You are on page 1of 8

23rd International Congress on Sound & Vibration

Athens, Greece 10-14 July 2016


ICSV23
IMPLEMENTATION OF THE VOLD-KALMAN ORDER
TRACKING FILTERS FOR ONLINE ANALYSIS
Martin ala and Petr Bene
Brno University of Technology, Department of Control and Instrumentation, Brno
email: xcalam00@stud.feec.vutbr.cz

The article presents the implementation of the Vold-Kalman order tracking lters of both rst
and second generation. The introduction examines the principle of both lters with the necessary
theoretical background needed to run the lter in the online manner. Attention is also focused
on the calculation optimizations in the Cholesky decomposition and preconditioned conjugate
gradients methods as well as an effective storage of large sparse matrices. The algorithms are
implemented in the LabVIEW environment. Feasibility of the presented implementation is veried
by the timing benchmarks, results of the order tracking using synthetic signals and additionally
the practical experiment on a common cooling fan with the ball bearing.

1. Introduction
Vold-Kalman lters are frequently used in vibrodiagnostics as tool for order analysis. The purpose
of such tool is to extract vibration signal of dened time-varying frequency called order for further
analysis. Theoretical base was presented by multiple authors [1], [2], or [3]; practical aspects,
especially lters bandwidth, were mentioned in [4], [5], or [6]. The lters can be applied on both
vibration or acoustics signals [7]. Additionally, lters need precise information about the instantaneous
rotor speed measured by either a tacho probe or an incremental encoder. Methods eliminating the
tacho probe also exist but have its limitations [8].

2. Vold-Kalman filters
Vold-Kalman lters are used as order tracking technique capable of extracting sinusoidal signals
of the time-varying frequencies called orders from the vibration signal of rotary machines. The lters
are of two generations. First one (VKF1), described by Vold and Leuridan [3], extracts waveform
corresponding with the order frequency. The second generation (VKF2) considers signal to be
modulated and is capable of extracting order envelope. Both lters do not add phase delay to the
extracted order which is benecial in diagnostics [4].
Structure of the Vold-Kalman lters is inspired by the Kalman lter, there are used structural
and data equations. The structural equation describes how the rotating motion is modulates into the
sinusoidal oscillations resulting in vibrations or an acoustic signal to be measured. Such signal might
contain more orders from more sources including the noise or other non-harmonic signals. This is
expressed by the data equation. The equations of VKF1 and VKF2 will be discussed together for
compactness.

1
The 23rd International Congress of Sound and Vibration

2.1 Structural equation


For VKF1, equation is represented as the difference equation of K-th order, where K = 2. Eq. (1)
describes how the order x is modulated in frequency and amplitude

xn + cn xn1 + xn2 = n , (1)

where n designates data samples up to N. Left side of Eq. (1) is xed and depends on the order phase
cn . To enable slight changes in amplitude and frequency, error parameter n is added on the right side.
Solution of Eq. (1) is harmonic function. The order phase cn uses order instantaneous frequency f and
sampling period Ts
c = [c1 , c2 , . . . cn ] = 2 cos(2fTs ) . (2)
VKF2 has variable lter order expressed by difference equation
K  
X n
K
xn = (1) i
xni = n , (3)
i=0
i

for example when K = 2, Eq. (3) becomes

2 xn = xn 2xn1 + xn2 = n . (4)

2.2 Data equation


Data equation for VKF1 simply sums all P orders xp to the measurement signal y, expresses
both measurement noise and all signals not originating from any order
P
X
y= xp + . (5)
p=1

In case of VKF2, orders are additionally modulated


P
X
y= xp ejp + , (6)
p=1

where p is cumulative phase calculated from particular order frequency fp using cumsum

p = 2Ts cumsum(fp ) . (7)

2.3 Filters solution


Both previously mentioned equations explains nature of the Vold-Kalman lters. Goal of the
Vold-Kalman ltration is to extract one or more orders xp from measured signal y and minimize error
terms in both structural and data equations using formula

