Vous êtes sur la page 1sur 13

101 WAYS TO EXTRACT MODAL PARAMETERS

WHICH IS THE ONE FOR ME?

Dr. Peter Avitabile


Modal Analysis & Controls Laboratory
University of Massachusetts Lowell
Lowell, Massachusetts

ABSTRACT
Modal parameter estimation has always been a source of difficulty in the development of an experimental modal model.
Much time can be wasted in this process of extraction of parameters especially when the data collected is not optimal. A
brief review of the history of modal parameter estimation usage from a user perspective is presented related to the
frequency response testing approach. More importantly, the lessons learned over the years on how to best extract
parameters and guidelines for reduction of data are presented. Several case examples are presented to illustrate some of
the common pitfalls that have been observed over the years. Also several cases that illustrate interpretation of data are
presented.

INTRODUCTION
The modal parameter estimation process has evolved over the years. A brief overview of the history of the techniques and
their utilization are presented. Tools associated with the interpretation and reduction of the data are discussed from a user
perspective but in a historical fashion as they were developed. The intent is to provide an overview of the development of
the techniques but from a user perspective on the reduction of data and use of the tools. The intent is not to show every
modal parameter estimation method but to explain the typical common approaches employed and most reasonable
approach for their effective use.

MODAL PARAMETER ESTIMATION TECHNIQUES


Each of the different techniques that have evolved over the years have always had a history of development and then a
history related to their implementation and use. Some of the more popular techniques are described with some of the
variations and implementations as well as their practical use (Some of the techniques are still used today so their
description and use is important). The detailed theoretical development is not the intent of this paper nor are references
provided to the numerous theoretical developments made over the years.
Peak Pick
One of the most obvious first techniques employed was the peak pick technique. The implementation is very
straightforward and is based on the fact that the peak of the frequency response function is directly related to the mode
shape of the system. The peak amplitude is a good indicator provided that there is sufficient separation between each of

the modes of the system. The peak pick process is shown schematically in Figure 1. The implementations depend on the
individual software packages but typically either a single cursor or banded cursor is used; a banded cursor finds the peak
amplitude over the band. Peak pick estimation is very fast and straightforward and is an excellent tool for first estimation
attempts. However, frequency resolution, leakage, coherence dropout at resonance and other factors make the peak
amplitude susceptible to error and does not provide the most accurate data.

MODE 1 CONTRIBUTION

MODE 2 CONTRIBUTION

Figure 1: Schematic Representation for Peak Pick Estimation

Circle Fit
The circle fit was one of the first mathematical approaches to estimating parameters. The basis of circle fitting lies in the
fact that the Nyquist plot of the frequency response function looks very much like a circle. This very early approach was
implemented on many commercial systems and is still used to some degree today in modal parameter estimation.
Concerns with the technique center on the effects of adjacent modes and their effect on extracted parameters. Adjustment
terms can be added to account for some of the out of band effects to some degree but as the modes become too close, these
adjustments cannot provide adequate compensation. Another issue related to the lack of good coherent measurement
quality at the resonant frequency and some implementations allowed the user to deselect some of the data points around
resonance to improve the estimation process. However, this tended to become a very tedious process.

Figure 2: Schematic Representation for Circle Fitting Estimation and Effects of Out of Band Modes
(Source Left Figure: Heylen, Modal Analysis - Theory and Testing) (Source Right Figure: Ewins - Modal Testing, 2nd Edition)

Single Degree of Freedom Polynomial


The Single Degree of Freedom Polynomial (SDOF Poly) is a frequency domain implementation of the system transfer
function which is fit to estimate modal parameters. A schematic of the approach is shown in Figure 3. The polynomial
representation was modified to add additional numerator terms to account for the effects of out of band modes. But as
with the Circle Fit only so much can be expected of this adjustment when the modes become very close. Often attempts
were made to bias the curvefit band so as to minimize the effects of adjacent modes but this is always a tedious and
questionable approach for estimating parameters.
SDOF FIT

