Content-Transfer-Encoding: 8bit
Message-ID: <5.1.0.14.0.20021015135147.00a8df98@gemini.herts.ac.uk>
Date: Tue, 15 Oct 2002 15:35:20 +0100
To: rsgb_lf_group@blacksheep.org
From: "James Moritz" <j.r.moritz@herts.ac.uk>
In-reply-to: <Pine.LNX.3.95.1021015125639.216C-100000@fitec.omskcity.com >
References: <Chameleon.1034346317.Girardi@tlvk7v>
MIME-Version: 1.0
Subject: LF: Re: Non-linear signal processing
Content-Type: text/plain; charset=us-ascii; format=flowed
Sender: <owner-rsgb_lf_group@blacksheep.org>
Precedence: bulk
Reply-To: rsgb_lf_group@blacksheep.org

Dear Alex, LF Group,

At 13:26 15/10/2002 +0000, you wrote:
>So it is very simple to realize! The only addition is input convertion
>of the incoming signal
>
>S1(t) = S(t)/[A + S(t)*S(t)]
>
>where S(t) if signal from RX. Parameter A shold be controlled to get
>best reciption. Than signal S1(t) is processed by conventional procedure
>(say FFT).
>
>It is very interesting to try such an algorithm.
>
>73 de RA9MB/Alex
>http://www.qsl.net/ra9mb

As was seen some time ago with the experimental "Hendrixizer" clipping 
feature added to Jason, non-linear processing of the signal can certainly 
be an improvement under some conditions. The conditions seem to be that the 
noise within the receiver passband is wide band in nature; if strong 
narrow-band signals are also present, intermodulation and/or blocking 
effects occur. Due to the non-linear term in Alex's function above, I 
expect the same would apply. But if there is nothing but weak signals and 
noise within the RX passband, simply clipping the signal gives a 
significant improvement. The effects of clipping were found by accident, so 
I would not be surprised if Alex's more scientific approach also gave 
benefits - it is certainly true that most of the noise on LF is nothing 
like white noise.

Looking at the function Alex suggests, and assuming S(t)*S(t)] means 
S(t)^2, when S(t) is small, S1(t) is approximately S(t)/A, ie. linearly 
proportional. When S(t) is very large, S1(t) is approximately 1/S(t), ie. 
~=0. The gradient "turns over" when S(t)^2 = A. So the function is like a 
software noise blanker, removing the high amplitude parts of the input 
signal, with the value of A setting the threshold. One can intuitively see 
how this could work with LF noise consisting mainly of spikes of QRN

Well, now all I have to do is sit back and relax while the spectrogram 
software guys get it working... :-)

Cheers, Jim Moritz
73 de M0BMU