A DSP Version of



                            Coherent-CW (CCW)



           

               Obtain the advantages of coherent detection

                       of Morse signals using DSP

              

                   Coherent CW (CCW) is an old technique for
digging weak CW   signals out of the noise. The system
synchronizes the receiver     to the sender's keying. Thereafter,
due to the rules of Morse code (each DAH equals exactly three
"dit-times" and spaces between marking elements are always an
integral multiple of one dit time)the receiver knows precisely
when the carrier is allowed to turn on or off. This is useful
information that would otherwise have to be transmitted at the
expense of an additional power and/or bandwidth. Of course, the
sender must have an absolutely rock-steady rhythm for the scheme
to work. For this reason, CCW stations employ special KEYERS or
keyboards which can generate perfectly timed code. Traditionally,
CCW uses a keying rate of 12WPM (1OO milliseconds per dit) and a
received audio tone of 800 Hz.  The receiver divides its time
into 100-ms windows, or frames. Once synchronized, there is
complete certainty that the incoming tone is either present or
absent for the entire 100-ms window the rules of Morse code
prevent it from switching somewhere in the middle. This knowledge
makes it possible to use a matched filter(sometimes called an
integrate-and-dump filter) to reduce the received bandwidth down
to about 9 Hz (main lobe), eliminating most of the noise while
letting the CW tone pass through unscathed (see Fig 1). All
coherent 800-Hz energy received during the window is integrated
(accumulated), and only at the very end of each window does the
receiver make its decision and the question is: during the
previous 100 milliseconds, was the key at the transmitter up or
down?  The question is answered by comparing the measured
amplitude of the800-Hz tone to some threshold value more received
energy means the key probably was down, less energy means it
probably was up. If the receiver decides the key was down, it
sounds a local side tone oscillator for the next100 ms. Apart
from the 100-ms delay between the incoming audio and the
regenerated tone, the side tone signal follows the received
code,faithfully,and of course there is absolutely no electrical
noise present from that point on since the signal has been
completely rebuilt by the receiver. It is  like a repeater on a
fiber-optic cable.  At each stage, hardware makes a firm decision
based on a marginal situation then announces confidently whether
the window was marking or spacing.  Noise (at least in its common
form) is completely removed from the incoming signal, so a human
operator can copy it without stress. That is, as long as the
hardware makes the right decision at the end of each window! If
the signal-to-noise ratio is so awful that the matched filter
gets a wrong answer, a different kind of noise enters the
picture: the system still generates perfect code at its output,
but it's no longer the same message the transmitter originated.
Then the human operator can step in and say, "Hey, that didn't
make sense," and try to figure out what the real message must
have been based on context, prior knowledge, etc.  It has been
argued that a seasoned, skilled CW operator can perform the same
filtering operation as CCW performs with his brain, concentrating
on the incoming code while blissfully ignoring all the other
stuff in the receiver's passband at the time. Maybe so, but it's
very hard work. An equally narrow(9-Hz) conventional audio
filters wouldn't help either: the circuit's response time would
be so slow it would spread out the leading(attack; and falling
(decay) edges of the keyed waveform in time, blurring the
decision (lnhrtwsf~n markR and February 1994 25 spaces. After all
the votes are in, it would seem that coherent CW carries an
advantage of some 20 dB over regular CW at 12 WPM. So why doesn't
everybody do it that way?  Answer: Up until now it has been quite
difficult to get a CCW  station on the air. Once synchronization
had been established, in order to keep the transmitter and
receiver from drifting apart,expensive frequency standards were
needed at each end of the link. Transceivers had to be
stabilized, usually by phase-locking their master oscillators to
some common external standard such as WWVB. And that
integrate-and-dump filter circuit was no piece of cake to build
and align. These technical challenges have kept all but the most
dedicated devotees away from the mode until now.  Can DSP Help? 
