THE ROLE OF SINGULAR VALUES OF MEASURED FREQUENCY RESPONSE FUNCTION MATRIX IN MODAL DAMPING ESTIMATION ( PART II : INVESTIGATIONS )

Original scientific paper The singular value decomposition of the measured frequency response function matrix, as a very effective tool of experimental modal analysis is used over the last twenty-five years. The complex mode indication function has become a common numerical tool in processing experimental data. There are many references on the development of complex mode indication function including the enhanced mode indication function and its use together with the enhanced frequency response function to form spatial domain modal parameter estimation methods. Another amendment of the enhanced mode indicator function method is the extension of the single degree-of-freedom aspects of the complex mode indication function method to include a limited number of modes. In the paper, methods for estimation of damped eigenfrequencies, modal damping and mode shapes are presented that are based on singular value decomposition of frequency response function matrix. It is shown how to obtain phase information for the complex mode indication function, in order to use the standard single degree-of-freedom modal parameter estimation methods. New aggregated frequency response function is introduced. A leastsquares approximation will be presented for eliminating the error caused by frequency discretization. Analytical models and examples taken from vehicle industry are also used to demonstrate applications of the aggregated frequency response function method and estimation of modal damping.


Analytical investigations
The properties and the applicability of the wellknown eFRF, the vector tracking method and the newly developed aggregated aFRF have been investigated via computer simulations.A 2DOF and a 15DOF model were examined with proportional and non-proportional damping.
The modal parameters (undamped eigenfrequency and modal damping) were estimated based on the eFRF and aFRF, then they were compared to the analytical solution of the original model.The mode shapes were not the subject of this investigation.The simulations were made with MathCAD, MatLAB and LabVIEW.

2DOF analytical models
The model used in 2DOF simulation is shown in Fig. 1.For non classical modes the damping coefficients c 1,2 = c 2 = 0,01 N•s/m should be chosen.
For classical normal mode computations let the proportional damping matrix be: .01 The eFRF diagrams of the 2DOF system are drawn in Fig. 3.The least-square-method regression approximates the eigenvalues very well.
The newly developed aggregated aFRF can be seen in Fig. 4. Note that this is the only global function which gives the damped eigenfrequencies at their peaks.

Figure 4
The aFRF of classical damped 2DOF system Four points were taken from the vicinity of each peak, then the complex eigenfrequencies were calculated using the circle fitting formula detailed in Part 1.The results can be seen in Tab. 1 and Tab. 2.
The estimation of the complex eigenvalues was eligible, the approximations are almost the same as the exact values determined from the solution of the eigenvalue problem.The effect of additional noise was not examined.This method -beyond expectations -also approximates well the modal parameters of the 2DOF system with non-classical damping.Probably the reason for this is that the method does not use the left singular vector but only its phase angle.

15DOF Analytical Model
In this section we introduce the usage and applicability of estimation methods based on the eFRF

Lightly-damped system with classical normal mode
The left and right singular vectors have special structure, when SVD is performed on the FRF of a system with classical normal modes.Measurement with rough frequency resolution does not provide enough points on the modal semicircle for the approximation of the modal parameters with circle fitting or the half-bandwidth method.In this example a measured FRF is simulated with a particular frequency resolution, when only one measured point falls on a modal semicircle, all other measured points are outside of the half-power bandwidth.The accuracy of the approximation was examined for this case.The model of the lightly-damped structure was built using the mass and stiffness data from Fig. 5 the damping matrix was constructed as , since we assumed proportional damping.The function of the singular values of the FRF matrix of this model can be seen in Fig. 6.Note that the curve of the first singular value (red curve) detects all eigenfrequencies of the 15 modes.Mode #4 and mode #5 are close eigenfrequencies.The vector tracking method described in Part I gives the ordered singular value curves, as it is shown in Fig. 7.
The vector tracking method put the singular values in the right order.It has to be noted that all the modes in this model were separated.Therefore, investigation has to be made for structures having overlapped modes.The approximations based on eFRF and aFRF were made for the damped eigenfrequencies of mode #1, mode #4 and mode #5 (see Fig. 8).
The exact eigenfrequencies and those extracted from the eFRF and aFRF are compared in Tab. 3. The parameters determined with different methods show close resemblance.

MDOF System with General Viscous Damping
The behavior of the aFRF method has also been investigated for the case when the structure has damping matrix C that does not satisfy the C KM K CM As it can be seen in Fig. 9, the peak which belongs to mode #5 is not detectable on the curve of the 1 st singular value (red curve).This "lost" mode becomes visible on the plot of the 2 nd singular value (see Fig. 10).
The first singular value was used to estimate the eigenfrequency of mode #1 and mode #4, the second singular value was used for mode #5.The results of the approximations are in Table 4.
The extracted eigenfrequencies show close resemblance with the exact values.

