This is only a preview of the September 2024 issue of Practical Electronics. You can view 0 of the 80 pages in the full issue. Articles in this series:
Articles in this series:
Articles in this series:
Articles in this series:
Articles in this series:
Articles in this series:
Articles in this series:
|
Circuit Surgery
Regular clinic by Ian Bell
Topics in digital signal processing –
Filter Requirements
W
e are looking at various
topics related to digital signal
processing (DSP).
DSP covers a wide range of electronics
applications where signals are manipulated, analysed, generated, stored or
displayed as digital data. However, they
usually originate from and/or end up as
real-world signals for interaction with
humans or other parts of the physical world.
Fig.1 shows the key elements of a
generic DSP system with a signal path
from an analogue input via digital processing to an analogue output. This
does not necessarily represent every
DSP system (not all have all the parts
shown), but it serves are as reference for
the various subsystems we will look at.
Last month, we covered some key
topics in DSP relating to the frequency
domain view of sampled signals. To
depict the frequency domain, we used
graphs of signal strength versus frequency
– the signal’s spectrum.
We also plotted parameters such as
the gain of a circuit against frequency
as a frequency response graph. Looking at frequency responses and signal
spectra together shows us the effect
circuits or signal processing will have
on a signal and helps us understand
design requirements.
This month, we will look at some
aspects of how the spectra of sampled
signals informs the requirements for the
filters required at the inputs and outputs
of DSP systems – the antialiasing and
reconstruction filters shown in Fig.1.
Frequency Domain Recap
Our discussion last month concluded
with a graph of the spectrum of a generic sampled signal. A similar example
Analogue
In
Antialiasing
filter
Sample and
hold
is shown in Fig.2. In this example, we
show a signal where the original spectrum of the signal extends from DC to
20kHz (Fig.2a).
This shape is used to represent a general signal which has frequency content
from DC to 20kHz. It could be audio,
which has frequencies in this range,
but here we are not being specific about
what the signal is.
The spectrum of a 50kHz sampling
signal (an impulse train) is shown in
Fig.2b. As discussed last month, the
spectrum contains single-frequency
peaks (impulses) at integer multiples
of the sampling frequency.
Mathematically, the sampled signal’s
spectrum can be obtained by convoluting the sampled and sampling signals’
spectra –the result is shown in Fig.2c.
This is because the two signals are multiplied together (in the time domain),
which corresponds to convolution in
the frequency domain.
Convolution has a reputation for being
a difficult mathematical concept and, as
we have said, DSP in general is a mathsheavy topic. We are avoiding getting
into too much detailed mathematics
in this series, but will be highlighting
a few of the key ideas.
Last month, we referenced a YouTube
video that shows some graphical representations of convolution. Another
video, which readers with some mathematical background (functions, calculus)
who are interested in convolution might
find useful is from Professor Iain Collings (https://youtu.be/RmePGKWOSMQ).
It considers the continuous case,
rather than the simplified discrete example we looked at. The video starts
with the impulse response – another
important concept in signal processing.
Digital
ADC
Digital
processing
Analogue
DAC
Fig.1: A generic digital signal processing (DSP) system structure.
Practical Electronics | September | 2024
Reconstruction
filter
Out
We previously discussed impulses,
which are idealised zero-length pulses
with specific areas under the pulse. An
impulse response is the output a circuit
produces with single impulse applied
to the input, something we will mention again later.
Sampled Signal Spectra
Examples
For considering general, straightforward cases of sampled signals in the
frequency domain, you do not have to
perform detailed convolution calculations. As can be seen in Fig.2c, sampling
a signal at a constant sampling rate (fs)
results in the spectrum of the sampled
signal being replicated at integer multiples of f s . These are referred to as
replica spectra or spectral images.
We can use this fact to visualize spectra of sampled signals under different
scenarios, and look at the implications
for the design of DSP systems.
Consider what happens if we lower
the sampling frequency in the Fig.2 example to 40kHz and 35kHz. The 40kHz
xc(f)
a)
–100
–50
–2fs
–fs
–100
–50
–2fs
–fs
–100
–50
0
s(f)
50
100 f/kHz
fs
2fs
50
100 f/kHz
fs
2fs
50
100 f/kHz
b)
0
xs(f)
c)
0
Fig.2: Frequency domain views of a
sampled signal: (a) original continuous
time signal; (b) sampling signal
spectrum; (c) sampled signal spectrum.
41
–2fs
–100
xs(f)
–fs
–50
0
fs
2fs
50
100 f/kHz
Fig.3: A variation of the spectrum in
Fig.2(c), where the sampling frequency is
reduced to 40kHz.
sampling frequency case is shown in
Fig.3, where we see that the replicated
spectra (spectral images) of the original
signal in the sampled signal abut one
another in spectrum but are separate.
