Académique Documents
Professionnel Documents
Culture Documents
389
13th IFAC Symposium on System Identification
390
13th IFAC Symposium on System Identification
effects of shear deformation and rotary inertia, and is Laplace variable s (result: Young’s modulus E( jω)
accurate for thin beams only [12]. In practice (2) can and its uncertainty in the frequency band of interest).
be used to model the first few resonance frequencies These four main steps are explained in detail in the
corresponding to the simple mode shapes. sequel of this section, and each step is illustrated on
Timoshenko’s theory [12] considers the effects of the longitudinal vibrations of a plexiglass beam.
shear deformation and rotary inertia, giving the
following implicit relationship between the poles s k 3.1. Measurement of the frequency response function
and Young’s modulus E The frequency response function (FRF) of the system
is measured using random phase multisines [14, 15].
cosh(b 1(s k)L)cos(b 2(s k)L) – 1 + These are periodic signals consisting of the sum of
harmonically related sine waves with user defined
b 22(s k) – b 12(s k) (4)
----------------------------------- sinh(b 1(s k)L)sin(b 2(s k)L) = 0 amplitudes and random phases. The measurement
2b 1(s k)b 2(s k) procedure of [16] is followed to estimate the FRF and
its uncertainty. It consists of the following basic steps:
where
1 Choose the amplitude spectrum and the
frequency resolution of the random phase
b i2(s) = ( – 1 ) ( i + 1 ) c(s) + c 2(s) + a(s) multisine.
2 Make a random choice of the phases of the
ρ
c(s) = ------------- ( 1 + γ (s) )s 2 (5)
random phase multisine, and calculate the
2E(s) corresponding time signal.
ρ A 2 ρ 2 γ (s) 4
a(s) = – ------------s 3 Apply the excitation to the system and measure
+ --------------
-s P ≥ 2 periods of the steady state response u(t) ,
E(s)I E 2(s)
y(t) .
4 Repeat steps 2 and 3 M ≥ 6 times.
A is the cross section area of the beam, and γ (s) 5 Calculate the DFT spectra of the input u(t) , and
depends on the shape of the cross section and output y(t) signals for each period of each
Poisson’s coefficient ν(s) ( γ (s) = ( 12 + 11 ν(s) ) ⁄ 5 experiment at the excited DFT frequencies.
for a rectangle, and γ (s) = ( 7 + 6 ν(s) ) ⁄ 3 for a circle
[13]). From these M × P sets of noisy input/output spectra,
one can calculate for each experiment the average
2.3. Discussion FRF Ĝ [ m ] over the P periods and its sample variance
Equations (1), (2) and (4) establish a relationship σ̂ Ĝ
2
[ m ] ( m = 1, …, M ). An additional averaging over
between the poles s k of the longitudinal and flexural m gives the final FRF Ĝ of the whole measurement
vibration experiments and Young’s modulus E(s k) procedure
when the beam ( L , A , I , J ) and the material ( ρ ,
ν(s) ) properties are known. Fortunately, the cross Ĝ [ m ] Ĝ [ m ] – Ĝ 2
∑m = 1 ----------- ∑m = 1 ---------------------------
M M
Ĝ = , σ̂ Ĝ
2 = (6)
section dimensions of the beam can always be chosen M M (M – 1)
such that the terms in (1) and (4) depending on
Poisson’s coefficient ν(s) are correction terms. together with its sample variance σ̂ Ĝ
2 . If the system is
Hence, only a rough guess of ν(s) is required. This is linear, then σ̂ Ĝ should be equal to the mean value of
2
the basic idea used for identifying Young’s modulus σ̂ Ĝ
2
[ m ] divided by M
E(s) in Section 3.
1
σ̂ Ĝ - ∑M σ̂ 2 [ m ]
2 = ------- (7)
n M 2 m = 1 Ĝ
3. THE IDENTIFICATION PROCEDURE
The procedure consists of the follow main steps. (i) (the extra factor M copes with the averaging over the
Choice of the broadband periodic excitation signals M experiments). Indeed, in the linear case the
and measurement of the frequency response function variability of the FRF measurement over the M
(result: FRF with its uncertainty). (ii) Approximation experiments with different excitation signals can be
of the measured FRF by a rational form in the Laplace explained by the disturbing input/output noise only. In
variable s (result: poles and their uncertainty). (iii) that case σ̂ Ĝ
2 (6) and σ̂ 2 (7) are both a measure of
Ĝ n
Selection of the poles of the rational form the influence of the disturbing input/output noise on
corresponding to the longitudinal or flexural vibration the FRF. If σ̂ Ĝ
2 (6) is larger than σ̂ 2 (7), then this is
Ĝ n
modes and calculation of Young’s modulus E(s k) via an indication that the systems behaves nonlinearly.
eq. (1), (2) or (4) (result: Young’s modulus E(s k) and Indeed, the difference
its uncertainty at the values of the poles s k ). (iv)
Approximation of E(s k) by a rational form in the σ̂ ĜS
2 = σ̂ 2 – σ̂ 2
Ĝ Ĝ n
(8)
391
13th IFAC Symposium on System Identification
40 40 40
0 0 0
Amplitude (dB)
Amplitude (dB)
Amplitude (dB)
-40 -40 -40
Fig. 3: Force-to-acceleration FRF (bold black line) Fig. 5: Comparison between the measured and
and its standard deviation (thin black line: (7) modelled FRF of the longitudinal vibration
noise errors only, and gray line: (6) noise errors experiment on the plexiglass beam. Bold black
+ nonlinear distortions) of the longitudinal line: Ĝ( jω) and G( jω, θ̂) , gray line: the 95%
vibration experiments. Left: plexiglass confidence bound of the FRF measurement
( L = 1.983 m ), right: copper ( L = 2.209 m ). ( 3σ̂ Ĝ ), and black line: the complex error
Ĝ( jω) – G( jω, θ̂) .
