JANUS: frequency hopping pattern is incorrect

Hello to the UnetStack team! This is Emil, a PhD student at NTNU. I am working with UnetStack-based modems and experimenting with manufacturer interoperability via JANUS. I believe I have discovered a bug in your implementation of JANUS, which is linked to the frequency hopping pattern.

What I observed

I produced a JANUS packet with the JANUS C 3.0.5 implementation and transmitted it using another manufacturer's modem, which supports JANUS. The Subnero RE modem that received it, which ran UnetStack 3.2.0, could detect this packet, but produced bad frame notifications with seemingly random contents. I sent a JANUS packet with identical bits from the Subnero modem, with similar results on the other end (packet detected, random contents).

When I investigated the transmitted signals by plotting their spectrograms, I noticed that the frequency hopping patterns were different. It appeared that the hopping pattern from UnetStack was offset with respect to the C implementation. When I looked closer immediately after the preamble, I noticed that the UnetStack hopping pattern had reset. This pattern offset was also observed when I loopback-recorded a JANUS packet with Unet Audio 3.3.0 and viewed its spectrogram.

How to reproduce

Spectrogram of UnetStack JANUS packet

Refer to this answer. Zoom in to the first 0.4 seconds in the spectrogram. The first 0.2 seconds should be very similar to the next 0.2 seconds. In fact, this can be seen in the spectrogram in the linked answer.

Spectrogram of reference JANUS packet

Find the transmission script that is produced after compiling the JANUS C implementation and execute it as follows.
./janus-tx --pset-file etc/parameter_sets.csv --pset-id 1
Launch a Python environment from the same folder, then run the following piece of code.

import scipy.io.wavfile as wav
import scipy.signal as dsp
import matplotlib.pyplot as plt
import numpy as np

fs, x = wav.read('./janus.wav')
fr, ts, spec = dsp.spectrogram(x, fs=fs, nperseg=256, nfft=512)
plt.pcolormesh(ts, fr, spec, shading='nearest')
plt.show() # can be skipped if using Jupyter notebook

Zoom in to the first 0.5 seconds in the spectrogram. The first half should not be similar to the next half.

org-arl / unet-contrib