Due to the replicated spectra being
closer together, we can now see part of
the copies at +3fs and -3fs in the range
plotted in Fig.3.
The 35kHz sampling frequency case
is shown in Figs.4 & 5. Fig.4 shows the
replicas of the original spectra individually, with the ranges where they overlap
highlighted. In the overlap ranges, the
contributions of two the spectra add up,
so the overall spectrum will look something like the one in Fig.5 – this is a
sketch, it was not accurately calculated.
The labelled overlap in Fig.4 ranges
from fs – B, where B is the bandwidth
(maximum frequency) of the input
signal, to B, which is 15-20kHz in this
example. Other overlaps in the spectrum
are at the same positions relative to the
various replicas of the sampled signal.
Looking at the spectra in Figs. 2c & 3,
we can imagine that it is straightforward
to obtain the original spectrum of the
continuous signal, and hence an exact
copy of the original signal, by applying
a filter that can remove all the frequencies not in the original signal.
This is the role of the reconstruction filter shown in Fig.1. There are
some additional considerations due to
the DAC output not being an ideal impulse train. We will discuss that in a
later article; for now, assume that we
are filtering an ideal sampled signal in
–2fs
–fs
–100
–50
xs(f)
–fs/2
0
fs/2
–100
100 f/kHz
–50
fs
0
–50
0
–100
xs(f)
–50
–fs/2
0
fs/2
50
100 f/kHz
–100
–50
0
0
2fs
50
100 f/kHz
t
Fig.6: A time domain (waveform) view of
a sampled signal.
signal frequency of 20kHz – there is no
headroom, and we definitely need an
ideal filter to achieve this.
Ideal filters have infinitely fast cut-off
(sometimes referred to as “brick wall” filters for this reason), which is not possible
in reality. We will discuss the implications of this in more detail shortly.
If we attempt to recover the sampled
signal shown in Fig.5 using an ideal low
pass filter with a cutoff of fs/2 (17.5kHz
in this case), we fail to obtain the original
continuous signal. This is shown in Fig.9,
where the difference in the shape of the
spectrum from Fig.2a can be observed.
A different spectrum means a different
continuous time waveform and hence
failure to reconstruct the sampled signal.
The overlapping parts of the spectrum
with the replicas produces additional
frequency content in range of 15kHz
to 17.5kHz, where the overlap region
is inside the filter’s pass band.
The original signal bandwidth was
20kHz, so perhaps we should use an
ideal filter with this cutoff – however,
that will not help reconstruct the signal
correctly, as the extra bandwidth is also
in the overlap range and will contain
Ideal filter
response
fs
2fs
50
–50
fs
Sampled signal xs(t)
An ideal low pass filter with a cutoff
of fs/2 can do a perfect job of recovering
or reconstructing the original continuous signal from a correctly sampled
signal, as shown in Figs.7 & 8
For the situation in Fig.7, the cutoff
frequency is 25kHz and there is some
‘headroom’ between this and maximum
signal frequency of 20kHz. For the situation in Fig.8, the cutoff frequency is
20kHz, which is equal to the maximum
–fs
Overlap
xs(f)
Fig.5: A sketch of an actual spectrum
resulting from the situation in Fig.4, with
the overlapping signals combining.
Reconstruction Filtering
Examples
–2fs
–100
–fs
–2fs
100 f/kHz
–100
–50
xc(f)
Fig.7: Filtering the sampled signal spectrum
from Fig.2c with an ideal filter to recover the
original continuous signal (Fig.2a).
42
100 f/kHz
form of a weighted impulse train –
see Fig.6. For more detail on this,
refer to the discussion relating
to Fig.10 in last month’s article.
The fact that it is possible (at
least in theory) to perfectly recover a correctly sampled continuous
signal from the sampled data is an important concept in DSP.
That applies to the simple cases
shown here, where the there is no
actual processing (we are just looking
at the pure sampled signals), but also
to digitally-processed (eg, filtered) and
digitally-created (synthesised) sampled
signals, as long as the requirements discussed below are met.
Of course, absolutely perfect reconstruction of a continuous signal from
samples requires ideal circuitry and will
not be achieved in practice; however,
with good design, very high performance is possible.
xc(f)
–100
50
–fs
–2fs
2fs
Fig.4: A variation of the spectrum in
Fig.2(c), where the sampling frequency
is reduced to 35kHz. The overlapping
ranges are highlighted.