SDOF FIT

WELL SEPARATED

NOT WELL SEPARATED

Figure 3: Single DOF Polynomial Schematic


Prior to the introduction of the multiple degree of freedom curvefitting approaches, attempts were made to compensate for
the out of band effects through an iterative process. First all the modes were fit using the SDOF technique. Then an
iterative process of synthesizing all the modes except for one mode of interest was performed. This was then subtracted
from the original frequency response function which was then fit as a better approximation of the SDOF; the intent was
that effects of others modes were almost completely removed. This is shown schematically in Figure 4. This also proved
to be a tedious process but was used as a reasonable alternative before the multiple degree of freedom techniques were
introduced.

FRF
FRF - SYNTH
SYNTH

Synthesize all modes


except mode of interest

Figure 4: Single DOF Polynomial Approximation with Out of Band Modes Removed

Time Domain Complex Exponential


The time domain complex exponential was one of the first multiple degree of freedom curvefitting methods that became
extremely popular and is still widely used today. This curvefitting technique is robust, numerically stable and can handle
many modes in the extraction process. A little historical background on its use and implementation is warranted since this
technique exists today in almost every commercial package available.

Amplitude

The basic idea behind the technique is to process data in the time domain where the characteristic response is nothing more
than the sum of sinusoidal damped exponentially decaying signals represented by the response of each of the modes of the
system due to an impulsive excitation. The damped exponential response for a single mode is shown in Figure 5 to
identify the key portions of characteristics to be determined. In the early days of this technique, the main use was in
support of time domain data acquired by impulsive excitation, most notably by Ibrahim and his time domain technique
(ITD). In order for the numerical algorithm to function properly, more modes need to be specified than that which actually
exist in the data. This is not a serious concern but the user must be aware of the extra computational modes that result.

h(t)

1
md

e t

sin d t

Damping Decay

Period

Figure 5: Schematic of Single DOF Impulse Response

As frequency response testing became more popular, this extraction technique became a viable possibility for estimating
parameters. This approach required that the frequency response functions acquired be transformed back to the time
domain in order to utilize the technique. This required the inverse FFT to be performed over a specified bandwidth where
modal parameters were to be extracted. This inverse process caused some difficulties since the inverse transform is only
exact if the entire frequency response over all frequencies was included. If not, then a time domain leakage effect resulted.
This is schematically shown in Figure 6. The exponential decaying response could be distorted by this time domain
leakage effect and in many software packages the user is allowed to identify the portion of the time response to be used by
the extraction process in order to minimize this inverse FFT artifact.

DESIRED MODE RESPONSE

IFT

UNDESIRED TIME DOMAIN LEAKAGE


Figure 6: Schematic of Time Domain Leakage Effect of the Inverse FFT Process
In the early years of using this technique, there were no interpretive tools available to sort the good modes from the so-

called computational modes that resulted. These computational modes arise from the numerical algorithm used as well as
from noise and poor measurement quality. It was a significant chore to select the true modes of the system and deselect
the computational modes. Many times the user based this deselection process on poor phase representation, unusually high
damping estimates and other practical aspects that were based on experience in many cases. Eventually tools developed to
help the user sort through this difficult process. The Least Squares Error Chart was one tool that assisted the user in
determining approximately how many modes were expected to exist in the band evaluated; this graphical representation of
the Covariant Matrix typically had a distinct knee of the chart to identify the approximate number of modes. Other tools
followed such as the Mode Confidence Factor (MCF) which helped the user to identify potential modes easier.
Typically the user needed to identify some order of model to be fit to the data. Parameters were extracted and potential
good modes were selected. The resulting parameters were used to synthesize the frequency response function and
compared to the actual measurement to determine if suitable parameters were identified. This process was continued until
an appropriate set of parameters were identified. This also was a very tedious process. Many times some modes were
questionable as to their suitability for the extracted process and generally the user struggled with this part of the parameter
estimation process when the data was not of the best quality.

