You are on page 1of 11

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/266559270

Practical Applications of Singular Value Decomposition in Rotordynamics

Article  in  Australian Journal of Mechanical Engineering · January 2005

CITATIONS READS

12 1,284

5 authors, including:

William Foiles
BP
10 PUBLICATIONS   155 CITATIONS   

SEE PROFILE

All content following this page was uploaded by William Foiles on 26 December 2020.

The user has requested enhancement of the downloaded file.


Practical Applications of Singular Value Decomposition in Rotordynamics
C. Hunter Clouda, William C. Foilesb, Guoxin Lia, Eric H. Maslena and Lloyd E. Barretta
a
Department of Mechanical and Aerospace Engineering, University of Virginia, Charlottesville, Virginia 22904 USA
b
GE Energy Services, General Electric Company, Atlanta, Georgia 30339 USA

Singular value decomposition (SVD) is used extensively in the controls community to examine the
dynamic behavior of systems. SVD is one component of linear systems theory that has developed into a
very mature mathematical tool for assessing systems. One objective of this paper is to illustrate the
manner in which that large base of analysis can be brought to bear on both classical and emerging
rotordynamics problems. This paper reviews the mathematical fundamentals of SVD and addresses its
physical implications with respect to rotordynamics. To illustrate these physical concepts, simple rotor
systems are examined in terms of forced response and stability margin characteristics. Additional
examples are presented in which SVD is applied to more complex rotor systems for balancing, model
reconciliation, and model reduction objectives.

Keywords: Rotordynamics, Singular value decomposition, Balancing, Model reduction,


Model reconciliation, Stability Margin

1. Introduction

The field of rotordynamics has advanced greatly over the past thirty years. During the same period, linear systems
theory has also experienced significant growth, largely due to the controls community’s efforts toward modeling many
types of dynamic systems. Singular value decomposition (SVD) has emerged as a fundamental linear systems tool for
analyzing and optimizing complex systems. The controls community has identified its utility in determining the extreme
(worst and best) dynamics of multiple degree of freedom systems, an ability which exceeds the capabilities of normal
eigenvalues and eigenvectors. This ability also makes it useful for analyzing uncertainties in models, a factor of primary
concern in all fields. The modal analysis community has leveraged these capabilities as well.1 Unfortunately, it has
received little application within the rotordynamics community even though the SVD (1889)2 is older than the landmark
rotor analyses of Foppl (1895) and Jeffcott (1919).
This paper is intended to demonstrate the applicability of SVD to practical rotordynamics problems. By describing
its physical implications with respect to rotors, this paper hopes to introduce this powerful tool, which can improve the
analysis and understanding of rotor behavior while also streamlining the design process.

2. Mathematical Properties and Physical Implications

Before introducing the SVD, it is useful to review a few important mathematical definitions. A unitary matrix
satisfies the following relationship:

U TU = I for real U
(1)
U U =I
H
for complex U

where the superscript H refers to the Hermitian or complex conjugate transpose. This unitary property also implies that
the vector columns of the matrix U, u, be orthogonal and that each column have unit magnitude or

2
u = ∑ ui =1 (2)
i

where ⋅ denotes the Euclidean vector norm (or 2 norm). Another important vector norm is the infinity norm ⋅ ∞ ,
which establishes the vector’s element with largest magnitude. For a vector of n elements, the Euclidean norm and
infinity norm are related by

u ∞≤ u ≤ n u ∞
(3)
In interpreting the SVD, assumptions about vector norms are very important. For a rotordynamicist, the “right”
vector norm is neither the infinity norm nor the 2 norm. Examining a single axial plane with x and y response, the
plane’s radial clearance requires that x2 + y2 < clr2. In vector form, z = {x y}T, therefore, ⎜⎜z⎜⎜< clr becomes the
requirement, and it’s clear that the 2 norm is the natural engineering norm. However, if z contains several axial planes
where z = {x1 y1 x2 y2 … xN yN}T, then one desires a combination of the 2 and infinity norms, with the requirement
being xi2 + yi2 < clri2. This can be translated into an infinity norm through a scaled function, ri = (xi2 + yi2)/clri2.
Unfortunately, this mixed norm is not mathematically tractable, and so the most tractable norm remains the 2 norm.
Therefore, one must recognize that restricting the 2 norm to unit magnitude, as in equation (2), might be overly
restricting the sizes of individual elements. This rotordynamics dilemma with norms additionally suggests that a rotor’s
rotational degrees of freedom should be neglected when lateral motion is the vibration of interest.
With these properties reviewed, the SVD of a complex M × N matrix A of rank r is given by

A = UΣV H (4)

where U and V are each unitary and whose columns are designated left and right singular vectors, respectively.3 The
matrix Σ is always positive real and takes the form

⎡Σ 0⎤
Σ=⎢ r
0⎥⎦
(5)
⎣0
where Σr is a diagonal matrix containing the non-zero, positive singular values σi. These singular values are arranged in
descending order along the diagonal, or
σ1 ≥ σ 2 ≥ L ≥ σ r (6)