Ideal filter
response
fs
2fs
50
–fs
–2fs
Overlap
xs(f)
xs(f)
–fs/2
fs
0
fs/2
Ideal filter
response
2fs
50
100 f/kHz
50
100 f/kHz
xc(f)
50
100 f/kHz
Fig.8: Filtering the sampled signal spectrum
from Fig.3 with an ideal filter to recover the
original continuous signal (Fig.2a).
–100
–50
0
Fig.9: Filtering the sampled signal spectrum
from Fig. with an ideal filter fails to recover
the original continuous signal (Fig.2a).
Practical Electronics | September | 2024
xs(f)
19kHz
0
19kHz
35kHz
16kHz 19kHz
50
f/kHz
Fig.10: A 19kHz component in the original
signal produces an unwanted 16kHz signal
due to the limited 35kHz sampling rate.
additional frequency content. It is not
possible to define a filter that will recover the signal correctly in this case.
Nyquist and Aliasing in the
Frequency Domain
In the first article in this series (May
2024), we discussed the NyquistShannon sampling theorem and the
concept of aliasing. The theory shows
that if the input does not contain any
frequency components at or beyond
half the sampling frequency, it is possible to perfectly reproduce the original
signal from the sample data.
The scenarios shown in Figs.7 & 8
meet the Nyquist sampling requirement.
When we look at the spectra, we see that
sampled signal can be reconstructed
correctly if a suitable reconstruction
filter is used.
If the Nyquist sampling requirement
is not met, the sampled data is ambiguous; that is, there could be more than
one signal that results in the same set
of sample values. Therefore, signals reconstructed from the sampled data will
not match the original input.
This ambiguity is called aliasing –
one signal is an alias of the other, and
they cannot be distinguished. The discussion in the May article just looked
at the time domain signals. Now we
have seen how this problem manifests
in the frequency domain.
For the scenario in Fig.9, the Nyquist
Fig.11: An LTspice schematic for simulating aliasing examples.
sampling requirement is not met, so the
signal is aliased.
The labelled overlap region shown in
Figs.4 & 5 runs from 15kHz, which is
the sample frequency minus the maximum input signal frequency of 20kHz
(35kHz – 20kHz = 15kHz). This overlap range ends at 20kHz (the maximum
input signal frequency).
Fig.10 shows a zoom-in on part of Fig.4
in which one aliased frequency from the
original signal (19kHz) is highlighted.
For the replica spectrum centred on
fs, the 19kHz component of the original signal (which has components at
±19kHz due to the symmetry of the
spectrum in negative frequencies) produces replicas at fs ±19kHz, that is at
35kHz - 19kHz = 16kHz and 35kHz +
19kHz = 54kHz.
Here we consider the 16kHz signal,
which is inside the original sampled
signal bandwidth of 20kHz – it is therefore an alias signal.
If we filter the sampled signal with
an ideal low-pass filter with a cut-off
frequency 20kHz, the alias of 19kHz at
16kHz would not be removed by the
filter, but it should not be there – it
was not in the original sampled signal.
Fig.11 shows an LTspice circuit thatcan demonstrate the aliasing situation
Fig.12: The results of simulating the circuits shown in Fig.11.
Practical Electronics | September | 2024
just discussed in the time domain. This
is similar to the example in the May
issue, but with signals relevant to the
current example.
Sinewaves at 19kHz and 16kHz are
generated. The 19kHz waveform is
sampled at 35kHz. This is achieved by
multiplying the sinewave from source
V3 by the sample pulse train from source
V2 using behavioural source B1.
The top trace in Fig.12, plotted by the
simulation of Fig.11, shows the 19kHz
waveform and the samples obtained from
it. 19kHz is beyond the Nyquist rate of
17.5kHz for 35kHz sampling, so aliasing will occur, with an alias at 16kHz.
The bottom trace in Fig.12 shows
the 16kHz waveform together with the
samples from the 19kHz waveform.
All the samples are exactly on points
on the 16kHz waveform, showing that
the same set of samples could have
been obtained from either a 16kHz or
19kHz waveform.
Reconstruction Filter
Requirements
If we sample a signal with frequency
content right up to the Nyquist rate (just
less than half the sampling frequency,
as in Fig.8), no aliasing occurs in the
sampling process. However, the signal
can only be recovered using an ideal
“brick wall” reconstruction filter (as
shown in Fig.8), which is impossible
to do in practice.
If the sampling frequency is above the
Nyquist rate, there is some headroom, as
shown in Fig.7. However, this does not
guarantee that the signal can be perfectly reconstructed. Ideally, the filter must
completely remove the replica spectra,
but the closer the sampling rate is to the
Nyquist rate, the more demanding the
filtering requirements become.
