On Wed, 05 Oct 2011 04:03:48 -0500, glias wrote:

> Hello all,
> I finally found an article from TI to help me to calculate the noise
> factor of my analog chain.
> 
> http://www.ti.com/lit/an/slyt094/slyt094.pdf
> 
> For sum up, to reduce the noise in my bandwidth, I need to optimize the
> noise factor of my analog chain. The oversampling process with averaging
> permit to reduce the white noise in the band of interest which come from
> the analog chain AND quantization noise from the ADC.
>  
> For my final board, I have 8 voices to place on the PCB, and it will be
> difficult to keep my actual design (need more space). So I thought that
> I could remove the Bessel filter and make the filtering with the FPGA.
> Do you think it is a good idea ?

It's certainly an approach that I'd want to investigate.  Whether I'd 
take it in the end depends on what I found, but that's usually how things 
fall out.

> For a low pass FIR filter, how I can do to estimate the number of cells
> that the filter will take in the FPGA ? (The FPGA used would probably be
> an
>  ALTERA Arria II GZ.

Consider alternatives to a whomping big FIR filter in the FPGA.  If 
you're oversampling, you may want to start by implementing a CIC filter 
(which can be implemented quite efficiently) and knocking down the 
sampling rate, then following that with a FIR or an IIR filter.

-- 
www.wescottdesign.com

On Oct 5, 2:03&#4294967295;am, "glias" <glias37@n_o_s_p_a_m.hotmail.com> wrote:
> Hello all,
> I finally found an article from TI to help me to calculate the noise factor
> of my analog chain.
>
> http://www.ti.com/lit/an/slyt094/slyt094.pdf
>
> For sum up, to reduce the noise in my bandwidth, I need to optimize the
> noise factor of my analog chain. The oversampling process with averaging
> permit to reduce the white noise in the band of interest which come from
> the analog chain AND quantization noise from the ADC.
>
> For my final board, I have 8 voices to place on the PCB, and it will be
> difficult to keep my actual design (need more space). So I thought that I
> could remove the Bessel filter and make the filtering with the FPGA.
> Do you think it is a good idea ?
> For a low pass FIR filter, how I can do to estimate the number of cells
> that the filter will take in the FPGA ? (The FPGA used would probably be an
> &#4294967295;ALTERA Arria II GZ.
>
> PS :I have some difficulties to read your message : I mean that some
> special characters doesn't seem to pass on the forum, is it normal ?
>
> regards



Re filtering:

Based on what I think I know re your system, I would suggest that you
keep your DC block and a functional Nyquist filter in front of the A/
D.  Implementation could be simple R-C filters, just make sure you
meet your attenuation spec at Nyquist or more importantly the
frequency at which stuff starts to fold back into your analysis
window.  This is especially important when making tradeoff&#4294967295;s for
space.

Your final analysis filter certainly could be done in the FPGA.  I
would suggest that you visit the Altera or Xilinx sites to answer your
questions re cell count.

My other suggestion is to pay attention to your layout --- especially
if you&#4294967295;re placing your high gain amp in a digital environment.

Re the weird stuff:  Could be my text editor. I don&#4294967295;t see what you&#4294967295;re
talking about.

Regards

On Oct 5, 5:03&#4294967295;am, "glias" <glias37@n_o_s_p_a_m.hotmail.com> wrote:
> Hello all,
> I finally found an article from TI to help me to calculate the noise factor
> of my analog chain.
>
> http://www.ti.com/lit/an/slyt094/slyt094.pdf
>
> For sum up, to reduce the noise in my bandwidth, I need to optimize the
> noise factor of my analog chain. The oversampling process with averaging
> permit to reduce the white noise in the band of interest which come from
> the analog chain AND quantization noise from the ADC.
>
> For my final board, I have 8 voices to place on the PCB, and it will be
> difficult to keep my actual design (need more space). So I thought that I
> could remove the Bessel filter and make the filtering with the FPGA.
> Do you think it is a good idea ?
> For a low pass FIR filter, how I can do to estimate the number of cells
> that the filter will take in the FPGA ? (The FPGA used would probably be an
> &#4294967295;ALTERA Arria II GZ.
>
> PS :I have some difficulties to read your message : I mean that some
> special characters doesn't seem to pass on the forum, is it normal ?
>
> regards

be aware, most practical noise issues do not originate from the noise
sources listed in that paper, noise from nearby switching power
supplies, external EMI, ground loops, clocks etc all require different
solutions, mostly guidelines are used to minimized these problems
during design (guardbands, EMI doghouses, seperate power planes etc) ,
but they, for the most part, cannot be quantified or simulated or
predicted (interactions are too complex).

