Forums

Bidirectional equalization

Started by alberto.fuggetta June 5, 2010
Hello,

I'm simulating a Decision Feedback Equalizer for MSK signal, using the well
know LMS algorithm. In order to get better performance I tried to equalize
each packet in two stages.
The structure of the packet is as follows:
[Payload1 TrainingSeq Payload2]
Since the training sequence is placed in the middle (GSM like), I'd like to
equalize the second part in forward direction and the first part in
backward direction.
To accomplish the backward equalization, I tried to feed the equalizer with
the flipped version of the incoming signal (just payload1 + training
sequence). I also flipped the training sequence only, so that the DFE has
the right one.
The system works well in presence of noise. However, when I simulate the
multipath channel, the performance is under expectation.
Actually, with bidirectional equalization I get worse results compared to
unidirectional equalization.
I observed that the BER distribution along different packet in the
simulation loop has a different behavior for backward and forward
equalization.
Backward equalization has BER a couple of order worse than forward
equalization.
Has anyone experienced a similar problem?

Thanks a lot,

Alberto
alberto.fuggetta wrote:
> Hello, > > I'm simulating a Decision Feedback Equalizer for MSK signal, using the well > know LMS algorithm. In order to get better performance I tried to equalize > each packet in two stages. > The structure of the packet is as follows: > [Payload1 TrainingSeq Payload2] > Since the training sequence is placed in the middle (GSM like), I'd like to > equalize the second part in forward direction and the first part in > backward direction. > To accomplish the backward equalization, I tried to feed the equalizer with > the flipped version of the incoming signal (just payload1 + training > sequence). I also flipped the training sequence only, so that the DFE has > the right one. > The system works well in presence of noise. However, when I simulate the > multipath channel, the performance is under expectation. > Actually, with bidirectional equalization I get worse results compared to > unidirectional equalization. > I observed that the BER distribution along different packet in the > simulation loop has a different behavior for backward and forward > equalization. > Backward equalization has BER a couple of order worse than forward > equalization. > Has anyone experienced a similar problem? > > Thanks a lot, > > Alberto
Why, specifically, are you wanting to do this? Unless the equalizer is symmetric then why would you expect the results to be the same? Why not pass Payload1 through the equalizer in the forward direction? Latency? I'm having a hard time visualizing that. Fred
On 6/6/10 5:58 PM, Fred Marshall wrote:
> alberto.fuggetta wrote: >> Hello, >> >> I'm simulating a Decision Feedback Equalizer for MSK signal, using the >> well >> know LMS algorithm. In order to get better performance I tried to >> equalize >> each packet in two stages. >> The structure of the packet is as follows: >> [Payload1 TrainingSeq Payload2] >> Since the training sequence is placed in the middle (GSM like), I'd >> like to >> equalize the second part in forward direction and the first part in >> backward direction. >> To accomplish the backward equalization, I tried to feed the equalizer >> with >> the flipped version of the incoming signal (just payload1 + training >> sequence). I also flipped the training sequence only, so that the DFE has >> the right one. >> The system works well in presence of noise. However, when I simulate the >> multipath channel, the performance is under expectation. >> Actually, with bidirectional equalization I get worse results compared to >> unidirectional equalization. >> I observed that the BER distribution along different packet in the >> simulation loop has a different behavior for backward and forward >> equalization. >> Backward equalization has BER a couple of order worse than forward >> equalization. >> Has anyone experienced a similar problem? >> >> Thanks a lot, >> >> Alberto > > Why, specifically, are you wanting to do this? > > Unless the equalizer is symmetric then why would you expect the results > to be the same? > > Why not pass Payload1 through the equalizer in the forward direction? > Latency? I'm having a hard time visualizing that.
Payload1 is normally passed backward (in GSM) because you normally have a very good estimate of the channel because of the training seq. If you pass it forward, the training seq is at the end of the payload. The channel estimate at the beginning of the payload isn't very good and that degrades performance of DFSE's. I haven't done this in years, but some care must be taken when running the equalizer backwards. You can't just reverse the data and pass it through the equalizer. For one, the channel taps are in the reverse order, I think. Ray
Ray,

thanks for your answer. The channel taps in reverse order could be a likely
explanation for the issue I currently see.

I borrowed this technique from some GSM receiver employing MLSE
equalization. However, I found something similar in a Matlab example where
a DFE is used instead.

I think I will try some different technique to improve my equalizer
performance.

Alberto 


>On 6/6/10 5:58 PM, Fred Marshall wrote: >> alberto.fuggetta wrote: >>> Hello, >>> >>> I'm simulating a Decision Feedback Equalizer for MSK signal, using the >>> well >>> know LMS algorithm. In order to get better performance I tried to >>> equalize >>> each packet in two stages. >>> The structure of the packet is as follows: >>> [Payload1 TrainingSeq Payload2] >>> Since the training sequence is placed in the middle (GSM like), I'd >>> like to >>> equalize the second part in forward direction and the first part in >>> backward direction. >>> To accomplish the backward equalization, I tried to feed the equalizer >>> with >>> the flipped version of the incoming signal (just payload1 + training >>> sequence). I also flipped the training sequence only, so that the DFE
has
>>> the right one. >>> The system works well in presence of noise. However, when I simulate
the
>>> multipath channel, the performance is under expectation. >>> Actually, with bidirectional equalization I get worse results compared
to
>>> unidirectional equalization. >>> I observed that the BER distribution along different packet in the >>> simulation loop has a different behavior for backward and forward >>> equalization. >>> Backward equalization has BER a couple of order worse than forward >>> equalization. >>> Has anyone experienced a similar problem? >>> >>> Thanks a lot, >>> >>> Alberto >> >> Why, specifically, are you wanting to do this? >> >> Unless the equalizer is symmetric then why would you expect the results >> to be the same? >> >> Why not pass Payload1 through the equalizer in the forward direction? >> Latency? I'm having a hard time visualizing that. > >Payload1 is normally passed backward (in GSM) because you normally have >a very good estimate of the channel because of the training seq. If you >pass it forward, the training seq is at the end of the payload. The >channel estimate at the beginning of the payload isn't very good and >that degrades performance of DFSE's. > >I haven't done this in years, but some care must be taken when running >the equalizer backwards. You can't just reverse the data and pass it >through the equalizer. For one, the channel taps are in the reverse >order, I think. > >Ray >