J = T + T . (8)

Following equations are generalized for extraction of multiple orders P . If P = 1 lter extracts
only one order. Extraction using VKF1 describes equation

B1 E E x1 y
E B2 E x2 y
.. .. . . .. .. = .. , (9)

. . . . . .
E E BP xP y

2 ICSV23, Athens (Greece), 10-14 July 2016


The 23rd International Congress of Sound and Vibration

where
Bp = ATp RTp Rp Ap + E . (10)
Matrix Ap has dimension of N 2 by N and contains 3 non-zero diagonals, E is identity matrix and
Rp has one non-zero diagonal containing factors r for adjusting lter bandwidth (higher r results in
narrower bandwidth)

1 cp,1 1 0 0
0 1 cp,2 1 0
Ap = .. .. . . . .. , Rp = diag(rp,1 , rp,2, . . . , rp,N 2) . (11)

. . . . . . . . .
0 0 1 cp,N 2 1
Vector cp is calculated using Eq. (2).
VKF2 uses formally the same matrix equation as Eq. (9)

B1 CH1 C 2 C H
1 C P x 1 C H
1 y
CH C1 B2 CH CH y
2 2 CP x 2 2
.. .. .. .. .. = .. , (12)
. . . . . .
H H
CP C1 CP C2 BP xP CHPy

where Bp has the same form as Eq. (10), only matrix Ap uses Eq. (3) instead. Matrix Cp is calculated
using cumulative phase from Eq. (7)

Cp = diag ejp . (13)




Solutions of both lter generations in Eqs. (9) and (12) have common scheme with the sparse
matrix
Bx = v , (14)
dimensions of the matrix B are NP by NP , while dimensions of the vectors x and v are NP by 1.
Primary challenge is to invert the system matrix B. Such matrix system is commonly solved by the
Cholesky decomposition if the extraction of one order is needed. Decomposition requires B to be
positive denite. This is satised for matrices used in Vold-Kalman ltration, therefore it is possible
to nd upper triangular matrix U, which meets the equation
Bx = UT Ux = y . (15)
Decomposition of the matrix leads also to the decomposition of Eq. (15) using
UT z = y , Ux = z , (16)
which preserves matrix sparsity and is computationally efcient since only diagonal elements are
stored in the computer memory.
Situation is much more complicated when extraction of multiple orders is required, usually with
crossing orders. Cholesky decomposition cannot be used since the matrix B contains additional
elements outside main diagonal (visible in Eqs. (9) and (12) as elements E and CH x Cy respectively)
and calculation would not preserve sparsity of matrix B, it would become denser. One of the
usually employed method for sparse systems like Eq. (14) is a family of iterative algorithms based
on conjugate gradients [9]. Suitable variant, especially for solving extraction including crossing
orders, is preconditioned conjugate gradient method [10]. This method in each iteration partially
solves the system of Eqs. (9) or (12) directly using only matrices Bp similar to the extraction of
orders separately. Then the whole system is solved with physical representation of splitting the energy
between the crossing orders. Although the preconditioning increases computational requirements of
the one iteration, number of iterations as well as the total execution time are signicantly reduced.
Additional information brings [4].

ICSV23, Athens (Greece), 10-14 July 2016 3


The 23rd International Congress of Sound and Vibration

3. Online algorithm
An idea of using Vold-Kalman lters as an online tool already exists since Pan described adaptive
method of the ltration [11]. The method is implemented only for VKF2 and also the execution speed
is rather problematic for online analysis since an order from 5 seconds long signal sampled with the
frequency 2 kHz was extracted even more than 20 seconds.
Proposed online method do not process measured data in sample-by-sample basis but divides
continuous signal into data blocks of predened size with overlapping like showed in Fig. 1. Then one
or more orders are extracted from every block separately and results of adjacent blocks are merged
together. Since lters does not introduce phase shift to the extracted order, simple combination of
results is possible.

Figure 1: Scheme of separating the measured signal into data blocks, processing and merging using
overlapping.