Hello all,
I finally found an article from TI to help me to calculate the noise factor
of my analog chain.

http://www.ti.com/lit/an/slyt094/slyt094.pdf

For sum up, to reduce the noise in my bandwidth, I need to optimize the
noise factor of my analog chain. The oversampling process with averaging
permit to reduce the white noise in the band of interest which come from
the analog chain AND quantization noise from the ADC.
 
For my final board, I have 8 voices to place on the PCB, and it will be
difficult to keep my actual design (need more space). So I thought that I
could remove the Bessel filter and make the filtering with the FPGA.
Do you think it is a good idea ?
For a low pass FIR filter, how I can do to estimate the number of cells
that the filter will take in the FPGA ? (The FPGA used would probably be an
 ALTERA Arria II GZ.

PS :I have some difficulties to read your message : I mean that some
special characters doesn't seem to pass on the forum, is it normal ?

regards

On 10/3/2011 5:26 PM, Tim Wescott wrote:
> On Mon, 03 Oct 2011 13:40:59 -0700, steve wrote:
>
>> On Oct 3, 1:06 pm, Tim Wescott<t...@seemywebsite.com>  wrote:
>>
>>> The oversample-and-average technique actually doesn't reduce
>>> quantization noise by itself: if you had an infinitely quiet ADC front
>>> end then the quantization noise would come through unscathed.
>>
>> that's only true for DC inputs, which the OP states is not under
>> consideration
>>
>> quantization noise is most often modeled as just broad band noise
>> extending way past fs/2, with no oversampling most of the noise power is
>> folded back into the analysis band (1.5k to 350k ), oversampling just
>> increases the folding frequency, thus reducing the amount of high
>> frequency aliased noise (from any source) folded back into the analysis
>> band
>
> The OP states a bandwidth, but doesn't say where his signals usually lie.
>
> If the nature of the input is unknown, or if you're operating in closed
> loop, the conservative way to estimate quantization noise is as a square
> wave concentrated at the worst possible frequency.  This is more valid
> for closed-loop operation -- because the system will often oscillate at
> or very close to that frequency -- but it also covers your bases for a
> general system that might see a small tone at any frequency.
>
> "That could be a problem but it's very likely that it won't happen" is,
> in my book with "This will pass all qualification tests and then crop up
> as a problem when the World's Pickiest Customer is using the system".

It is appropriate again to explain Murphy's Law. "If anything can go 
wrong, it will" is not the creed of a pessimist, but a warning from an 
experienced engineer. He could have said at greater length, "If the 
system allows something bad to happen, rest assured that that bad thing 
will eventually happen.

Jerry
-- 
Engineering is the art of making what you want from things you can get.

On Oct 3, 2:26=A0pm, Tim Wescott <t...@seemywebsite.com> wrote:
> On Mon, 03 Oct 2011 13:40:59 -0700, steve wrote:
> > On Oct 3, 1:06=A0pm, Tim Wescott <t...@seemywebsite.com> wrote:
>
> >> The oversample-and-average technique actually doesn't reduce
> >> quantization noise by itself: if you had an infinitely quiet ADC front
> >> end then the quantization noise would come through unscathed.
>
> > that's only true for DC inputs, which the OP states is not under
> > consideration
>
> > quantization noise is most often modeled as just broad band noise
> > extending way past fs/2, with no oversampling most of the noise power i=
s
> > folded back into the analysis band (1.5k to 350k ), oversampling just
> > increases the folding frequency, thus reducing the amount of high
> > frequency aliased noise (from any source) folded back into the analysis
> > band
>
> The OP states a bandwidth, but doesn't say where his signals usually lie.
>
> If the nature of the input is unknown, or if you're operating in closed
> loop, the conservative way to estimate quantization noise is as a square
> wave concentrated at the worst possible frequency. =A0This is more valid
> for closed-loop operation -- because the system will often oscillate at
> or very close to that frequency -- but it also covers your bases for a
> general system that might see a small tone at any frequency.
>
> "That could be a problem but it's very likely that it won't happen" is,
> in my book with "This will pass all qualification tests and then crop up
> as a problem when the World's Pickiest Customer is using the system".
>
> --www.wescottdesign.com- Hide quoted text -
>
> - Show quoted text -



I thought it was pretty clear that the OP had a non-DC band limited
signal well below Nyquist.  He=92s asking for approaches and techniques
to approach designs like this.

