adaptive complex coefficient filter

Hello,
 I am working on some fetal heart rate algorithm. To verify it,
I decide to do the reverse thing. I use a LPC model(similar to
the model of voiced speech, using some period impulse to excite an all
pole filter)

the input of this model are:
1. the original signal
2. fetal heart rate calculated out from this original signal

the output is my reconstructed signal, I want to use this signal attached
with the known result(the fetal heart rate) to test my algorithm. 
However, the test result shows these reconstructed signals are too easy to
be calculated, is this caused by the phase of the signal?
now I see all the coefficients of the LPC adaptive filter are real
numbers, 
do I need to use a complex number adaptive filter? If so, can anyone tell
me how to adapt these complex coefficients?
Thks
elliot



		
This message was sent using the Comp.DSP web interface on
www.DSPRelated.com

elliot wrote:
> Hello,
>  I am working on some fetal heart rate algorithm. To verify it,
> I decide to do the reverse thing. I use a LPC model(similar to
> the model of voiced speech, using some period impulse to excite an all
> pole filter)
>
> the input of this model are:
> 1. the original signal
> 2. fetal heart rate calculated out from this original signal
>
> the output is my reconstructed signal, I want to use this signal attached
> with the known result(the fetal heart rate) to test my algorithm.
> However, the test result shows these reconstructed signals are too easy to
> be calculated, is this caused by the phase of the signal?
> now I see all the coefficients of the LPC adaptive filter are real
> numbers,
> do I need to use a complex number adaptive filter? If so, can anyone tell
> me how to adapt these complex coefficients?

I am not sure I understand your question. Do you see that the results
from analyzing the modeled data are too good, it is too easy to
reconstruct the input parameters?

If so, this is completely consistent with my experience, analyzing
simulated data that comply perfectly to the assumptions behind the
analysis is of no value beyond verifying that the implementation
is correct. You need to introduce some sort of imperfection in the
simulated system.

For the heart-beat simulation, you could do this in several ways.
One would be to excite the system with "blurred" impulses, say,
narrow Gaussian curves. This might introduce enough "fuzz" in the
system for the thing not to work perfectly.

The second thing you could do, is to introduce some sort of
"non-minimum phase-ness" to the heartbeat waveforms. Pass the
pulses through some all-pass filter at the source to break
down the minimum phase property. After that, simulate some
internal reflections as if the pulse propagates through
layered tissue.

It could take a lot of time getting these pieces together if you
have to do everything from scratch, so look around to see if
anybody have done anything similar. What you want to do, is to
make the signal break with the assumptions your algorithm is
based on.

Rune

Thank u, Rune.
 That is exactly what I meant. The simulated signal is too good.
My algorithm could calculted a perfect fetal heart rate line from it,
exactly as the input parameter. Is there a easier method to make 
the output(the simulated signal) as close to input 1(the original signal)
as possible, and also has the pitch period according to input 2?
Someone told me maybe complex number adaptive filter could achieve that.




>elliot wrote:
>> Hello,
>>  I am working on some fetal heart rate algorithm. To verify it,
>> I decide to do the reverse thing. I use a LPC model(similar to
>> the model of voiced speech, using some period impulse to excite an all
>> pole filter)
>>
>> the input of this model are:
>> 1. the original signal
>> 2. fetal heart rate calculated out from this original signal
>>
>> the output is my reconstructed signal, I want to use this signal
attached
>> with the known result(the fetal heart rate) to test my algorithm.
>> However, the test result shows these reconstructed signals are too easy
to
>> be calculated, is this caused by the phase of the signal?
>> now I see all the coefficients of the LPC adaptive filter are real
>> numbers,
>> do I need to use a complex number adaptive filter? If so, can anyone
tell
>> me how to adapt these complex coefficients?
>
>I am not sure I understand your question. Do you see that the results
>from analyzing the modeled data are too good, it is too easy to
>reconstruct the input parameters?
>
>If so, this is completely consistent with my experience, analyzing
>simulated data that comply perfectly to the assumptions behind the
>analysis is of no value beyond verifying that the implementation
>is correct. You need to introduce some sort of imperfection in the
>simulated system.
>
>For the heart-beat simulation, you could do this in several ways.
>One would be to excite the system with "blurred" impulses, say,
>narrow Gaussian curves. This might introduce enough "fuzz" in the
>system for the thing not to work perfectly.
>
>The second thing you could do, is to introduce some sort of
>"non-minimum phase-ness" to the heartbeat waveforms. Pass the
>pulses through some all-pass filter at the source to break
>down the minimum phase property. After that, simulate some
>internal reflections as if the pulse propagates through
>layered tissue.
>
>It could take a lot of time getting these pieces together if you
>have to do everything from scratch, so look around to see if
>anybody have done anything similar. What you want to do, is to
>make the signal break with the assumptions your algorithm is
>based on.
>
>Rune
>
>


