Digitizing a signal with small and much larger voltages

[Nicholas Kinar]

Hello--

I'm working with an experimental acoustics sensor which will produce
voltages ranging between 0.5 microvolt (minimum) and 5 V (peak to peak).
What is the best way to digitize this signal? At first blush, it
appears that a 24-bit ADC would work quite well for this type of
application, since (5V)/(2^24) = 3E-7 Volts. However, would it be
better to do the following?

(1) Convert the voltage signal into a current using a transconductance
amplifier.

(2) Use a logarithmic amp to find the logarithm of the current. The
AD8304 from Analog Devices is a 160dB (100 pA to 10 mA)) logarithmic amp
used as a photo-diode detector:

http://www.analog.com/static/imported-files/data_sheets/AD8304.pdf

(3) Digitize the voltage output of the AD8304 log amp using an ADC.

BUT how do I determine the number of bits of my ADC? The logarithmic
amp will take the logarithm of the signal, but how do I choose the
number of bits of the ADC so that I can adequately measure the range of
voltages?

OR is it best to simply use a 24-bit ADC?


Nicholas

 

[David L. Jones]

Hello--

I'm working with an experimental acoustics sensor which will produce
voltages ranging between 0.5 microvolt (minimum) and 5 V (peak to
peak). What is the best way to digitize this signal? At first
blush, it appears that a 24-bit ADC would work quite well for this
type of application, since (5V)/(2^24) = 3E-7 Volts.

Take a look at Cirrus Logic's Geophysical products, designed for ultimate
performance in the acoustic arena:
http://www.cirrus.com/en/products/pro/techs/T18.html
Their 24bit Delta-Sigma converters are the ducks guts. Performance can be
had beyond the datasheet typical values too.

What's your bandwidth?

Dave.

 

[Nicholas Kinar]

Take a look at Cirrus Logic's Geophysical products

Great stuff, Dave. These ADCs seem to have excellent performance. I can
think of a number of applications for these particular ADCs which would
be extremely useful in environmental sensing applications and
environmental physics.

What's your bandwidth?

I've calculated that the maximum bandwidth for my particular sensor
would be 30 kHz. However, for time-of-arrival estimates and digital
filtering, the sample rate of the ADC must be much greater than Nyquist.
I would wonder if this ADC would fit the bill:

http://www.ti.com/lit/gpn/ads1672

Perhaps the 625 kSPS rate is too much, but I like the fact that the ADC
has an SPI bus, and that the maximum sampling rate is at least 5x to 10x
greater than what might be required.


Nicholas

 

[Nicholas Kinar]

What is the bandwidth?

I've calculated a maximum bandwidth of at least 30 kHz.

What kind of accuracy are you looking for?

At least 16-bit accuracy. But how would I deal with the range of
signals? Perhaps a log amp would be worthwhile to use?

None of the 24-bit converters deliver true 24 bit performance.

Of course, since there will always be noise. However, by oversampling
and decimation, it may be possible to obtain close to 24-bit accuracy.
This implies that I would need to choose a much larger sampling rate.


Nicholas

 

[Nicholas Kinar]

Caution: d-s ADCs often have analog bandwidths well below what you'd

expect from the rate at which they deliver data. They do a lot on
internal digital filtering.

Thanks, John. I took a look at a few 24-bit ADC datasheets and found
that the digital filter can greatly restrict the passband. This
information has made me take a more critical look at these ADCs.

 

[Fred Bartoli]

Nicholas Kinar a écrit :

Hello--

I'm working with an experimental acoustics sensor which will produce
voltages ranging between 0.5 microvolt (minimum) and 5 V (peak to peak).
What is the best way to digitize this signal? At first blush, it
appears that a 24-bit ADC would work quite well for this type of
application, since (5V)/(2^24) = 3E-7 Volts. However, would it be
better to do the following?

(1) Convert the voltage signal into a current using a transconductance
amplifier.

(2) Use a logarithmic amp to find the logarithm of the current. The
AD8304 from Analog Devices is a 160dB (100 pA to 10 mA)) logarithmic amp
used as a photo-diode detector:

http://www.analog.com/static/imported-files/data_sheets/AD8304.pdf

(3) Digitize the voltage output of the AD8304 log amp using an ADC.