We can probably agree that for given N, a flash converter with a
higher sample rate will have a lower noise density.

The OP wants to use this information to develop his system.  He now
knows that he can use this info to select his A/D for both sample rate
and N bits to optimize his noise floor.  He may not yet know exactly
how, but he knows it can be done.

What we really haven=92t talked about, although Tim may have been making
an attempt, is what will happen when he gets a large signal into his
acquisition.  Or he gets a small single tone signal into his
acquisition and it creates a bigger spur than the large signal =85 left
to the interested student.

Regards

On Mon, 03 Oct 2011 13:40:59 -0700, steve wrote:

> On Oct 3, 1:06&nbsp;pm, Tim Wescott <t...@seemywebsite.com> wrote:
> 
>> The oversample-and-average technique actually doesn't reduce
>> quantization noise by itself: if you had an infinitely quiet ADC front
>> end then the quantization noise would come through unscathed.
> 
> that's only true for DC inputs, which the OP states is not under
> consideration
> 
> quantization noise is most often modeled as just broad band noise
> extending way past fs/2, with no oversampling most of the noise power is
> folded back into the analysis band (1.5k to 350k ), oversampling just
> increases the folding frequency, thus reducing the amount of high
> frequency aliased noise (from any source) folded back into the analysis
> band

The OP states a bandwidth, but doesn't say where his signals usually lie.

If the nature of the input is unknown, or if you're operating in closed 
loop, the conservative way to estimate quantization noise is as a square 
wave concentrated at the worst possible frequency.  This is more valid 
for closed-loop operation -- because the system will often oscillate at 
or very close to that frequency -- but it also covers your bases for a 
general system that might see a small tone at any frequency.

"That could be a problem but it's very likely that it won't happen" is, 
in my book with "This will pass all qualification tests and then crop up 
as a problem when the World's Pickiest Customer is using the system".

-- 
www.wescottdesign.com

On Oct 3, 1:06=A0pm, Tim Wescott <t...@seemywebsite.com> wrote:

> The oversample-and-average technique actually doesn't reduce quantization
> noise by itself: if you had an infinitely quiet ADC front end then the
> quantization noise would come through unscathed. =A0

that's only true for DC inputs, which the OP states is not under
consideration

quantization noise is most often modeled as just broad band noise
extending way past fs/2, with no oversampling most of the noise power
is folded back into the analysis band (1.5k to 350k ), oversampling
just increases the folding frequency, thus reducing the amount of high
frequency aliased noise (from any source) folded back into the
analysis band

