07.12.2012 Views

MODAL ANALYSIS - the Dept. of Mechanical Engineering at - Vrije ...

MODAL ANALYSIS - the Dept. of Mechanical Engineering at - Vrije ...

MODAL ANALYSIS - the Dept. of Mechanical Engineering at - Vrije ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

3.2 Maximum Likelihood Estim<strong>at</strong>ion<br />

3.2.1 Gauss-Newton Optimiz<strong>at</strong>ion<br />

Assuming <strong>the</strong> FRFs to be uncorrel<strong>at</strong>ed, <strong>the</strong> (neg<strong>at</strong>ive) log-likelihood function reduces to<br />

NoN<br />

N i f Hˆ<br />

k ( , ω f ) − H k ( ω f )<br />

� ML(<br />

) = ∑∑<br />

(35)<br />

var{ H ( ω )}<br />

k=<br />

1 f = 1<br />

k f<br />

2<br />

T T T T<br />

The Maximum Likelihood (ML) estim<strong>at</strong>e <strong>of</strong> = [ 1 , �,<br />

No<br />

N , ] is obtained by minimizing (35).<br />

i<br />

This can be done by means <strong>of</strong> a Gauss-Newton optimiz<strong>at</strong>ion algorithm, which takes advantage <strong>of</strong><br />

<strong>the</strong> quadr<strong>at</strong>ic form <strong>of</strong> <strong>the</strong> cost function (35). The Gauss-Newton iter<strong>at</strong>ions are given by<br />

(a)<br />

(b)<br />

solve Re( J<br />

set<br />

p+1<br />

=<br />

H<br />

p<br />

p<br />

J<br />

p<br />

+<br />

)<br />

p<br />

p<br />

= − Re( J<br />

with r p = r(<br />

p ) , J = ∂r(<br />

) ∂ and<br />

p<br />

p<br />

H<br />

p<br />

r<br />

p<br />

)<br />

for<br />

p<br />

⎧ Hˆ<br />

⎫<br />

11(<br />

, ω1)<br />

− H11(<br />

ω1)<br />

⎪<br />

⎪<br />

⎪ var{ H11(<br />

ω1)}<br />

⎪<br />

⎪<br />

⎪<br />

r( ) = ⎨<br />

� ⎬<br />

(37)<br />

⎪ Hˆ<br />

N ( , ) − ( ) ⎪<br />

oN<br />

ω<br />

i N H<br />

f NoN<br />

ω i N f<br />

⎪<br />

⎪<br />

⎪ var{ H N ( )}<br />

⎩<br />

oN<br />

ω i N f ⎪⎭<br />

The Jacobian m<strong>at</strong>rix J p has <strong>the</strong> same structure as <strong>the</strong> m<strong>at</strong>rix J given in (28). Also here it is possible<br />

H<br />

H<br />

to form <strong>the</strong> normal equ<strong>at</strong>ions (i.e. Re( J p J p ) and Re( J p rp<br />

) ) in a similar time-efficient way as<br />

presented in Section 3.1. See 6.43.8.2 Estim<strong>at</strong>ion with Known Noise Model and 6.43.8.4 Estim<strong>at</strong>ion<br />

with Unknown Noise Model for more inform<strong>at</strong>ion about frequency-domain ML identific<strong>at</strong>ion.<br />

3.2.2 Confidence Intervals<br />

A good approxim<strong>at</strong>ion <strong>of</strong> <strong>the</strong> covariance m<strong>at</strong>rix <strong>of</strong> <strong>the</strong> ML estim<strong>at</strong>e ˆ<br />

ML is obtained by inverting <strong>the</strong><br />

Fisher inform<strong>at</strong>ion m<strong>at</strong>rix (see 6.43.8. Frequency Domain System Identific<strong>at</strong>ion)<br />

ˆ<br />

cov{<br />

ML<br />

} ≈ [ 2 Re( J J<br />

H<br />

∞<br />

∞<br />

)]<br />

−1<br />

with J ∞ <strong>the</strong> Jacobian m<strong>at</strong>rix evalu<strong>at</strong>ed in <strong>the</strong> last iter<strong>at</strong>ion step <strong>of</strong> <strong>the</strong> Gauss-Newton optimiz<strong>at</strong>ion.<br />

As one is mainly interested in <strong>the</strong> uncertainty on <strong>the</strong> modal frequencies and damping r<strong>at</strong>ios, only <strong>the</strong><br />

covariance m<strong>at</strong>rix <strong>of</strong> <strong>the</strong> denomin<strong>at</strong>or coefficients is in fact required. Starting from (38), one can<br />

show th<strong>at</strong> this m<strong>at</strong>rix is given by<br />

ˆ<br />

cov{<br />

ML<br />

}<br />

⎡<br />

⎢2<br />

⎣<br />

≈ ∑ i oN N<br />

k=<br />

1<br />

−1<br />

(36)<br />

(38)<br />

T −1<br />

⎤<br />

T k − Sk<br />

⋅ R k ⋅ Sk<br />

⎥<br />

(39)<br />

⎦<br />

with R k , S k , and T k as defined in Section 3.1.3 but now applied to Re( J ∞ J ∞ )<br />

H<br />

. Hence, it is not<br />

necessary to invert <strong>the</strong> full m<strong>at</strong>rix occurring in (38). From (39), it is possible to compute <strong>the</strong><br />

uncertainty on <strong>the</strong> modal frequencies and damping r<strong>at</strong>ios. For flight flutter testing – but also for<br />

applic<strong>at</strong>ions such as vibr<strong>at</strong>ion-based fault detection and oper<strong>at</strong>ional modal analysis – <strong>the</strong> availability<br />

<strong>of</strong> reliable estim<strong>at</strong>es toge<strong>the</strong>r with confidence intervals is important.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!