Early Mode Indicator Functions


In the estimation of modal parameters, several additional tools were developed to help the user identify where potential
modes might exist in the set of data acquired. One very important tool developed was the Summation Function. This
provided a sum of all the frequency response functions and helps to identify modes that exist in the data. Another very
important tool used was the Mode Indication Function (MIF) and the Multivariate Mode Indicator Function (MMIF) for
multiple referenced data. This tool is still widely used today to help identify modes that exist in the data. The MIF is
essentially a ratio of the real part to the magnitude of the frequency response function. As such, for real normal modes,
this function dips at resonant frequencies giving the user an indication of a mode. Both of these tools are still widely used
today in most commercial software packages. The Summation and MIF are shown schematically in Figure 7.

Figure 7: Schematic of Summation Function (left) and MMIF (right)

Stability Diagram
The introduction of the Stability Diagram revolutionized pole selection process. Basically, an iterative process is initiated
starting from some low order mode size and then is continued to some higher order model size. As the process continues,
indications are given as to the continued prediction of a particular pole and with some predetermined tolerance. As the
process continues, poles (or roots) that are inherent in the system will repeat and be persistent whereas computational
modes and poles due to poor measurements that are identified will be spotty and sporatic in the presentation of this data.
In addition, this data can be plotted along with one or more of the indicator tools to give a much clearer indentification of
the true poles of the system. This Stability Diagram is an extremely useful tool and widely used in the modal parameter
estimation process. This is shown schematically in Figure 8.

Figure 8: Schematic of Early Form of Stability Diagram

Complex Mode Indicator Function


The Complex Mode Indicator Function (CMIF) was another tool developed after some of the initial indicator tools. The
CMIF is based on the singular valued decomposition of the frequency response function matrix. This indicator helped to
identify modes of the system as well as repeated roots. The CMIF is not affected by a real normal mode assumption as the
MIF/MMIF and is generally more robust. This technique is also used as a curvefitting technique. The CMIF is
schematically shown in Figure 9.
4

10

10

10

10
CMIF

10

-1

10

-2

10
0

50

100

150

200
250
Frequency (Hz)

300

350

400

Figure 9: Schematic of CMIF Diagram

450

500

Multiple Degree of Freedom Polynomial


The Multiple Degree of Freedom Polynomial (MDOF Poly) is a frequency domain implementation of the system transfer.
Initially this was developed as a single degree of freedom technique and then evolved into a multiple degree of freedom
curvefitter. This is an excellent curvefitting technique. The only drawback to the implementation of this technique is that
typically the bands need to be selected such that a few modes are extracted over a reasonably narrow band. If not there can
be some computational effects due to the size of the polynomials that are generated. There are many variations of this
technique implemented in a variety of different commercially available software packages.
In many packages, the user has the ability to augment the polynomial numerator and denominator terms to provide
improved extracted parameters. Addition of extra poles as part of the extraction process is possible but often this is not
considered reasonable. The addition of numerator terms however, is considered to be a very good mechanism for
accounting for the effects of residual modes outside the band of interest. Typically three of four terms are sufficient to
approximate the effective mass and stiffness contributions from the out of band modes. More than this many terms is not
considered reasonable since the data may be distorted as a result of too many adjustment terms. Figure 10a shows a
comparison of parameters extracted where overspecification of the residual terms produces a visually good fit but the
estimated parameters have errors. Figure 10b shows the same measurement with extracted parameters which are much
more accurate with much less error and overspecification of the residual terms was not needed.