On Oct 3, 5:28=A0am, "glias" <glias37@n_o_s_p_a_m.hotmail.com> wrote:
> >On Sep 29, 10:41=3DA0am, Tim Wescott <t...@seemywebsite.com> wrote:
> >> On Thu, 29 Sep 2011 11:29:38 -0500, glias wrote:
> >> >>On Sep 29, 8:03=3D3DA0am, Jerry Avins <j...@ieee.org> wrote:
> >> >>> On 9/29/2011 9:31 AM, me0...@yahoo.com wrote:
>
> >> >>> > On Sep 29, 3:47 am, "glias"<glias37@n_o_s_p_a_m.hotmail.com>
> >> > =3D3DA0wrote:
> >> >>> >> Hello all,
> >> >>> >> I have IR preamp with a gain of 64dB, the output noise is about
> >> >>> >> 50mV
> >> > p=3D3D
> >> >>eak
> >> >>> >> to peak for a signal of 4V peak peak (max) and the bandwidth is
> >> > 350kHz=3D3D
> >> >>.
> >> >>> >> I have a AD7626 (16 bits 10MSPS ADC) which is largly suffisient
> fo=3D
> >r
> >> > my
> >> >>> >> aplpication.
> >> >>> >> I would want to know what would be the best method to filter th=
e
> >> > noise
> >> >>> >> (which is white noise and 1/f noise which come from the IR
> >> >>> >> detector)
> >> > i=3D3D
> >> >>n my
> >> >>> >> band of interest ...? is it possible ? Does the oversampling
> could
> >> > hel=3D3D
> >> >>p me
> >> >>> >> to reduce it ? or does it only improves the quantification nois=
e
> o=3D
> >f
> >> > th=3D3D
> >> >>e ADC
> >> >>> >> ?
> >> >>> >> Does it exist another (digital) way to help me to reduce the
> noise
> >> >>> >> ?
> >> > D=3D3D
> >> >>oes
> >> >>> >> the averaging could help me ?
>
> >> >>> >> I hope that you could help me.
> >> >>> >> Regards
>
> >> >>> > The way to approach this problem is to solve for system NF of th=
e
> >> >>> > system.
>
> >> >>> > Once you get that number you can solve for system dynamic range
> tha=3D
> >t
> >> >>> > can be used to calculate the noise floor.
>
> >> >>> > No =3D3D3D (KTBGF)
>
> >> >>> > F =3D3D3D ((s/n)i)/((s/n)o)
>
> >> >>> > The dynamic range of the A/D can be calculated using:
>
> >> >>> > 6db*N + 10*log(Fs) 10*log(Bwa) 1.25db
>
> >> >>> > Where: KT is boltzman constant, B is analog bandwidth, G is amp
> >> >>> > gain, F is noise figure, N is A/D bits, Fs is sample rate and Bw=
a
> i=3D
> >s
> >> >>> > the final analysis bandwidth.
>
> >> >>> > So calculate the NF of the amp, NF of the A/D and everything
> falls
> >> >>> > out.
>
> >> >>> > Regards,
> >> >>> > Hope this helps
>
> >> >>> Hardly! He asked how to filter. (Not that the question has an
> answer
> >> >>> with the information given.)
>
> >> >>> Jerry
> >> >>> --
> >> >>> Engineering is the art of making what you want from things you can
> >> >>> get.-
> >> > =3D3D
> >> >>Hide quoted text -
>
> >> >>> - Show quoted text -
>
> >> >>I have a hard time telling if this is a homework problem or not.
>
> >> >>I was attempting to stimulate some thought re dynamic range and how
> he
> >> >>might use the concept of noise to solve his problem.
>
> >> >>It looks to me like he=3D3D92s taking a measurement and reading the =
fuzz
> =3D
> >on
> >> >>a scope trace. =3DA0But I don=3D3D92t know.
>
> >> >>An interested student would ask where those equations come from and
> how
> >> >>he could use them to solve his problem.
>
> >> >>The concept of how to design a gain lineup, whether that be a op-amp
> an=3D
> >d
> >> >>A/D or a downconvert with a A/D behind and solve for noise is all th=
e
> >> >>same.
>
> >> >>Simple concepts such as where to place that narrowband filter and wh=
y
> >> >>elude many.
>
> >> >>Never did I say that he supplied enough information =3D3D85 but he d=
oes
> h=3D
> >ave
> >> >>access to the solution. =3DA0He just needs to know what to solve for=
.
>
> >> > Hello all,
> >> > Thanks a lot for your replies.
> >> > First, the application is not a homework, it's a real application an=
d
> a
> >> > real problem (for me).
> >> > The bandwidth of the analog front end (bias and pre amp of the
> detector
> >> > is have a high pass filter (a simple first order ac coupling) with
> 1,5H=3D
> >z
> >> > for the high pass filter and 350kHz 3rd order Bessel type low pass
> >> > filter. I'm sorry for the measurement I don't have access to spectru=
m
> >> > analyzer so
>
> >> So the bandwidth of your _filter_ is 350kHz, fin. =3DA0But what is the
> usef=3D
> >ul
> >> bandwidth of your _signal_?
>
> >> > Yes, the measurement of the "noise" is just a reading from a scope.
> But
> >> > it gives me an idea.
> >> > I'm not a specialist and just wanted to know what could be done to
> >> > reduce my noise (if it is possible). The oversampling technique
> permit