Some years ago I designed and built a DSP engine dedicated to
receiving CCW. It worked, but it was far too complicated some
fifty ICs on the board. I concluded that no one else would ever
build one of those things, and I was right. At the time, personal
computers were just coming on the scene and I didn't think there
was enough processing power in them to do the job. But then there
appeared ATs,386s, 486s, and I thought it might be worth another
look. Maybe, with very careful coding, we could do the DSP
function on existing ham-shack computers with a minimum amount of
external hardware. If so, the cost would be next to nothing and
many more people could get in on CCW. After doing an audio
spectrum analyzer project (see "A Receiver Spectral Display using
DSP," in Jan 1992 QST) I realized that a CCW program using the
same analog interface was feasible. This interface is a
Sigma-Delta analog-to-digital converter circuit that measures
about 4 inches by 1.8 inches and runs off a 9-V battery. It uses
a handful of common CMOS chips and can be built in an evening. It
also can be purchased assembled and tested,ready to hook up. This
circuit samples the received audio 7,200 times each second,
converts the measured voltages to numbers, and passes these to
the computer through one of its serial ports (eg, COM1) running
at 115kbaud. No modifications to the radio or the computer are
required. It is exactly the same board described in the QST
article and also in the '93 Handbook. Some of you will already
have one.  What Does the Software look like? The primary thing to
keep in mind is that we are going to be hard pressed for time. At
7,200 samples per second,a new sample point's numerical value is
fed into the computer every 139 microseconds. The software has to
service the UART (serial port) interrupt,read the measured
voltage, and perform the DSP filtering operation 7200times per
second. With real tight coding it can be done, and in the time
leftover we can do some other pretty neat things as well.  The
challenge is to reduce the time spent handling interrupts to the
absolute minimum. The computer has to be able to process the
incoming numerical samples at least as fast as they're acquired
(ie, in less than 139 ms per sample, on average).  For this
algorithm to work, we cannot afford to turn off the interrupts
while we do some heavy computing; the interrupts have to be
always enabled to guarantee that we don't  miss a single sample. 
DOS systems usually have several background tasks that become
active periodically, sometimes shutting off the system interrupts
for up to a millisecond at a time. So the program's first action
is to take over the DOS timer interrupt to make sure no other
program gets control of the machine even fora milliseconds while
the DSP algorithm is running.  In writing the program, I
concentrated on shaving cycles from the serial port interrupt
service routine. In real-time programming,what's crucially
important is to reduce the amount of time the computer spends on
things which are done often(eg, 7,200 times every second). Things
which happen less frequently can be coded a little more sloppily.
How to minimize the interrupt service time?  Well, we would like
very much to avoid any particularly long machine instructions,
such as multiply or divide. The classic integrate-and-dump(I&D)
filter works by first shifting the frequency of the incoming
800-Hz tone down to a baseband" (dc). This is done by mixing the
audio with an 800-Hz reference tone. Once at baseband, the energy
in the signal is split into two separate channels called In-phase
(I)and quadrature (Q). These channels are then independently
integrated over the 100-ms window. This is essentially an
averaging over time operation and is only feasible because at0 Hz
(dc) the I and Q channels don't change their values with time.
However, if the incoming tone is not exactly at 800 Hz, a beat
note will be generated and will cause problems for us, at least
to some degree. Let's take an example. At 800 Hz, a 100-ms frame
consists of exactly 80 cycles of that sinusoidal waveform. We mix
it with our 800-Hz reference tone and we get exactly 0 Hz, or dc,
which can be averaged. Now let's say the incoming frequency was
not 800 Hz, but rather790 Hz. The beat note will be 10 Hz.But if
you look at it over 100 ms, or a tenth of a second, there will be
only one complete cycle. If we take the average value of its
voltage over that 100-ms period, we come up with zero! It's
positive for half the time, equally negative for the other half,
and the average value is zero volts. It matters not one iota what
the starting amplitude of that 790-Hz tone was; the output of our
I&D filter will be zero. And likewise for any frequency that is
spaced away from 800 Hz by an any exact multiple of 10 Hz.