The number of singular values is equal to the rank, r, of the matrix. The singular values and singular vectors are all
related to the eigenproblems of the matrices AHA and AAH. The singular values are the positive square roots of the
eigenvalues of AHA. The left singular vectors, u, are the normalized eigenvectors of AAH, while the right singular
vectors, v, are the normalized eigenvectors of AHA.
Computation of the SVD is numerically more stable than the damped eigenvalue calculation for rotor systems. This
is attributed to the fact that the singular values are the real eigenvalues of the symmetric AHA problem, while a damped
rotor system is a non-symmetric eigen problem resulting in complex eigenvalues. Details of the SVD computation can
be found in Reference 4.
At first glance, it is difficult to conclude what physical meaning underlies the singular values and singular vectors.
Unlike the complex eigenvalues of a system that denote modes’ damping and natural frequencies, the singular values
are always positive real. To examine the physical implications of SVD, return to the original description of the SVD for
matrix A given by equation (4), where, for purposes of this discussion, A is a square matrix representing some system.
Invoking the unitary property of V, this can be rewritten

AV = UΣ (7)

For each column, this becomes the classic input-output relationship of a multiple degree of freedom system.

[A]{v i }= σ i {u i } (8)

From a rotordynamics perspective, the input vector {v} can be thought of as some force distribution like unbalance
along the rotor, and {u} is the response distribution. With both {u} and {v} having 2 norm unit magnitude, the singular
value σ represents the “gain” between the input and the output. Often designated σ , the largest singular value, σ1, has
the highest gain for any input direction (2 norm). In other words, out of all possible input directions (“force
distributions”), the input distribution v creates the largest possible response according to distribution u . Conversely,
out of all possible input directions, the distribution v produces the smallest response ( σ r = σ ) in the direction u.
Reference 5 presents a very clear, simple example of these principles:

Consider a shopping cart (supermarket trolley) with fixed wheels which we may want to move in three directions:
forwards, sideways, and upwards. . . The strongest direction, corresponding to the largest singular value, will clearly
be in the forwards direction. The next direction, corresponding to the second singular value, will be sideways.
Finally, the most “difficult” or “weak” direction, corresponding to the smallest singular value, will be upwards
(lifting up the cart).

Mathematically, we can summarize this by the following relationship:


Av (9)
σ≤ ≤σ
v

which states that the gain of the system for any input is bounded by the maximum and minimum singular values.
Often, systems are described as being ill-conditioned. This implies that the output amplitude or gain is strongly
dependent on the input direction. The condition number, γ , quantifies this by taking the ratio of the maximum and
minimum singular values. An ill-conditioned system has γ >> 1 , while a well-conditioned system has γ ≈ 1 . The shopping
cart example is an ill-conditioned system in that any slight perturbation of a sideways force will suddenly cause the cart
to go forwards.
It is logical to question the relationship between singular values and eigenvalues. Returning to the input-output
relationship, the eigenvalue problem takes the form
[A]{ φi } = λ i { φi } (10)
where λi is an eigenvalue of A and φi is its associated eigenvector, or mode shape. When comparing this to singular
values and singular vectors, one sees that the eigenvalues measure “gain” for the special case in which the input and
output have the same direction or distribution (i.e., the eigenvectors).5 Thus, the following important relationship
between singular values and eigenvalues exists:

σ( A ) ≤ λ i ( A ) ≤ σ ( A ) (11)

The maximum and minimum singular values bound the magnitude of all the eigenvalues.

3. SVD and Simple Rotor Systems

Familiar to many in the rotordynamics field, the point mass or rigid rotor mounted in a fluid film bearing has been a
classic rotor system example since its introduction by Kapitsa.6 For simplification purposes, it is assumed that the
stiffness and damping properties of the bearing are constant with speed and frequency. Whirling motion of the mass is
created due to forces in the horizontal and vertical directions. Similar to some bearing test rigs, these directional forces
can be independent in magnitude and phase, but in this simple analysis, their frequency is equal to the rotational speed
of the rotor. The dynamic equations of motion for this two degree of freedom system become

⎡m 0 ⎤ ⎧ &x&⎫ ⎡c xx c xy ⎤ ⎧ x& ⎫ ⎡ k xx k xy ⎤ ⎧ x ⎫ ⎧ f x ⎫
⎢ 0 m⎥ ⎨ &y&⎬ + ⎢c ⎨ ⎬+
c yy ⎥⎦ ⎩ y& ⎭ ⎢⎣k yx
⎨ ⎬=⎨ ⎬
k yy ⎥⎦ ⎩ y ⎭ ⎩ f y ⎭
⎣ ⎦ ⎩ ⎭ ⎣ yx (12)

Assuming harmonic forced response, the frequency response function (FRF) matrix of the system is simply