BUT how do I determine the number of bits of my ADC? The logarithmic
amp will take the logarithm of the signal, but how do I choose the
number of bits of the ADC so that I can adequately measure the range of
voltages?

OR is it best to simply use a 24-bit ADC?


Nicholas

What accuracy do you aim at on the 0.5uV signal?

One thing is good to have in mind: WRT your 30kHz BW, if one guess a
30kHz brick wall filter, which I read you don't want because you want to
do some time of arrival measurement, 0.5uV rms noise is less than
3nV/rtHz input referred noise. Doable but you begin to need to be careful.
Now if you want some meaningful measurement and you want 0.15uV rms
noise, so that your 0.5uV signal is detectable, that is less than
0.9nV/rtHz. Still doable but you have to be increasingly careful.

If now you throw in your increased BW requirement, or you want more
accuracy, then, well, hem...
Unless your signal is repetitive and you have a trigger to sync on, in
which case averaging is possible.

WRT your dynamic range, can't you have some PGA front end? That'll ease
your unrealistic ADC requirement.

One possibility, if you can't switch gains, is to have several gain
stages working in parallel, digitize all of them at the same time and
select the right one in digital domain, with "a bit" of calibration and
post processing.

I've seen much more acrobatic processing done with good results.

 

[Nicholas Kinar]

What you want, is a variable gain amplifier that takes a

gain control voltage so that voltage (or power) gain is
proportioinal to exp(Vprogram), and feed AC into the amplifier
and take AC at the correct range for digitization from the output.

Like a floating point number, the gain control voltage is the exponent
part,
and the ADC reading is the mantissa part.

Thanks, wit3rd! That's a great idea, and would work quite well to be
able to vary the gain based on the amplitude of the signals which are
measured.

Perhaps some sort of DSP could be done to predict when the time-varying
signal is increasing in amplitude, and the gain could be dynamically
adjusted, thereby eliminating clipping.

 

[Nicholas Kinar]

Thanks Fred!

What accuracy do you aim at on the 0.5uV signal?

Measuring the presence of a 0.5uV signal. Any better accuracy would be
great.

One thing is good to have in mind: WRT your 30kHz BW, if one guess a
30kHz brick wall filter, which I read you don't want because you want to
do some time of arrival measurement, 0.5uV rms noise is less than
3nV/rtHz input referred noise. Doable but you begin to need to be careful.
Now if you want some meaningful measurement and you want 0.15uV rms
noise, so that your 0.5uV signal is detectable, that is less than
0.9nV/rtHz. Still doable but you have to be increasingly careful.

If now you throw in your increased BW requirement, or you want more
accuracy, then, well, hem...
Unless your signal is repetitive and you have a trigger to sync on, in
which case averaging is possible.

WRT your dynamic range, can't you have some PGA front end? That'll ease
your unrealistic ADC requirement.

Sure, perhaps what could be done is to increase or decrease the gain as
the signal changes. But I suspect that there will also be components in
the signal which will be measurable at much larger voltages, and I would
also like to measure the smaller voltages at the same time.

One possibility, if you can't switch gains, is to have several gain
stages working in parallel, digitize all of them at the same time and
select the right one in digital domain, with "a bit" of calibration and
post processing.

I've seen much more acrobatic processing done with good results.

Very neat idea, Fred. I would wonder if it would be possible to then
combine all of these gain stages into one signal, rather than selecting
the right one.

 

[Nicholas Kinar]

Now if you want some meaningful measurement and you want 0.15uV rms
noise, so that your 0.5uV signal is detectable, that is less than
0.9nV/rtHz. Still doable but you have to be increasingly careful.

Would you know of an example schematic or circuit that I could look at?

 

[Nicholas Kinar]

WRT your dynamic range, can't you have some PGA front end? That'll ease
your unrealistic ADC requirement.

The signal consists of a number of frequency components. Some frequency
components have larger voltage amplitudes (5V p-p), whereas other
frequency components have smaller voltage amplitudes in the microvolt
range.

