. 2010, Wang et al. 2015, Wang et reduction (i.e., planar and complex modal transformation)
al. 2015). However, the computational efficiency in the for reducing the size of the self-adjoint and non-self-adjoint
repeated dynamic analysis of the shaft-disc-blade system rotordynamic systems, respectively. The authors stated that
has encountered difficulties because of the vast number of both methods provided almost the same level of accuracy.
degrees of freedom generated in the system. Therefore, it is Likewise, Nandi (2004) also employed the modal
almost impossible to complete the analysis to establish the superposition technique to reduce the FE equation for a
Campbell diagram for investigating the critical speed and rotor model on non-isotropic spring support in a rotating
stability. As a result, many reduced-order models (ROMs) frame and mentioned that it was an effective technique.
techniques have been developed and applied. Those are the However, the existing methods only presented the spin-
generalized substructure-based methods (Bladh 2001, Saito velocity-dependent equation in which the computational
et al. 2009, Wang et al. 2014) and some other effective cost is still high for the Campbell diagram analysis because
methods for specific structures such as the subset of the projection matrix is required to compute for each
nominal modes (SNM) (Yang and Griffin 2001), the applied rotational speed. Consequently, to the best of the
fundamental mistuning model (FMM) (Feiner and Griffin authors’ knowledge, no effort has been made to use the
2002), the component mode mistuning method (CMM) modal superposition (MS) method compared with the
(Lim et al. 2007) and the Krylov subspace-based model Krylov subspace-based model order reduction for the
(Han 2014). The component mode synthesis method was a Campbell diagram analysis in the 3D FEA.
more general approach for the reduced-order model, so this Therefore, this paper presents the Campbell diagram
method was not so efficient to analyze the shaft-disc-blade analysis of rotordynamic systems in the 3D FEA by using
system (Wang et al. 2019). To study such a system, SNM, the full order model (FOM) and the model order reduction
FMM, and CMM were more efficient methods; however, (MOR) techniques for determining the critical speeds,
the shafts were commonly rigid; consequently, its flexibility identifying the stability and reducing the computational
tended to be overlooked, which is also not recommended in time. Several MOR methods are taken into account, such as
rotordynamic analysis (Wang et al. 2019). the MS method (Khulief and Mohiuddin, 1997) and the
Alternatively, Han (2014) successfully applied the Krylov subspace-based MOR technique (e.g., Ritz vector
Krylov subspace-based Arnoldi process via moment- (RV) (Gu et al. 2000), quasi-static Ritz vector (QSRV) (Gu
matching to reduce the model size of the rotordynamic et al. 2000), multifrequency quasi-static Ritz vector
system in FEA using the combined Ansys and Matlab. The (MQSRV) (Gu et al. 2000, Yoon 2012, Yoon et al.
author stated that the method used based on the Krylov 2015), multifrequency/multi-spin-speed quasi-static Ritz
subspace effectively sped up the Campbell diagram analysis vector (MMQSRV), and combined Ritz vector and
because the projection matrix calculation was independent modal superposition (RV+MS) methods). The MQSRV
of rotation speed. Consequently, it is noticed that the used method was proposed based on the QSRV method and
method fell into the RV method owing to the found to be best for the solution of the frequency-dependent
Model order reduction for Campbell diagram analysis of shaft-disc-blade system in 3D finite elements 413

Table 2 The comparison of the use of the model order reduction methods for the Campbell diagram analysis of
rotordynamic system in 3D FEA
QSRV MQSRV Proposed Combined (RV+MS)
Authors, year FOM RV method MS method
method method MMQSRV method method
Han (2014)   - - - - -
Present analysis       

(x’,y’+v’,z’+w’)
u’
w’

v’ w
z
z’
y” Q(x’,y’,z’) v
Ωt
y’ y’
z’
z” wt
(X,Y,Z)
y

x,x’

Fig. 1 The displaced position of a point on the shaft cross-section fixed in a rotating reference frame

acoustic system (Yoon 2012). Consequently, the trial-and- subsections based on the works of Nandi and Neogy (2001),
error MMQSRV method in this study is extended from the Vollan and Komzsik (2012), Kirchgäßner (2016).
MQSRV method by incorporating the rotational-speed-
dependent stiffness matrices into the Krylov subspace 2.1 Acceleration of a point on the shaft in a co-
during the MOR (model order reduction) process. Thus, this rotating reference system
note aims to respond to the question of whether to use the
MS method or the Krylov subspace-based MOR technique Fig. 1 illustrates the displaced position of a point
for Campbell diagram analysis of the rotating shaft-disc- Q(x',y',z') on the rotating structure fixed in a rotating
blade systems in the 3D FEA. More precisely, Table 2 coordinate system (x',y',z') in which x' coincided with x. The
illustrates the comparison of the MOR methods for the particle P passes through the displaced position of this point
Campbell diagram analysis conducted by the previous at time t. The displacement components of the cross-section
studies and the present study. Therefore, this study provides along the y' and z' directions are respectively v' and w' as
an in-depth insight into the knowledge of the model order seen in Eq. (1).
reduction techniques by using the Krylov subspace and
𝑦′ = 𝑌 ⋅ cos[(Ω − 𝜔)𝑡] − 𝑍 ⋅ sin[(Ω − 𝜔)𝑡]
modal superposition. The numerical examples for (1)
𝑧′ = 𝑌 ⋅ sin[(Ω − 𝜔)𝑡] + 𝑍 ⋅ cos[(Ω − 𝜔)𝑡]
producing the Campbell diagram are presented to
demonstrate the validity of the MOR methods. Where, Ω and ω are the shaft spinning speed and speed
of the rotating frame about the centreline of the bearings,
respectively. So, the displacements of the particle P at time t
2. Rotordynamic in finite element method with the speed of the rotating frame ω are

The equation of motion of the rotating structures may be 𝑢𝑝 (𝑋, 𝑌, 𝑍, 𝑡) = 𝑢′(𝑥′, 𝑦′, 𝑧′, 𝑡) (2)
derived referring to either rotating or inertial frames.
𝑣𝑃 cos(𝜔 ⋅ 𝑡) − sin(𝜔 ⋅ 𝑡) 𝑣′(𝑥′, 𝑡)
However, in this study, a co-rotating reference system is {𝑤 } = [ ]{ }
contemplated due to the use of the 3D solid finite element 𝑃 sin(𝜔 ⋅ 𝑡) cos(𝜔 ⋅ 𝑡) 𝑤′(𝑥′, 𝑡)
cos(𝜔 ⋅ 𝑡) − sin(𝜔 ⋅ 𝑡) 𝑌 (3)
model (e.g., 20-node hexahedral element). Therefore, the 𝑌
+[ ]{ } − { }
equation of motion is briefly represented in the following sin(𝜔 ⋅ 𝑡) cos(𝜔 ⋅ 𝑡) 𝑍 𝑍
414 Ty Phuor and GilHo Yoon

By using the operator in Eq. (4), the acceleration of a 𝜕2 𝑢′


𝜕𝑡 2
0
point can be derived as in Eq. (5). 𝜕𝑤′
𝜕2 𝑣′
𝑑 𝜕 𝜕 𝜕 + {−2Ω } +
𝜕𝑡
= − (Ω − 𝜔) + (Ω − 𝜔) (4) 𝑓𝑥′ 𝜕𝑡 2 𝜕𝑣′
𝑑𝑡 𝜕𝑡 𝜕𝑦′ 𝜕𝑥′ 𝜕2 𝑤′ 2Ω
{𝑓𝑦′ } = 𝜌 ⋅ Δ𝑉 { 2 } 𝜕𝑡 (8)
𝜕𝑡
𝑎𝑥′ 𝑓𝑧′
0 Ω2 𝑢′ −Ω2 𝑢′
{𝑎𝑦′ } = 2
𝑎𝑧′ { −Ω 𝑣′ } + { 0 } + {−Ω2 𝑦′}
2
𝑑 2 𝑢𝑃 𝜕2 𝑢′ 𝜕2 𝑢′ 𝜕2 𝑢′ ( −Ω 𝑤′ 0 −Ω2 𝑧′ )
= + 2(Ω − 𝜔) (𝑦′ − 𝑧′ )−
𝑑𝑡 2 𝜕𝑡 2 𝜕𝑧′𝜕𝑡 𝜕𝑦′𝜕𝑡
𝜕𝑢′ 𝜕𝑢′ 𝜕𝛱 𝜕(𝑈 + 𝑊)
(Ω − 𝜔)2 (𝑦′ + 𝑧′ )+ = =0 (9)
𝜕𝑦′ 𝜕𝑧′ (5) 𝜕𝑟 𝜕𝑟
𝜕2 𝑢′ 𝜕2 𝑢′ 𝜕2 𝑢′
(Ω − 𝜔)2 × (𝑦′2 −2 + 𝑧′2 ) Where U and W are the total strain energy and the total
𝜕𝑦′2 𝜕𝑦′𝜕𝑧′ 𝜕𝑧′2
𝑑 2 𝑣𝑃 𝜕2 𝑣′ 𝜕𝑤′
external work done as seen in Eqs. (10) and (11),
= − 2𝜔 − 𝜔2 𝑣′ − Ω2 𝑦′ respectively. r is the nodal displacement vector.
𝑑𝑡 2 𝜕𝑡 2 𝜕𝑡
𝑑 2 𝑤𝑃 𝜕2 𝑤′ 𝜕𝑣′
= + 2𝜔 − 𝜔2 𝑤′ − Ω2 𝑧′ 𝜕 2 𝑢′
{ 𝑑𝑡 2 𝜕𝑡 2 𝜕𝑡 } 0
𝜕𝑡 2 𝜕𝑤′
Where the quantity ∂⁄∂t is the time rate of change in that 𝜕 2 𝑣′ −2Ω
quantity which is observed from a fixed point in the rotating + 𝜕𝑡
1 𝑢′ 𝑇 𝜕𝑡 2 𝜕𝑣′
coordinate system. From Eq. (5), it is seen that, for a 𝑊 = ∭ { 𝑣′ } ⋅ 𝜌 ⋅ 𝑑𝑉𝑒 (10)
𝜕 2 𝑤′ {

