This project uses real-world RF data captured with a software-defined radio (SDR).
The RF capture chain was validated for stability prior to any dataset construction or ML modeling.
- SDR: Nooelec NESDR SMArt v5 (RTL2832U + R820T)
- Host OS: macOS
- Sample Rate: 1.024 MS/s
- Gain Control: Manual (no AGC)
- IQ Format: Unsigned 8-bit interleaved I/Q
RF stability was verified by:
- Capturing long IQ recordings
- Computing power spectral density (PSD) from early and late time slices
- Comparing spectra to detect frequency drift or instability
A persistent PLL not locked warning was observed from the tuner, but its impact was evaluated empirically via spectral stability checks.
-
FM Broadcast (100 MHz)
Wideband FM energy remained time-stable across capture windows. -
NOAA Weather Radio (162.5 MHz)
Narrowband carrier remained sharply defined and fixed in frequency across time slices.
- PSD peaks align across time slices for both wideband and narrowband signals
- No observable frequency drift or spectral smearing
- PLL warnings confirmed non-impacting for spectral-feature-based ML tasks
The RF capture pipeline is therefore considered stable and suitable for downstream spectrogram generation and machine learning.

