KT/C Noise canceling/reducing in analog or digital

[alexgaas]

KT/C Noise canceling

Is it possible to cancel or reduce the KT/C noise in some analog or digital way??
Thanks in advance!

 

[keith1200rs]

KT/C Noise canceling

Could you explain more what you mean? What type of circuit?

Keith.

 

[alexgaas]

Re: KT/C Noise canceling

Sure!I have a track and hold amplifier with very high bandwidth (near 10 GHz to sample an UWB signal ) and i have to choose the capacitance to store the sampled signal...The problem is that a large capacitance reduce the bandwidth of the buffer of the THA. I would a large dynamic range near 60 dB and probably if i choose a small capacitance i won't reach this range.
So i'm searching other approaches to solve this problem...in analog or digital way.
I sampled the signal in an undersampling way so the distance between two samples are between 2 and 1 nanoseconds.
If you need other information tell me!
Thank you!

 

[keith1200rs]

KT/C Noise canceling

So is the noise you are concerned about charge injection from the switches? If so there are techniques for minimising that. One way I have used puts two switches in series with the second switch shorted and driven by the opposite clocks, as I recall. The idea is the charges cancel one another out as the switches change, rather than affecting the sampling capacitor.

Keith.

 

[alexgaas]

Re: KT/C Noise canceling

No charge injection is another problem. This is about the integrated noise in a capacitor...In this paper you can see this noise (KT/C)

https://web.mit.edu/klund/www/CMOSnoise.pdf

This is a noise source but is indipendent from rout of the switch stage...

 

[FvM]

The only way to reduce noise in signal processing is averaging, if it's a periodical signal. For achievable SNR at 10 GHz bandwidth,
you can refer to standard measurement equipment, e.g. fast oscilloscopes. 60 dB dynamic seems not impossible, but challenging
though.

 

[alexgaas]

Thanks!Yes the signal is periodical, it's a radar.
Averaging in digital?But if we use a digital signal processing the noise rms don't have to be less than 1/2 LSB of my ADC but more than this limit. Otherwise we will average equal data and there's no improvement in SNR!
I was thinking to a way to reduce noise in an analog way if could be more efficient...I could use any technique because it's an IC custom design!

 

[keith1200rs]

The problem is that it is thermal noise so you need to reduce R (and increase C to compensate). But that takes more power and presumes that you have not reached some other limit (such as power consumption).

I am not sure what solutions there may be. I have designed laser rangefinders and usually rely on averaging multiple pulses but in your case the time between samples of the same object may be too long (it will have moved).

Keith.

Added after 3 minutes:

One other thought, there is a technique where a periodic signal (e.g. triangle) is added to the signal going into the ADC to ensure the signal covers more than 1LSB and then averaged. It allows you to get more effective bits out of ADC. The periodic signal disappears provided you average over it's period.

Keith.

 

[watersky]

I know some adc designs use digital method to reduce noise. Some different arithmetic is used commonly.

 

[alexgaas]

" One other thought, there is a technique where a periodic signal (e.g. triangle) is added to the signal going into the ADC to ensure the signal covers more than 1LSB and then averaged. It allows you to get more effective bits out of ADC. The periodic signal disappears provided you average over it's period. "

Keith, do you have any references of this technique?

Reduce r could be a solution...but is not so easy...and for example if i double the capacitance (and reduce r in the same time) I have only 3 dB of improvement in SNR.

I was thinking to re-sample (or integrate) another time (with period of 1 ns or 2ns). In this way the noise equivalent bandwidth is 1/20 of the last one (1GHz). But it's only an idea...i don't know any about this technique!What do you think?

 

[biff44]

I guess I am confused about your circuit. You have a track and hold amplifier with an external capacitor. A small capacitor allows good bandwidth/rise time, which you need for sampling a short pulse, so you have to use a small capacitor.

But you are apparently doing equivalent time sampling, which means you are sampling every 100 uS or so. So why do you not buffer the output of your capacitor with a high impedance amplifer, and at the output of that amplifier you can use an R-C lowpass filter with a 100 KHz lowpass response?

That way you get a fast sample rise time, and effectively average out most of the noise before it hits your ADC. Then you can play around with the lowpass corner frequency, or add additional measurement samples at that time delay point.

 

[keith1200rs]

It's called dithering and is referred to here: https://en.wikipedia.org/wiki/Analog-to-digital_converter#Dither

I worked on a project where the dither was a triangle, not random.

There is a bit more info here: https://www.analog.com/library/analogdialogue/archives/40-02/adc_noise.html

Try searching for "dithering to increase ADC resolution" in Google.

I am not sure I understand the second part of your reply. It sounds like you are talking about box averaging - average groups of 10 samples - I am not sure. If you don't mind smearing the time resolution you can use a moving average, or exponential average of samples.

Keith.

 

[alexgaas]

"You have a track and hold amplifier with an external capacitor. A small capacitor allows good bandwidth/rise time, which you need for sampling a short pulse, so you have to use a small capacitor."

-The capacitor is not external I'm doing a custom IC design an so the capacitor is in the track and hold amplifier.

