68 4. MEASURED DATA ANALYSIS
4. Divide the result by the filter bandwidth, f , resulting in the autospectrum (or power
spectrum),
G
xx
.f / D
Nx
2
.f; f; t/
f
D
1
.
f T
/
T
Z
0
x
2
.f; f; t/dt: (4.10)
e (real-valued) autospectrum provides information on the frequency distribution of the
standard deviation of a time history record. It therefore has the following property:
2
D
1
Z
0
G
xx
.f /df : (4.11)
A modern digital, state-of-the-art method for calculating the autospectrum employs the
finite Fourier transform of the time history record. By subdividing the record into “n
d
” distinct
sub-records or “windows” (of duration, T ) the autospectrum is defined as,
G
xx
.f / D
2
.n
d
T /
n
d
X
iD1
X
i
.f / X
i
.f /: (4.12)
X.f / is the finite Fourier transform of the time domain series, x.t /, where the temporal sam-
pling rate is t. e autospectrum, G
XX
.f /, is calculated at discrete frequencies, f
k
D k f ,
where 0 k n
w
=2. Important guidelines for computation of an appropriate autospectrum are:
f
max
D
1
2 t
(Nyquist frequency); f D
1
T
(bandwidth resolution)
n
w
D
T
t
(discrete Fourier transforms “window” length)
T
max
D n
d
T (total record length for “n
d
” distinct averages):
(4.13)
In practice, the originally measured “analog” signal, x.t/, should be low-pass filtered with
a “cut-off” frequency (typically 0:8 f
max
) in order to avoid aliasing before conversion to a time
series with digital bandwidth, f
max
. In addition, a sufficiently long data record, T
max
, should be
recorded to estimate an autospectrum associated with a desired bandwidth resolution, f , and
number of distinct averages, n
d
(typically in excess of 10 to minimize contributions associated
with extraneous noise sources). It should finally be noted that specialized “windowing” functions
and correction factors (e.g., the Hanning window and “bow-tie” correction factors) and overlap
processing (non-distinct successive records) are commonly applied practices [1].