rtl_power, but for a USRP.
Stepped-tune power spectrum sweeps using the UHD Python API, writing rtl_power-compatible CSV. Built for a B210 doing 900 MHz band surveys, but nothing in it is B210-specific beyond the defaults.
Two scripts, no GNU Radio flowgraph required:
usrp_power.py— sweeps a frequency range, Welch PSD per step, CSV outplot_power.py— CSV to PNG: peak-hold trace, waterfall, band overlay
python3 usrp_power.py -f 850M:960M -g 40 -i 0.25 -c 10 -o survey.csv
python3 plot_power.py survey.csvThe output plot shades known allocations (ISM, cellular, paging) so you can tell at a glance what class of thing you're looking at, and prints a peak table with SNR.
- UHD with Python bindings (
import uhd) - NumPy, Matplotlib
NumPy 2.x will break UHD's Python bindings if libpyuhd was compiled
against NumPy 1.x — you'll see AttributeError: _ARRAY_API not found. Either
pin numpy<2, or check for a stray user-level install shadowing your system
one:
python3 -c "import numpy; print(numpy.__version__, numpy.__file__)"
pip uninstall numpy # if it resolves to ~/.localMatplotlib is only needed for plotting, and the two scripts communicate solely
through CSV — so if you're stuck in NumPy dependency hell you can capture in a
numpy<2 venv and plot with system Python.
rtl_power-compatible, so existing tooling (heatmap.py et al.) will also read
it:
date, time, Hz_low, Hz_high, Hz_step, num_samples, dB, dB, dB, ...
Values are dBFS, uncalibrated — a constant offset from true power, so relative comparisons across a sweep are valid but absolute levels are not. An on-bin full-scale sinusoid reads 0.00 dBFS.
The B210 is a direct-conversion (zero-IF) receiver, so LO leakage puts a spike at DC — dead center of every capture. Two mitigations:
- The LO is offset via a UHD tune request, moving the spike off center.
- Only the center
--keepfraction (default 0.75) of each capture is retained, which discards the spike along with the analog filter rolloff at the band edges.
If you see spikes at regular intervals equal to rate x keep, that's residual
LO leakage rather than signal.
Step dwell is --integrate seconds; frames are averaged (Welch) to trade time
for variance. Overflows are counted and reported — if you get them, lower
--rate. USB2 tops out well below USB3, check with lsusb -t.
python3 usrp_power.py -f 860M:955M --simulate -c 5 -o fake.csv
python3 plot_power.py fake.csvSynthesizes noise plus CW tones at 868.3 / 902.5 / 915.0 / 927.0 / 947.4 MHz. If the peak table doesn't report those back at the right frequencies and amplitudes, the DSP/CSV path is broken rather than your radio.
usrp_power.py
-f, --freq sweep range, e.g. 850M:960M
-g, --gain RX gain dB (B210: 0-76, typically 30-50 on a bare whip)
-r, --rate sample rate per step (default 5M; raise only on USB3)
-n, --nfft FFT size (default 1024)
-i, --integrate seconds of IQ averaged per step
-c, --count sweeps; 0 = until Ctrl-C
--keep fraction of each capture retained (default 0.75)
-a, --antenna RX2 or TX/RX
-A, --args UHD device args (default type=b200)
--agc hardware AGC instead of fixed gain
--simulate synthetic data, no radio
plot_power.py
-o, --output output PNG
--region us | eu | none (which band chart to overlay)
--top number of peaks to label
--min-snr dB above floor to count as a signal
--min-sep MHz guard band between reported peaks
--agc makes levels non-comparable across steps, so fixed gain is the right
default for survey work.
Passive spectrum observation is broadly fine. Decoding the content of communications you aren't a party to is not, in most jurisdictions — in the US that's the Wiretap Act and 47 USC 605. Signal identification, power levels, metadata, and your own devices are the safe territory.
MIT