Between 800 Hz and 790 Hz,the filter's response takes on
intermediate values, ranging all the way from0 dB at 800-Hz down
to minus  infinity dB (total attenuation) at 790 Hz.  At the end
of each 100-ms integration period, the classic hardware filter
freezes the instantaneous voltages on its integrating capacitors
and, based on the values of the I and Q channels averaged over
the interval, computes the average amplitude of the 800-Hz tone
during that period. (It could also compute the phase of the tone
relative to that of the 800-Hz local reference oscillator, but
that generally is not done. ) The hardware version of an I&D
filter then has to dump all the charge from those capacitors
instantly to start measuring the signal in the next window. This
is physically impossible with real hardware integrators, so in
practice we always waste a little of each frame right at the
beginning while the capacitors are discharged of the voltages
built up during the previous frame. (Well, not absolutely
impossible: one could set up two complete filters and switch
between them on alternate frames, I guess.).  So, how can we get
the same answer with a DSP algorithm, get around the old
problems, and hopefully avoid having to build the hardware filter
at all?  Our basic challenge is to estimate the amplitude of an
800-Hz sinusoid which is assumed to be unvarying throughout the
entire 100-ms window,and which is likely to be much weaker than
any number of interfering carriers present in  the receiver's
passband at the same time.  The classical DSP solution to this
would be to emulate the hardware  I&D filter in software.  We
multiply the incoming samples with a unit vector rotating at 800
Hz to get instantaneous values of the  I and Q components of the
baseband signal, then we integrate (add up) all these values
over100 ms, eventually dividing by the total number of samples
taken to obtain the averaged I and Q values.  Then we compute the
square root of the sum of the squares of the two mutually
orthogonal components, and that's the answer. (We could also
calculate the phase of the received  sinewave by computing the
arctangent of the ratio of the Q and I components. ) In this
particular case, at7,200 samples per second, each 100- ms window
would consist of 720 samples.  Which would mean720 x 2
multiplications, 720 x 2additions, plus a whole bunch of other
time consuming stuff at the end of each window (such as
calculating the square root, clearing the accumulators,etc). All
of this is mathematically equivalent to ~fitting" an800-Hz
sinusoid to the sampled data points by least-squares, then
solving for its amplitude and phase. That's alot of number
crunching to get done in tenth of a second, even for today's
faster home computers. And there isn't a whole tenth of a second
available either a good portion of the time February 1994 27 goes
for overhead (pushes, pops, etc)in servicing those 720 interrupts
the DSP algorithm has to operate in whatever time is left over
after all that. Soit's tough, eh? Well, here is where we start
getting lucky. Oh most computers a multiply instruction takes
much longer to execute than a simple addition, so we would be far
ahead if we could eliminate those two multiplies per sample and
it just so happens we can. We know that at 7,200 samples per
second, each individual cycle of an800-Hz tone takes exactly 9
samples to cover. In other words, during the time between each
sample and the following one, an 800-Hz sinusoid will have
advanced through 40 degrees of phase. And 40 x 9 = 360 degrees,
which is one complete revolution exactly. Eureka! After 9 samples
have been precessed, the coefficients we have to multiply the
samples with will start to repeat, taking on the same sequence of
values for the next 9 samples, and so on. So of our 720 samples
in total,80 of them will be multiplied by two particular numbers,
another 80 will be multiplied by a new two-number set,etc. This
good fortune is entirely attributable to the fact that our
sampling rate just happens to be an exact integral multiple of
the frequency of the sinusoid whose amplitude we want to measure.
We can take advantage of it by using something called the
distributive law (in algebra): AxB + AxC+ AxD equals Ax(B+C+D).
You get exactly the same answer in the end,but one way needs
three multiplies and two adds, the other way needs only one
multiply and two adds. We will use this to solve for the average
I and Q component values with only 18 multiples per 720 samples
instead of 1440! But there is another trick available.  It turns
out that the cosine of 40 degrees (one of our sample phases) has
the same value as the cosine of 320 degrees (another one of our
points).  Likewise, of the remaining seven coefficient, six of
them can be paired up in this way, the seventh is unity, and it's
real easy to multiply by one! On the sine side, 8 of the 9
coefficients can be paired (allowing for sign changes),the other
coefficient is zero (it's even easier to multiply by that, hi!).