[
H (ω ) = − ω 2 M + jωC + K ]
−1
(13)

Since the FRF is frequency dependent, the singular values and singular vectors of the system will also have frequency
dependence. This dependence is shown in Figure 1, which is essentially a Bode plot of singular values, σ(ω) and σ (ω ) ,
often called a sigma plot. Much of the same information
10
0 contained in a Bode plot is revealed in the sigma plot.
Max Singular Value The peaks of σ (ω ) correspond directly to the natural
Min Singular Value
frequencies of the system. Furthermore, the asymmetry
of the system is revealed by the split natural frequencies.
The upper curve represents the worst possible major axis
RIGID ROTOR SYSTEM PARAMETERS
m = 17,527 kg response due to any unit force distribution. In other
words, there is no possible unit force distribution that
Displacement (mm)

-1
10 c = c = 5.25 N-sec/mm xx yy
c =c =0 xy yx

k = 105 N/mm
will cause the response orbit’s major axis to exceed the
bound established by σ (ω ) . Applying a few sample
xx
k = 175 yy
k = -k = 26.25 xy xy

force distributions to this rotor system will illustrate this


concept. The following four unit force distributions were
applied (Figures 2-5 present the rotor responses):
-2
10

⎧1⎫ ⎧0⎫ ⎧ .707 ⎫


Fh = ⎨ ⎬ Fv = ⎨ ⎬ Fc = ⎨ ⎬ Fmax = v (ω )
⎩0⎭ ⎩1⎭ ⎩.707 j ⎭
1 2 3 4 5 6 7 8 9 10
Excitation Frequency (rad/s)

Figure 1. Sigma plot of rigid rotor system


100 100

M
M Major Axis M M Major Axis
M
M Max SV M
Max SV
M
X M X
M
M Y Y
M
M
M
Displacement (mm) 10-1 10-1

Displacement (mm)
M M M
M
M M M
M
M M M
M M
M M
M M
M M M
M M
M M
M
M M
M M
M M
M
M M
M M
M M
M M
10-2 M
MM 10-2 M
M
MM MM
MM MM
MM MM
MM MM
MM MM
MM MM
MM MM
MM MM
MM MM
MM MM
MM M
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Excitation Frequency (rad/s) Excitation Frequency (rad/s)

Figure 2. Response due to Fh Figure 3. Response due to Fv


100 100

M
M
M Major Axis M
M M Major Axis
M Max SV M Max SV
M
M X X
M Y M Y
M
M
M M
10-1 10-1
Displacement (mm)

Displacement (mm)
M M M
M
M
M M M
M M
M M
M M M
M M
M M M
M M
M M
M M
M M
M
M M
M M
M M
-2 M -2 M
10 M 10 M
M
M MM
M MM
M
M MM
MM MM
MM MM
MM MM
MM MM
MM MM
MM MM
MM MM
M
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Excitation Frequency (rad/s) Excitation Frequency (rad/s)

Figure 4. Response due to Fc Figure 5. Response due to Fmax


For purely horizontal and vertical force distributions, Fh and Fv, the associated natural frequency is excited in that
direction. For the circular force, Fc, which has equal horizontal and vertical magnitudes that are 90° out of phase like
unbalance, both natural frequencies are excited. However, in all three cases, the major axis of the response orbit does
not exceed the maximum singular value bound, a convenient fact, since this is the primary amplitude of interest for the
rotor designer. Figure 5 shows that when the worst case force distribution, Fmax, is applied, the response achieves the
highest possible amplitudes throughout the frequency range. This is
1
expected because Fmax corresponds to the worst input direction as
0.75 determined by the SVD. Looking at the SVD for a specific frequency,
0.5 one can more clearly examine its implications and relationships. Figure 6
0.25
displays the “orbit” for the unit forcing function defined by the
maximum singular input vector, v , or Fmax at 2.6 rad/s. The forcing is
vy = Fy (N)

0
predominantly in the horizontal direction. The corresponding maximum
-0.25 singular output vector, u , is shown in its orbital form in Figure 7. Here
-0.5 the response is dimensionless, yet the “direction” of the response is
defined. The response orbit is elliptical, as expected due to the stiffness’s
-0.75
asymmetry. Furthermore, the predominant response is horizontal, which
-1
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1 is also expected due to the forcing displayed in Figure 6. The actual
v = F (N)
displacement response is shown in Figure 8. This is obtained by
x x

Figure 6. Fmax(2.6 rad/s) forcing “orbit” multiplying the orbital response in Figure 7 by the maximum singular
1 0.5

0.4
0.75
0.3
Vertical Displacement Y (mm)

0.5
0.2
0.25
0.1
uy (dim)

0 0

-0.1
-0.25
-0.2
-0.5
-0.3
-0.75
-0.4

-1 -0.5
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
ux (dim) Horizontal Displacement X (mm)