 

[Nicholas Kinar]

The signal consists of a number of frequency components. Some
frequency components have larger voltage amplitudes (5V p-p), whereas
other frequency components have smaller voltage amplitudes in the
microvolt range

But I would also suspect that the frequency components can take on
larger and smaller voltage amplitudes, so using separate analog
band-pass filters for specific frequency components may not work.

 

[Nicholas Kinar]

Here's another post placed elsewhere which deals with saturation at
higher amplitude levels:

http://www.seismicnet.com/psnlist/090207_145244_1.html

Here's an excerpt from the post located at the above address:

"Two reasons come to mind. You will see more quakes with a more
sensitive instrument. You can not really compensate with more
amplification because the dynamic range is smaller with the low bit
count devices. Yes, you can raise the amplification level so that an 8
bit device will respond to the same voltage signal that the 24 bit
device will see, but the 8 bit device will be saturated after only 256
counts while the 24 bit device would have only recorded 256 counts out
of 16.7 million possible counts."


I think that this might explain that I need to use a greater number of
bits to be able to detect small and large signals.


What is the best way to do this, given the bandwidth and sampling rate?

 

[Nicholas Kinar]

Here's another post placed elsewhere which deals with saturation at
higher amplitude levels:

http://www.seismicnet.com/psnlist/090207_145244_1.html

Here's an excerpt from the post located at the above address:

"Two reasons come to mind. You will see more quakes with a more
sensitive instrument. You can not really compensate with more
amplification because the dynamic range is smaller with the low bit
count devices. Yes, you can raise the amplification level so that an 8
bit device will respond to the same voltage signal that the 24 bit
device will see, but the 8 bit device will be saturated after only 256
counts while the 24 bit device would have only recorded 256 counts out
of 16.7 million possible counts."


I think that this might explain that I need to use a greater number of
bits to be able to detect small and large signals.


What is the best way to do this, given the bandwidth and sampling rate?

Here's maybe some other aspects of the required circuit:

(1) Bandwidth = 30 kHz (maximum)
(2) Sampling rate: appropriate sampling rate required to resolve arrival
time down to 1.6 microseconds
(3) Amplitudes: 5V (p-p) to 0.5 microvolt

I would wonder if there's a good ADC or other type of circuit to be able
to do these types of processing.

 

[Nicholas Kinar]

There's an ADC converter board offered by this company:

http://www.kenda.co.uk/pr/ke1162.htm

Here's what the system can do:

"The new PS2400 analogue-to-digital converter (ADC) board from
Earth Data Ltd, part of the Southampton-based Kenda Group,
provides three independent channels of 24-bit conversion and
offers true 24-bit dynamic range. It can resolve signals from
1µV to ±8V in its normal gain mode or 50nV to ±400mV in its high
gain mode."

But how to attain true 24-bit accuracy? Are there some ADCs capable of
attaining true 24-bit accuracy?

 

[Nicholas Kinar]

A DLVA (detector/log video amplifier) might work, if your accuracy
requirements are modest (say 0.5 dB).

Cheers

Phil Hobbs

Thanks, Phil. I'm going to have to take a look again at my accuracy
requirements.

 

[David L. Jones]

Great stuff, Dave. These ADCs seem to have excellent performance. I
can think of a number of applications for these particular ADCs which
would be extremely useful in environmental sensing applications and
environmental physics.


I've calculated that the maximum bandwidth for my particular sensor
would be 30 kHz. However, for time-of-arrival estimates and digital
filtering, the sample rate of the ADC must be much greater than
Nyquist. I would wonder if this ADC would fit the bill:

30KHz is very high, tough to design high performance with 24bit d-s, but if
you don't need the full performance range then much easier and possible.
Performance of d-s converters drops drastically as the bandwidth increases.
Analog Devices do a couple of suitable bandwidth 24bit converters I believe.
They also do competing seismic chipsets to Cirrus Logic, but they ain't as
good (last time I evalulated them).
Seismic specific market parts are optimised for 2KHz and under, and more
importantly for extremely low current consumption, due to the need to have
several thousand of them stringed together in serial cable systems.

http://www.ti.com/lit/gpn/ads1672

Perhaps the 625 kSPS rate is too much, but I like the fact that the
ADC has an SPI bus, and that the maximum sampling rate is at least 5x
to 10x greater than what might be required.