general case, the finite element formulation becomes 2 𝑉𝑒 𝜕𝑡 }
𝑤′ { 𝜕𝑡 2 }
cumbersome. Consequently, the simpler case can be done
by setting Ω=ω. So, Eq. (5) becomes 0 Ω2 𝑢′
2
+ { −Ω 𝑣′ } + { 0 }
𝜕2 𝑢′ ( −Ω2 𝑤′ )
0
𝑎𝑥′ 𝜕𝑡 2
𝜕2 𝑣′ 𝜕𝑤′ 1 1
{𝑎𝑦′ } = − 2Ω − Ω2 𝑣′ − Ω2 𝑦′ (6)
𝑈 = ∭ 𝑟 𝑇 [𝐵]𝑇 ⋅ [𝐷] ⋅ [𝐵] ⋅ 𝑑𝑉𝑒 ⋅ 𝑟 = 𝑟 𝑇 ⋅ [𝐾𝑒 ] ⋅ 𝑟 (11)
𝜕𝑡 2 𝜕𝑡
𝑎𝑧′ 𝜕2 𝑤′ 𝜕𝑣′ 2 𝑉𝑒 2
2 2
{ 𝜕𝑡 2 + 2Ω 𝜕𝑡 − Ω 𝑤′ − Ω 𝑧′}
From Galerkin's method (Genta 2005), the
displacements at a point fixed in the rotating frame can be
2.2 Equation of motion of the rotating structures expressed in Eq. (12).
𝑢′
From Eq. (6), the total forces acting on the particle may 𝑣′
{ 𝑣′ } = [𝑁] ⋅ 𝑟, { } = [𝑁1 ] ⋅ 𝑟, {𝑢′} = [𝑁2 ] ⋅ 𝑟 (12)
be obtained by using Newton’s second law, which can be 𝑤′
𝑤′
written as in Eq. (7).
Where [N], [N1 ] and [N2 ] are the shape functions.
𝜕 2 𝑢′ Thus, Eq. (9) becomes
0
𝜕𝑡 2 𝜕𝑤′ 𝜕2 𝑟
𝜕 2 𝑣′ −2Ω [𝑁]𝑇 [𝑁] +
𝜕𝑡 2
+ 𝜕𝑡 +
𝑓𝑥′ 𝜕𝑡 2 𝑇 0 −1 𝜕𝑟
𝜕𝑣′ [𝐾𝑒 ] ⋅ 𝑟 + ∭𝑉 2Ω[𝑁1 ] [1 0 ] [𝑁1 ] 𝜕𝑡 − 𝜌 ⋅
{𝑓𝑦′ } = 𝜌 ⋅ Δ𝑉 𝜕 2 𝑤′ {
2Ω (7) 𝑒 (13)
𝜕𝑡 } Ω2 [𝑁1 ]𝑇 [𝑁1 ] ⋅ 𝑟 +
𝑓𝑧′ { 𝜕𝑡 2 }
−Ω2 𝑢′ (Ω2 [𝑁2 ]𝑇 [𝑁2 ] ⋅ 𝑟 )
0 Ω2 𝑢′
2
{ −Ω 𝑣′ } + { 0 } + {−Ω2 𝑦′} 𝑑𝑉𝑒 = 0
2
( −Ω 𝑤′ 0 −Ω2 𝑧′ ) Moreover, if the damping and the bearing system are
considered, the general equation of motion of the rotating
For bending mode, when integrated over the cross- structures can be
section element, the forces ρΔVΩ2y' and ρΔVΩ2z' produce
zero resultant forces and zero resultant moments (Nandi and 𝑀 ∙ 𝑟̈ (𝑡) + [𝐶𝑑 + 𝐶𝑐 (Ω) + 𝐶𝑏 (Ω)] ∙ 𝑟̇ (𝑡) +
(14)
Neogy 2001). Alternatively, for the axial mode, the [𝐾𝑒 + 𝐾𝑐 (Ω) − 𝐾𝑠 (Ω) + 𝐾𝑏 (Ω)] ∙ 𝑟(𝑡) = 𝑓(𝑡)
magnitude of this resultant moment is dependent on the
displacement u'. However, this resultant moment can be Where,
neglected when the displacement u' is small. Consequently, 𝑀 = ∭𝑉 [𝑁]𝑇 [𝑁]𝜌 ⋅ 𝑑𝑉𝑒 : The mass matrix
𝑒
considering the critical speeds for the bending modes, the 𝐶𝑐 (Ω) = 2Ω ∭𝑉 [𝑁1 ]𝑇 [
0 −1 [𝑁 ]𝜌
] 1 ⋅ 𝑑𝑉𝑒 : The speed-
last term of Eq. (7) is assumed to have vanished in this 𝑒 1 0
study. dependent skew-symmetric Coriolis matrix
The equation of motion of the rotating structures can [𝐾𝑒 ] = ∭𝑉 [𝐵]𝑇 ⋅ [𝐷] ⋅ [𝐵] ⋅ 𝑑𝑉𝑒 : Elastic stiffness
𝑒
also be derived from a state of stable equilibrium in which matrix
the computation of the minimum stored potential energy for [𝐾𝑠 (Ω)] = Ω2 ∭𝑉 [𝑁1 ]𝑇 [𝑁1 ]𝜌 ⋅ 𝑑𝑉𝑒 : The speed-
all the nodal displacements is in need. 𝑒
Model order reduction for Campbell diagram analysis of shaft-disc-blade system in 3D finite elements 415

dependent spin stiffening matrix ̃ + 𝜆 ∙ 𝐶̃ + 𝐾


(𝜆2 ∙ 𝑀 ̃) ∙ 𝜙 = 0 (22)
[𝐾𝑐 (Ω)] = Ω2 ∭𝑉 [𝑁2 ]𝑇 [𝑁2 ]𝜌 ⋅ 𝑑𝑉𝑒 : The speed-
𝑒 Because of the difficulty in solving the QEP of the Eq.
dependent centrifugal stiffening matrix (22) in the displacement space, Eq. (21) is transformed into
𝐶𝑏 (Ω) and 𝐾𝑏 (Ω) are the damping and stiffness the state space equation, which is expressed
matrices induced by the bearing system. Cd and f(t)
represent the structural damping and vector of the external 𝐼 0 𝑟̇ 0 −𝐼 𝑟
[ ̃ ] {𝑟̈ } + [⏟𝐾 ]{ } = 0 (23)
forces. 𝑟̈ (𝑡), 𝑟̇(𝑡) and 𝑟(𝑡) denotes the vector of the
⏟0 𝑀 ̃ 𝐶̃ 𝑟̇
𝐸 𝐴
acceleration, velocity, and displacement responses,
𝑟
respectively Where, 𝑦 = { } is the 2N-dimensional state vector
𝑟̇
owing to the inclusion of the displacement and velocity
2.3 Establishment of the Campbell diagram for the full vectors. As a result, Eq. (21) can be transformed into the
order model (FOM) first-order eigenvalue problem as below
Generally, the Campbell diagram is set up to understand (𝜆 ⋅ 𝐸 + 𝐴) ⋅ 𝜂 = 0 (24)
the dynamic behavior of a rotating shaft-disc-blade system, 𝜙
in particular, to determine its critical speed (Kirchgäßner where, 𝜂 = { }, E and A are the 2N×2N matrices.
𝜆∙𝜙
2016, Nandi and Neogy 2001, Nelson 1980, 2007, Vollan They are real but indefinite and unsymmetrical. From Eq.
and Komzsik 2012). This diagram is drawn by the repeat of (24), it is observed that this formulation doubles the size of
the quadratic eigenvalue analyses with the multiple steps the system matrices. So, when the FE model has a huge
corresponding to different values of spinning velocity (Ω). number of DOFs, the computational cost is significantly
More practically, the diagram is established by tying the increased. Therefore, an appropriate MOR technique should
successive value of the frequency or angular frequency be chosen to reduce the size of the system.
corresponding to the succession of the applied rotating
speed (Ω). Accordingly, the solution of the repetition of the 2.4 Establishment of the Campbell diagram for the
Eq. (14) is necessary. Consequently, the linear differential reduced order model (ROM)
equation of motion of the rotating structure with respect to
the spinning speed (Ω) in the rotating coordinate system Similar to Eq. (15), the equation of motion of the ROM
(Eq. (14)) can be expressed in more general form as in rotordynamic system can be written as in Eq. (25).
̃ ∙ 𝑟̈ (𝑡) + 𝐶̃ ∙ 𝑟̇ (𝑡) + 𝐾
𝑀 ̃ ∙ 𝑟(𝑡) = 𝐹̃ (𝑡) (15)
𝑀𝑟 ∙ 𝑧𝑧̈(𝑡) + 𝐶𝑟 ∙ 𝑧𝑧̇(𝑡) + 𝐾𝑟 ∙ 𝑧𝑧(𝑡) = 𝐹𝑟 (𝑡) (25)
̃=𝑀
𝑀 (16) Where

𝐶̃ = 𝐶𝑑 + 𝐶𝑐 (Ω) + 𝐶𝑏 (Ω) (17) ̃ ∙𝑄


𝑀𝑟 = 𝑄 𝑇 ∙ 𝑀 (26)

̃ = 𝐾𝑒 + 𝐾𝑐 (Ω) − 𝐾𝑠 (Ω) + 𝐾𝑏 (Ω)


𝐾 (18) 𝐶𝑟 = 𝑄 𝑇 ∙ 𝐶̃ ∙ 𝑄 (27)
From a mathematical perspective, to compute the
̃∙𝑄
𝐾𝑟 = 𝑄 𝑇 ∙ 𝐾 (28)
eigenfrequency of a rotating structure in the bending modes,
the right-hand side of Eq. (15) is set to be zero. Thus, the
general form of the solution of the free vibration for the 𝐹𝑟 = 𝑄 𝑇 ∙ 𝐹̃ (29)
damped second-order equation in Eq. (15) can be written as where, 𝑄 ∈ ℜ 𝑁×𝑛
is a low-dimensional subspace or the
(Han 2014) projection matrix in relation to a reduced vector (𝑧𝑧 ∈
𝑟(𝑡) = 𝜙 ⋅ 𝑒 𝜆𝑡 (19) ℜ𝑛 / 𝑛 ≪ 𝑁), and 𝑟 ≅ 𝑄 ∙ 𝑧𝑧. Therefore, the computation
of Q is required to obtain the reduced system (𝑛 ≪ 𝑁)
Where ϕ is the vector of the amplitude and λ is the using the modal superposition method and the Krylov
complex frequency of the response which can be expressed subspace-based model order reduction, which will be
𝜆 = 𝜎 ± 𝑗 ∙ 𝜔𝑎𝑓 . The real part of the eigenfrequency (σ) is presented in the following sections. Subsequently, the QEP
the damping constant, and when it is positive, the mode is of the reduced-order model in the displacement space can
unstable and vice versa. The imaginary part (𝜔𝑎𝑓 ) is the be expressed
damped circular frequency of whirling in the rotor
dynamical system. Consequently, the critical speed and the (𝜆̂2 ∙ 𝑀𝑟 + 𝜆̂ ∙ 𝐶𝑟 + 𝐾𝑟 ) ∙ 𝜙̂ = 0 (30)
stability can be determined based on λ. So, Eq. (30) can be transformed into a state-space
By applying the first and second derivatives on Eq. (19), equation (Eq. (31)), and then the 2n-dimensional first-order
one obtains eigenvalues problem (Eq. (32)).
𝑟̇ (𝑡) = 𝜙 ∙ 𝜆 ∙ 𝑒 𝜆𝑡 (20) 𝐼 0 𝑧𝑧̇ 0 −𝐼𝑛 𝑧𝑧
[𝑛 ]{ } + [ ]{ } = 0
⏟0 𝑀𝑟 𝑧𝑧̈ 𝐾
⏟𝑟 𝐶𝑟 𝑧𝑧̇ (31)
𝑟̈ (𝑡) = 𝜙 ∙ 𝜆2 ∙ 𝑒 𝜆𝑡 (21) 𝐸𝑟 𝐴𝑟