All lters including VKF have number of basic properties, namely bandwidth and settling time. For
overlapping method, mainly the settling time is critical (highlighted with grey gradients on extracted
orders edges) because it must be signicantly shorter than the data block otherwise the merged signal
from multiple data blocks will not be smoothly connected and data distortion occurs. Both parameters
are handled by values of factors r described in Eq. (11). With increasing value of r, the bandwidth
is narrower and settling time becomes longer. Settling times can be signicantly suppressed by
overlapping, in Fig. 1 is overlapping 50 % which is maximum reasonable value resulting in doubled
computational requirements compared to standard (offline) lter.
Limitation in having rather wider bandwidth with shorter settling time seems signicant but when
the order amplitude changes rapidly, long settling time would distort extracted order shape. This
limitation is valid also for other order tracking methods.

3.1 Implementation details


There are multiple implementation of VKF lters, but mainly VKF2 is implemented, commonly
with ability to extract multiple orders [1], [4], or [10]. Matlab environment as data analysis tool
is usually employed. Since the ability to measure real-world signals using Matlab environment is
limited, combination of LabVIEW and congurable acquisition hardware was used.
The matrices contain only few non-zero elements and therefore it is effective to save them in some
sparse matrix storage formats. Since the matrices are strictly diagonal therefore the storage format
should be also diagonal to preserve effectiveness. This is very important since the matrix sizes are
huge (usually more than thousand). Ratio describing the number of non-zero elements divided by the
number of all matrix elements in matrix Bp is
(2K + 1)N
dnonzero = , (17)
N2

4 ICSV23, Athens (Greece), 10-14 July 2016


The 23rd International Congress of Sound and Vibration

numerically sparse storage format for VKF1 needs to store 5000 elements for N = 1000 samples
while dense matrix needs 1 million. Along with huge memory effectiveness comes also computational
efciency because numerical operations with zeros are not executed in sparse matrices.

3.2 Tests on synthetic signals

Both implemented VKF algorithms are veried by multiple signals. First one uses one linear sine
sweep from 50 Hz to 100 Hz with constant amplitude. Signal 2 contains additional 2nd and 3rd orders
(i.e. 100 Hz to 200 Hz and 150 Hz to 300 Hz). Third signal contains two crossing orders constant
sine with 100 Hz and sine sweep from 50 Hz to 150 Hz. Amplitude of all orders is 1 g. Every signal
contains 10 kS sampled at frequency 1 kHz. Factor r = 1000 and VKF2 of the second order were
selected for extraction.
Figure 2 shows proles of used synthetic signals. Figures 3 show results of extraction on signals 2
and 3 using online method with VKF2. It is clear that amplitude differences in the range 1 % are
only visible at the beginning and the end of the signal, the rest is correct. Such behaviour is typical for
standard VKF and shows that online method works the same. VKF1 was not included in tests because
the output signal in form of less readable waveform. However, both lters lead to comparable results.

300 150
order 1 order 1
order 2 140 order 2
order 3
250 130

120
rotor speed [Hz]

rotor speed [Hz]

200 110

100

150 90

80

100 70

60

50 50
0 2 4 6 8 10 0 2 4 6 8 10
time [s] time [s]

Figure 2: Speed proles of test signals 1 and 2 (left), and 3 (right).

1.01 1.01
order 1 order 1
1.008 order 2 1.008 order 2
order 3
1.006 1.006

1.004 1.004
acceleration [g]

acceleration [g]

1.002 1.002

1 1

0.998 0.998

0.996 0.996

0.994 0.994

0.992 0.992

0.99 0.99
0 2 4 6 8 10 0 2 4 6 8 10
time [s] time [s]

Figure 3: Extracted order envelopes of test signals 2 (left) and 3 (right) for VKF2.