Figure 7. Unit response orbit at 2.6 rad/s Figure 8. Actual response orbit at 2.6 rad/s
value, σ (2.6) = .0192 . Thus, the singular value simply becomes a gain or scaling factor for determining the actual
response level as shown in equation (8). Comparing Figures 5 and 8, the responses agree. It is important to remember
that this information is a direct product of the SVD; no exhaustive search to determine the worst-case response and its
associated worst-case forcing input has been necessary.
Such benefits of the SVD become even more significant when applied to more complex rotor systems. The three
mass rotor supported on fluid film bearings is a simple but realistic model for demonstrating this point. Figure 9 shows
this rotor system. While the rotational displacements are not included in the dynamics of this simple system, their
influence should be included in a real rotor’s FRF relationship. For simplification purposes, it is assumed that the
bearings are identical and have constant stiffness and damping properties. As in the rigid rotor model, any applied
forces can be independent in phase and magnitude but are synchronous with rotor speed.
Shaft Characteristics : 10-1
D = 89 mm L = 122 cm E = 207 GPa
Masses:
W1 = W3 = 11.3 kg W2 = 45.4 kg y3
Bearing Stiffness :
y2 Max Singular Value
Min Singular Value
Kxx = Kyy = 11 MN/m 10 -2

Kxy = -Kyx = 17.5 N/m


Bearing Damping: D

Displacement (mm)
Cxx = Cyy = 3.5 kN-sec/m
Cxy = Cyx = 0 x3 -3
10
y1 ω

x2 10
-4

x1
L
10-5
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Excitation Frequency (rad/s)

Figure 9. Three mass rotor on flexible bearings Figure 10. Three mass rotor sigma plot

Upon performing an SVD on this rotor system’s FRF, H(ω), another sigma plot can be generated versus frequency.
Figure 10 displays the maximum singular value of H(ω) versus excitation frequency. The highly amplified first natural
frequency is clearly apparent, along with the well-damped second and third resonances. Because multiple axial planes
are present (not just a single plane as with the rigid rotor), Figure 10 is a conservative estimate of the worst-case major
axis response. It provides the designer with a conservative upper bound such that no plane’s major axis response will
exceed σ ( H(ω)) . In other words,

σ ( H (ω )) ≥ max( Major Axis i ) (14)


N

where N equals the total number of planes (in this case, three). By linearly scaling the input vector, v , to the
appropriate force level, such a property allows a quick evaluation of whether any vibration limits are being exceeded at
any point along the rotor. This can streamline the design process by eliminating the production of numerous Bode plots
for varying force distributions; instead, a single sigma plot suffices. To determine the actual worst-case major axis
response for a particular plane of motion, a transformation must be completed.

H r (ω ) = [Re{H (ω )} − Im{H (ω )}]2 N × 4 N


⎡ Hp x (ω ) ⎤ ⎡ H r (ω )i ⎤
Hpi (ω ) = ⎢ i y ⎥ =⎢ ⎥
⎢⎣ Hpi (ω )⎥⎦ 2 × 4 N ⎣ H r (ω )i +1 ⎦ 2 × 4 N
Each Hpi(ω) represents the individual rotor plane’s FRF (x and y) due to arbitrary forcing in any of the N planes.
Details behind this transformation can be found in References 7, 8. Performing an SVD on each particular plane’s
Hpi(ω) highlights the major axis response once again by examining the singular values. However, for practical
applications with many planes, an SVD is easier to interpret and analyze if the response and forcing are scaled to unity
prior to performing the SVD. This scaling demands engineering decisions concerning the acceptable amplitudes due to a
vibration specification, radial clearance limitation, or residual unbalance limits. By scaling these to unity, the SVD
{ }
identifies that these requirements are achieved when σ hˆ(ω ) < 1 , where ĥ(ω ) is the appropriately scaled FRF.
Demonstrating this scaling for the three mass rotor example, we assume a vibration limit at the bearings of 0.0254
mm p-p and 0.254 mm p-p at the center mass, and prescribe a force limit of 44.48 N in any plane in any direction.
Therefore, we define the following limit matrices:

Z1lim = Z 3lim = .0127[ I ]2× 2 Z 2lim = .127[ I ]2× 2


Flim = 44.48[ I ]12×12
The scaled FRF for the bearing plane and center mass plane becomes

Hˆ p1 (ω ) = [Z1lim ]−1 Hp1 (ω )[Flim ]


Hˆ p2 (ω ) = [Z 2lim ]−1 Hp2 (ω )[Flim ]

Figure 11 displays the maximum singular values for the individual planes. Each sigma plot curve represents the worst-
case scaled major axis displacement for that plane due to worst-case arbitrary forcing in any of the N planes. This
interpretation is identical to the rigid rotor situation except that other planes can introduce forcing for the multiple plane
rotor. For the bearings, the relative displacement exceeds 1 at the first natural frequency, signifying that the worst-case
forcing limited by Flim would exceed the vibration limits by a factor of three. For the second natural frequency, the
limits are not exceeded at the bearings. With a larger vibration limit at the center mass, no force combination based on
Flim will exceed the 0.254 mm p-p limit at any frequency.
The earlier discussion on vector norms should be kept in
10
1
mind especially when assessing individual plane sigma
SCALED OVERALL COMPLIANCE
plots. In this case, a 2 norm for the output is appropriate for
BEARINGS (Planes 1 & 3) the individual plane. However, the SVD also imposes a 2
10 0 norm on the force input vector. If the engineer believes that
the forces should be infinity norm bounded, then a scaling
Relative Displacement (dim)

factor of n should be applied to each curve.