Looks fairly suitable, although you have to go into a lot of detail with
these things. Front end amp design is also another rather exotic area you
have to get right.

Dave.

 

[Nicholas Kinar]

Looks fairly suitable, although you have to go into a lot of detail with

these things. Front end amp design is also another rather exotic area you
have to get right.

Thank you once again for your reply, Dave. BTW, I'm also thinking about
using another similar ADC, but with a lower sampling rate. The AD7766
from Analog Devices is a 24-bit ADC with a 109.5 dB dynamic range at 128
kSPS, which may also be okay for my application:

http://www.analog.com/static/imported-files/data_sheets/AD7766.pdf


IMHO, apparently the best way to deal with front end amp design is by
studying the evaluation kit that the company puts out. In general, if
the company doesn't have an evaluation kit for the ADC (most do), then
it probably isn't a good idea to use the part.

 

[Martin Brown]

Here's another post placed elsewhere which deals with saturation at
higher amplitude levels:

http://www.seismicnet.com/psnlist/090207_145244_1.html

Here's an excerpt from the post located at the above address:

"Two reasons come to mind. You will see more quakes with a more
sensitive instrument. You can not really compensate with more
amplification because the dynamic range is smaller with the low bit
count devices. Yes, you can raise the amplification level so that an 8
bit device will respond to the same voltage signal that the 24 bit
device will see, but the 8 bit device will be saturated after only 256
counts while the 24 bit device would have only recorded 256 counts out
of 16.7 million possible counts."


I think that this might explain that I need to use a greater number of
bits to be able to detect small and large signals.


What is the best way to do this, given the bandwidth and sampling rate?

Probably the same way autoranging voltmeters do it. Sit with the system
at high gain when nothing much is going on so you can digitise the noise
floor with good resolution and decrease the gain by a factor of ten each
time it (nearly) clips and vice versa when the signal drops below 1/10
of fullscale. You will lose a few samples in the switchover.

You probably do not want any funny log or A-law amplifiers in the signal
path if there will be a spectrum of frequencies present with some high
level fundamentals and low level components also of interest. If this is
the raw data for some scientific experiment ask the experimenters what
sort of signal they want and what if any conditioning is acceptable.

Linear time domain data is a lot easier to analyse for faint frequency
components.

Regards,
Martin Brown

 

[Fred Bartoli]

Nicholas Kinar a écrit :

The signal consists of a number of frequency components. Some frequency
components have larger voltage amplitudes (5V p-p), whereas other
frequency components have smaller voltage amplitudes in the microvolt
range.

Well, we still don't know what your signal is and what you want to do
with it.
Do you want to process it in the time domain or frequency domain?
That's not the same thing. For one, in frequency domain, the total noise
is spread over the whole bandwidth, and detecting/measuring 0.5uV
spectral lines is much more easier, since it'll be way over the noise floor.

Now, one other 'small' thing is you say you have at the same time, 0.5uV
and 5V signals. That means you need better than -140dB THD for the full
processing chain. At a full 30kHz BW? Hmmm,... still doable, but this
requires being increasingly cautious, and if you're to ask all those
questions, you're in some trouble there.

I think it's time for you to think more carefully about your signals,
your needs and what you want to do with this.
Engineering has to do with the limits of physics and just piling up
requirements generally won't give you any sensible answer.

 

[Nicholas Kinar]

Probably the same way autoranging voltmeters do it. Sit with the system
at high gain when nothing much is going on so you can digitise the noise
floor with good resolution and decrease the gain by a factor of ten each
time it (nearly) clips and vice versa when the signal drops below 1/10
of fullscale. You will lose a few samples in the switchover.

You probably do not want any funny log or A-law amplifiers in the signal
path if there will be a spectrum of frequencies present with some high
level fundamentals and low level components also of interest. If this is
the raw data for some scientific experiment ask the experimenters what
sort of signal they want and what if any conditioning is acceptable.

Linear time domain data is a lot easier to analyse for faint frequency
components.

Thahks, Martin. I'm going to have to take a look at what is required.
The autoranging voltmeter idea may be very useful.

Nicholas