The bottom line is that by combining terms we can get by with
just 8 multiplication instead of the 18 that would otherwise be
needed. This is starting to look doable! What we have to obtain
during each100-ms window are the end values of nine (9)
accumulators, and on each of the 720 interrupts we only have to
add the current sample into one of those nine accumulators. For
each interrupt we sum into another accumulator, and after the
ninth sample has been processed we start over with the first
accumulator. Thus, aside from the housekeeping associated with
servicing the interrupt, checking for overrun and/or clipping,
figuring out which accumulator to address and such mundane
things, we are left with a simple addition to perform. For sure,
at the end of the 100-ms window we must do some further
calculations, but we have all the time in the world (relatively
speaking) to get them done in the 100 ms during which the data
for the next-following window are being acquired.  Now, there's
one little complication: I would like to run not one, but three
(3) I&D filters concurrently. These three filters should overlap
in time so that by comparing the three outputs I can decide if
the window phase is drifting and it would be advantageous to make
a slight adjustment to the phasing cycle that determines when
each window starts and ends. The most economical way to run three
such I&D filters concurrently is never to clear the various
accumulators. Instead, I keep a "delay line" in memory,
consisting of the last 720samples taken. For each new sample, I
add it into the appropriate accumulator, then subtract off the
value of the sample taken 720 time slots earlier. In this way the
accumulators are guaranteed not to overflow, because in the long
run we subtract out as many counts as we add in. Furthermore,at
the end of every 9 samples (one cycle of the 800-Hz audio
signal)the nine accumulators always hold exactly the same numbers
you'd get if you started 100 ms ago with cleared accumulators and
integrated through one complete window's worth of samples. Try
doing that with analog integrators! It works on the computer
because digital fixed point arithmetic is absolute there are no
errors such as would arise in an analog integrator due to charge
leaking off a capacitor,component values changing slightly with
temperature, etc. Any analog integrator would eventually saturate
at one rail or the other due to such errors. The digital
integrator can run for hours (or days) with absolutely no
accumulated long-term error. We still must ~sample and hold" the
values in those nine accumulators whenever we have to compute the
amplitude of the measured component over that particular window.
This involves moving nine 16-bit numbers to secondary storage
positions. It is done with a single block move instruction and
has no impact whatsoever on the processing of subsequent samples,
soothe computer version can start to process each new window
immediately(not throwing away any of the incoming energy) as
opposed to the analog version, which has to wait for the
capacitors to discharge fully before starting a new integration
cycle. At the end of each window, we must then compute the
amplitude of the measured 800-Hz component. This involves those
eight multiplications mentioned above, then taking the square
root of the sum of the squares.  This is done in tightly coded
assembly language in order to execute in the shortest possible
time. As a fine-tuning aid for the operator, the program also
measures the frequency of that 800-Hz component (to the nearest
tenth of a Hz) and displays it on the computer screen, updated at
the end of each 100-ms window.  There is a trick to this: the
frequency has to be measured after the DSP filtering operation.