MIDSPAN (Plane 2)
10 -1 Figure 11 also displays the sigma plot of the scaled
overall FRF, Hˆ (ω ) . The conservative designer could
streamline the analysis process simply by determining this
10
-2
curve which bounds the individual planar curves as well.
Bearings Major Axis Limit = 0.0254 mm Once again, if the either the input forcing or output
Midspan Major Axis Limit = 0.254 mm
Force Limit per Direction = 44.48 N response are more likely infinity norm bounded vectors, a
10 -3 scaling factor of n can easily be applied to this single
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Excitation Frequency (rad/s) curve.
Figure 11. Scaled sigma plots for individual planes Providing a wealth of information about the system
and overall 3 mass system dynamics, the SVD provides an efficient method of
analyzing the forced response of the machine without the
typical multiple case runs currently done during rotor analyses. The SVD directly provides the rotordynamicist with the
primary elements of the forced response picture: natural frequency locations, the force distributions necessary to excite
these natural frequencies to their highest levels, and the magnitudes and directions of this worst-case response. If
engineering limits are applied, as in the 3 mass rotor, quick conclusions can be reached regarding the acceptability of
the machine’s forced response.

4. Balancing

As shown in the preceding section, the SVD provides many benefits with respect to forced response performance.
This translates directly into a useful tool for the balancing of machinery. An example application is presented here.
The balance influence coefficients for a 75 MW gas turbine were developed using a regression model of both
factory and field test data. This estimate for the influence coefficients at running speed is used to balance these
turbines. The two columns of the influence coefficient matrix correspond to balance planes on either end of the turbine.
The balance radius does not vary on either plane. Therefore, it is customary to use units of mass instead of mass-length
for the coefficients. Both shaft relative data and seismic data were used to generate the coefficients, but only the shaft
relative information is shown. Table 1 presents coefficients in polar coordinates; their amplitudes represent the
magnitude of the response in µm to 1 kg of mass at the respective balance radius.
The singular values for this 4 × 2 complex valued matrix of influence coefficients are 1354.49 and 222.10. The
singular value decomposition yields two balance weight distributions corresponding to the singular input vectors.
Because of the relative magnitudes of the singular values, the response of a unit balance weight (2 norm on the weights
in all planes) corresponding to the first singular value will have a response 610% more than the response of a unit
weight in the direction corresponding to the second singular value. Table 2 presents the balance weight distributions
determined by the SVD. The responses due to V1 and V2, when corrected for the probe angles, appear as mostly in-
phase and out-of-phase response on the shaft relative probes. Thus, these weights have similarities with modal balance
weights (in particular “static” and “couple” pairs that are used for balancing). While V1 is an in-phase pair, the
magnitudes are unequal on the two ends. V2 is an out-of-phase pair.
Table 1 Table 2
Gas Turbine Balance Influence Coefficients (µm/kg) SVD Balance Weight Distributions
Bearing Balance Plane 1 Balance Plane 2 Balance Plane V1 V2
1X 747.68 ∠ 61° 227.39 ∠ 50° 1 0.9763 ∠ 0° 0.2164 ∠ 180°
1Y 791.48 ∠ 145° 246.86 ∠ 133° 2 0.2164 ∠ 359° 0.9763 ∠ 359°
2X 449.57 ∠ 274° 72.14 ∠ 346°
2Y 437.87 ∠ 343° 75.39 ∠ 65°

Because of the magnitudes of the singular values, the practical implication of using these weight combinations is
that a weight with shape V1 does six times the work of a weight in the shape of V2 of equal magnitude (2 norm). This
situation is not unusual when performing multiplane balancing of flexible rotors.
Vibration with significant components matching the weakest output singular vector, u, cannot be easily balanced.
Recall the example of lifting up the shopping cart. These planes can require large balance weights to correct for even
relatively small magnitudes of vibration. This could imply that one has chosen an inappropriate or insufficient set of
balance planes. It could also mean that the vibration does not result from unbalance.

