Dear all, I work with low frequency oscillations of the brain while a human subject is cognitively engaged in a task. From the off-line analysis (using zero-phase band pass FIR filters), I discovered that task-related cognitive signals are located in the range of [0.2 0.3]Hz and in [0.6 0.8]Hz. The fluctuations/oscillations below 0.2Hz is a REAL devil due to 1/f nature ( the noise power is ~100 times higher than signal's power). Now as for my experimental demands, I need to estimate online in real-time the signals mentioned in the above range and manipulate stimulus presentation. This eventually means I need to have a very sharp high pass filter with almost zero group/phase delay. Which sounds impossible! However, I should come up with a decent trade of between SNR and phase-delay. Do you have any suggestions? All suggestions ranging from signal processing/machine learning/hardware to solve this problem is most well come! Thanks in advance! Sincerely, Gangadhar GARIPELLI Doctoral student, EPFL Switzerland
How to deal with 1/f noise in low frequency oscillations?
Started by ●November 5, 2010
Reply by ●November 5, 20102010-11-05
> Which sounds impossible!"Time-bandwidth" product comes to mind, but let's see what the others say.
Reply by ●November 5, 20102010-11-05
On Nov 5, 4:46�am, "ganges.g" <gangadhar.garipelli@n_o_s_p_a_m.epfl.ch> wrote:> Dear all, > I work with low frequency oscillations of the brain while a human subject > is cognitively engaged in a task. From the off-line analysis (using > zero-phase band pass FIR filters), I discovered that task-related cognitive > signals are located in the range of [0.2 0.3]Hz and in [0.6 0.8]Hz. The > fluctuations/oscillations below 0.2Hz is a REAL devil due to 1/f nature ( > the noise power is ~100 times higher than signal's power). > > Now as for my experimental demands, I need to estimate online in real-time > the signals mentioned in the above range and manipulate stimulus > presentation. This eventually means I need to have a very sharp high pass > filter with almost zero group/phase delay. Which sounds impossible! > > However, I should come up with a decent trade of between SNR and > phase-delay. Do you have any suggestions? All suggestions ranging from > signal processing/machine learning/hardware to solve this problem is most > well come! > > Thanks in advance! > Sincerely, > Gangadhar GARIPELLI > Doctoral student, > EPFL SwitzerlandSo the signal is already contaminated with 1/f noise before you start sensing/amplifying/filtering? Or might the electrode[s]/amplifiers be contributing to the noise? Is there any way to acquire the signal that would reduce the noise? What are you doing with the signal to be extracted that requires "real time" processing? Maybe if you expanded a bit on what you are going to do with the signal we might have a better idea. As it is, it does seem difficult. Good luck!
Reply by ●November 5, 20102010-11-05
On Nov 5, 8:24�am, cassiope <f...@u.washington.edu> wrote:> On Nov 5, 4:46�am, "ganges.g" > > > > > > <gangadhar.garipelli@n_o_s_p_a_m.epfl.ch> wrote: > > Dear all, > > I work with low frequency oscillations of the brain while a human subject > > is cognitively engaged in a task. From the off-line analysis (using > > zero-phase band pass FIR filters), I discovered that task-related cognitive > > signals are located in the range of [0.2 0.3]Hz and in [0.6 0.8]Hz. The > > fluctuations/oscillations below 0.2Hz is a REAL devil due to 1/f nature ( > > the noise power is ~100 times higher than signal's power). > > > Now as for my experimental demands, I need to estimate online in real-time > > the signals mentioned in the above range and manipulate stimulus > > presentation. This eventually means I need to have a very sharp high pass > > filter with almost zero group/phase delay. Which sounds impossible! > > > However, I should come up with a decent trade of between SNR and > > phase-delay. Do you have any suggestions? All suggestions ranging from > > signal processing/machine learning/hardware to solve this problem is most > > well come! > > > Thanks in advance! > > Sincerely, > > Gangadhar GARIPELLI > > Doctoral student, > > EPFL Switzerland > > So the signal is already contaminated with 1/f noise before you start > sensing/amplifying/filtering? �Or might the electrode[s]/amplifiers be > contributing to the noise? �Is there any way to acquire the signal > that > would reduce the noise? > > What are you doing with the signal to be extracted that requires "real > time" > processing? �Maybe if you expanded a bit on what you are going to do > with > the signal we might have a better idea. �As it is, it does seem > difficult. > > Good luck!- Hide quoted text - > > - Show quoted text -Is your acquisition system accessible? Can you break it into to pieces and determine what element is the biggest offender? I.E, if you short out the input to your A/D is it good to go? Like another poster said, it would be helpful to know the source of the noise. Are you also saying you can't use any ave?