Conclusion of Analytical Experiments
Analyzing the result of the computer simulations leads to the following statements: • the vector tracking method has only been verified for structures having separated modes, further investigation is necessary for overlapped modes; • both the eFRF and aFRF can be used for parameter estimation; • the least squares approximations only use a few points on the first singular value function in the vicinity of each peak instead of using all the measured data; • a diagram plotting all singular values has proven to be useful to detect all peaks; • when two modes overlap, the second singular value can be used to estimate the modal parameters.

Experimental Investigations
The frequency response of a train's brake arm was measured to demonstrate the use of the aFRF.The brake arm and its place in the brake system can be seen in the pictures below (see Fig. 11).

Experimental Setup
A total of 20 key points are selected in order to obtain full information about the dynamics of brake-arm, and their distribution is shown in Fig. 12b.Single-input single-output approach is employed during the modal test.
Each time response vibration acceleration location 1Y and 2Z, while impacted by a hammer with force transducer is roved from point 1 ÷ 20 in directions Y and Z. N i = 2, N 0 = 40.Measurements of the EMA (Experimental Modal Analysis) brake-arm, SISO (Single Input -Single Output), and FRF (Frequency Response Function) were performed (Fig. 13) and (Fig. 14).A schematic diagram of the experimental setup is shown in the illustration.The excitation and response measurements were performed horizontally (Y) and vertically (Z).The total number of the measured locations was 62. Frequency range: 0 ÷ 1505 Hz, frequency resolution df = 0,04 Hz.

Parameter Identification
The measured point-to-point FRF are plotted on a waterfall diagram in Fig. 15.N m = 7 modes were detected on the aFRF in the investigated frequency range (Fig. 16).Modal damping and the damped eigenfrequency were estimated by applying the Eq. ( 1) formula to the aFRF.
Then the parameter estimations were performed by circle fitting detailed in the Part I in order to approximate the regression curve (Fig. 16).By performing SVD on the measured FRF the damped eigenfrequencies can easily be determined.Not every mode can be detected on every local FRF.For example mode#3 cannot be detected on the FRF of 2Z1Y location (see Fig. 17).
Using the estimated global parameters of the system the residues can be approximated with linear regression.In Fig. 17 the result of the approximation is plotted for location 2Z1Y.The eligibility of using the aFRF for parameter estimation is verified by this experiment.Note that the coefficients of these functions depend on the support of the structure.For reasons of brevity, this article does not deal with the effect of separating the structure from the support.

Conclusions
In this article the use of the singular value decomposition of the FRF in order to estimate modal parameters was investigated.We found that the peaks of the singular value function estimate the damped natural frequency well.A new aggregated frequency response aFRF was derived from the eFRF known from the literature.The properties of the aFRF function are: • it is one complex function of frequency for the whole MDOF system, • each peak on the absolute value function indicates a mode, and the corresponding frequency of a peak approximates the damped natural frequency, • this function contains phase information, thus every mode is represented as a modal circle on the Nyquist plot, • it smoothes noise, • the modal damping can be estimated with linear regression, not necessarily all points have to fall in the half-power bandwidth, • also successful when only data with rough frequency resolution is provided.
In the Part I of this paper a new complex valued aggregator frequency response function (aFRF) is introduced.The definition of the aFRF function is ) jω) the first singular value of FRF matrix vs. frequency ω, singular vector of FRF matrix vs. frequency ω.

Figure 1
Figure 1 MDOF=2 test model The vector tracking method succeeded to put the left and right singular vectors in the right order when it was applied to the 2DOF model.The results of the vector tracking are shown on the bottom side of Fig. 2. The vector tracking method was able to sort the singular vectors independently of the damping model.Parameters: m 1 = m 2 = 0,01 kg; k 1 = k 2 = 4 N/m.

Figure 2 Figure 3
Figure 2 Singular values of classical damped 2DOF model and aFRF and the use of vector tracking on a 15DOF model.The structure of the 15DOF model is shown in Fig.5.Simulations were made for two typical cases:• Classical normal modes with light damping let the damping matrix:

Figure 6 Figure 7
Figure 6 Singular values of lightly-damped system with classical normal modes

Figure 9
Figure 9 Non-tracked singular values of generally damped 15DOF model

Figure 10
Figure 10The aFRF of the generally damped 15DOF system (Left: Amplitude-frequency diagrams of aFRF1 and aFRF2 Right: Nyquist-Plot)

Figure 11
Figure 11 Location of investigated part in the brake system (Left: Bogie equipment, Middle: Compact wheel brake caliper, Right: Brake arm www.knorr-bremse.hu

Figure
Figure 14 Experimental setup view

Table 3
eFRF and aFRF estimates of light damped 15DOF system

Table 4
eFRFand aFRF estimates of generally damped 15DOF system