40 40
0 0
estimate θ̂ together with its covariance matrix
Cov(θ̂) . Finally, from θ̂ and Cov(θ̂) the poles and
Amplitude (dB)
Amplitude (dB)
-40
-40
their uncertainty are calculated.
-80 -80
-120 -120
Fig. 5 illustrates the procedure on the longitudinal
0 1000 2000 3000 4000
0 1000 2000 3000
frequency (Hz)
4000
frequency (Hz) vibration experiment of a plexiglass beam. Using a
rational form (10) of order n a = n b = 34 , the
Fig. 4: Force-to-velocity FRF (bold black line) and
approximation error is about at the level of the
its standard deviation (6) (gray line) of the
uncertainty of the FRF measurement (between -50 dB
flexural vibration experiments. Left: plexiglass
and -60 dB).
( L = 1.372 m ), right: copper ( L = 2.209 m ).
3.3. Calculation of Young’s modulus at the poles
quantifies the contribution of the stochastic non-linear The first step consists in selecting the poles s k of
distortions to the FRF measurement; while the non- G(s, θ̂) (10) corresponding to the longitudinal or
linear bias contributions G B are bounded by flexural vibration modes. This can easily be done by
comparing s k ⁄ ( 2π ) to the resonance frequencies of
1 the FRF. Next the index k of each observed
--- Mσ̂ ≤ G B ≤ γ Mσ̂ (9)
γ ĜS ĜS resonance peak (pole) is determined. This is done by
comparing the first few peaks of the measured FRF to
where γ typically lies between two and ten (see [15, the resonance frequencies predicted using eq. (1), (2)
17 and 18] for the details). or (4), the beam parameters, and a rough guess of
Young’s modulus and Poisson’s coefficient. Finally
The measurement procedure is illustrated in Figures 3 Young’s modulus is calculated at the values of the
and 4. For each measurement P = 10 and M = 25 . poles s k via eq. (1), (2) or (4), where (1) and (4) need
It can be seen that the stochastic nonlinear distortions a rough guess of ν(s k) ( ν = 0.33 for plexiglass and
are mostly dominant, and that the signal-to-noise ratio ν = 0.3 for copper). By an appropriate choice of the
Ĝ ⁄ σ̂ Ĝ of the FRF measurement (6) varies between beam cross section dimensions, the correction term in
40 to 60 dB. (1) depending on ν(s k) can often be neglected.
Equation (4) is a nonlinear algebraic equation in
3.2. Approximation of the FRF by a rational form E(s k) , and is solved via the Newton-Raphson root
finding algorithm [20]. As starting value we use the
According to the Mittag-Leffler theorem [19] the Euler solution (2) or the previous solution E(s k – 1) of
infinite dimensional transfer functions shown in (4). Uncertainty bounds on E(s k) are obtained from
Figures 3 and 4 can be approximated arbitrary well by the uncertainty of the poles through a first order
a rational form of finite order in a particular frequency sensitivity analysis of eq. (1), (2) and (4).
band
Fig. 6 gives the result for the longitudinal vibration
nb
B(s, θ) ∑ n=0 n
b sn experiments. The following standard deviations are
G(s, θ) = ---------------- = ---------------------------- (10) found
A(s, θ) na
∑ n=0 n
a sn
5
plexiglas: std(E(s k)) = 1 ×10 N/m 2
The parameters θ of the rational form are found by 6
(11)
minimizing the sample maximum likelihood cost copper: std(E(s k)) = 1 ×10 N/m 2
function (see [15] for the details). The result is an
392
13th IFAC Symposium on System Identification
3
E
194.8 195
DI
E E E
4
I II I I IEI IEI I EI
E E IE E DIDIDI
Amplitude (dB)
Amplitude (dB)
E E
III DIEDI
IEIII D D
194.6 2.5 194.5 3
Phase (°)
Phase (°)
E E DI DIDIDIEDIDI DIEDIDI DI
E E E D EDI DIEDI DIEDI DI EDI
194.4
E 2 E 194 IIDDDDDD D DDDD D 2 DI
IDIDD
D DD
DD
DI
194.2 1.5 193.5 1
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
|poles|/(2π) (kHz) |poles|/(2π) (kHz) |poles|/(2π) (kHz) |poles|/(2π) (kHz)
221.49 0.04 222 0.1
E E
Amplitude (dB)
Amplitude (dB)
E E I EIIIIIEI I I DI
DIDIDIDIDIDIDIDIDIDIDIIIIIIIII I
0.03
Phase (°)
221.5
Phase (°)
E E DDDDDD DI
221.48
E E E E DDDDDD
0.05
DIDIDIDIDIDIDIDIE DI
DDDDD DI DIDIDIDIDIDIEDIDIDIDIDIDIDIDIDIDI DI DI
0.02 221
D E E
221.47 0.01 220.5 0
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
|poles|/(2π) (kHz) |poles|/(2π) (kHz) |poles|/(2π) (kHz) |poles|/(2π) (kHz)
Fig. 6: Comparison between the measured and the Fig. 7: Comparison between the elastic moduli (N/
modelled elastic modulus (N/m2) at the poles s k m2) obtained from the longitudinal (o: Love
of the longitudinal vibration experiments: model (1)) and the flexural (+: Euler model
plexiglass (row 1), and copper (row 2). Circles: (2), x: Timoshenko model (4)) vibration
measurement E(s k) (1), and solid line: model experiments.
E(s k, θ̂) (row 1 (10), row 2 (12))
393
13th IFAC Symposium on System Identification
394