5. Model Reduction and Reconciliation

A rotor model is usually developed by finite element or transfer matrix method. The resulting model size is often
large. A lower order model is sometimes preferred to reduce the computation cost, especially in the design stage, when
optimization is performed. In addition, the model’s higher modes can possess inaccuracies due to the inherent
assumptions in the model and the numerical difficulties involved. With these factors in mind, the objective of model
reduction is to obtain a lower order model to preserve the essential properties of the original model for the design and
analysis. Singular values of certain transfer functions and matrices are employed as an important measurement to
achieve these objectives. The model reduction error can be measured in different transfer function matrix norm
e = G o − G r = ∆(ω) (15)

where Go is the transfer matrix of original full size model and Gr is the reduced size model. We determine error using
the H∞ norm: the largest maximum singular value of this function over all frequencies is defined by

∆(ω) ≡ max[σ{∆(ω)}] (16)


∞ ω
10
-3 Many model reduction techniques have been developed in the
linear system theory. Modal truncation is also used in the
-5
rotordynamic community. The basic assumption is that the response
10
contribution of truncated high modes is small. Modal truncation
Hankel Singular Value

usually works well if enough modes are retained. The number of


10
-7
modes to retain often is determined only by the engineer’s
experience. Use of more systematic approaches such as the balanced
10
-9 reduction9 and Hankel norm approximation10 provide an error
bound. In these methods, singular values are severed, providing the
engineer with a measurement criterion for reduction. To perform
-11
10
balanced reduction or Hankel norm approximation, the second order
matrix system is reformulated into first order state space form. With
10
0 10 20
-13
30 40
certain responses chosen, the singular values, in this case called
Mode Number Hankel singular values, are determined for each mode of the system.
Figure 12. Hankel singular values for a Figure 12 shows the Hankel singular values of a typical rotor.
rotor’s modes Figure 12 clearly shows that the first four rotor modes are the
most influential given the magnitude of their Hankel singular
values. Similarly, it indicates that the fourth mode is just as important as the third mode. With this information at hand,
the engineer can select the number of modes to retain, p. If the reduced model has order p, the reduction error is
bounded by the singular values of the truncated modes according to:
e = Go − G p ≤ 2(σ p +1 + σ p + 2 + ... + σ N )

The balance truncation based on Hankel singular values may not provide the best error bound when a certain order
of reduced model is obtained. The Hankel norm technique can reduce the error. Figure 13 provides an interesting
comparison of the three reduction techniques for a high-speed rotor with significant gyroscopics. The original model
contains 24 modes and was used for comparison. Instead of using four Bode plots for each channel to show the
-3
10 Full Model
Balanced
reduction error, a single sigma plot was chosen to illustrate the
Hankel
reduction errors for this rotor. Modal truncation, balanced reduction
Modal

10
-4
and Hankel norm approximation were used to obtain a reduced order
model of 16 modes at 20,000 rpm. Modal truncation does not provide
a good solution because the gyroscopic matrix is truncated according
σ1(ω)

10
-5
to the undamped orthogonal eigenvectors. Balanced reduction gives
much better results over a certain frequency range, but one higher
mode is missing. The Hankel norm reduction matches many natural
10
-6
frequencies of the modes but not the response between them.
In assessing the veracity of a model, experimental data from the
actual machine is compared to predictions from the model.
10
-7
5000 10000 Typically, frequency response functions are obtained by striking or
Frequency (rad/s) shaking the machine at specific locations, fi and then measuring the
Figure 13. Reduction error comparison for a response at other locations, zIDi, using accelerometers or
high-speed gyroscopic rotor displacement probes. To evaluate the model, the error between the
experimental and predicted response is formed using the
corresponding model and experimental frequency response functions:

{e(ω )} = {zm (ω )} − {z ID (ω )} = [H m (ω ) − H ID (ω )]{ f (ω )} (17)

Note that HID may indicate an actual extracted matrix function, or it may instead simply represent the data at the
frequency ω. Therefore, it isn't critical to this analysis to extract the transfer matrix from the data (by curve fitting11)
although there may be some analysis advantages to doing so.
To evaluate this error, we want to find the magnitude of the worst-case error, the largest magnitude of {e} for all
feasible force inputs, {f}. One possibility would be to simply try a series of feasible arrangements of magnitude and
phase of the vector f to see which produces the largest {e} or one might even conduct a Monte Carlo test where
thousands of random feasible forces are constructed, and the error assessed for each. However, this is not necessary
since the largest magnitude of {e} for all feasible {f} is a maximum singular value problem.
Assuming that force bounds |fi|<Fi can be established, define a new variable pi = fi/Fi so that |pi|<1. As done earlier,
applying the diagonal force limit matrix F, the scaled error becomes
[
{e(ω )} = ]
H m (ω ) − H ID (ω ) [ F ] p
(18)

Further, assume that the maximum acceptable predictive error at each measured location is of the order Zi; that is, if |zm,i
- zID,i|<Zi, then the model is acceptably accurate. Thus, if |ei/Zi|<1, then the model is acceptable: |Z-1e|<1 where, again,
the matrix Z is a diagonal matrix of the Zi. Thus, similar to the 3 mass rotor example, the desired outcome is

