Serial Data Quality Comparator Ideas

[kf4hcw]

On 12/18/2017 12:48 PM, Joe Leikhim wrote:
>
> Thoughts. If I am reinventing the wheel, please let me know!
>
> This is to be a device, so using test equipment is not the solution.

I'm fond of reinventing wheels -- it teaches us about wheels and we need 
folks that can invent them instead of just buying them... plus, there's 
a lot more to know about wheels than appears in the brochure.

This device essentially becomes a piece of test equipment.

I spent some time thinking about this kind of thing when working on the 
DMDL concept -- where the first stages of the demodulator for a DMDL 
chain would have to recover the carrier and then make predictions about 
what has happened to it.

You need to define a bit more closely what characteristics you want to 
determine about signal quality.

1. Is there enough of it that you can capture the carrier/clock with a 
PLL, and how much confidence do you have?

It sounds like you've been thinking of this part because you talk about 
measuring jitter. Jitter is one part of measuring clock recovery 
confidence. Another would be signal strength.

2. Deliberate phase shift vs unintended phase shift. Specifically, in 
most HF cases anyway, you're going to expect that a deliberate phase 
shift in the carrier will be some multiple of 90 degrees. Other vhf/uhf 
cases might include finer shifts than that... but the goal is to measure 
it precisely so:

A good way to think about this is to determine at your PLL whether 
you've seen a 90, 180, or 270 degree phase shift. If you have then your 
PLL should be able to remain locked on the "original" clock. If your PLL 
is driving an oscillator at 4x the desired frequency and you divide that 
down into quadrature then you can build the phase detection logic to use 
the best match and determine the confidence of that match. Choose the 
quad with the highest confidence by freely switching to it for your PLL 
loop, make note of the shift if it occurs, and keep track of the 
confidence value.

This gets you a couple of useful things... An SSB signal will typically 
switch phase by 180 degrees at zero crossings of the modulated signal -- 
but otherwise shouldn't jitter much. Similarly bpsk - except that the 
bpsk signal's amplitude should not change much at all between phase 
shifts -- also qpsk. You shouldn't start to see amplitude shifts very 
much until you get to qam type signals.

FM signals will tend to show a more continuous phase shift that you 
should be able to track -- but that's more specialized.

FSK signals will tend to show high confidence an one of several 
predicted frequencies but not elsewhere.

3. Amplitude consistency. If you're looking at a digital modulation then 
you should have some expected amplitudes that you can determine from the 
median maximum amplitude. Signal quality measured in this aspect will 
depend on the kind of signal you're receiving and the amount of 
deviation from the expected amplitudes.

If you're receiving qpsk, bpsk, you should expect the amplitude of the 
highest quality signal to remain almost constant at the median 
amplitude. The more it varies the lower the quality of the signal. If 
qam then there should be some discrete amplitudes and the amplitudes 
that you see should not vary much from the median discrete values seen 
and those calculated from the maximum median amplitude. The more 
variability the lower the quality of the signal.

If you're looking at an am, dsb, or ssb signal then you can decode the 
amplitude and examine it's spectra. If the spectra of the amplitude 
modulation is a reasonable match for voice traffic and tends to remain 
self consistent then the quality is good. If the spectra departs from 
this model then the quality is reduced. In particular, the more random 
noise that is present the lower the quality of the signal. Within a 
given 2-3K bandwidth there should be some fairly specific peaks and 
valleys --- if the spectra looks closer to flat that means there is less 
intelligible data present (more like random noise).

===

All of this boils down to making measurements that compare 
characteristics of the apparent signal to the characteristics predicted 
by some model. Most of these models are fairly easy to construct (at 
least logically).

You could output the scalar values of the model matches individually or 
you could combine them into a single scalar value to represent signal 
quality.

Decibels may be helpful here -- if you define a perfect match as 0db, 
and reduce the signal quality in each domain based on the percentage 
match for that model as a ratio then a signal that matches the model 
less will have a negative score associated with that characteristic.

For example, if the model predicts that the phase of the recovered clock 
will be either 0 degrees or 180 degrees then the model's match 
characteristic should be 0 at 90 degrees (as far as possible from 
matching). So your ratio will be between 1/1 and 0/1 --- If your 
recovered clock jitters by 10 degrees (say from 80 to 100) then that's 
roughly 10 % of your available margin giving you a signal quality of 
10log (9/10) or -0.46 dB. Get off a bit more, say 20% either side and 
now you can say you have only (90-20)/90 == 7/9 or -1.1 dB... and so forth.

You can get even more accurate about phase if you also model the baud 
clock... Presume that a particular phase state should persist for at 
least n amount of time and that if it changes each change must occur 
after cn time. An easier way to do this is to start a clock at the last 
detected phase change and run the clock until either resetting it at the 
expected baud clock (phase didn't change, but aligns with the baud 
clock) or it changes again. Essentially measuring jitter in the 
recovered baud clock.

The number of clock ticks that occur before the phase change divided by 
the number of clock ticks that can occur in a perfectly received 
symbol... Let's say 100ms and a 1ms clock giving you a maximum of 100 
ticks per phase state.

If the detected phase lasts for 80 tics out of 100, but remember that 
once you get to 50% you're no better than flipping a coin so you've got 
a maximum loss of 50/100... so do your calculation ends up being more 
like 30/50 and for that measurement you get about -5.74 dB... integrate 
those in a sliding average and you get some idea of how accurately you 
can detect the bits that are encoded on the apparent signal.

-- these are all just suggestions that I've played with; but hopefully 
you get the idea of how they might be implemented:

Start with a collection of PLLs (as you've suggested); adjust their 
phase detection logic to accommodate predicted phase shifts and measure 
jitter. That gives you a scalar value for the instantaneous 
"unpredictability" of carrier phase angle.

Use the synchronized clock to help detect the amplitude of the apparent 
carrier. This might even come easily from a quadrature phase detection 
system in the first step.

Having detected the amplitude compare the amplitude signal(s) to the 
expected model. BPSK and QPSK amplitudes should be very consistent so 
measuring them against a low pass (integrated) version will give you a 
scalar value of the "unpredictability" of the amplitude.

Go a step further and for digital signals where you can predict the baud 
rate measure the predictability of system to detect new signals (for QAM 
include discrete amplitude values, for AM, DSB, SSB use spectral 
characteristics).

===

Let me know if any of this was useful... and if you decide to build some 
of it, as I've not had time yet but I've been dying to put it together.

Your signal quality measurement device could be the same hardware 
(mostly) as my DMDL detector.

_M

PS: One of the other things I've toyed with regarding amplitude 
characterization is the idea of measuring the slope of any shifts in 
amplitude. For audio type signals the slopes (normalized dv/t) will have 
an upper limit associated with the highest frequency expected in the 
audio channel. In a digital modulation scheme the slopes will have a 
bipolar limit since symbol shifts tend to create either no shift or a 
high frequency shift. A generalized DMDL detector might measure this as 
a matter of course and then actual signal recovery would occur down 
stream by appropriately integrating and/or differentiating these slopes. 
In signal quality detection the problem is one simply of allowing slope 
ranges that are appropriate for the expected signal and considering 
slopes outside of that envelope to be misses (unpredicted). The ratio of 
predicted changes to unpredicted changes tells you the signal quality.

-- 
kf4hcw
Pete McNeil
lifeatwarp9.com/kf4hcw