Forums

Need help on LMS based equalizer for GSM

Started by richard_zhang June 9, 2008
Hi all,

I am doing GSM software defined radio and trying to use normalized LMS
algorithm to implement an equalizer, since LMS has low complexity. But now
I find that the BER is not improved at all after adding the equalizer
block.

I just simply take the normalized LMS algorithm and use it, and don't have
deep investigation on it. My questions are:

1) typically, how much the BER performance can be improved by LMS
algorithm?
2) how to verify that the equalizer is working? The equalizer output has
10% to 20% difference when compared with the local ideal reference signal
(actually the training sequence), i.e.,  (eq_out - reference)/reference =
10%~20%, can equalizer work?
3) how to determine the filter tap numbers in the equalizer?


any suggestion is appreciated.

thank you


Richard
Hi Richard,

LMS is not usually used in GSM.  MLSE (viterbi) equalizer is usually
used. And there are quite a few reasons why.
Regarding your specific questions, I don't think you would get data
(even it were available) without specifying your operating conditions
i.e. SNR,  power delay profile, doppler etc.
For the last question, the number of taps for the adaptive filter is
usually equal to the delay spread (length of impulse response).

Regards
Piyush

On Jun 10, 6:47&#2013266080;am, "richard_zhang" <hardhear...@yahoo.com.cn> wrote:
> Hi all, > > I am doing GSM software defined radio and trying to use normalized LMS > algorithm to implement an equalizer, since LMS has low complexity. But now > I find that the BER is not improved at all after adding the equalizer > block. > > I just simply take the normalized LMS algorithm and use it, and don't have > deep investigation on it. My questions are: > > 1) typically, how much the BER performance can be improved by LMS > algorithm? > 2) how to verify that the equalizer is working? The equalizer output has > 10% to 20% difference when compared with the local ideal reference signal > (actually the training sequence), i.e., &#2013266080;(eq_out - reference)/reference = > 10%~20%, can equalizer work? > 3) how to determine the filter tap numbers in the equalizer? > > any suggestion is appreciated. > > thank you > > Richard
Hi Piyush,

thanks for the reply.

Is MLSE the only equalizer which can be used in GSM? The reason that I
don't want to adopt MLSE is its high complexity, and I did see some
articles saying that LMS can be used in GSM, but actually I am not sure.




>Hi Richard, > >LMS is not usually used in GSM. MLSE (viterbi) equalizer is usually >used. And there are quite a few reasons why. >Regarding your specific questions, I don't think you would get data >(even it were available) without specifying your operating conditions >i.e. SNR, power delay profile, doppler etc. >For the last question, the number of taps for the adaptive filter is >usually equal to the delay spread (length of impulse response). > >Regards >Piyush > >On Jun 10, 6:47=A0am, "richard_zhang" <hardhear...@yahoo.com.cn> wrote: >> Hi all, >> >> I am doing GSM software defined radio and trying to use normalized LMS >> algorithm to implement an equalizer, since LMS has low complexity. But
now=
> >> I find that the BER is not improved at all after adding the equalizer >> block. >> >> I just simply take the normalized LMS algorithm and use it, and don't
have=
> >> deep investigation on it. My questions are: >> >> 1) typically, how much the BER performance can be improved by LMS >> algorithm? >> 2) how to verify that the equalizer is working? The equalizer output
has
>> 10% to 20% difference when compared with the local ideal reference
signal
>> (actually the training sequence), i.e., =A0(eq_out -
reference)/reference =
>=3D >> 10%~20%, can equalizer work? >> 3) how to determine the filter tap numbers in the equalizer? >> >> any suggestion is appreciated. >> >> thank you >> >> Richard > >
>Hi Richard, > >LMS is not usually used in GSM. MLSE (viterbi) equalizer is usually >used. And there are quite a few reasons why. >Regarding your specific questions, I don't think you would get data >(even it were available) without specifying your operating conditions >i.e. SNR, power delay profile, doppler etc. >For the last question, the number of taps for the adaptive filter is >usually equal to the delay spread (length of impulse response). >
%%%%% Hi Can you please tell the reasons why MLSE is used in GSM. According to my knowledge MLSE is very complex. And in GSM channel can will be time varying, so you need adaptive equaliser at receiver, thats why in the frame format we transmitt training sequence. Assuming that channel has got 5 taps, and QPSK modualtion is used, the complexity will be 4^5. Where using LMS or RLS significant performance can be obtained. Regards, Chintan

richard_zhang wrote:
> Hi all, > > I am doing GSM software defined radio and trying to use normalized LMS > algorithm to implement an equalizer, since LMS has low complexity.
For GMSK, they typically use either MLSE or DFE. LMS is not very optimal.
> But now > I find that the BER is not improved at all after adding the equalizer > block.
> I just simply take the normalized LMS algorithm and use it,
The things never work this way.
> and don't have > deep investigation on it. My questions are:
Then do a deep investigation on it, and you will not have questions.
> 1) typically, how much the BER performance can be improved by LMS > algorithm?
Completely depends on the channel and SNR.
> 2) how to verify that the equalizer is working? The equalizer output has > 10% to 20% difference when compared with the local ideal reference signal > (actually the training sequence), i.e., (eq_out - reference)/reference = > 10%~20%, can equalizer work?
How big is the difference before EQ compared to the difference after EQ?
> 3) how to determine the filter tap numbers in the equalizer?
Should be about BT plus the length of the multipath spread, i.e. 10..20 microseconds.
> any suggestion is appreciated.
How much is the appreciation? Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com