Z −1e(ω ) = Z −1[H m (ω ) − H ID (ω )]Fp (19)

to be less than 1 for all possible p: |p|<1. The right side defines the induced norm of the matrix function Z-1[Hm(ω)-
HID(ω)]F which, as described earlier, is the maximum singular value of this matrix function. Thus, if

{
σ Z −1[H m (ω ) − H ID (ω )]F < 1 } (20)

for all of the frequencies represented in the test data set, then the model is acceptably accurate for these frequencies.
Apparently, we must evaluate the complex matrix Hm at each frequency for which we have data for HID , take the
indicated difference, perform the indicated F and Z scaling, and then compute the singular values of the resulting
complex matrix. If the largest singular value exceeds 1 at any test frequency, then the model is not acceptably accurate.
However, if HID is available in a functional form and if the range of test frequencies is very wide, then we can evaluate
the largest maximum singular value of this function over all frequencies with a single computation, the H∞ norm of the
matrix operator per equation (16).

6. Stability Margin

The concept of stability margin has been used in many contexts within rotordynamics. For this discussion, stability
margin refers to the ratio of the threshold level of destabilizing cross-coupled stiffness to the expected level. The newly
adopted API stability analysis requirements incorporate this ratio.12 While this stability margin and the associated
threshold level are often predicted analytically, it is difficult to actually verify this margin in the field or on the test
stand. The SVD provides a potential solution to this problem. To illustrate this stability margin analysis using SVD, the
classic Jeffcott rotor is employed. With rigid bearings and a centrally located mass on a flexible rotor, the model is
modified with asymmetric viscous damping, C, acting on the mass, along with forward whirl destabilizing skew
symmetric cross-coupled stiffness, Q. This rotor system takes the form
⎡m 0 ⎤ ⎡.75c 0⎤ ⎡k shaft 0 ⎤ ⎡ 0 q⎤
M =⎢ ⎥ C=⎢ ⎥ K=⎢ ⎥ Q=⎢ ⎥
⎣ 0 m⎦ ⎣ 0 c⎦ ⎣ 0 k shaft ⎦ ⎣− q 0 ⎦
Through a damped eigenvalue analysis, a stability sensitivity curve can be determined by varying the magnitude of
Q present on the rotor. This is a typical design analysis procedure on machines and allows the stability margin to be
assessed. Figure 14 shows such a sensitivity plot for a specific Jeffcott rotor where the threshold level of cross coupling
is 646.3 N/mm.
0.2

646.3 N/mm

0.15
Stability Threshold
Forward Mode Log Decrement

0.1

0.05

UNSTABLE

-0.05
0 200 400 600 800
Cross Coupled Stiffness q (N/mm)

Figure 15. Cross coupled stiffness as feedback


Figure 14. Asymmetric Jeffcott Rotor
Forward Mode Stability Sensitivity

Therefore, our objective is to verify this threshold level based on test stand or field measurements obtained when
some lower (and stable) level of cross coupling is present. First, frequency response functions of the rotor need to be
obtained utilizing a nonsynchronous excitation device similar to those used to measure the stability level of
machinery.13,14 These measured FRF’s represent the closed loop performance of the system, where Q provides a
positive feedback mechanism as shown in Figure 15. Representing destabilizing phenomena in rotors as feedback
mechanisms has been a successful approach not only for cross coupling13 but also for synchronous instabilities.15 In
terms of system parameters, this closed loop frequency response is described by

Z (ω ) Displacement
H (ω ) = = = G (ω )[I − QG (ω )]−1 where G (ω ) = (−ω 2 M + jωC + K ) −1
F (ω ) Ext. Forcing (21)

Upon performing a modal analysis curve fit on the H(ω) test data,11 the next step in the procedure requires a
determination of Q. In reality, Q would be estimated through the use of either a physical model of destabilizing
stiffness like a seal code, an empirical technique,12 or perhaps through a more advanced model reconciliation
technique.16 With Q and H(ω) available, the loop transfer function QG(ω) can be extracted using system identification
techniques. The loop transfer function is of primary interest because of the requirements of the small gain theorem.17 A
more general case of Bode’s gain margin concept,18 this theorem requires that the “magnitude” of QG(ω) be less than 1
for the system (closed loop) to remain stable. Mathematically, this stability requirement is expressed as the following:
Small Gain Theorem : QG (ω ) < 1 for all ω
10
1 Physically, this theorem implies that the any forces should not
JEFFCOTT ROTOR PARAMETERS be amplified due to the presence of the feedback Q, i.e., the
m = 49.9 kg
kshaft = 9060 N/mm
c = 1.75 N-sec/mm
gain must be kept small. Applied to a matrix, not a vector, this
q = 87.5 N/mm
“gain” or “magnitude” of QG(ω) must be an induced matrix
100
norm5 providing the largest possible amplifying power. From
Stability Margin our previous discussion, the maximum singular value provides
LOOP GAIN (dim)

0.156 just such a measure.


