radiosonde_auto_rx/auto_rx/test/notes/2019-03-02_performance_base...

1.7 KiB

2019-03-02 auto_rx Performance Baseline

Mark Jessop - 2019-03-02

The scripts in the auto_rx/test directory provide the ability to measure the various decoders performance with a set of radiosonde signal samples, with calibrated Eb/N0 (SNR-per-bit) values.

A few notes:

  • All measurements have been performed using the demodulator (rtl_fm, filtering) settings as per the 2019-02-27 auto_rx release.
  • All decoders (rs41, rs92, etc...) have any available error-correction options enabled.
  • Packet Error Rates have been calculated based on a comparison with the known number of packets in the high-SNR source file.

FM Demodulation is performed with the following rtl_fm sample rates:

  • RS41: 15 kHz
  • RS92: 12 kHz
  • DFM09: 20 kHz
  • M10: 22 kHz

PER Performance 2019-03-02

The traces for the error-corrected decoders (RS41, RS92, DFM), exhibit the typical 'cliff' response, with the PER going from 1.0 to 0 in only a few dB.

A few things to look into:

  • This is a plot of packet-error-rate. Can we get a measure of the number of bits in a 'packet' (whatever that may be for each sonde; it may mean multiple frames) to be able to estimate a bit-error-rate? How does this compare to theoretical FSK modem performance?
  • Why is ths RS92 'cliff' offset from the others? Are the packets bigger? Is there something sub-optimal in the demodulation chain?
  • Further investigation into what the rtl_fm samples rates mean for pre-FM-demod filter bandwidth is required, to ensure they are set optimally.
  • Double and triple-checking of the Eb/N0 calculations in generate_lowsnr.py, to be sure we're generating our low-Eb/N0 samples correctly.