Substituting Eqs. (19)-(21) into Eq. (16), the quadratic


(𝜆̂ ⋅ 𝐸𝑟 + 𝐴𝑟 ) ⋅ 𝜂̂ = 0 (32)
eigenfrequency problem (QEP) can be written as in Eq. (22)
416 Ty Phuor and GilHo Yoon

can be employed as the orthogonalization algorithm (Gu et


Start
al. 2000, Ma and Wang 2007).
+ Create and mesh the model 𝜅𝑛 (𝐴, 𝑏) = 𝑠𝑝𝑎𝑛{𝑏, 𝐴 ⋅ 𝑏, 𝐴2 ⋅ 𝑏, . . . , 𝐴𝑛−1 ⋅ 𝑏} (33)
+ Compute the component matrices in the
equation of motion Therefore, the algorithm of the Ritz vector method is
Apply RV/QSRV/MQSRV/MMQSRV in the 𝜓1 = 𝐾𝑒−1 ⋅ 𝐺 (34)
Krylov subspace with the spin-speed-
dependent stiffness matrix and the new force 𝜓𝑖𝑖∗ = 𝐾𝑒−1 ⋅ 𝑀 ⋅ 𝜓𝑖𝑖−1 , 𝑖𝑖 = 2,3, . . . , 𝑛 (35)
vector (the inverse of the diagonal stiffness
matrix) Where 𝜓1 is the starting vector, G is the vector of
loading pattern, and the mass-orthonormalization for each
All bases (Q) are combined from the Krylov
subspace bases and the eigenmodes with the step is
application of the orthogonalization and
mass normalization processes
𝜓̄𝑖𝑖 = 𝜓𝑖𝑖∗ − ∑𝑖𝑖−1 𝑇 ∗
𝑗=1 (𝜓𝑗 𝑀𝜓𝑖𝑖 )𝜓𝑗 (36)

All reduced matrices are computed and ̄ 𝑖𝑖


𝜓
transformed them into the state-space
𝜓𝑖𝑖 = 1/2 (37)
̄ 𝑇 𝑀𝜓
(𝜓 ̄ 𝑖𝑖 )
𝑖𝑖
equation
Subsequently, on the basis of the quasi-static recurrence
No Eigenfrequencies are computed and
procedure of the Ritz bases, the QSRV method was
verified? proposed to let the Ritz vectors span the configuration space
at the desired frequency and efficiently capture the possible
Yes
dynamic deformations for the desired frequency range (Gu
To establish the Campbell diagram et al. 2000). Thus, the algorithm of the QSRV method can
be written as follows.
𝜓1∗ = (𝐾𝑒 − 𝜔𝑐2 ⋅ 𝑀)−1 ⋅ 𝐺 (36)
Finish
𝜓1∗
Fig. 2 The diagram of the methodology used in this study 𝜓1 = 𝑇 1/2 (37)
((𝜓1∗ ) ⋅𝑀⋅𝜓1∗ )

𝜙̂ 𝜓𝑖𝑖∗ = (𝐾𝑒 − 𝜔𝑐2 𝑀)−1 ⋅ 𝑀 ⋅ 𝜓𝑖𝑖−1 , 𝑖𝑖 = 2,3, . . . , 𝑛 (38)


where, 𝜂̂ = { }.
̂𝜆 ∙ 𝜙̂
Consequently, the expectation of the MOR techniques 𝜓̄𝑖𝑖 = 𝜓𝑖𝑖∗ − ∑𝑖𝑖−1 𝑇 ∗
𝑗=1 (𝜓𝑗 𝑀𝜓𝑖𝑖 )𝜓𝑗 (39)
can be accomplished when 𝜆𝑘 ≅ 𝜆̂𝑘 and 𝜙𝑘 ≅ 𝑄𝑛 ∙ 𝜙̂𝑘 for ̄ 𝑖𝑖
𝜓
some order k, where 𝑘 ≤ 𝑛. After solving Eq. (31), the 𝜓𝑖𝑖 = 1/2 (40)
̄ 𝑇 𝑀𝜓
(𝜓 ̄ 𝑖𝑖 )
modal assurance criterion (MAC) (Mogenier et al. 2014) 𝑖𝑖

should be applied for the mode shape tracking in order to Where 𝜔𝑐 is the centering frequency which is at the
draw the Campbell diagram. Fig. 2 illustrates the midpoint of the frequency range of interest. 𝜓1∗ is the first
methodology flowchart for the Campbell diagram analysis vector of the QSRV which depends on the loading pattern
using the MOR in rotordynamic system in this study. From G. More precisely, the first vector is mass-normalized by
this flowchart, several possible MOR methods can be Eq. (37), and then the quasi-static recurrence procedure is
implemented for establishing the Campbell diagram by given in Eq. (38) to Eq. (40).
adjusting the values of the central frequencies and central In the QSRV procedure, only one centering frequency is
spinning frequencies. Those methods are the Krylov employed. Thus, the MQSRV method was proposed by
subspace-based model order reduction, modal superposition considering more than one centering frequency in order to
method, and the combined approach. make the techniques more robust for general cases (Gu et
al. 2000, Yoon 2012). It is noted that, in the MQSRV
method, the constant stiffness and the frequency-dependent
3. Krylov subspace-based model order reduction stiffness are employed by Gu et al. (2000), Yoon (2012),
respectively. Thus, the quasi-static recurrence procedure of
3.1 The RV, QSRV, MQSRV methods the MQSRV method, for example, if three centering
frequencies are used, can be extended by
For mathematical perspective, the Ritz vector method
employed the Krylov subspace to generate the vector bases 𝜓1∗ = (𝐾 − 𝜔𝑐1
2
⋅ 𝑀)−1 ⋅ 𝐺, for 𝑖𝑖 = 1 (41)
which is dependent on the elastic stiffness matrix, mass
𝜓𝑖𝑖∗ = (𝐾 − 𝜔𝑐1
2
⋅ 𝑀)−1 ⋅ 𝑀 ⋅ 𝜓𝑖𝑖−1 , for 𝑖𝑖 = 2,3, . . . , 𝑙 (42)
matrix, and load vector (Gu et al. 2000, Koh et al. 2020,
Wilson et al. 1982,Yoon 2010). In this sense, with a matrix 𝜓𝑖𝑖∗ = (𝐾𝑒 − 𝜔𝑐2 𝑀)−1 ⋅ 𝑀 ⋅ 𝜓𝑖𝑖−1 , 𝑖𝑖 = 2,3, . . . , 𝑛 (43)
A and a vector b, the linear subspace of the n order Krylov
subspace can be generated as in Eq. (33). The computed 𝜓𝑖𝑖∗ = (𝐾 − 𝜔𝑐2
2
⋅ 𝑀)−1 ⋅ 𝑀 ⋅ 𝜓𝑖𝑖−1 , for 𝑖𝑖 = 𝑙 + 2, 𝑙 +
Krylov subspace bases are found to be almost linearly (44)
3, . . . , 𝑚
dependent on each other; consequently, Arnoldi iteration
Model order reduction for Campbell diagram analysis of shaft-disc-blade system in 3D finite elements 417

(a) MMQSRVM with rectangular form (b) MMQSRVM with diagonal form
Fig. 3 The diagram of the multiple central frequencies and the multiple central spinning frequencies
for MMQSRVM with (a) rectangular form and (b) diagonal form

∗ 2 𝑠 𝑠
𝜓𝑚+1 = (𝐾 − 𝜔𝑐3 ⋅ 𝑀)−1 ⋅ 𝐺, for 𝑖𝑖 = 𝑚 + 1 (45) 𝜔𝑎𝑓,𝑠𝑡𝑎𝑟𝑡 and 𝜔𝑎𝑓,𝑒𝑛𝑑 are the starting and ending angular
frequencies in the s-th domain, respectively.
𝜓𝑖𝑖∗ = (𝐾 − 𝜔𝑐2
2
⋅ 𝑀)−1 ⋅ 𝑀 ⋅ 𝜓𝑖𝑖−1 , for 𝑖𝑖 = 𝑚 + Thus, the quasi-static recurrence procedure of the
(46)
2, 𝑚 + 3, . . . , 𝑛 MMQSRV method, for example, if three centering
Where l, m-l, and n-m are the numbers of Ritz vectors in frequencies and three centering spinning frequencies are
the recurrence group corresponding to 𝜔𝑐1 , 𝜔𝑐2 and 𝜔𝑐3 , used as in Fig. 3(b), can be written as follows. The starting
respectively. It is worth mentioning that after each vector of the MMQSRV method is computed by
recurrence step, the orthogonality and mass normalization For 𝑖 = 1 ⟹ 𝜓1∗ =
are applied by using Eq. (39) and Eq. (40), respectively. −1 (48)
̃ (𝜔𝑐,𝑠𝑓 (1)) − (𝜔𝑐,𝑎𝑓 (1))2 ∙ 𝑀)
(𝐾 ∙𝐹
3.2 The trial and error MMQSRV methods Where
𝜔𝑐,𝑎𝑓 = [𝑦𝑎1,1 𝑦𝑎1,2 𝑦𝑎2,1 𝑦𝑎3,1 𝑦𝑎3,2 ]
As mentioned above, Gu et al. (2000) proposed the
QSRV method, which is equivalent to the Ritz vector (RV) 𝜔𝑐,𝑠𝑓 = [𝑥𝑎1,1 𝑥𝑎1,2 𝑥𝑎2,1 𝑥𝑎3,1 𝑥𝑎3,2 ]
method when the zero-central frequency is applied. On the (49)
𝜓1∗
basis of this QSRV method, Yoon (2012) developed and 𝜓1 = 𝑇 1/2
successfully applied the MQSRV method for solving the ((𝜓1∗ ) ⋅𝑀⋅𝜓1∗ )