ICSV23, Athens (Greece), 10-14 July 2016 5


The 23rd International Congress of Sound and Vibration

3.3 Timing benchmarks


The benchmark measures execution time of all operations needed to extract particular order(s)
from signal using both lters generations. Signals from previous subsection were used. Resulting
execution time is averaged from 100 trials. Extraction of multiple orders using PCG method is stopped
if the solution error is less than 1E-8. Common computer with Core i5-4460 processor, Windows 7
and LabVIEW 2014 is used.
Table 1 shows benchmark results. It is clear that single order extraction is the least time consuming
since it uses efcient Cholesky decomposition. The most time consuming is the extraction of crossing
orders which is also greatly dependent on the value of the factor r. Lower value of r means wider lter
bandwidth which complicates the energy splitting between orders. To run any mentioned extraction
in the online manner with 50% overlapping, the execution time must not be greater than 500 ms. This
condition is met for all scenarios.

Table 1: Benchmarks of average execution times for different signal scenarios.


Scenario Description average time [ms]
1 VKF1 single order 2.17
2 VKF2 single order 2.31
3 VKF1 multiple orders 22.5
4 VKF2 multiple orders 58.1
5 VKF1 crossing orders 16.0
6 VKF2 crossing orders 146.0

4. Experiment
To show Vold-Kalman lters extraction results on real-world signals, experiment with common
cooling fan was conducted. There were measured two signals by NI 9234 acquisition card acceler-
ation using a piezoelectric accelerometer and speed prole using a tacho probe. Sampling frequency
was 2048 Hz, online method using 1024 samples with 512 samples overlap was employed.
Fan has 4 coils and 7 blades, thus expectable orders are multiples of 4 and 7 of the rotational speed.
Figure 4 shows the acceleration and speed prole signals. Orders 16, 20, 24, 28, and 32 indicated the

35
1

0.8 30

0.6
25
rotor speed [Hz]

0.4
acceleration [g]

20
0.2

0 15

0.2
10
0.4
5
0.6

0.8 0
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
time [s] time [s]

Figure 4: Signals measured during fan run-up acceleration on the fan frame (left), fan rotor speed
(right).

6 ICSV23, Athens (Greece), 10-14 July 2016


The 23rd International Congress of Sound and Vibration

0.16 0.16
32th order 32th order
0.14 28th order 0.14 28th order
24th order 24th order
0.12 20th order 0.12 20th order
16th order 16th order
acceleration [g]

acceleration [g]
0.1 0.1

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02

0 0
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
time [s] time [s]

Figure 5: Extraction of selected orders from run-up of a PC fan using VKF1 (left) and VKF2 (right).
5 4
x 10 x 10
2 1.5
32th order 32th order
1.5 28th order 28th order
24th order 1 24th order
1 20th order 20th order
16th order 16th order
absolute error [g]

absolute error [g]

0.5
0.5

0 0

0.5
0.5
1
1
1.5

2 1.5
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
time [s] time [s]

Figure 6: Envelopes extraction errors of the online method compared to the standard one using VKF1
(left) and VKF2 (right).

most energy during trial run, especially in the run-up phase. Therefore all 5 mentioned orders were
included in the extraction. VKF2 of the 1st order was selected, as well as factors r = 40 (VKF1) and
r = 100 (VKF2).
Results of the extraction using VKF1 and VKF2 are depicted in Fig. 5. VKF1 extracts waveforms
with the extracted order frequency (gure on the left; only signal above time axis is showed), while
results of VKF2 are waveform envelopes (gure on the right). Extracted signals are very similar.
Visible bumps during the rst second of VKF1 extraction in Fig. 5 are result of inaccurately
determined rotor speed. VKF2 is much more immune against such inaccuracies. Distorted rotor speed
during the rst second (Fig. 4 on the right) should smoothly follow rest of the curve. Fortunately,
beginning of the signal is not important for the analysis in this experiment. Error of the online method
originating from distortion of the data blocks compared to the standard (offline) method is expressed
as the difference of the extracted order waveforms showed in Fig. 6. The error is not greater than
2E-5 g for VKF1 and 1.5E-4 g for VKF2. These values are affected by the lters bandwidth (factor r)
as well as the window size and an overlapping size used.
It is clearly visible that the most of changes in Fig. 5 is situated within rst 5 seconds where the
rotor speed changes the most. Extracted signals thus contain reaction to various resonances present
in the examined system. The rest of the extracted signal shows stabilized response.