> >> > to reduce the noise from the quantification noise from the ADC.
> (correc=3D
> >t
> >> > me if I'm wrong !).
>
> >> More or less correct, yes. =3DA0_If_ the ADC has enough intrinsic nois=
e
> tha=3D
> >t
> >> it shows up in the ADC output, then yes, oversampling and averaging ca=
n
> >> help you to overcome quantization noise.
>
> >> > But it doesn't reduce the noise which come from my
> >> > analog signal chain, isn't it ?
>
> >> Correct. =3DA0It can't do a thing about that.
>
> >> > I guess that the only solution to my problem is to improve the analo=
g
> >> > chain . ?
>
> >> Unless your analog low-pass filter is wider than necessary, or your
> noise
> >> is being introduced after it, yes.
>
> >> > So using the digital filter just permit to reduce the analog filter
> >> > (anti aliasing) ?
>
> >> Using a filter in digital-land offers a multitude of advantages, about
> >> the only thing it _doesn't_ do is undo the effects of aliasing -- that
> >> would be like a blender that can un-scramble eggs.
>
> >> --www.wescottdesign.com-Hide quoted text -
>
> >> - Show quoted text -
>
> >Which topology are you using:
> >Sensor--->Filter--->op-amp--->A/D
> >or
> >Sensor---> Op-Amp--->Filter--->A/D
>
> >For any amp string the noise performance of the system, i.e., the
> >noise floor will be set by the first amp. =A0It can never be better than
> >that. =A0Another way to say it is, if you create noise in your amp
> >string, you can=3D92t see a signal from your input source better than
> >that.
>
> >Any lossy components, (i.e. resistors, non-ideal filters), in front of
> >an amplifier adds to the noise.
> >You will never be able to see below the noise floor set by your system
> >NF.
>
> >What process gain and averaging allow you to do, is get down to the
> >noise floor of your gain chain. =A0So you want your narrowest filter to
> >be the digital filter. =A0Noise Out of your gain stage =3D3D KTBGF.
> >Consider your A/D as just another gain stage with a NF.
>
> >There is hope that things could get better than what you see on your
> >scope. =A0Especially, if your topology is the first one shown above.
> >I suspect you are making the noise measurement with a scope bandwidth
> >wider than your filter bandwidth. =A0Correct? =A0If true, that adds to t=
he
> >fuzz on the trace because of the increased bandwidth. =A0Also, are you
> >making the measurement with the sensor connected?
>
> >Let your digital filter set your acquisition bandwidth. =A0If you aren=
=3D92t
> >already, use the second topology above.
> >Terminate the input to your amp/filter with the appropriate impedance
> >and make the measurement with your A/D and digital filter to determine
> >your true system noise floor. =A0If it goes up when you add your sensor,
> >well you know.
>
> >It should be with a few dB of your calculation or you=3D92ve got other
> >problems. =A0You could also sweep your system using a sine source at the
> >input to verify measurement bandwidth. =A0Just suck the data into MatLab
> >if you don=3D92t have a canned FFT. =A0Painful, but it works.
>
> >All of the above applies to thermal noise and not the 1/f stuff.
>
> >As Tim said, the digital filter will not fix any aliased stuff =3D85 you
> >have to stop that before it gets to the converter. =A0Your analog filter
> >should do that. =A0However, beware, depending on the self-resonance of
> >your filter components, you could very well have re-entrant modes at
> >Nyquest. =A0So use your scope and a signal generator to sweep the A/D
> >pre-filter.
>
> >Hope this helps.
>
> Hello,
> thanks a lot for your help and sorry for my late answer.
>
> The bandwidth of the signal is 50Hz to 314kHz.
> For the topology, I have :
> Sensor---> Op-Amp--->Filter--->A/D
>
> I though that the "oversampling" method permit to reduce white noise
> provided by op amp, detector resistor and ADC quantification noise... but
> if I understand well, It only permits to reduce quantification noise.
> I would want to know how can I calculate the noise factor of my analog
> chain. Since the noise factor more used in RF components, there is no val=
ue
> of it in datasheet in low noise op amp like the ADA4898 (this is this op
> amp that I use for all my stages : pre-amp and filter stages since that I
> have low impedance sensor =3D> I have to use a low noise voltage op amp).
> Could you please help me how I can do to calculate the NF with this op am=
p
> ?
>
> regards- Hide quoted text -
>
> - Show quoted text -