10-1 Performing an SVD on the test data derived QG(ω), one
can plot the maximum singular value versus frequency in a
sigma plot. Figure 16 displays the sigma plot for the specific
10
-2 Jeffcott rotor example. Based on the small gain theorem, the
distance from the peak in the sigma plot to 1.0 provides a
measure of the relative stability of the system, i.e., the stability
10
-3
margin. The stability margin is therefore
200 400 600 800 1000
q 1 1
SM = 0 = =
Excitation Frequency (rad/s)

Figure 16. Sigma plot of QG(ω) for asymmetric q max[σ (QG (ω )] QG (ω ) ∞


ω
Jeffcott rotor
Using the peak loop gain from Figure 16, the stability margin for the asymmetric Jeffcott rotor is 6.41. With the applied
level of q equal to 87.5 N/mm, the estimated threshold level q0 is 6.41 × 87.5 = 560.88 N/mm which is less than the
analytically predicted threshold of 646.3 N/mm. Because only a matrix norm is considered, not the structure of Q, the
small gain theorem always results in a conservative estimate of the stability margin. More advanced analysis techniques
such as the structured singular value19 enable a more accurate estimation of stability margin and threshold. Currently,
research is underway to verify these experimentally and determine their applicability to stability margin testing on
industrial turbomachinery.

7. Conclusions

The SVD is a powerful mathematical tool being leveraged by many fields which, physically, yields the extremes of
a system’s input and output. The applications described above show its utility with respect to many rotordynamics
issues. In particular, the SVD can streamline forced response analysis of rotor systems by providing critical speed
locations as well as worst-case unbalance distributions and responses without the need of multiple analysis iterations.

8. References

1. Maia NMM. Fundamentals of singular value decomposition. Proceedings of the Ninth International Modal
Analysis Conference 1991; 2: 1515-1521.
2. Sylvester. Messenger of Mathematics 1889; 19: 42.
3. Eckart C and Young G. A principal axis transformation for non-Hermitian matrices. Bulletin of the American
Mathematical Society 1939; 45: 118-121.
4. Press WH, Teukolsky SA, Vetterling WT and Flannery BP. Numerical Recipes in Fortran: The Art of Scientific
Computing. New York, NY: Cambridge University Press, 1992.
5. Skogestad S, Postlethwaite I. Multivariable Feedback Control Analysis and Design. Chichester, UK: John Wiley
& Sons, 1996.
6. Kapitsa PL. Stability and transition through the critical speed of fast rotating shafts with friction. Journal of
Theoretical Physics 1939; 9(2).
7. Little RM. The application of linear programming techniques to balancing flexible rotors. Ph.D. Dissertation.
School of Engineering & Applied Science, University of Virginia, 1971.
8. Foiles WF, Allaire PE and Gunter EJ Advanced techniques for balancing and forced response of rotor bearing
systems. ROMAC Laboratories Annual Meeting, Charlottesville, Virginia, 1999.
9. Moore BC. Principal components analysis in linear systems: controllability, observability and model reduction.
IEEE Transactions on Automatic Control 1981; AC-26: 17-31.
10. Glover K. All optimal Hankel-norm approximations of linear multivariable system and their L∞ error bounds.
International Journal of Control 1984; 39: 1115-1193.
11. Richardson MH and Formenti DL. Global curve fitting of frequency response measurements using the rational
fraction polynomial method. Proceedings of the Third International Modal Analysis Conference, Orlando, 1985:
390-397.
12. API 617. Axial and Centrifugal Compressors and Turboexpanders for Petroleum, Chemical and Gas Industry
Services, 7th Edition. Washington D.C.: American Petroleum Institute, 2002.
13. Bently DE and Muszynska A. Stability evaluation of rotor/bearing system by perturbation tests. NASA CP-2250
1982: 307-322.
14. Baumann U. Rotordynamic stability tests on high-pressure radial compressors. Proceedings of the Twenty-Eighth
Turbomachinery Symposium, Texas A&M University, 1999: 115-122.
15. Keogh PS and Morton PG. The dynamic nature of rotor thermal bending due to unsteady lubricant shearing within
a bearing. Proceedings of the Royal Society, London 1994; 445(1924): 273-290.
16. Sortore CK. Observer based critical response estimation in rotating machinery. Ph.D. Dissertation. School of
Engineering & Applied Science, University of Virginia, 1999.
17. Zames G. On the input-output stability of time-varying nonlinear feedback systems-part I: Conditions derived
using concepts of loop gain, conicity, and positivity. IEEE Transactions on Automatic Control 1966; AC-11: 228-
238.
18. Bode HW. Relations between attenuation and phase in amplifier design. Bell System Technical Journal 1940; 19:
421-454.
19. Doyle JC. Analysis of feedback systems with structured uncertainties. IEEE Proceedings 1982; 129(6): 242-250.

View publication stats

You might also like