problem of the frequency-dependent acoustic system. And


for the rotordynamic system, Han (2014) successfully 𝐹𝑜𝑟 𝑖 = 2,3, … , 𝑑1 ⟹ 𝜓𝑖∗ =
employed the RV method to compute the projection matrix −1 (50)
̃ (𝜔𝑐,𝑠𝑓 (1)) − (𝜔𝑐,𝑎𝑓 (1))2 ∙ 𝑀)
(𝐾 ∙ (𝑀 ∙ 𝜓𝑖−1 )
for the MOR. However, it is found that the error increased
when the externally applied rotational speed increased. ∗
𝐹𝑜𝑟 𝑖 = 𝑑1 + 1 ⟹ 𝜓𝑑1+1 =
Therefore, a trial-and-error MMQSRV method is proposed
−1 (51)
in this study by extending the MQSRV method by ̃ (𝜔𝑐,𝑠𝑓 (2)) − (𝜔𝑐,𝑎𝑓 (2))2 ∙ 𝑀)
(𝐾 ∙𝐹
introducing the use of the multiple central frequencies and
multiple central spinning frequencies for generating the 𝐹𝑜𝑟 𝑖 = 𝑑1 + 2, 𝑑1 + 3, … , 𝑑2 ⟹ 𝜓𝑖∗ =
projection matrix in the Krylov subspace as depicted in Fig. −1 (52)
3. As already seen in the Krylov subspace-based MOR ̃ (𝜔𝑐,𝑠𝑓 (2)) − (𝜔𝑐,𝑎𝑓 (2))2 ∙ 𝑀)
(𝐾 ∙ (𝑀 ∙ 𝜓𝑖−1 )
method, the bases are generated with only the central
frequencies of interest. But, for the rotordaynamic problem, ∗
𝐹𝑜𝑟 𝑖 = 𝑑2 + 1 ⟹ 𝜓𝑑2+1 =
the MMQSRV method computes the bases with both −1 (53)
multiple central frequencies and multiple central spinning ̃ (𝜔𝑐,𝑠𝑓 (3)) − (𝜔𝑐,𝑎𝑓 (3))2 ∙ 𝑀)
(𝐾 ∙𝐹
frequencies (see Eq. (48)). And the magnitudes of the
central frequency and central spinning frequency in the s-th 𝐹𝑜𝑟 𝑖 = 𝑑2 + 2, 𝑑2 + 3, … , 𝑑3 ⟹ 𝜓𝑖∗ =
domain should be calculated first by −1 (54)
̃ (𝜔𝑐,𝑠𝑓 (3)) − (𝜔𝑐,𝑎𝑓 (3))2 ∙ 𝑀)
(𝐾 ∙ (𝑀 ∙ 𝜓𝑖−1 )
𝑠 𝑠 𝑠 𝑠
𝜔𝑠𝑓,𝑠𝑡𝑎𝑟𝑡 +𝜔𝑠𝑓,𝑒𝑛𝑑 𝜔𝑎𝑓,𝑠𝑡𝑎𝑟𝑡 +𝜔𝑎𝑓,𝑒𝑛𝑑
𝑠
𝜔𝑐,𝑠𝑓 = 𝑠
and 𝜔𝑐,𝑎𝑓 = (47)
2 2 ∗
𝐹𝑜𝑟 𝑖 = 𝑑3 + 1 ⟹ 𝜓𝑑3+1 =
Where 𝑠
𝜔𝑠𝑓,𝑠𝑡𝑎𝑟𝑡 and 𝑠
are respectively the
𝜔𝑠𝑓,𝑒𝑛𝑑 −1 (55)
̃ (𝜔𝑐,𝑠𝑓 (4)) − (𝜔𝑐,𝑎𝑓 (4))2 ∙ 𝑀)
(𝐾 ∙𝐹
starting and ending spinning frequencies in the s-th domain.
418 Ty Phuor and GilHo Yoon

(a) without the orthogonality and mass normalization


(a) without the orthogonality and mass normalization

(b) with the orthogonality and mass normalization


Fig. 5 The surface plot of Eq. (61) for 3D rotating shaft

(b) with the orthogonality and mass normalization


MS method may reduce the size of the system by using only
Fig. 4 The surface plot of Eq. (61) for Rayleigh beam the mass and elastic stiffness matrices. Therefore, for
computational accuracy and time reduction of the Campbell
diagram analysis in this study, the real modal analysis uses
𝐹𝑜𝑟 𝑖 = 𝑑3 + 2, 𝑑3 + 3, … , 𝑑4 ⟹ 𝜓𝑖∗ = the real undamped eigenvalues easily calculated from the
−1 (56) mass and elastic stiffness matrices employing the standard
̃ (𝜔𝑐,𝑠𝑓 (4)) − (𝜔𝑐,𝑎𝑓 (4))2 ∙ 𝑀)
(𝐾 ∙ (𝑀 ∙ 𝜓𝑖−1 ) eigenvalue problem (see Eq. (60)). Obviously, the computed

modal matrix is independent of the spinning velocity;
𝐹𝑜𝑟 𝑖 = 𝑑4 + 1 ⟹ 𝜓𝑑4+1 = consequently, the modal matrix is just computed once and
−1 (57) preserved as a parameter throughout the analysis. Moreover,
̃ (𝜔𝑐,𝑠𝑓 (4)) − (𝜔𝑐,𝑎𝑓 (4))2 ∙ 𝑀)
(𝐾 ∙𝐹 it is worth mentioning that the eigenvector is typically
mass-normalized, as shown in Eq. (61). A modal matrix
𝐹𝑜𝑟 𝑖 = 𝑑4 + 2, 𝑑4 + 3, … , 𝑑5 ⟹ 𝜓𝑖∗ = (𝑄𝑀𝑆 ) whose columns are the eigenvectors normalized
−1 (58) with respect to the mass matrix, provide a reduced bases
̃ (𝜔𝑐,𝑠𝑓 (5)) − (𝜔𝑐,𝑎𝑓 (5))2 ∙ 𝑀)
(𝐾 ∙ (𝑀 ∙ 𝜓𝑖−1 ) from the N to n in the system, where 𝑛 ≪ 𝑁.
Where d1, d2-d1, d3-d2, d4-d3, and n-d4 are the [𝐾𝑒 ]𝜓𝑀𝑆 = 𝜆𝑀𝑆 [𝑀]𝜓𝑀𝑆 (60)
numbers of Ritz vectors in the recurrence group
corresponding to 𝜔𝑐,𝑠𝑓 (1) 𝑎𝑛𝑑 𝜔𝑐,𝑎𝑓 (1) ; 𝑇
𝑄𝑀𝑆 ⋅ [𝑀] ⋅ 𝑄𝑀𝑆 = [𝐼] (61)
𝜔𝑐,𝑠𝑓 (2) 𝑎𝑛𝑑 𝜔𝑐,𝑎𝑓 (2) ; 𝜔𝑐,𝑠𝑓 (3) 𝑎𝑛𝑑 𝜔𝑐,𝑎𝑓 (3) ;
𝜔𝑐,𝑠𝑓 (4) 𝑎𝑛𝑑 𝜔𝑐,𝑎𝑓 (4) ; and 𝜔𝑐,𝑠𝑓 (5) 𝑎𝑛𝑑 𝜔𝑐,𝑎𝑓 (5) , It is noticed that Eq. (61) holds true for the beam
element, i.e., Rayleigh beam, as seen in Fig. 4. However,
respectively. It is worth mentioning that after each
this orthogonality condition does not complete for 3D solid
recurrence step, the orthogonality and mass normalization
FE, as illustrated in Fig. 5(a). For this situation, we suggest
must be applied by using Eq. (39) and Eq. (40),
using a procedure to improve the orthogonality bases of the
respectively. Therefore, the final bases for the Krylov
modal superposition method by applying the orthogonality
subspace-based MMQSRVM can be expressed as in Eq.
and mass normalization to the modal matrix. Consequently,
(61). The equation
this proposed algorithm might be useful for the MOR
method to its practical implementation in a 3D
𝑄=[ 𝜓
⏟1 … 𝜓𝑑1 , 𝜓 ⏟𝑑4+1 … 𝜓𝑑5 ] (59)
⏟𝑑1+1 … 𝜓𝑑2 , … , 𝜓 rotordynamic system, especially for design optimization.
𝑡ℎ𝑒 1𝑠𝑡 𝑑𝑜𝑚𝑎𝑖𝑛 𝑡ℎ𝑒 2𝑛𝑑 𝑑𝑜𝑚𝑎𝑖𝑛 𝑡ℎ𝑒 5𝑡ℎ 𝑑𝑜𝑚𝑎𝑖𝑛 The procedure to improve the orthogonality of the
Where, 𝑄 ∈ ℜ 𝑁×𝑛
is the projection matrix. projection matrix in the modal superposition method is
1. Find the n bases eigenvectors of Eq. (60)
2. Apply the orthogonality
4. Modal superposition method 𝜓̄𝑖𝑖 = 𝜓𝑖𝑖∗ − ∑𝑖𝑖−1 𝑇 ∗
(62)
𝑗=1 (𝜓𝑗 𝑀𝜓𝑖𝑖 )𝜓𝑗 , 𝑖𝑖 = 1,2, . . . , 𝑛

An alternative to the Krylov subspace-based MOR, the 3. Apply the mass normalization
Model order reduction for Campbell diagram analysis of shaft-disc-blade system in 3D finite elements 419

Fig. 6 The 3d finite element mesh model of a rotating


simply supported beam (Number of nodes: 709, number of
elements: 160, degree of freedom: 2127, Young’s modulus:
200 GPa, Density: 7800 kg/m3, Poisson’s ratio: 0.3),
Length: 1 m and slenderness ratio: radius/(2×Length)

Fig. 7 The comparisons of the non-dimensional critical


̄ 𝑖𝑖
𝜓 speeds for a uniform, simply supported shaft produced by
𝜓𝑖𝑖 = 1/2 (63) the present analysis and the existing solutions
̄ 𝑇 𝑀𝜓
(𝜓 ̄ 𝑖𝑖 )
𝑖𝑖