Reply by ●November 5, 20102010-11-05
On 11/05/2010 04:46 AM, ganges.g wrote:> Dear all, > I work with low frequency oscillations of the brain while a human subject > is cognitively engaged in a task. From the off-line analysis (using > zero-phase band pass FIR filters), I discovered that task-related cognitive > signals are located in the range of [0.2 0.3]Hz and in [0.6 0.8]Hz. The > fluctuations/oscillations below 0.2Hz is a REAL devil due to 1/f nature ( > the noise power is ~100 times higher than signal's power). > > Now as for my experimental demands, I need to estimate online in real-time > the signals mentioned in the above range and manipulate stimulus > presentation. This eventually means I need to have a very sharp high pass > filter with almost zero group/phase delay. Which sounds impossible! > > However, I should come up with a decent trade of between SNR and > phase-delay. Do you have any suggestions? All suggestions ranging from > signal processing/machine learning/hardware to solve this problem is most > well come!When you say "real time", what sort of response time are you thinking about? With the signal buried that far down in the noise it'll take time to dig it out conclusively. Don't ignore the suggestion about avoiding the noise in the first place -- if you can arrange to take your data without the noise that'll help immensely. -- Tim Wescott Wescott Design Services http://www.wescottdesign.com Do you need to implement control loops in software? "Applied Control Theory for Embedded Systems" was written for you. See details at http://www.wescottdesign.com/actfes/actfes.html
Reply by ●November 5, 20102010-11-05
One more thought (see "averaging" above): Can you get multiple measurements where signal and noise aren't fully correlated? There are many tricks to achieve this, for example using multiple sensors or chopping the signal before the noise source.
Reply by ●November 5, 20102010-11-05
mnentwig <markus.nentwig@n_o_s_p_a_m.nokia.com> wrote:> One more thought (see "averaging" above): > Can you get multiple measurements where signal and noise aren't fully > correlated? There are many tricks to achieve this, for example using > multiple sensors or chopping the signal before the noise source.I was assuming that the noise was from the source. Consider thermal noise from a resistor. One can cool it down to reduce it, but otherwise there isn't much else that can be done. One can't cool the brain down to reduce its low frequency noise. -- glen
Reply by ●November 5, 20102010-11-05
Some more 1/f noise from my brain: If the measurement is somehow electrical, you could try to modulate the load impedance. If the wanted signal and the noise have different origins, they would get affected in different ways and you can separate them.
Reply by ●November 6, 20102010-11-06
On 11/05/2010 01:46 PM, glen herrmannsfeldt wrote:> mnentwig<markus.nentwig@n_o_s_p_a_m.nokia.com> wrote: > >> One more thought (see "averaging" above): >> Can you get multiple measurements where signal and noise aren't fully >> correlated? There are many tricks to achieve this, for example using >> multiple sensors or chopping the signal before the noise source. > > I was assuming that the noise was from the source. Consider > thermal noise from a resistor. One can cool it down to reduce > it, but otherwise there isn't much else that can be done. > One can't cool the brain down to reduce its low frequency noise.My understanding of EKGs (mostly garnered from this group) is that there's a lot of noise from voluntary muscles and other localized sources -- so there's often physical things you can do to reduce noise, before you get down to swallowing the bitter pill of a one dimensional vector of signal+noise that you have to sort out somehow. One hopes that EEG is the same. -- Tim Wescott Wescott Design Services http://www.wescottdesign.com Do you need to implement control loops in software? "Applied Control Theory for Embedded Systems" was written for you. See details at http://www.wescottdesign.com/actfes/actfes.html
Reply by ●November 15, 20102010-11-15
Dear all, Here are my clarifications: 1. Acquisition : I can not avoid 1/f noise, as the noise is basically originating from the brain and I work with non-invasive methods (EEG) acquire the signals. I believe, it is the same even if one goes invasive. The signals and noise have same origins. This is quite common with biological signals. Also, all the other noise ( eye movement, or muscular etc) is (and will be) avoided as much as possible during the experiment. 2. Multiple measurement : Yes, indeed I have many electrodes recoding simultaneously across different spatial locations. Eventually, spatial filtering does help increase SNR. But this is not enough to remove 1/f noise in the low frequency range. And I can not afford to do multiple measurements in time, as per the experimental needs. 3.Real-time: I would like to detect within 2s with respect to a visual- stimuli which results in fluctuations in the [0.2 0.8] Hz with respect to cognition of the subject. 4. Using Choppers : This may help do amplitude modulation of my signals to a chopper-frequency (or carrier frequency, say 200Hz) where IIR filter design is easier. But, designing of filter with sharper transition between 200.1Hz and 200.2Hz is still a problem! Looking forward for more suggestions! :-) Thanks a lot for your attention! G. On Nov 6, 4:52�pm, Tim Wescott <t...@seemywebsite.com> wrote:> On 11/05/2010 01:46 PM, glen herrmannsfeldt wrote: > > > mnentwig<markus.nentwig@n_o_s_p_a_m.nokia.com> �wrote: > > >> One more thought (see "averaging" above): > >> Can you get multiple measurements where signal and noise aren't fully > >> correlated? There are many tricks to achieve this, for example using > >> multiple sensors or chopping the signal before the noise source. > > > I was assuming that the noise was from the source. �Consider > > thermal noise from a resistor. �One can cool it down to reduce > > it, but otherwise there isn't much else that can be done. > > One can't cool the brain down to reduce its low frequency noise. > > My understanding of EKGs (mostly garnered from this group) is that > there's a lot of noise from voluntary muscles and other localized > sources -- so there's often physical things you can do to reduce noise, > before you get down to swallowing the bitter pill of a one dimensional > vector of signal+noise that you have to sort out somehow. > > One hopes that EEG is the same. > > -- > > Tim Wescott > Wescott Design Serviceshttp://www.wescottdesign.com > > Do you need to implement control loops in software? > "Applied Control Theory for Embedded Systems" was written for you. > See details athttp://www.wescottdesign.com/actfes/actfes.html