Figure 10a : Estimated Parameters Distorted by Overspecification of Residual Terms

Figure 10b : Estimated Parameters Improved with Appropriate Use of Residual Terms

Residual Modes
The modal parameter estimation process is typically employed over small bands where a collection of modes exist.
However, there are also effects of modes outside this band of interest. Many times, the modal parameter estimation
algorithm allows the user to add compensation terms to account for these adjustment terms as part of the modal parameter
estimation process. This may be in the form of adding extra modes, or extra polynomial terms, or mass and stiffness
adjustments to the extraction process depending on the particular algorithm implemented. These adjustment terms result in
extracted parameters commonly referred to as residual modes. These residual modes are often similar in shape to one or
more of the lower or upper modes outside the band of interest. The effects of these out of band modes are shown
schematically shown in Figure 11.

RESIDUAL
EFFECTS

RESIDUAL
EFFECTS

Figure 11: Schematic Showing Residual Effects


If the modes outside the band are far removed from the band of interest, then the use of residual terms may not be
necessary. However, as the effects of the out of band modes become closer to the band of interest, then these residual
terms become an important factor in the estimation of good parameters. In the case where the out of band modes become
extremely close to the band of interest, then the residual terms may not be adequate to compensate for these effects. In this
case, it may be reasonable to consider alternate bands for extraction of modes. This is why the modal parameter estimation
process is typically some iterative manipulation of the data the user never knows for sure which band may be the most
appropriate for the identification of suitable modal parameters.

Modal Parameter Extraction Considerations


The modal parameter estimation process is typically employed over small bands where a collection of modes exist. The
modes may appear to be well separated distinct modes or may be closely spaced as shown in Figure 12. The user must
identify which estimation schemes are to be employed utilizing single or multiple DOF strategies as well as either time or
frequency domain techniques as appropriate.

MDOF

SDOF

Figure 12: Schematic Showing Potential Grouping of Modes


The user may identify different bands but often it is not clear as to how to group different bands over different frequency
ranges. Many times the user will investigate many different strategies and overlapping bands may be considered in order
to extract the best possible parameters. The iterative process is shown schematically in Figure 13.

Figure 13: Schematic Showing Potential Bands for Extraction


Validation of Parameters Extracted
Once all the modal parameters are extracted, a synthesized function can be compared to the actual measurement made.
This is the best way to assure that proper parameters have been extracted. Usually all of the measurements are made and
then compared using an approach such as the Amplitude Weighted Assessment Function (AWAF). The AWAF is nothing
more than a comparison of the summation and mode indicator functions for the measured and synthesized functions. A
typical synthesized frequency response function is shown schematically in Figure 14 and the results from an AWAF
assessment are shown in Figure 15. The modes extracted need to be able to reproduce (synthesize) the measured functions
reasonably well.

Figure 14: Schematic Showing Synthesized Comparison for Validating Results

MMIF 1

SUMMATION
MMIF 2
MMIF 2

Figure 15: Schematic Showing AWAF Synthesized Summation and MMIF for Validating Results
Multiple Referenced Measurements and Modal Participation Factors
Many times multiple reference data is collected from a structure. Often the parameters are extracted from all of the
measured data and very often the extracted parameters cannot adequately recreate (synthesize) the measured data correctly.
An example of this is shown in Figure 16. The estimated parameters are questionable especially at the higher frequencies.

Figure 16 - Typical Synthesized FRF using All References

The extraction of modal parameters is heavily dependent on the adequacy of the measurement acquired. Obviously, poorly
measured responses are much more difficult to extract than well measured response. However, this is easy to state but
very difficult to achieve in a realistic measurement situation especially when the structure to be characterized has dominant
directional modal information. The process of extracting poles then becomes much more difficult. Even when multiple
referenced shaker excitation is used, there is always the possibility that some of the measured functions will not be excited
well by every one of the reference shakers. When this is the case, then the modal parameter estimation process becomes
much more difficult.
When a large set of data is collected and multiple references are utilized, the process of extracting parameters becomes
more difficult especially when directional modes exist in the data set. Typically, these types of situations require
significant user intervention in the modal parameter estimation process. When processing multiple reference data, a
function called the Modal Participation Factor (MPF) results as part of the scaling and sorting process. This MPF contains
very useful information for the user to further manipulate the data set.
Typically, not all modes are excited by all the references employed. Usually, only one or two of the multiple references
show significant modal participation and often the other references show almost no participation whatsoever. This case is
shown schematically in Figure 17 (for the data case shown at the beginning of this section) where only two of the
references adequately excite most of the lower order modes of the system. In these cases, the frequency response matrix
needs to be sorted and sifted to provide an optimum set of measurements for the extraction of poles of the system. When
this painstaking operation is performed, the modal parameters extracted are generally much better than those extracted
using all of the data acquired in one set.