To the best of the authors' knowledge, we would like to


emphasize that, despite this algorithm is easy to implement 6. Verification of the finite element rotordynamic
with the modal superposition method, it has not been either model
proposed or used in the literature before.
It is interesting to note that the Krylov subspace bases The verification of the 3D rotordynamic model in FEA is
and the eigenmodes can also be combined performed by the computation of the non-dimensional critical
𝑄=[ ⏟
𝜓1 … 𝜓𝑑1 , ⏟
𝜓𝑑1+1 … 𝜓𝑑2 , … , speeds of a rotating simply supported beam, as shown in Fig.
𝑡ℎ𝑒 1𝑠𝑡 𝑑𝑜𝑚𝑎𝑖𝑛 𝑡ℎ𝑒 2𝑛𝑑 𝑑𝑜𝑚𝑎𝑖𝑛
6. Nelson (1980) provided the expression for computing the
(64) non-dimensional critical speed by [𝜌𝑙 2 Ω2𝐶𝑟 /(𝐸 × (𝑅/
⏟ ⏟𝑒𝑖𝑔,1 … 𝜓𝑒𝑖𝑔,𝑚 ]
𝜓𝑑𝑠+1 … 𝜓𝑛 , 𝜓 2𝑙)2 )]1/4 , where 𝜌, 𝑙, Ω𝐶𝑟 , 𝐸 and R are the density, length,
𝑡ℎ𝑒 𝑠−𝑡ℎ 𝑑𝑜𝑚𝑎𝑖𝑛 𝑚 𝑒𝑖𝑔𝑒𝑛𝑚𝑜𝑑𝑒𝑠 critical speed, Young's modulus, and radius of the beam. The
analysis of the rotating shaft is conducted with the beam
slenderness ratio of 0.02 to 0.1 with an increment of 0.02. Fig.
5. Boundary conditions and finite element technique 7 illustrates the comparisons for the first and second forward
(FW1-2) and backward (BW1-2) whirling modes of the non-
The 3D FEA of the rotating structures is carried out by dimensional critical speeds produced by Nelson (1980) using
employing the in-house code in the framework of Timoshenko beam theory and the present analysis using the 3D
MATLAB. Both ends of the shaft are applied with the FEA. The graph revealed that the solutions of the present
simply supported boundary conditions by allowing the analysis are in close agreement with those computed by Nelson
structure to rotate freely above the x-axis. To avoid a rank (1980) for all cases. Moreover, the non-dimensional critical
deficiency in the element stiffness matrix, all the speeds of the nonrotating shaft (NRS for zero rotating speed)
component matrices (e.g., mass, elastic stiffness) are are also computed in the present analysis and compared to
computed with a full integration technique using a Gaussian those calculated by Nelson (1980). It is observed that the
quadrature of 9 by 9 by 9 for 20-hexahedral elements and of differences have occurred and increased with the increase in
6 by 3 by 6 for 15-node pentahedral elements (Gruttmann the slenderness ratio in conjunction with the number of modes.
and Wagner 2004, Phuor 2020, Phuor et al. 2021a, Phuor et These differences might be due to the employment of the
al. 2021b). The elements used in this study are assumed to different methods and the assumptions made in each technique.
behave as linear isotropic material. Moreover, the parallel However, it is found that the solutions have good trends to
“for loop” in conjunction with the fast assembly of the FE each other, and consequently, the two solutions are in reliable
matrices technique is also implemented based on the agreement. Therefore, the computed results in the 3D FEA in
method developed by Rahman and Valdman (2013). The this study are reliable.
viscous damping for the rotor system to increase the values
of the damping exponents is assumed to be a Rayleigh
damping with 𝛽 = 0.217𝑒 −8 𝜇𝑠. In this study, the force 7. Results and discussion
vector is assumed to be the inverse of the diagonal elastic
stiffness matrix (see Eq. (65)) because it has more physical The numerical examples for the Campbell diagram
meaning than the unit force vector which was proposed by analysis are presented to demonstrate the validity and
Han (2014). effectiveness of the MOR techniques used in this study. The
angular velocity is applied in a range of 0 to 6283.1853
𝐹 = 1/𝑑𝑖𝑎𝑔(𝐾𝑒 ) (65) (rad/s)/1000 (Hz) with the increment of 157.08 (rad/s)/25
420 Ty Phuor and GilHo Yoon

(a) (b)
Fig. 8 (a) The geometry and a 3D FE mesh of a shaft-disc-blade assembly system (Number of nodes: 45685, number of
elements: 9504, Young’s modulus: 207.8 GPa, Density, 7806 kg/m3, Poisson’s ratio: 0.3), and (b) the Campbell diagram
established with the first 14 modes by FOM

Table 3 The geometrical properties of the shaft, disc, and


(Hz) to determine the critical speeds of the forward and blade for example 1
backward whirling (FW and BW) modes of the system. The Dimension (m)
accuracy is checked with the calculation of the norm of the Structure
Length Inner radius Outer radius
relative error (Er) by A 0.4 0 0.02
B 0.1 0.02 0.04
𝐸𝑟 = ‖∑(𝑓𝑅𝑂𝑀𝑖 − 𝑓𝐹𝑂𝑀𝑖 )⁄𝑓𝐹𝑂𝑀𝑖 × 100‖ (66)
C 0.01 0.04 0.15
where 𝑓𝑅𝑂𝑀 and 𝑓𝐹𝑂𝑀 are the frequency produced by D 0.05 0.15 0.16
the reduced-order model (ROM) and the full order model 12@0.05 with the
E 0.16 0.225
revolution of 6o
(FOM), respectively, and i denotes the number of modes.
All simulations are performed in the framework of Matlab
on an Intel® CoreTM i7-4770 CPU@3.4 GHz processors
and 32GB RAM. rotating speed. The angular frequency and the rotating
speed of interest are respectively from 0 to 12566.4 rad/s
7.1 Example 1: Rotordynamic system and 0 to 6283.2 rad/s for this example. The comparisons of
the ERs are carried out in such a way that all techniques
For the first example, the 3d finite element model of a contain the same total number of bases.
shaft-disc-blade assembly system in Fig. 8(a) is considered. Fig. 9(a) shows the comparisons of the ERs produced by
The total number of the element is 9504 with 45685 nodes. the MQSRV and the MMQSRV methods. MQSRV60_2caf
The total degree of freedom (DOF) is 137055. Table 3 gives and MQSRV60_3caf represent the MQSRV method having
the geometrical properties of the rotating structure. The 60 total bases with 2 and 3 central angular frequencies (caf),
Campbell diagram established with the first 14 (i=14 in Eq. respectively. The central frequencies are set 𝜔𝑐,𝑎𝑓 =
(66)) modes by FOM is shown in Fig. 8(b). These first 14 [0 1000] (Hz) for MQSRV60_2caf and 𝜔𝑐,𝑎𝑓 =
modes are selected in order to compare the norm of the [0 666.67 1333.33] (Hz) for MQSRV60_3caf. In
relative errors produced by all MOR techniques and these addition, MMQSRV60_3caf_3csf (rectangular form) and
modes should be sufficient for the actual practical works, MMQSRV60_3caf_3csf (diagonal form) are the MMQSRV
while the existing work just took the first ten modes (Han, method containing 60 total bases with three central angular
2014). As a result, from this diagram, several critical speeds frequencies and three central spinning frequencies for the
are found. By intersecting with the 2P line (𝜔𝑠𝑓 = 2𝜔𝑎𝑓 ), rectangular and diagonal forms (see Fig. 3), respectively.
the calculated backward whirl critical speeds are found at Hence, 𝜔𝑐,𝑎𝑓 =[0 0 0 666.67 666.67 666.67 1333.33
116, 220, 243, 500, 700, and 825 (Hz) for BW1, BW2, 1333.33 1333.33] (Hz) and 𝜔𝑐,𝑠𝑓 =[0 666.67 1333.33 0
BW3, BW4, BW5, and BW6 modes, respectively. 666.67 1333.33 0 666.67 1333.33] (Hz) are set for
Similarly, by intersecting with the horizontal zero-axis, the MMQSRV60_3caf_3csf for the rectangular form.
FW2 critical speed is found at 213 (Hz) or 12780 (rpm). 𝜔𝑐,𝑎𝑓 =[0 0 666.67 1333.33 1333.33] (Hz) and 𝜔𝑐,𝑠𝑓 = [0
Under this case, this model demonstrates the instability at 666.67 1333.33 0 666.67] (Hz) are for
213 (Hz). MMQSRV60_3caf_3csf with the diagonal form. The
To demonstrate the effectiveness of the MOR Campbell diagrams are compared by computing the ERs as
techniques, Fig. 9 provides several Campbell diagrams, depicted in Fig. 9(a). From this figure, it is seen that the ER
which are analyzed and compared in the form of calculating is significantly increased with the increase in the applied
the norm of relative errors (ERs) with respect to the applied rotating speed. For instance, when the spinning velocity
Model order reduction for Campbell diagram analysis of shaft-disc-blade system in 3D finite elements 421

(a) (b)

(c) (d)

(e) (f)

Fig. 9 The comparisons of the norm of the relative errors produced by (a) MQSRV and MMQSRV methods, (b) QSRV
and MQSRV methods, (c) QSRV and RV methods, (d) RV and the combined RV and MS methods, (e) MS and the
combined RV and MS methods, and (f) RV and MS methods

Ω=1000 (Hz), the Ers are up to 4.98%, 3.60%, 38.55%, form). Consequently, the MQSRV method is dominant
29.71%, 2.24% and 25.29% for MQSRV60_2caf, 3caf, compared to the MMQSRV method in terms of the
MMQSRV60_2caf_2csf (rectangular form), 3caf_3csf Campbell diagram analysis. However, the results produced
(rectangular form), 3caf_3csf (diagonal form) and 4caf_4csf by these two methods are still unacceptable because the
(diagonal form), respectively. It is found that magnitudes of the Ers are out of the allowable tolerance
MMQSRV60_3caf_3csf (diagonal form) yields the smallest value (1%).
error; however, from the computational point of view, it is Fig. 9(b) illustrates the comparisons of the Campbell
not a suitable method for Campell diagram analysis because diagram analysis in the form of the ERs produced by the
its ER are not consistent. For example, the errors of QSRV and the MQSRV methods. The comparisons are
MMQSRV60_3caf_3csf (diagonal form) are smaller than performed by many cases, in which QSRV60_1caf
those of MMQSRV60_4caf_4csf and 2caf_2csf (diagonal represents the QSRV method with one central angular
422 Ty Phuor and GilHo Yoon