ICSV23, Athens (Greece), 10-14 July 2016 7


The 23rd International Congress of Sound and Vibration

5. Conclusion
Presented paper briefly summarizes the nature of calculation both Vold-Kalman lters for the
extraction of single or multiple orders from the signal. Special attention drew to the optimizations
using sparse matrices to speed up the calculation. Implemented algorithms has proved to be effective
according to the timing benchmarks. Online method with overlapping enables to extract one or multiple
orders from the signal using both lter generations at the cost of up to doubled computational cost and
only a little distortion of the result, namely 2E-5 for VKF1 and 1.5E-4 for VKF2 at most. The online
method is suitable for condition monitoring techniques, where extracted orders can be immediately
examined for potential machine faults. Compared to other condition monitoring methods, Vold-
Kalman lters are suitable for inspection of the machines with changing rotor speed. Differences
between the extracted orders using VKF1 and VKF2 are smaller than 0.01 g. This value is influenced
by the different properties of both lters (VKF1 is bandpass lter, VKF2 lowpass lter) and also
dependent on different bandwidth settings set by factor r.

ACKNOWLEDGMENT
Authors gratefully acknowledge nancial support from the Ministry of Education, Youth and
Sports of the Czech Republic under projects No. LO1210 Energy for Sustainable Development
(EN-PUR) solved at the Centre for Research and Utilization of Renewable Energy.

REFERENCES
1. Pan, M.-Ch. and Lin, Y.-F. Further exploration of Vold-Kalman-ltering order tracking with shaft-speed
informationI: Theoretical part, numerical implementation and parameter investigations, Mechanical Sys-
tems and Signal Processing, 20(5), 1134-1154, (2006).
2. Gade, S. Ed., Characteristics of the Vold-Kalman order tracking filter. Nrum, Denmark, (1999).
3. Vold, H. and Leuridan, J. High resolution order tracking at extreme slew rates using Kalman tracking lters,
Shock and vibration, 2(6), 507515, (1995).
4. Tma, J., Vehicle gearbox noise and vibration: measurement, signal analysis, signal processing and noise
reduction measures, John Wiley & Sons, Chichester, England (2014).
5. Tma, J. Setting the passband width in the Vold-Kalman order tracking lter, Proceedings of the 12th
International Congress on Sound and Vibration, Lisbon, Portugal, 1114 July, (2005).
6. Pan, M.-Ch. and Lin, Y.-F. Further exploration of Vold-Kalman-ltering order tracking with shaft-speed
informationII: Engineering applications, Mechanical Systems and Signal Processing, 20(6), 1410-1428,
(2006).
7. Brandt, A. Noise and Vibration Analysis: Signal Analysis and Experimental Procedures. John Wiley &
Sons, Chichester (2011). ISBN 978-0-470-74644-8.
8. Pedersen, T. F. Ed., Order tracking in vibro-acoustic measurements: A novel approach eliminating the tacho
probe, Conference & Exposition on Structural Dynamics, Denmark, (2005).
9. Hestenes, M. R and Stiefel, E. Methods of Conjugate Gradients for Solving Linear Systems, Journal of
Research of the National Bureau of Standards, 49(6), 409-436, (1952).
10. Feldbauer, Ch. and Hldrich, R. Realization of a Vold-Kalman tracking lter - a least squares problem,
Proceedings of the COST G-6 Conference on Digital Audio Effects, Verona, Italy, 7-9 December, (2000).
11. Pan, M.-Ch. and Wu, C.-X. Adaptive Vold-Kalman ltering order tracking, Mechanical Systems and Signal
Processing, 21(8), 2957-2969, (2007).

8 ICSV23, Athens (Greece), 10-14 July 2016

You might also like