Otherwise, any nearby strong carrier could disrupt the
measurement and give an erroneous reading. In fact, during each
marking window (once we have decided that the 800-Hz tone was
indeed present during that window) we also measure its phase
(averaged over the entire100-ms window) and save this for later
reference. On the next marking window, as long as it is not too
far in time away from the last measured one, we measure the phase
again. If there is any slight discrepancy between the frequency
of the incoming 800-Hz tone(from the receiver) and our
precise800-Hz reference tone (which is actually determined by the
1.8432-MHZcrystal oscillator on the Sigma-Delta board), there
will be some phase shifting the detected signal (in the same way
the amplitude of the "beat note" varies regularly whenever there
is a slight difference between two compared frequencies). If we
know the amount of phase shift as well as the time interval over
which this phase shift accumulated, we can figure out how much
the received frequency differs from the nominal 800-Hz value, and
there you have it. The operator can set a software switch to
enable this phase comparison to also occur across an intervening
space frame (or not). If the transmitter is phase-coherent from
one key-down to the next, then it is feasible to use those two
key down periods to measure his frequency. If not, then the
software only uses phases measured in consecutive marking
frames(eg, during a "dah"), when the key is presumably held down
for 300 ms continuously and the transmit carrier could not change
its phase during that period.  That is the essence of a DSP 
version of the integrate-and-dump filter. The algorithm runs on
just about any IBM-compatible computer. The program incorporating
this algorithm is called Coherent. There is enough time left over
after the DSP calculations to allow for lots of other CCW
goodies.  For instance, Coherent has an auto-tune feature. It is
important to keep the incoming audio tone centered in the
filter's rather narrow passband. Coherent tracks the incoming
signal's frequency. If it deviates more than half a hertz from
the nominal 800-Hzvalue, Coherent issues a pulse on one of two
RS232 control lines to make the receiver tune up or down by 1
Hz.Many modern rigs can tune in precise1-Hz steps by pressing the
MIC up/down buttons, so Coherent makes the signals needed to do
this automatically. Once the signal has been tuned in initially,
the operator can sit back,put his feet up, and leave the driving
to the computer. Even if his receiver drifts in frequency (or if
the transmitter drifts) it's no problem because the software will
retune the radio as necessary to maintain the CW tone at800
Hz.Coherent also has a frame-phasing tracking loop. After each
100-msframe is processed, the program looks at whether the SNR
would have been better had the window ended 1 cycle(1.25 ms)
earlier or 1 cycle later. If there is consistent evidence that
going to a slightly earlier window would improve the SNR, then
the program does this automatically. What this means is that once
synchronization has been achieved, the operator can let the
program track the incoming signal and adjust the phasing as
necessary to maximize the SNR advantage the mode is capable of.
With this system, special frequency standards and rig
stabilization are no longer needed. The only equipment you need
to operate CCW is a reasonably stable transceiver, the little
Sigma-Delta interface board,and a computer. And, of course, the
Coherent program also lets you send CCW just by typing on the
keyboard. That should go without saying.  As well, the program
has a "beacon"mode, where a prestored CCW message can be
scheduled to go out at precise time intervals based on the
computer's clock.  And Now,Something to think about... We have
seen that by synchronizing our receiver to the keying at the
remote transmitter it is possible to shrink the passband of our
receiving filter down to a rather astonishing nine hertz or so,in
the process eliminating much of the noise that would otherwise
render the signal unreadable. That is fine for coherent CW
stations, but what about ordinary CW.where the guy at the other
end is sending by hand and his carrier can turn on or off at any
arbitrary time? After all, the overwhelming majority of amateur
stations around the world don't use coherent CW.Is there anything
we can do with DSP to help dig these weak signals out of the mud?
Consider a train of RF pulses, where a carrier is switched on
for, say 100milliseconds,then switched off for the next
100milliseconds. Let's assume this signal is received by a normal
amateur side band rig with its local oscillator tuned800 Hz away
from the incoming carrier.  Looking at the audio coming out of
the speaker, what frequency components are present? Well,that
depends on your point of view!  On the one hand,if we take the
position that for a frequency component to exist it must be
present always with unvarying amplitude and phase,then we must
presume many frequencies, all adding up to make the on/off pulsed
waveform. On the other hand,if we examine the waveform on an
oscilloscope, we see 80 cycles of a pure(800 Hz) sinewave inside
each pulse with no other frequencies present during either the
pulses or the silent period. Common sense tells us there is but
one frequency(800 Hz), and that its only there some of the time.