Table 4 The comparisons of damped frequency between FOM and ROMs (RV40, RV60 MS40, and MS60) at
Ω=60000 rpm for example 1
Frequency (Hz) Relative error (%)
Mode
Whirl FOM RV40 RV60 MS40 MS60 RV40 RV60 MS40 MS60
No.
(N=137055) (n=40) (n=60) (n=40) (n=60) (n=40) (n=60) (n=40) (n=60)
1 FW2 931.329 977.356 934.870 936.870 933.798 4.942 0.380 0.595 0.265
2 BW2 1019.377 1019.394 1019.388 1019.390 1019.389 0.002 0.001 0.001 0.001
3 FW1 -785.059 -785.032 -785.058 -785.056 -785.056 -0.004 0.000 0.000 0.000
4 BW1 1213.464 1213.564 1213.472 1213.480 1213.480 0.008 0.001 0.001 0.001
5 FW3 993.344 995.437 995.398 995.446 995.443 0.211 0.207 0.212 0.211
6 BF3 1069.952 1070.135 1070.113 1070.144 1070.094 0.017 0.015 0.018 0.013
7 Shear 1129.147 1129.505 1129.378 1130.037 1130.013 0.032 0.020 0.079 0.077
8 FW4 1202.405 1205.642 1205.580 1205.715 1205.302 0.269 0.264 0.275 0.241
9 BW4 1299.188 1300.698 1300.145 1300.262 1299.807 0.116 0.074 0.083 0.048
10 FW5 1477.230 1481.537 1481.284 1482.255 1481.551 0.292 0.274 0.340 0.292
11 BW5 1554.032 1557.490 1556.821 1558.354 1556.216 0.223 0.180 0.278 0.141
12 FW6 1682.264 1687.558 1686.273 1688.599 1684.992 0.315 0.238 0.377 0.162
13 BF6 1722.183 1727.197 1725.692 1728.152 1723.588 0.291 0.204 0.347 0.082
14 Shear 1767.712 1773.352 1772.862 1776.696 1776.575 0.319 0.291 0.508 0.501
Norm of the relative error (ER) (%) 4.998 0.744 1.096 0.756

frequency (ω𝑐,𝑎𝑓 = 1000 (Hz)) and 60 total bases. From


this plot, when Ω=1000 (Hz), it is observed that the ERs are
respectively 5%, 4.96%, 0.83%, 15.78%, 5.66%, 4.99%,
20.10%, 4.37%, and 3.60% for the QSRV and MQSRV
methods having 1, 2 and 3 central frequencies with the 40,
50 and 60 total bases. Thus, from the computational
perspective, it is found that the QSRV method is dominant
over the MQSRV method.
In addition, the Campbell diagram analysis produced by
the QSRV method is also compared with those computed by
the Ritz vector (RV) method, as seen in Fig. 9(c).
Similarly, the ER is drastically increased with a large
number of spinning velocities. For instance, when Ω=1000
(Hz), the errors are found at 5%; 4.99%, 4.95%; 4.56% and
0.83%; 0.744% by QSRV and RV for total bases of 40, 50 Fig. 10 The Campbell diagram established with the first 14
and 60, respectively. It is observed that acceptable results modes by the MS60
are obtained at 60 bases for both methods. However, based
on the trends of the ERs, the RV method is rather better.
Similarly, Fig. 9(d) also demonstrates the analogy of the method is dominant at low spinning velocity and the MS
solutions in terms of the ERs produced by the RV method method is dominant at high spinning velocity for the
and the combined RV and modal superposition (MS) Campbell diagram analysis in this example. More precisely,
methods. The RV10+MS50_1caf represents the Table 4 also provides the comparisons of the damped
combination of the 50 eigenmodes from the MS method and frequency, relative and norm of the relative error produced
the 10 Krylov subspace bases at zero central frequency in by the FOM and the ROMs for RV40, RV60, MS40 and
the RV method. From Fig. 9(d), it is observed that, when MS60 when Ω=60000 rpm. It is observed that the trend of
Ω=1000 (Hz), the computed ERs are 4.99%; 1.24%, 4.56%; the MS method is rather better. Consequently, the Campbell
1.05% and 0.74%; 0.88% by the RV method and the diagram produced by MS60 can be depicted in Fig. 10.
combined methods for the total bases of 40, 50, and 60, From this diagram, it is found that the critical speeds remain
respectively. Thus, it is found that the ER produced by the the same as those produced by the FOM (see Fig. 8(b)).
combined method is more consistent. In addition, Fig. 9(e) Table 5 compares the time computing in total for the
depicts the comparisons of the errors calculated by the MS Campbell diagram analysis produced by the FOM and
method with the combined methods. The computed results ROMs. It is seen that the ROMs can reduce the
reveal that, for Ω=1000 (Hz), the errors are 1.1%; 1.23%, computational time up to 96% as compared to that of FOM
0.91%; 1.05% and 0.76%; 0.88% given by MS and the (9501 s). This total computation time is taken into account
combined methods for 40, 50, and 60 bases, respectively. for the computations of the matrix components (i.e., mass
Thus, from the plot, the MS method provides smaller errors. matrix, elastic stiffness matrix), the generation of the
Besides, Fig. 9(f) illustrates the comparison of the errors Krylov bases, the calculation of the eigenmode, and the
between the MS and RV methods. It is noticed that the RV solution of the quadratic eigenvalue problems in Matlab.
Model order reduction for Campbell diagram analysis of shaft-disc-blade system in 3D finite elements 423

(a) (b)
Fig. 11 (a) The geometry and a 3D FE mesh of a shaft-disc-blade assembly system (Number of nodes: 63017, number of
elements: 14208, Young’s modulus: 207.8 GPa, Density, 7806 kg/m3, Poisson’s ratio: 0.3, and (b) The Campbell diagram
established with the first 14 modes by FOM

Table 5 The comparison of the total time computing for Table 6 The geometrical properties of the shaft, disc, and
Campbell diagram analysis produced by the FOM and blade for example 2
ROMs for example 1 Dimension (m)
Structure
Methods Computation time (s) Length Inner radius Outer radius
FOM (𝑁 = 137055) 9501 s 100% A 0.4 0 0.02
RV60 85 s 0.90% B 0.1 0.02 0.04
RV50 72 s 0.77% C 0.01 0.04 0.15
RV40 60 s 0.63% D 0.05 0.15 0.16
QSRV60 87 s 0.92% 12@(Starting=0.01 and
QSRV50 73 s 0.77% E Ending=0.005) with the 0.16 0.225
QSRV40 60 s 0.64% revolution of 20o
MQSRV60_2caf 133 s 1.41%
MQSRV60_3caf 133 s 1.41%
MMQSRV60_3caf_3csf (rect form) 363 s 3.82% 7.2 Example 2: Rotating shaft-disc-blade assembly
MMQSRV60_3caf_3csf (diag form) 224 s 2.36% system
MMQSRV60_4caf_4csf (diag form) 319 s 3.36%
Fig. 11(a) depicts the 3D FE shaft-disc-blade assembly
RV10+MS30_1caf 245 s 2.59%
model used in this example which has 63017 nodes, 14208
RV10+MS40_1caf 273 s 2.88%
elements, and 189051 degrees of freedom. The geometrical
RV10+MS50_1caf 282 s 2.98%
properties of this model are tabulated in Table 6. Similar to
MS40 230 s 2.42%
example 1, the Campbell diagram was established with the
MS50 235 s 2.48%
first 14 modes by FOM, and several critical speeds can be
MS60 288 s 3.04% determined, as seen in Fig. 11(b). Obviously, the backward
whirl (BW) critical speeds are respectively found at 223,
126, 242, 424, 493, and 518 (Hz) for BW1, BW2, BW3,
More precisely, by taking the computational time of the BW4, BW5, and BW6 modes. In contrast, a single second
FOM as a reference, the time consuming by the ROMs are forward whirl (FW2) critical speed is found at 223 (Hz) or
respectively 0.90%, 0.77%, 0.63%, 0.92%, 0.77%, 0.64%, 13380 (rpm) by the intersection with the horizontal zero-
1.41%, 1.41%, 2.12%, 3.82%, 2.36%, 3.36%, 2.59%, axis. Thus, this rotordynamic system shows instability at
2.88%, 2.98%, 2.42%, 2.48% and 3.04% for the RV60, 223 Hz under this circumstance.
RV50, RV40, QSRV60, QSRV50, QSRV40, Similar to the first example, the computation of the ER,
MQSRV60_2caf, MQSRV60_3caf, MMQSRV60_2caf_2csf as in Fig. 12, can determine the effectiveness of the MOR
(rect form), MMQSRV60_3caf_3csf (rect form), techniques for the Campbell diagram analysis. The angular
MMQSRV60_3caf_3csf (diag form), frequency and the rotating speed of interest are respectively
MMQSRV60_4caf_4csf (diag form), RV10+MS30_1caf, set from 0 to 7539.84 rad/s and from 0 to 6283.2 rad/s for
RV10+MS40_1caf, RV10+MS50_1caf, MS40, MS50 and this case. The comparisons of the ERs are carried out in
MS60. As expected, it is found that the higher bases of such a way that all techniques contain the same total
ROM take a longer time for all methods. Thus, from the number of bases.
engineering perspective, the MOR methods effectively The comparisons of the ERs produced by the MQSRV
reduce the time for the Campbell diagram analysis. and MMQSRV methods are depicted in Fig. 12(a). In this
424 Ty Phuor and GilHo Yoon

(a) (b)

(c) (d)

(e) (f)

Fig. 12 The comparisons of the norm of the relative errors produced by (a) MQSRV and MMQSRV methods, (b) QSRV
and MQSRV methods, (c) QSRV and RV methods, (d) RV and the combined RV and MS methods, (e) MS and the
combined RV and MS methods, and (f) RV and MS methods

case, the central frequencies are set 𝜔𝑐,𝑎𝑓 =[ 0 600] (Hz) applied rotating speed. For example, the ERs are
for MQSRV60_2caf and 𝜔𝑐,𝑎𝑓 =[0 400 800] (Hz) for respectively up to 2.77%, 6.92%, 1.57%, 2.17%, 1.13% and
MQSRV60_3caf. Also, 𝜔𝑐,𝑎𝑓 =[0 0 0 400 400 400 800 800 5.58% for MQSRV60_2caf, 3caf, MMQSRV60_2caf_2csf
(rectangular form), 3caf_3csf (rectangular form), 3caf_3csf
800] (Hz) and 𝜔𝑐,𝑠𝑓 =[0 333.33 666.67 0 333.33 666.67 0
(diagonal form) and 4caf_4csf (diagonal form), when the
333.33 666.67] (Hz) are set for MMQSRV60_3caf_3csf spinning velocity Ω=1000 (Hz). It is observed that the
with the rectangular form. Likewise, 𝜔𝑐,𝑎𝑓 =[0 0 400 800 smallest error is given by MMQSRV60_3caf_3csf (diagonal
800] (Hz) and 𝜔𝑐,𝑎𝑓 =[0 666.67 333.33 0 666.67] (Hz) are form); however, due to the inconsistency of its results, it is
for MMQSRV60_3caf_3csf with the diagonal form. Thus, not an appropriate method. As a result, the MQSRV method
the comparisons of the Ers of the Campbell diagrams are is more influential for establishing the Campbell diagram.
illustrated as in Fig. 12(a). From this figure, it is seen that However, magnitudes of the ER produced by these two
the ERs are significantly increased with the increase in the methods are unacceptable because they are greater than the
Model order reduction for Campbell diagram analysis of shaft-disc-blade system in 3D finite elements 425