That is why it’s common for audio to
be sampled at either 44.1kHz or 48kHz
when the highest frequency that’s intended to be reproduced is 20kHz. That
gives 4.1kHz or 8kHz of headroom, allowing the reconstruction filters to have
very little attenuation up to 20kHz, but
43
xs(f)
–fs
–B
+B
Real low-pass
filter response
fs
f
xc(f)
–1/T
–B
–1/2T
x(f)
Ideal low-pass filter
removes frequencies > fs/2
which will cause aliasing
+B
1/2T
1/T
xc(f)
f
–B
–fs
f
+B
–fs/2
x(f)
Real low-pass filter will
need cut-off below fs/2
to sufficiently attenuate
signals at frequencies
above fs/2
fs/2
fs
fs/2
Bf
fs
x(f)
Unwanted content
in filter output
–fs
–B
+B
fs
f
Fig.13: Filtering a sampled signal with a
non-ideal reconstruction filter may result in
unwanted frequency content in the output
due to imperfect reconstruction.
significant attenuation from 22.05kHz
or 24kHz and up.
The problem with trying to reproduce
frequencies right up to the Nyquist limit
is illustrated in Fig.13, which shows a
case where a realistic reconstruction
filter fails to remove part of the spectral
images. The result is unwanted frequency
content in the output, which means the
output waveform will have a different
shape from that implied by the sample
data – the reconstruction is imperfect.
Antialiasing Filter
Requirements
Similar challenges exist with the requirements for antialiasing filters. If we
had an ideal filter, we could simply set
the cut-off to be equal to the Nyquist
rate, and that would remove any possibility of aliasing, as shown in Fig.14.
That figure shows a continuous signal
spectrum which is to be sampled (like
Fig.2a), with a bandwidth (B) extending beyond the Nyquist rate. Hence, it
will cause aliasing. The filter removes
all frequencies above half the sample
rate (fs/2) to produce a signal with a
new bandwidth (BF = fs/2), which will
not cause aliasing.
We cannot use an ideal filter in real
–fs
–fs/2
–Bf
fs/2
Bf
fs
f
Fig.14: Filtering a continuous signal with an
ideal antialiasing filter.
–fs
–fs/2
–Bf
f
Fig.15: Filtering a continuous signal with a
realistic antialiasing filter.
being sampled to guide the antialiasing
designs, but we can use a cutoff frefilter design.
quency below the Nyquist rate so that
Both antialiasing and reconstruction
the attenuation at the Nyquist frequency
filters require trade-offs. Filters with
is sufficient not to cause any problems.,
very sharp cut-offs are larger, more
as illustrated in Fig.15.
complex and more expensive than lowerHowever, one problem here is that
performance filters, but provide better
part of the signal that can be sampled
signal quality into and out of the DSP
correctly at the chosen sample rate is
system.
attenuated – the sampled signal will not
Increasing the sample frequency (for
be aliased, but it will not be a perfect
a given input signal bandwidth) reducrepresentation of the original signal in
es the demand on the filtering, as there
the ±fs/2 bandwidth range.
is larger range up to the Nyquist rate,
The simplified example in Fig.15
over which the filter’s gain can decrease
shows all the unwanted content rewithout affecting the wanted signal.
moved, but a real filter may not be able
This means lower-performance filto achieve that, resulting in some level
ters can be used, or a filter cut-off rate
of aliasing. Changing the filter charactersuitable for a lower sampling rate will
istics will control how much the parts
give better quality. However, a higher
of the signal that can be correctly samsample rate increases the demands on
pled are unfortunately removed, and
other parts of the system (faster ADCs
how much aliasing occurs.
and DACs, more data to handle at a
If the signal content above the Nyfaster rate) so such an approach has its
quist frequency (particularly close to the
own limitations.
Nyquist frequency) is not strong, the filPE
tering requirements
are less demanding.
Simulation files
So it is important (if
Most, but not every month, LTSpice is used to support
possible) to have a
descriptions and analysis in Circuit Surgery.
good understandThe examples and files are available for download from
ing of the frequency
the PE website: https://bit.ly/pe-downloads
content of the signal
Terrington
Components
• Project boxes designed and manufactured in the UK.
• Many of our enclosures used on former Maplin projects.
• Unique designs and sizes, including square, long and deep
variaaons of our screwed lid enclosures.
• Sub-miniature sizes down to 23mm x 16mm, ideal for
IoT devices.
MADE IN BRITAIN
www.terrington-components.co.uk | sales<at>terrington-components.co.uk | Tel: 01553 636999
44
Practical Electronics | September | 2024
|