"But you are apparently doing equivalent time sampling, which means you are sampling every 100 uS or so."

-Yes I use equivalent time sampling but I sample every 1 or 2 ns. I have strong requirements also in time of acquiring...in 1 ms or less I have to acquire a scattering signal of 10 Ghz with 60-70 dB SNR and the pulse could be in a window time of 10 ns...

"So why do you not buffer the output of your capacitor with a high impedance amplifer, and at the output of that amplifier you can use an R-C lowpass filter with a 100 KHz lowpass response?"

I coudn't do this because i don't have such time. After 1 ns I have to take another sample. Maybe you are trying to say what i had already say in the last answer. Is it right?

If I didn't explain well tell me!

 

[keith1200rs]

It sounds like you have some scope to do more averaging - if you have 1ns samples but only need 10ns resolution.

Keith.

 

[alexgaas]

"I am not sure I understand the second part of your reply. It sounds like you are talking about box averaging - average groups of 10 samples - I am not sure."

Which part?

Keith probably I don't understand your answer. I'm going to explain well..
I'm doing an UWB radar receiver (short range) that is used for medical imaging.
An UWB signal is transmitted and penetrate in the body and is reflected by skin and internal tissues. We need a 60-70 dB SNR over 10GHz bandwidth and I have to view a 10 ns of this reflected signal.
So I'm designing an oscilloscope IC to digitize this signal. I used a coherent sampling and to reduce jitter (I have a PLL with 1 ps of jitter but it's not enough) I need to reacquire samples and average about 100 samples.
Probably this effect of dither is already created by jitter...

Thanks a lot guys!

 

[keith1200rs]

"I am not sure I understand the second part of your reply. It sounds like you are talking about box averaging - average groups of 10 samples - I am not sure."

Which part?

This bit:

I was thinking to re-sample (or integrate) another time (with period of 1 ns or 2ns). In this way the noise equivalent bandwidth is 1/20 of the last one (1GHz). But it's only an idea...i don't know any about this technique!What do you think?

It might still be worthwhile thinking about dither. Also, a search for high speed digitising oscilloscope techniques might be worthwhile. Tektronix and LeCroy in particular.

Keith.

 

[alexgaas]

If you take the formula of KT/C noise of the link you can see that KT/C noise is indipendent from r...ok that's true. But it's the result of the integration of a single pole
RC network an this constant RC is very short for the very high bandwidth of signal.
So my idea is to re-sample (or integrate but the series switch with a capacitance is like an integrator) this storage voltage over a bandwidth less large.
In this way, probably my noise is filtered but now my new capacitance (of the second sample and hold) could be 10 or 20 time the first one (KT/C noise reduced by 13 dB if we filter with 2 ns).
Probably i'm doing some mistakes...what do you think?


I have already searched about this topic...and obviously le croy and others don't say details about that...also i'm using a standard CMOS nanometer library and not others exotic technologies.

 

[keith1200rs]

If you take the formula of KT/C noise of the link you can see that KT/C noise is indipendent from r...ok that's true. But it's the result of the integration of a single pole
RC network an this constant RC is very short for the very high bandwidth of signal.
So my idea is to re-sample (or integrate but the series switch with a capacitance is like an integrator) this storage voltage over a bandwidth less large.
In this way, probably my noise is filtered but now my new capacitance (of the second sample and hold) could be 10 or 20 time the first one (KT/C noise reduced by 13 dB if we filter with 2 ns).
Probably i'm doing some mistakes...what do you think?

Yes, that sounds OK. What you will be doing is reducing the bandwidth, but it sounds like you don't need 10GHz bandwidth.

By the way, from the work I did on laser rangefinders the optimum receiver bandwidth was 1/(2.pi.PW) where PW is the pulse width and pi=3.14. I cannot remember the source of that information, but a wider bandwidth than that increases the noise without carrying any extra useful information. The signal might be a little larger with a wider bandwidth, but the noise starts to increase more than the signal increase, I think.

Keith.

EDIT: just checking that 1/(2.pi.PW) and it doesn't seem quite right. However, I do remember that the optimum bandwidth is smaller than you might think. I will see if I can find the original source, but it was quite a while ago.

 

[alexgaas]

"Yes, that sounds OK. What you will be doing is reducing the bandwidth, but it sounds like you don't need 10GHz bandwidth. "

Unfortunately I need 10GHz of bandwidth in the front end to capture the signal.
I need all information about the UWB signal because in all 60 dB of dynamic range there are all scattering information about biological tissues.

"By the way, from the work I did on laser rangefinders the optimum receiver bandwidth was 1/(2.pi.PW) where PW is the pulse width and pi=3.14. I cannot remember the source of that information, but a wider bandwidth than that increases the noise without carrying any extra useful information. The signal might be a little larger with a wider bandwidth, but the noise starts to increase more than the signal increase, I think. "

In this project you have to detect only the time between two pulses...in my project i'm also interested in the shape. Anyway my idea is to detect the signal but "look" it when i have time and over a large time. So the noise equivalent bandwidth could be very narrow -> increase resolution!
I will try...thanks again!

Obviously any criticisms are welcomed!