Table 7 The comparisons of damped frequency between FOM and ROMs (RV40, RV60 MS40, and MS60) at
Ω=60000 rpm for example 2
Frequency (Hz) Relative error (%)
Mode
Whirl FOM RV40 RV60 MS40 MS60 RV40 RV60 MS40 MS60
No.
(N=189051) (n=40) (n=60) (n=40) (n=60) (n=40) (n=60) (n=40) (n=60)
1 FW2 927.519 1000.091 941.144 944.058 932.477 7.824 1.469 1.783 0.535
2 BW2 1021.668 1021.672 1021.671 1021.672 1021.671 0.000 0.000 0.000 0.000
3 FW1 -776.046 -776.044 -776.044 -776.042 -776.044 0.000 0.000 -0.001 0.000
4 BW1 1222.498 1222.504 1222.504 1222.507 1222.502 0.000 0.000 0.001 0.000
5 FW3 1041.398 1042.375 1042.210 1042.364 1041.987 0.094 0.078 0.093 0.057
6 BF3 1071.600 1071.626 1071.622 1071.628 1071.619 0.002 0.002 0.003 0.002
7 Shear 1127.129 1127.183 1127.141 1127.600 1127.527 0.005 0.001 0.042 0.035
8 FW4 1194.591 1195.464 1194.982 1195.041 1195.037 0.073 0.033 0.038 0.037
9 BW4 1211.609 1211.643 1211.633 1211.653 1211.651 0.003 0.002 0.004 0.003
10 FW5 1278.900 1279.207 1279.116 1279.158 1279.072 0.024 0.017 0.020 0.013
11 BW5 1284.620 1284.641 1284.636 1284.678 1284.672 0.002 0.001 0.004 0.004
12 FW6 1311.948 1312.046 1312.023 1312.043 1312.020 0.007 0.006 0.007 0.005
13 BF6 1313.773 1313.797 1313.792 1313.806 1313.803 0.002 0.001 0.002 0.002
14 Shear 1321.243 1321.266 1321.266 1321.277 1321.272 0.002 0.002 0.003 0.002
Norm of the relative error (ER) (%) 7.825 1.472 1.787 0.540

Table 8 The comparison of the total time computing for


allowable limit value (1%). Campbell diagram analysis produced by the FOM and
Likewise, the ERs computed by the MQSRV method are ROMs for example 2
also compared with those produced by the QSRV method Methods Computation time (s)
with many different scenarios, as seen in Fig. 12(b). When FOM (𝑁 = 189051) 57075 s 100%
reaching Ω=1000 (Hz), the errors are found at 7.83%,
RV60 278 s 0.49%
2.22%, 1.49%, 8.0%, 7.75%, 2.77%, 7.87%, 7.6% and
RV50 248 s 0.43%
6.91% by the QSRV and the MQSRV methods having 1, 2
RV40 224 s 0.39%
and 3 central frequencies with the 40, 50 and 60 total bases,
QSRV60 278 s 0.49%
respectively. Thus, from the computational perspective, the
QSRV50 251 s 0.44%
QSRV method is dominant over the MQSRV method.
Also, Fig. 12(c) compares the ERs calculated by the QSRV40 227 s 0.40%
QSRV and the Ritz vector (RV) methods. It is observed that MQSRV60_2caf 477 s 0.84%
the ER is drastically increased with the increase in the MQSRV60_3caf 666 s 1.17%
magnitude of the spinning velocities. Specifically, the errors MMQSRV60_2caf_2csf (rect form) 905 s 1.59%
are respectively reached to 7.83%; 7.83%, 2.22%; 1.59% MMQSRV60_3caf_3csf (rect form) 2065 s 3.62%
and 1.49%; 1.47% for the QSRV and RV method for the MMQSRV60_3caf_3csf (diag form) 1134 s 1.99%
total bases of 40, 50 and 60, when Ω=1000 (Hz). MMQSRV60_4caf_4csf (diag form) 1763 s 3.09%
Consequently, it is observed that the RV method provides RV10+MS30_1caf 896 s 1.57%
better results for the Campbell diagram analysis. RV10+MS40_1caf 1000 s 1.75%
Furthermore, Fig. 12(d) portrays the analogy of the RV10+MS50_1caf 1025 s 1.80%
calculated errors produced by the RV method and the MS40 1047 s 1.84%
combined RV and modal superposition (MS) methods. MS50 1115 s 1.95%
From Fig. 12(d), it is found that the errors are respectively MS60 1186 s 2.08%
7.82%; 2.83%, 1.59%; 1.55% and 1.47%; 0.52% computed
by the RV method and the combined methods for the total
bases of 40, 50, and 60 when Ω=1000 (Hz). So, the ER spinning velocity. But, in general, the MS method is more
produced by the combined method is more stable than those suitable owing to its more consistent responses (see Table
produced by the RV method. Additionally, the comparisons 7). Consequently, Fig. 13 depicts the Campbell diagram
of the errors calculated by the MS method with the established by the MS method with 60 bases (MS60). It is
combined method are also provided in Fig. 12(e). The observed that the diagram produced by the MS60 is likely
computed results reveal that the errors are respectively the same as that produced by the FOM (see Fig. 11(b)).
1.78%; 2.83%, 0.56%; 1.55% and 0.54%; 0.52% given by Table 8 presents the comparisons of the computational
the MS method and the combined methods for 40, 50, and time in total for establishing the Campbell diagram
60 total bases when Ω=1000 (Hz). Therefore, the MS produced by the FOM and the various cases of the
method provides smaller errors. Besides, Fig. 12(f) also MMQSRV, MQSRV, QSRV, RV, MS, and the combined
compares the ERs calculated by the MS method and the RV methods. It is found that the FOM takes about 57075 s as
method. It is found that the MS method is dominant at high the ROMs consume respectively only 278 s, 248 s, 224 s,
spinning velocity and the RV method is dominant at low 278 s, 521 s, 227 s, 477 s, 666 s, 905 s, 2065 s, 1134 s, 1763
426 Ty Phuor and GilHo Yoon

increases, the calculated ER produced by the RV method is


significantly increased. On the contrary, the ER produced
by the MS method is smaller and more consistent
throughout the range of the applied rotational speed by
maintaining the values of the critical speed and the
stabilities.
In a nutshell, the proposed MMQSRV method is also
provided by introducing the rotational-speed-dependent
stiffness matrix into the Krylov subspace using the
multifrequency and multi-spinning frequency. It is found
that this proposed method is unsuitable for the Campbell
diagram analysis because the number of bases is
insufficient. However, this study provides an in-depth
insight into the knowledge of the model order reduction
techniques by using the Krylov subspace and modal
Fig. 13 The Campbell diagram established with the first 14 superposition. Obviously, for one or a few frequencies of
modes by the MS60 specific interest, the accuracy of the solution produced by
the Krylov subspace-based model order reduction is more
dominant. On the other hand, to solve the eigenproblem for
s, 896 s, 1000 s, 1025 s, 1047 s, 1115 s and 1186 s for the establishing the Campbell diagram with several frequencies
RV60, RV50, RV40, QSRV60, QSRV50, QSRV40, (typically greater than 10) of interest, the MS method is
MQSRV60_2caf, MQSRV60_3caf, MMQSRV60_2caf_2csf more dominant. Moreover, this MS method is easily
(rect form), MMQSRV60_3caf_3csf (rect form), implemented in the FE program. It is worth mentioning that
MMQSRV60_3caf_3csf (diag form), no attempt has been reported to compare the calculated
MMQSRV60_4caf_4csf (diag form), RV10+MS30_1caf, results using the MS method and the Krylov subspace-based
RV10+MS40_1caf, RV10+MS50_1caf, MS40, MS50 and MOR technique for the Campbell diagram analysis of the
MS60 methods. Likewise, it is observed that the 3D FE rotordynamic system in the literature.
computational time can be reduced up to around 96% as
compared to the FOM. Thus, from the engineering point of
view, all MOR methods are effective in terms of time
Acknowledgments
computing.
This work was supported by Korea Institute of Energy
Technology Evaluation and Planning (KETEP) grant funded
8. Conclusions by the Korea government (MOTIE) (2021202080026D,
Development of platform technology and operation
The Campbell diagram analyses of the 3D FE management system for design and operating condition
rotordynamic system are presented and compared in the diagnosis of fluid machinery with variable devices based on
form of calculating the norms of the relative errors AI/ICT).
produced by the full order model (FOM), the Krylov
subspace-based reduced order model (ROM), and the modal
superposition techniques. The computational algorithms for References
MOR are also provided, such as the modal superposition
(MS), the Ritz vector (RV), quasi-static Ritz vector Bladh, R., Castanier, M.P. and Pierre, C. (2001), “Component-
(QSRV), the multifrequency quasi-static Ritz vector mode-based reduced order modeling techniques for mistuned
(MQSRV), and the combined methods. The verification of bladed disks-Part 1: Theoretical models”, J. Eng. Gas Turbin.
the 3D FE rotordynamic is also presented, and reliable Power, 123(1), 89-99. https://doi.org/10.1115/1.1338947.
agreements are made between the current and the existing Castanier, M.P. and Pierre, C. (2002), “Using intentional mistuning
in the design of turbomachinery rotors”, AIAA J., 40(10), 2077-
results. All simulations are carried out by using the in-house
2086. https://doi.org/10.2514/2.1542.
codes in the framework of MATLAB. Chatelet, E., D’Ambrosio, F. and Jacquet-Richardet, G. (2005),
The numerical examples for producing the Campbell “Toward global modelling approaches for dynamic analyses of
diagram are presented to demonstrate the validity of the rotating assemblies of turbomachines”, J. Sound Vib., 282(1-2),
MOR methods. Based on the numerical facts presented in 163-178. https://doi.org/10.1016/j.jsv.2004.02.035.
this paper, one can conclude that the MOR techniques are Crawely, E.F., Ducharme, E.H. and Mokadam, D.R. (1986),
effective with respect to time computing and response “Analytical and experimental investigation of the coupled
accuracy because the projection matrix is calculated once bladed disk/shaft whirl of a cantilevered turbofan”, J. Eng. Gas
and preserved as a parameter for the whole MOR process. Turbin. Power, 108, 567-575. https://doi.org/10.1115/1.3239948.
Feiner, D.M. and Griffin, J.H. (2002), “A fundamental model of
Moreover, it is found that the accurate numerical results are
mistuning for a single family of modes”, J. Turbomach., 124(4),
subsequently decreased by the MS, RV, QSRV, MQSRV, 597-605. https://doi.org/10.1115/1.1508384.
and MMQSRV methods. In this sense, the RV and MS Genta, G. (2005), Dynamics of Rotating System, Springer Science
methods are dominant at low and high rotating speeds, & Business Media, New York, USA.
respectively. More precisely, when the spinning velocity Genta, G. and Gugliotta, A. (1988), “A conical element for finite
Model order reduction for Campbell diagram analysis of shaft-disc-blade system in 3D finite elements 427