The OP-Amp guys spec noise in terms of input referred voltage and
current of their op-amp, the active device

In order to calculate NF you need those numbers, your circuit topology
and the values of those components.

The analysis goes like this:

Use the input referred voltage and current noise along with the total
input referred noise from your circuit topology to calculate the input
spectral density.

Apply the noise to the definition of noise factor, F, and solve for NF
which is the decibel equivalent.

You can now use that NF number in calculating your system noise, noise
floor and hence your dynamic range as function of frequency.

These concepts maybe seem strange, but once you get it, they are very
useful.  Build yourself a lib of circuit topologies and just plug in
component values.

Years ago TI put out some papers on this subject.  They did a pretty
good job.  Search their site.

Also a couple of old text books, that maybe helpful, I don=92t know if
they=92re still in print:

One by a guy named Mervin Frerking, (I think I spelled his name
correctly), called Digitial Signal Processing in Radio Comm System, or
something like that.  Covers RF gain line up including the A/D.  His
terminology may seem strange but the concepts are there.

Also a pure RF book, Introduction to Radio Frequency Design by Wes
Hayward.  Great book for getting the basics down re elements in a gain
stage that effect dynamic range.  No A/D stuff.

And of course I suspect there are several newer books that cover this
stuff, others may be of some help.

Regards

On Mon, 03 Oct 2011 07:28:32 -0500, glias wrote:

>>On Sep 29, 10:41=A0am, Tim Wescott <t...@seemywebsite.com> wrote:
>>> On Thu, 29 Sep 2011 11:29:38 -0500, glias wrote:
>>> >>On Sep 29, 8:03=3DA0am, Jerry Avins <j...@ieee.org> wrote:
>>> >>> On 9/29/2011 9:31 AM, me0...@yahoo.com wrote:
>>>
>>> >>> > On Sep 29, 3:47 am, "glias"<glias37@n_o_s_p_a_m.hotmail.com>
>>> > =3DA0wrote:
>>> >>> >> Hello all,
>>> >>> >> I have IR preamp with a gain of 64dB, the output noise is about
>>> >>> >> 50mV
>>> > p=3D
>>> >>eak
>>> >>> >> to peak for a signal of 4V peak peak (max) and the bandwidth is
>>> > 350kHz=3D
>>> >>.
>>> >>> >> I have a AD7626 (16 bits 10MSPS ADC) which is largly suffisient
> fo=
>>r
>>> > my
>>> >>> >> aplpication.
>>> >>> >> I would want to know what would be the best method to filter
>>> >>> >> the
>>> > noise
>>> >>> >> (which is white noise and 1/f noise which come from the IR
>>> >>> >> detector)
>>> > i=3D
>>> >>n my
>>> >>> >> band of interest ...? is it possible ? Does the oversampling
> could
>>> > hel=3D
>>> >>p me
>>> >>> >> to reduce it ? or does it only improves the quantification
>>> >>> >> noise
> o=
>>f
>>> > th=3D
>>> >>e ADC
>>> >>> >> ?
>>> >>> >> Does it exist another (digital) way to help me to reduce the
> noise
>>> >>> >> ?
>>> > D=3D
>>> >>oes
>>> >>> >> the averaging could help me ?
>>>
>>> >>> >> I hope that you could help me.
>>> >>> >> Regards
>>>
>>> >>> > The way to approach this problem is to solve for system NF of
>>> >>> > the system.
>>>
>>> >>> > Once you get that number you can solve for system dynamic range
> tha=
>>t
>>> >>> > can be used to calculate the noise floor.
>>>
>>> >>> > No =3D3D (KTBGF)
>>>
>>> >>> > F =3D3D ((s/n)i)/((s/n)o)
>>>
>>> >>> > The dynamic range of the A/D can be calculated using:
>>>
>>> >>> > 6db*N + 10*log(Fs) 10*log(Bwa) 1.25db
>>>
>>> >>> > Where: KT is boltzman constant, B is analog bandwidth, G is amp
>>> >>> > gain, F is noise figure, N is A/D bits, Fs is sample rate and
>>> >>> > Bwa
> i=
>>s
>>> >>> > the final analysis bandwidth.
>>>
>>> >>> > So calculate the NF of the amp, NF of the A/D and everything
> falls
>>> >>> > out.
>>>
>>> >>> > Regards,
>>> >>> > Hope this helps
>>>
>>> >>> Hardly! He asked how to filter. (Not that the question has an
> answer
>>> >>> with the information given.)
>>>
>>> >>> Jerry
>>> >>> --
>>> >>> Engineering is the art of making what you want from things you can
>>> >>> get.-
>>> > =3D
>>> >>Hide quoted text -
>>>
>>> >>> - Show quoted text -
>>>
>>> >>I have a hard time telling if this is a homework problem or not.
>>>
>>> >>I was attempting to stimulate some thought re dynamic range and how
> he
>>> >>might use the concept of noise to solve his problem.
>>>
>>> >>It looks to me like he=3D92s taking a measurement and reading the
>>> >>fuzz
> =
>>on
>>> >>a scope trace. =A0But I don=3D92t know.
>>>
>>> >>An interested student would ask where those equations come from and
> how
>>> >>he could use them to solve his problem.
>>>
>>> >>The concept of how to design a gain lineup, whether that be a op-amp
> an=
>>d
>>> >>A/D or a downconvert with a A/D behind and solve for noise is all
>>> >>the same.
>>>
>>> >>Simple concepts such as where to place that narrowband filter and
>>> >>why elude many.
>>>
>>> >>Never did I say that he supplied enough information =3D85 but he
>>> >>does
> h=
>>ave
>>> >>access to the solution. =A0He just needs to know what to solve for.
>>>
>>> > Hello all,
>>> > Thanks a lot for your replies.
>>> > First, the application is not a homework, it's a real application
>>> > and
> a
>>> > real problem (for me).
>>> > The bandwidth of the analog front end (bias and pre amp of the
> detector
>>> > is have a high pass filter (a simple first order ac coupling) with
> 1,5H=
>>z
>>> > for the high pass filter and 350kHz 3rd order Bessel type low pass
>>> > filter. I'm sorry for the measurement I don't have access to
>>> > spectrum analyzer so
>>>
>>> So the bandwidth of your _filter_ is 350kHz, fin. =A0But what is the
> usef=
>>ul
>>> bandwidth of your _signal_?
>>>
>>> > Yes, the measurement of the "noise" is just a reading from a scope.
> But
>>> > it gives me an idea.
>>> > I'm not a specialist and just wanted to know what could be done to
>>> > reduce my noise (if it is possible). The oversampling technique
> permit
>>> > to reduce the noise from the quantification noise from the ADC.
> (correc=
>>t
>>> > me if I'm wrong !).
>>>
>>> More or less correct, yes. =A0_If_ the ADC has enough intrinsic noise
> tha=
>>t
>>> it shows up in the ADC output, then yes, oversampling and averaging
>>> can help you to overcome quantization noise.
>>>
>>> > But it doesn't reduce the noise which come from my analog signal
>>> > chain, isn't it ?
>>>
>>> Correct. =A0It can't do a thing about that.
>>>
>>> > I guess that the only solution to my problem is to improve the
>>> > analog chain . ?
>>>
>>> Unless your analog low-pass filter is wider than necessary, or your
> noise
>>> is being introduced after it, yes.
>>>
>>> > So using the digital filter just permit to reduce the analog filter
>>> > (anti aliasing) ?
>>>
>>> Using a filter in digital-land offers a multitude of advantages, about
>>> the only thing it _doesn't_ do is undo the effects of aliasing -- that
>>> would be like a blender that can un-scramble eggs.
>>>
>>> --www.wescottdesign.com- Hide quoted text -
>>>
>>> - Show quoted text -
>>
>>
>>Which topology are you using:
>>Sensor--->Filter--->op-amp--->A/D
>>or
>>Sensor---> Op-Amp--->Filter--->A/D
>>
>>For any amp string the noise performance of the system, i.e., the noise
>>floor will be set by the first amp.  It can never be better than that. 
>>Another way to say it is, if you create noise in your amp string, you
>>can=92t see a signal from your input source better than that.
>>
>>Any lossy components, (i.e. resistors, non-ideal filters), in front of
>>an amplifier adds to the noise.
>>You will never be able to see below the noise floor set by your system
>>NF.
>>
>>What process gain and averaging allow you to do, is get down to the
>>noise floor of your gain chain.  So you want your narrowest filter to be
>>the digital filter.  Noise Out of your gain stage =3D KTBGF. Consider
>>your A/D as just another gain stage with a NF.
>>
>>There is hope that things could get better than what you see on your
>>scope.  Especially, if your topology is the first one shown above. I
>>suspect you are making the noise measurement with a scope bandwidth
>>wider than your filter bandwidth.  Correct?  If true, that adds to the
>>fuzz on the trace because of the increased bandwidth.  Also, are you
>>making the measurement with the sensor connected?
>>
>>Let your digital filter set your acquisition bandwidth.  If you aren=92t
>>already, use the second topology above. Terminate the input to your
>>amp/filter with the appropriate impedance and make the measurement with
>>your A/D and digital filter to determine your true system noise floor. 
>>If it goes up when you add your sensor, well you know.
>>
>>It should be with a few dB of your calculation or you=92ve got other
>>problems.  You could also sweep your system using a sine source at the
>>input to verify measurement bandwidth.  Just suck the data into MatLab
>>if you don=92t have a canned FFT.  Painful, but it works.
>>
>>All of the above applies to thermal noise and not the 1/f stuff.
>>
>>As Tim said, the digital filter will not fix any aliased stuff =85 you
>>have to stop that before it gets to the converter.  Your analog filter
>>should do that.  However, beware, depending on the self-resonance of
>>your filter components, you could very well have re-entrant modes at
>>Nyquest.  So use your scope and a signal generator to sweep the A/D
>>pre-filter.
>>
>>Hope this helps.
>>
>>
> 
> Hello,
> thanks a lot for your help and sorry for my late answer.
> 
> The bandwidth of the signal is 50Hz to 314kHz. For the topology, I have
> :
> Sensor---> Op-Amp--->Filter--->A/D
> 
> I though that the "oversampling" method permit to reduce white noise
> provided by op amp, detector resistor and ADC quantification noise...
> but if I understand well, It only permits to reduce quantification
> noise. I would want to know how can I calculate the noise factor of my
> analog chain. Since the noise factor more used in RF components, there
> is no value of it in datasheet in low noise op amp like the ADA4898
> (this is this op amp that I use for all my stages : pre-amp and filter
> stages since that I have low impedance sensor => I have to use a low
> noise voltage op amp). Could you please help me how I can do to
> calculate the NF with this op amp ?