Since there is only one frequency in the signal, it would make a
lot of sense to use an arbitrarily narrow filter (ie, 0-Hzwide)
centered on that 800-Hz tone.   Such a filter would eliminate all
the noise (QRM, QRN)except that which happened to be on exactly
the same frequency. The way we usually design highly selective
(narrow) filters is to take many samples spaced over a long
interval of time and combine them mathematically to isolate the
contribution of all the individual frequencies. Analog filters
use the same techniques, storing energy in reactive components.
The narrower the filter (the higher the "Q") the longer the
energy from any given cycle stays around inside the tank circuit.
The drawback with all these filters is that it takes along time
for them to attain their final output value after a step change
in the input signal (as is the case when a CW carrier is keyed).
The sharper a filter is in frequency, the longer it takes (in
time) for it to respond. This is why experienced CW operators
will tell you it doesn't pay to use IF filters much narrower than
about 250 Hz when trying to copy code by ear. Narrower filters
actually make it harder to copy because they obliterate (smear)
the sharp leading edges of the keyed tones which the ear needs to
recognize code patterns. But the characteristic time spreading of
such filters is not a result of some insurmountable law of
physics! It follows entirely from the particular way they were
designed: they observe a signal over a long time span to make
fine distinctions in frequency in order to realize the narrow
response. The ideal filter for copying ordinary CW would be 0-Hz
wide and have an instantaneous response time. When the key went
down at the transmitter,the output of the "sliver" filter at the
receiver would reflect the amplitude change immediately. What
approach can we take in designing such a filter? A good question
is: for any given signal, how much of it do we need and how long
do we have to observe it before we can break it down into its
component frequencies and state what the amplitude at any
specific frequency must be? Could we take just a momentary
snippet out of a waveform, analyze it extensively on a fast
computer and figure out its complete spectral content just from
that tiny portion we looked at? The answer will surprise you. The
answer is yes!  In fact, there is no minimum amount of time for
which we need to observe a waveform in order to completely
characterize it. In theory, we could sample a complex signal for
just one instant and immediately know the amplitude of an 800-Hz
sinusoid in it regardless of whatever other frequencies might be
present. The calculation gets a30 QEX whole lot more complicated
when many other frequencies are present,but it still can be done.
The most straightforward case is when we know there is only one
sinusoid, say at800 Hz, and we want to ascertain its amplitude
with just one instantaneous glance at the waveform.  Hmmm,if we
knew the phase, it would be easy. With only a single voltage
measurement taken at a known point along a sine curve (phase), we
can determine its amplitude. Not knowing the phase in advance, we
have to solve for it. Which means we need at least two (2)
independent measurements, both taken at the same instant in time.
The actual sampled voltage will do for one of them. For the other
we can use the first derivative the rate the voltage is changing
at that particular moment. This derivative can be obtained
without taking anytime: convert the voltage to a current,run it
through an inductor, and measure the instantaneous voltage across
the inductor. Here we have an implementation of our "ideal CW
filter for the simplest case where there is only one frequency
component to resolve.  When there are many frequency components
to separate, we will obviously need more information, but it is
all available in that same instant; the higher order derivatives
are mutually orthogonal, hence independent, and they're there for
the measuring.  So it is possible (at least in principle) to
design a CW receiving filter with arbitrarily narrow bandwidth
(approaching 0 Hz) and virtually instantaneous response time.
What's needed is hardware to differentiate a signal repeatedly
and a very fast computing machine to crunch the numbers. 
Obtaining the software. The following can be ordered from the
author: Coherent CCW software package,$20 Bare circuit board for
constructing Sigma-Delta interface, $24 Assembled and tested
Sigma-Delta board, ready to hook up, $95.  All prices in US
dollars, and please include $5 for airmail shipment to anywhere
on the planet. For more information on CCW, contact: CCW interest
group  Peter Lumb,G3IRM  2 Briarwood Ave  Bury St Edmunds 
Suffolk IP33 3QF  England  0029 Sommet VertSt. Adolphe d'Howard,
QCJOT 2BO Canada24-hr BBS: 514 226 7796                           
 

                   by Bill de Carle, VE21Q