element rotor dynamics”, J. Sound Vib., 120(1), 175-182. Okabe, A., Otawara, Y., Kaneko, R., Matsushita, O. and Namura,
https://doi.org/10.1016/0022-460X(88)90342-2. K. (1991), “An equivalent reduced modelling method and its
Gruttmann, F. and Wagner, W. (2004), “A stabilized one-point application to shaft-blade coupled torsional vibration analysis of
integrated quadrilateral Reissner-Mindlin plate element”, Int. J. a turbine-generator set”, Proc. Inst. Mech. Eng., Part A: J. Pow.
Numer. Meth. Eng., 61, 2273-2295. Energy, 205, 173-181.
https://doi.org/10.1002/nme.1148. https://doi.org/10.1243/PIME_PROC_1991_205_026_02.
Gu, J., Ma, Z.D. and Hulbert, M.G. (2000), “New load-dependent Petrov, E.P., Zachariadis, Z.I., Beretta, A. and Elliott, R. (2013),
Ritz vector method for structural dynamics analyses: Quasi- “A study of nonlinear vibrations in a frictionally damped turbine
static Ritz vectors”, Finite Elem. Anal. Des., 36, 261-278. bladed disk with comprehensive modeling of aerodynamic
https://doi.org/10.1016/S0168-874X(00)00036-6. effects”, J. Eng. for Gas Turbin. Power, 135(3), 032504-11.
Han, J.S. (2014), “Krylov subspace-based model order reduction https://doi.org/10.1115/1.4007871.
for Campbell diagram analysis of large-scale rotordynamic Phuor, T. (2020), “Development and application of three-
systems”, Struct. Eng. Mech., 50(1), 19-36. dimensional finite element interface model for soil-jack-up
https://doi.org/10.12989/sem.2014.50.1.019. interaction during preloading”, PhD Thesis, Universiti
Huang, S.C. and Ho, K.B. (1996), “Coupled shaft-torsion and Teknologi Petronas.
blade-bending vibrations of a rotating shaft-disk-blade unit”, J. Phuor, T., Harahap, I.S., Ng, C.Y. and Al-Bared, M.A.M. (2021a),
Eng. Gas Turbin. Power, 118, 100-106. “Development and of the skew boundary condition for soil-
https://doi.org/10.1115/1.2816524. structure interaction in three-dimensional finite element
Khulief, Y.A. and Mohiuddin, M.A. (1997), “On the dynamic analysis”, Comput. Geotech., 137, 104264.
analysis of rotors using modal reduction”, Finite Elem. Anal. https://doi.org/10.1016/j.compgeo.2021.104264.
Des., 26(1), 41-55. https://doi.org/10.1016/S0168-874X(96)00070-4. Phuor, T., Harahap, I.S.H. and Ng, C.Y. (2021b), “Bearing
Kirchgäßner, B. (2016), “Finite elements in rotordynamics”, capacity factors for rough conical footing by viscoplasticity
Procedia Eng., 144, 736-750. finite element analysis”, ASCE Int. J. Geomech., 22(1),
https://doi.org/10.1016/j.proeng.2016.05.079. 04021266. https://doi.org/10.1061/(ASCE)GM.1943-5622.0002256.
Koh, H.S., Kim, J.H. and Yoon, G.H. (2020), “Efficient topology Rahman, T. and Valdman, J. (2013), “Fast MATLAB assembly of
optimization of multicomponent structure using substructuring- FEM matrices in 2D and 3D: Nodal elements”, Appl. Math.
based model order reduction method”, Comput. Struct., 228, Comput., 219(13), 7151-7158.
106146. https://doi.org/10.1016/j.compstruc.2019.106146. https://doi.org/10.1016/j.amc.2011.08.043.
Lazarus, A., Prabel, B. and Combescure, D. (2010), “A 3D finite Saito, A., Castanier, M.P. and Pierre, C. (2009), “Effects of a
element model for the vibration analysis of asymmetric rotating cracked blade on mistuned turbine engine rotor vibration”, J.
machines”, J. Sound Vib., 329(18), 3780-3797. Vib. Acoust., Trans., ASME, 131(6), 061006-9.
https://doi.org/10.1016/j.jsv.2010.03.029. https://doi.org/10.1115/1.4000458.
Li, C., She, H., Tang, Q. and Wen, B. (2019), “The coupling She, H., Li, C., Tang, Q. and Wen, B. (2018), “The investigation of
vibration characteristics of a flexible shaft-disk-blades system the coupled vibration in a flexible-disk blades system
with mistuned features”, Appl. Math. Model., 67, 557-572. considering the influence of shaft bending vibration”, Mech.
https://doi.org/10.1016/j.apm.2018.09.041. Syst. Signal Pr., 111, 545-569.
Lim, S.H., Bladh, R., Castanier, M.P. and Pierre, C. (2007), https://doi.org/10.1016/j.ymssp.2018.03.044.
“Compact, generalized component mode mistuning Stephenson, R.W. and Rouch, K.E. (1993), “Modeling rotating
representation for modeling bladed disk vibration”, AIAA J., shafts using axisymmetric solid finite elements with matrix
45(9), 2285-2298. https://doi.org/10.2514/1.13172. reduction”, J. Vib. Acoust., Trans., ASME, 115(4), 484-489.
Ma, O. and Wang, J.G. (2007), “Model order reduction for impact- https://doi.org/10.1115/1.2930376.
contact dynamics simulations of flexible manipulators”, Sternchüss, A. (2009), “Multi-level parametric reduced models of
Robotica, 25, 397-407. rotating bladed disk assemblies”, École Centrale Paris.
https://doi.org/10.1017/S026357470600316X. Vest, T.A. and Darlow, M.S. (1990), “A modified conical beam
Martel, C. and Sánchez-Álvarez, J.J. (2018), “Intentional element based on finite element analysis: Experimental
mistuning effect in the forced response of rotors with correlations”, J. Vib. Acoust., Trans., ASME, 112(3), 350-354.
aerodynamic damping”, J. Sound Vib., 433, 212-229. https://doi.org/10.1115/1.2930515.
https://doi.org/10.1016/j.jsv.2018.07.020. Vollan, A. and Komzsik, L. (2012), Computational Techniques of
Mogenier, G., Baranger, T., Ferraris, G., Dufour, R. and Durantay, Rotor Dynamics with the Finite Element Method, CRC Press.
L. (2014), “A criterion for mode shape tracking: Application to Wagner, M.B., Younan, A., Allaire, P. and Cogill, R. (2010),
Campbell diagrams”, J. Vib. Control, 20(2), 179-190. “Model reduction methods for rotor dynamic analysis: A survey
https://doi.org/10.1177/1077546312463714. and review”, Int. J. Rotat. Mach., 2010, Article ID 273716.
Nandi, A. (2004), “Reduction of finite element equations for a https://doi.org/10.1155/2010/273716.
rotor model on non-isotropic spring support in a rotating Wang, S., Bi, C.X. and Zheng, C.J. (2019), “A reduced-order
frame”, Finite Elem. Anal. Des., 40(9-10), 935-952. model for the vibration analysis of mistuned blade-disc-shaft
https://doi.org/10.1016/S0168-874X(03)00121-5. assembly”, Appl. Sci., 9(22), 4762.
Nandi, A. and Neogy, S. (2001), “Modelling of rotors with three- https://doi.org/10.3390/app9224762.
dimensional solid finite elements”, J. Strain Anal. Eng. Des., Wang, S., Wang, Y., Zi, Y. and He, Z. (2015), “A 3D finite
6(4), 359-371. https://doi.org/10.1243/0309324011514539. element-based model order reduction method for parametric
Nelson, F.C. (2007), “Rotor dynamics without equations”, Int. J. resonance and whirling analysis of anisotropic rotor-bearing
COMADEM, 10(3), 2-10. systems”, J. Sound Vib., 359, 115-135.
Nelson, H.D. (1980), “A finite rotating shaft element using https://doi.org/10.1016/j.jsv.2015.08.027.
Timoshenko beam theory”, J. Mech. Des., Trans., ASME, Wang, S., Wang, Y., Zi, Y., Li, B. and He, Z. (2015), “Reduced-
102(4), 793-803. https://doi.org/10.1115/1.3254824. order modeling for rotating rotor-bearing systems with cracked
Nelson, H.D. and McVaugh, J.M. (1976), “The dynamics of rotor- impellers using three-dimensional finite element models”, J.
bearing systems using finite elements”, J. Mech. Des., Trans., Sound Vib., 355, 305-321. https://doi.org/10.1016/j.jsv.2015.06.037.
ASME, 98(2), 593-600. https://doi.org/10.1115/1.3438942. Wang, S., Zi, Y., Li, B., Zhang, C. and He, Z. (2014), “Reduced-
428 Ty Phuor and GilHo Yoon

order modeling for mistuned centrifugal impellers with crack


damages”, J. Sound Vib., 333(25), 6979-6995.
https://doi.org/10.1016/j.jsv.2014.07.009.
Wilson, E.L., Yuan, M.W. and Dickens, J.M. (1982), “Dynamic
analysis by direct superposition of Ritz Vectors”, Earthq. Eng.
Struct. Dyn., 10(6), 813-821. https://doi.org/10.1002/eqe.4290100606.
Yang, M.T. and Griffin, J.H. (2001), “A reduced-order model of
mistuning using a subset of nominal system modes”, J. Eng.
Gas Turbin. Power, 123(4), 893-900.
https://doi.org/10.1115/1.1385197.
Yoon, G.H. (2010), “Structural topology optimization for
frequency response problem using model reduction schemes”,
Comput. Meth. Appl. Mech. Eng., 199(25-28), 1744-1763.
https://doi.org/10.1016/j.cma.2010.02.002.
Yoon, G.H. (2012), “Toward a multifrequency quasi-static Ritz
vector method for frequency-dependent acoustic system
application”, Int. J. Numer. Meth. Eng., 89, 1451-1470.
https://doi.org/10.1002/nme.3301.
Yoon, G.H., Kim, J.H., Jung, K.O. and Jung, J.W. (2015),
“Transient quasi-static Ritz vector (TQSRV) method by Krylov
subspaces and eigenvectors for efficient contact dynamic finite
element simulation”, Appl. Math. Model., 39(9), 2740-2762.
https://doi.org/10.1016/j.apm.2014.10.059.
Yuan, J., Scarpa, F., Allegri, G., Titurus, B., Patsias, S. and
Rajasekaran, R. (2017), “Efficient computational techniques for
mistuning analysis of bladed discs: A review”, Mech. Syst.
Signal Pr., 87, 71-90. https://doi.org/10.1016/j.ymssp.2016.09.041.

PL

You might also like