The oversampling and averaging technique does, indeed work to reduce the 
effect of any white noise before the ADC conversion (actually, any 
bandlimited noise that's above the Nyquist rate of the sampling -- true 
white noise would be of infinite power at the ADC, and would screw up 
your measurement infinitely).  This applies to noise both inside and 
outside of the ADC.  You can work this out on paper: for noise that's of 
significantly higher bandwidth than your sampling rate, the "after-
sampling" noise magnitude is always the same.  Because faster sampling 
spreads out the after-sampling spectrum, the noise power spectral density 
for that component of noise gets spread out.

You _should_ be able to bandlimit the op-amp noise, and leave yourself 
with just the noise on the output of your filter (of which there will be 
some, if it's an active filter) and the (generally) quite considerable 
noise at the input of the ADC.

The oversample-and-average technique actually doesn't reduce quantization 
noise by itself: if you had an infinitely quiet ADC front end then the 
quantization noise would come through unscathed.  What reduces the 
quantization noise is the contribution of the high-bandwidth noise that 
'scrambles' the quantization noise, making it whiter.  Then the out-of-
band components of the quantization noise can be filtered out.  If you 
had an ADC that was too quiet you could fix it up by dithering the input 
or otherwise adding noise.  Sigma-delta converters, in fact, are set up 
to inherently generate this "dithering" in such a way that the noise PSD 
is shaped away from the desired signal PSD, allowing the filtering to 
work much better than it would with random noise.

Really, working this out on paper would help to make sense of it all.

-- 
www.wescottdesign.com