		
This message was sent using the Comp.DSP web interface on
www.DSPRelated.com

elliot wrote:
> Thank u, Rune.
>  That is exactly what I meant. The simulated signal is too good.
> My algorithm could calculted a perfect fetal heart rate line from it,
> exactly as the input parameter. Is there a easier method to make
> the output(the simulated signal) as close to input 1(the original signal)
> as possible, and also has the pitch period according to input 2?
> Someone told me maybe complex number adaptive filter could achieve that.

In my experience, the easy-to-use data simulation tools generate
data that fit the assumptions of the signal processing so well
that it is difficult to test the robustness of the methods.

The simplest ways to mess up the signal would probably be some
allpass filter or convolving the signal with a Gaussian pulse
or something. I can't see how complex numbers can make any
difference. 

Rune

elliot wrote:
> Hello,
>  I am working on some fetal heart rate algorithm. To verify it,
> I decide to do the reverse thing. I use a LPC model(similar to
> the model of voiced speech, using some period impulse to excite an all
> pole filter)

Heart rate varies.  Are you working on heart sound or ECG?

> 
> the input of this model are:
> 1. the original signal
> 2. fetal heart rate calculated out from this original signal
> 
> the output is my reconstructed signal, I want to use this signal attached
> with the known result(the fetal heart rate) to test my algorithm. 
> However, the test result shows these reconstructed signals are too easy to
> be calculated, is this caused by the phase of the signal?
> now I see all the coefficients of the LPC adaptive filter are real
> numbers, 
> do I need to use a complex number adaptive filter? If so, can anyone tell
> me how to adapt these complex coefficients?
> Thks
> elliot
> 
> 
> 
> 		
> This message was sent using the Comp.DSP web interface on
> www.DSPRelated.com

Hi Stan,
 I work on ultrasound doppler signals, any suggestions? Thks.
 elliot
 

>elliot wrote:
>> Hello,
>>  I am working on some fetal heart rate algorithm. To verify it,
>> I decide to do the reverse thing. I use a LPC model(similar to
>> the model of voiced speech, using some period impulse to excite an all
>> pole filter)
>
>Heart rate varies.  Are you working on heart sound or ECG?
>
>> 
>> the input of this model are:
>> 1. the original signal
>> 2. fetal heart rate calculated out from this original signal
>> 
>> the output is my reconstructed signal, I want to use this signal
attached
>> with the known result(the fetal heart rate) to test my algorithm. 
>> However, the test result shows these reconstructed signals are too easy
to
>> be calculated, is this caused by the phase of the signal?
>> now I see all the coefficients of the LPC adaptive filter are real
>> numbers, 
>> do I need to use a complex number adaptive filter? If so, can anyone
tell
>> me how to adapt these complex coefficients?
>> Thks
>> elliot
>> 
>> 
>> 
>> 		
>> This message was sent using the Comp.DSP web interface on
>> www.DSPRelated.com
>


		
This message was sent using the Comp.DSP web interface on
www.DSPRelated.com

Thank u, I will spend some time to try these method.
elliot
>
>elliot wrote:
>> Thank u, Rune.
>>  That is exactly what I meant. The simulated signal is too good.
>> My algorithm could calculted a perfect fetal heart rate line from it,
>> exactly as the input parameter. Is there a easier method to make
>> the output(the simulated signal) as close to input 1(the original
signal)
>> as possible, and also has the pitch period according to input 2?
>> Someone told me maybe complex number adaptive filter could achieve
that.
>
>In my experience, the easy-to-use data simulation tools generate
>data that fit the assumptions of the signal processing so well
>that it is difficult to test the robustness of the methods.
>
>The simplest ways to mess up the signal would probably be some
>allpass filter or convolving the signal with a Gaussian pulse
>or something. I can't see how complex numbers can make any
>difference. 
>
>Rune
>
>


		
This message was sent using the Comp.DSP web interface on
www.DSPRelated.com