|Subject:||Re: LF: False Decode or Real? - Opera vs opds|
|From:||Markus Vester <[email protected]>|
|Date:||Sat, 23 May 2015 18:30:13 -0400|
|Dkim-signature:||v=1; a=rsa-sha256; c=relaxed/relaxed; d=mx.aol.com; s=20140625; t=1432420213; bh=Q3yboWlHiWmJ9yvNNjA+R+4ldspHPD3nQwWaajg3qgs=; h=From:To:Subject:Message-Id:Date:MIME-Version:Content-Type; b=QJ9z5rohv80G1Z2f/zewY+7CBwzLjK2abwWaxzLUS10gF6jYNybpGd3r1N5+TUYP/ ZhBys46YhSi1/HD0iIxEMu1XKazfHuVXEqWLG2KLIFW97l6y1hSIsQLNijRbvPnhn8 P2MrfRWXXblyUNefb6bX4TJN+uIsPTAMDu8cj+Pw=|
For the record, I should probably correct a couple of minor errors in my last post:
The Opera software version used in the test was 1.5.4 (ie. the last one before dynamic), not 1.5.6.
Scaling from Op-05 to Op-32 was of course by 18 dB, not 24 dB. I had tested it with Op-8 setting first, and then switched to 05 to aquire a reasonable number of results for statistics. After scaling, the results of the slow and quick tests were comparable so there was no indication of different decoding techniques at work.
Looking at the detectability graph, to be fair I should probably not claim a -49 dB limit but rather -48 dB, which would then be only 3 dB below -45 dB for Dynamic-DS. What is noticeable is that the decoder falls off pretty sharply between -39 dB for "all" and -41 dB for "nothing" (probably typical for any FEC coding scheme with checksums), whereas the correlation process degrades more softly in an "analog" fashion with decreasing success rate (probably also true for "dynamic").
LF: False Decode or Real? - Opera vs opds
Von: Markus Vester <[email protected]>
An: rsgb_lf_group <[email protected]>
Datum: Sa, 23 Mai 2015 8:24 pm
> OPDS makes use of Wolfs some what excellent spectrum software as DSP
Spectrum Lab is used as the frontend for opds, the only signal processing done there is a straight high-resolution FFT before exporting the data. Opds then searches for spectral peaks, and transforms a 1024 bin channel around each peak back to time domain (0.5 Hz wide, 4x oversampled). For coherent signals, the central carrier is extracted (including possible drift and fading), and synchroneous demodulation is performed. Then the resulting real waveform is Fourier transformed again for a CPU-efficient crosscorrelation to each of the callsign templates.
> both systems reliably produce false detections when subjected to noise ..
Opds-32 has been running on LF here continuously, typically analyzing about 20 QRM peaks every 10 minutes in a 115 Hz wide band. The searchlist contains currently around 50 entries. With a correlation threshold of 15 dB, 28 false detections have occured in 20 weeks since year 2015:
In a pdf file distributed with Opera 1.5.7 software, the author EA5HVK mentions that he tested opds and got 11 false detections in only 4 hours - an observation which is obviously not corroberated by my statistics.
> hence ultra stable TX and RX can give advantage
By using synchroneous demodulation rather than power detection, opds can detect coherent and stable signals that about 4 dB weaker. Attached are two plots, showing a side-by-side comparison of detection probability and SNR output from the Opera 1.5.6 decoder and the opds2h5c detector.
SpecLab's digimode terminal was used to generated coherent and perfectly timed Opera signals, and white noise from the test signal generator was added with variable power density (dB/Hz). The very same output was analyzed within SpecLab to feed opds, and played to the Opera software using VAC. To speed up the experiments, all testing was done at Op-05 speed (30 seconds), and SNR values were then scaled down by 24 dB to Opera-32. During the test, no false detections were observed in the output from either program.
In the attached plot successrate.png, the solid lines with squares show detection probability (0 to 100%) against average SNR in 2.5 kHz. The classic Opera decoder (red) achieved 50 % detections at -40 dB. Opds correlation (blue) goes down to -49 dB, showing a 4 dB advantage for these ideal signals. The blue crosses indicate correlation dB output from opds - note that only hits above the standard 15 dB threshold (dashed line) were counted as successful detections.
SNR_output.png shows indicated SNR values versus actual SNR from both programs. Opera 1.5.6 seemed to consistently read 1 dB high, whereas opds reads approximately 1 or 2 dB low, with a larger scatter. Part of this negative offset is because I had originally assumed the "dBOp" scale to be referenced 4 dB (instead of 3 dB) below PEP.
I have not yet tested the dynamic deep search in Opera 1.5.7, but the claimed -45 dB threshold (ie. 5 dB better than the decoder) seems quite plausible. I believe opds is also around 4 dB less sensitive for non-coherent signals, which would then put both programs in the same ballpark.
> Now mines a pint or are we onto shorts now ?
All the best,
|<Prev in Thread]||Current Thread||[Next in Thread>|