FSV detail (from IEEE EMC Symposium 2005)
The structure of the FSV involves
reading the two data files to be compared and interpolating them over the
common window (often common frequency range) so that the data-points to be
compared are coincident. This approach
ensures that like is being compared with like and will not affect the overall
results unless the data are severely under sampled. It must be remembered that the purpose of the
FSV is to mimic a visual comparison and so long as any interpolation does not
produce visually different results, the approach is perfectly acceptable.
The actual comparison is based on
decomposing the original data into trend information and feature
information. So the next step is to
Fourier Transform the data and window the transformed data to separate out the
lower and higher portions of the data.
The high and low portions are then inverse transformed back into the
original domain. Combinations of these
filtered data sets and their derivatives are used to compute the Amplitude
Difference Measure (ADM) and the Feature Difference Measure (FDM): which can be
combined into the Global Difference Measure (GDM).
More specifically, the procedure
is:
- Read data, obtain the overlap window and
interpolate the data, if necessary, in the overlap region to ensure
coincident pairs of data points.
This ensures that that the two data sets to be compared have the
same number of data points located at the same positions on the
independent (x) axis.
- Fourier Transform both data sets. Depending on whether a Fast or Discrete
approach is used, this may require sufficient zero padding to enable
conversion.
- Calculate the ‘low’ data sets using the
transformed data.
- Ignore the first four data points in the
transformed set (in order to avoid DC and very low frequency components)
and sum the intensities of the remaining data.
- Obtain a 40% location by summing the data from the
fifth point (i.e. ignoring the near-DC data) until the total reaches 40%
of the total value calculated in step 3a.
The ‘40%’ location used by the FSV is the lowest of the two
resulting numbers (from the two original data sets). A ‘break-point’ five
data points above this is returned – a value that allows a comfortable
transition window between the low and the high results.
- Window the transformed data for both data sets by
taking a linearly decreasing envelope from three (or five) points below
the break-point to three (or five) points above it. Essentially, low-pass filtering the
transformed data
- Inverse transform the windowed data to give the
‘low region’ data for both original data sets. (Lo1(f) and Lo2(f) it is assumed for ease of
representation that the original data is frequency domain based, although
it could equally well be time domain or space domain based)
- Calculate the ‘high’ data sets using the
transformed data. Repeat the
process from 3c, by applying the opposite envelope to the transformed
data: essentially high pass filtering it.
This data is then inverse transformed to give the ‘high region’
data for both of the original data sets. (Hi1(f) and Hi2(f)
again assuming frequency domain for ease of representation)
- Calculate the ADM on a point-by-point basis.
Given that each data set has N points and, at an arbitrary data point f,
the ADM is calculated from the following equation. This point-by-point representation can
be abbreviated ADMi
(1)
- Calculate the single value of ADM. A mean value of the ADM(f) gives an
overall, single figure, goodness-of-fit.
It is obtained from the following equation, noting that fmin
and fmax are the lower and upper limits for the data
range. Note: a median value, rather
than a mean value has demonstrated some improvements in agreement with
visual interpretations.
(2)
- Calculate the ADM confidence histogram. The range of values for the ADM, and, in
fact, the FDM and GDM can be
divided into six categories, each with a natural language descriptor:
Excellent, Very Good, Good, Fair, Poor, Very Poor. These are the terms that are most often
used in descriptions of the quality of comparisons. The confidence histogram, like a
probability density function, provides some intelligence as to how much
emphasis can be placed on the single figure of merit. There is some evidence to show that this
mirrors the overall group assessment of any data pair by a number of
engineers. The determination of the
histogram is simply a case of counting the proportion of points that fall
into one of the categories, according to the rule base in the following
table.
-
FSV interpretation
scale
|
FSV value (quantitative)
|
FSV interpretation
(qualitative)
|
|
Less than 0.1
|
Excellent
|
|
Between 0.1 and 0.2
|
Very good
|
|
Between 0.2 and 0.4
|
Good
|
|
Between 0.4 and 0.8
|
Fair
|
|
Between 0.8 and 1.6
|
Poor
|
|
Greater than 1.6
|
Very poor
|
- Calculate derivatives in preparation for the FDM
calculation. The following
components need to be calculated: The
first derivatives of the Lo(f) and
Hi(f) data sets and the second derivatives of the Hi(f) data sets. The derivatives accentuate the high
rate-of-change features in the original data and differences based on the
derivatives are combined in the determination of the FDM. The first derivatives are currently
obtained by a simple difference approach (i.e. ignoring the x-axis data)
so, for example, Lo’(f) = Lo(f+2) – Lo(f-2). The second derivatives of the Hi data
sets are obtained from the first derivatives using a simple difference
scheme again, in this case, Hi”(f)
= Hi’(f+3) – Hi’(f-3). This
simple difference approach was used in the original formulation and in the
current free-standing application produced by the authors.
- Calculate the point-by-point FDM. The FDM is
formed from three parts based on the derivatives calculated in step 8.
(3)
(4)
(5)
(6)
- Calculate the single value of FDM. This is done in exactly the same way as
for the ADM.
- Calculate the FDM confidence histogram. This is dome in exactly the same way as
was done for the ADM.
- Obtain the point-by-point GDM value. The GDM is premised on the ADM and FDM being
largely independent, which means that:
(7)
- Calculate the overall GDM value and the GDM
confidence histogram. This
follows the same procedure as the ADM and FDM.
- Determine the equivalent visual scale values for
ADM, FDM, and GDM. The FSV
values can be scaled to a visual six point scale (where Excellent = 1 and
Very poor = 6). The piece-wise
approach for this is given in Table II, where X is the ADM, FDM or GDM
(total or point-by-point):
Piecewise visual
conversion
|
If X £
0.1 Then V = 1 +10X
|
|
If X > 0.1 and X £ 0.2 Then V = 2 + 10(X –
0.099)
|
|
If X > 0.2 and X £ 0.4 Then V = 3 + 5(X –
0.199)
|
|
If X > 0.4 and X £ 0.8 Then V = 4 + 2.5(X
– 0.399)
|
|
If X > 0.8 and X £ 1.6 Then V = 5 + 1.25(X
– 0.799)
|
|
If X > 1.6 Then V = 6
|
Back to front page