Figure 17: Modal Participation Matrix Showing Mode Contribution


As an illustration for this data set, the painstaking task of sifting and sorting was performed to extract the best possible
poles of the system using the best possible sets of measurements for the best sets of references for particular sets of modes.
Once this effort is expended then typically better parameters are extracted. A typical set of improved, synthesized
frequency response functions for this same data case are shown in Figure 18. Clearly the effort expended produces visibly
superior results.

Figure 18 - Typical Synthesized FRF using Selective References

PolyMAX A New Modal Parameter Estimation Approach


The evolution of the modal parameter estimation process has been long. In the sections above, the care and pain involved
in developing, implementing and utilizing the many techniques developed over the years was described for the user to
understand how modal parameter estimation processes have evolved over the years. A more recent technique referred to
as PolyMAX has been recently developed and implemented which revolutionizes the modal parameter estimation process.
The PolyMAX technique in its early introduction and use appears to have overcome all of the barriers of the modal
parameter estimation process described above.
Traditionally, there are several areas that always pose difficulties when extracting modal parameters. Lightly damped
systems vs heavily damped systems always proves to be difficult and users often question whether time domain or
frequency domain approaches should be applied to extract parameters. Generally, frequency domain methods are not
optimum when selecting wide frequency bands that have numerous modes in the band but have an advantage in residual
compensation. Time domain methods typically can handle wide bands with many modes but residual effects and
computational modes always present difficulties. And as in the last example, pole selection becomes a tedious process
when all of the references do not adequately represent all the modes of the system as is typically the case in structures that
have either directional modes or an abundance of local modes. The PolyMAX technique appears to overcome many of
these difficulties.
For the multiple referenced data set mentioned in the last section, the complete set of measured frequency response
functions were processed using this new PolyMAX approach. A typical Stability Diagram from a lightly damped system
is shown in Figure 19 for reference. Clearly, the pole selection process is vastly improved and an extremely wide band,
with numerous modes has been used in this reduction. While this data set has proved troublesome in any of the modal
parameter estimation approaches traditionally employed in any of the commercial software packages, PolyMAX has no
difficulty in processing this complicated data set and provide clear indication of the poles related to the system.

Figure 19: Typical PolyMAX Stability Diagram Showing Clear Pole Identification over Wide Frequency Band

SUMMARY
A brief overview of some of the modal parameter estimation techniques that have evolved over the years was presented.
This was done from a user perspective to illustrate some of the difficulties, pitfalls and obstacles that have been endured
over the years. Current modal parameter estimation processing has made significant strides over the past several decades.
Numerous efforts and contributions and experiences were presented from an application perspective rather than a
theoretical development approach. The hope is that users of current modal parameter estimation algorithms gain an
appreciation of the development and application of these techniques.

ACKNOWLEDGEMENT
Specific references are not provided since there is wealth of papers on modal parameter estimation approaches. The intent
of the paper is to provide some insight into the historical use of the various modal parameter estimation approaches from a
practical standpoint. Experiences described come from an array of various commercially available software packages but
in particular ModalPlus, SDRC Ideas Test, SMS Structural Analysis Software, SMS STAR, Vibrant Technology
MEscope, and LMS Modal Analysis are typical packages that have incorporated techniques mentioned in this paper.

Vous aimerez peut-être aussi