OzLabs/rf-spectral-trajectories

RF Spectral Trajectories

Temporally ordered sequences of RF (radio frequency) spectrograms from simulated wideband communication environments. Each trajectory is a sliding window of 16 consecutive STFT observations from a continuous RF scene, capturing how multiple signals appear, disappear, drift in frequency, and vary in power over time.

Dataset Structure

Files

File	Split	Trajectories	Size
train.h5	train	13,841	10 GB
val.h5	validation	2,938	1.1 GB
test.h5	test	2,999	1.2 GB

HDF5 Schema

Each file contains:

Key	Shape	Dtype	Description
observations	[N, 16, 256, 51, 2]	float16	STFT magnitude (real, imaginary channels)
timestamps	[N, 16]	float64	Time in seconds for each timestep
source_ids	[N]	string	Scene identifier (for provenance)
sequence_ids	[N]	string	Unique trajectory identifier

Observation tensor dimensions:

• N — number of trajectories
• 16 — timesteps per trajectory (80ms each, 1.28s total)
• 256 — frequency bins (STFT, covering ±75 kHz)
• 51 — time bins within each 80ms STFT window
• 2 — real and imaginary components

Splits

Split by source scene (not by trajectory) to prevent temporal leakage from overlapping sliding windows.

Split	Scenes	Ratio
train	280	70%
validation	60	15%
test	60	15%

Loading

With HuggingFace datasets

from datasets import load_dataset
import numpy as np

ds = load_dataset("ozlabs/rf-spectral-trajectories", split="train")

# Metadata columns are directly accessible
print(ds[0]["regime"])        # e.g. "bursty"
print(ds[0]["snr_db"])        # e.g. 14
print(ds[0]["source_id"])     # e.g. "scene_0042"

# Decode observation tensor from binary
obs = np.frombuffer(ds[0]["observations"], dtype=np.float16).reshape(16, 256, 51, 2)

With h5py (for direct HDF5 access)

import h5py

with h5py.File("train.h5", "r") as f:
    obs = f["observations"][0]      # [16, 256, 51, 2] float16
    ts = f["timestamps"][0]         # [16] float64
    src = f["source_ids"][0]        # bytes

With PyTorch

from lewm_pipeline.dataset import LeWMDataset

ds = LeWMDataset("train.h5")
item = ds[0]
# item["observations"]: torch.float32 tensor [16, 256, 51, 2]
# item["timestamps"]: torch.float64 tensor [16]
# item["source_id"]: str
# item["sequence_id"]: str

Generation Parameters

Signal Simulation

Parameter	Value
Sample rate	150 kHz
Timestep duration	80 ms (12,000 samples)
Scene duration	5.2 s (65 timesteps)
Trajectory length	16 timesteps (1.28 s)
Sliding window stride	1 timestep
Modulation types	BPSK, QPSK, 8PSK, 16QAM, 64QAM
Signals per scene	2–4 (randomly placed in frequency)
Channel models	Rayleigh, Rician (with evolving fading)
SNR range	-8 to +30 dB
Doppler speeds	0–12 m/s

STFT Parameters

Parameter	Value
Window	Hamming, 256 samples
Overlap	16 samples
FFT size	256
Sided	Two-sided (complex input)

Activity Regimes

Each scene follows one of 8 activity regimes (50 scenes each, 400 total):

Regime	Description
quiet	Few active signals, low duty cycle, long silences
dense	Many signals active simultaneously, high overlap
bursty	Rapid on/off transitions, short bursts
ramp_up	Signals appear progressively through the scene
interference_event	Stable signals, then a disruptive signal appears mid-scene
correlated_alternating	Signal pairs alternate: A on when B is off
correlated_leader_follower	Signal B appears 1–3 timesteps after signal A
random	Independent random burst patterns

Dynamic Features

• Frequency drift: Signals slowly wander in frequency (random walk, bounded to ±50% of signal bandwidth)
• Power variation: Smooth per-timestep power levels (0.3–1.0 when active), gradual fade-in/out, correlated power drift
• Channel fading: Rayleigh/Rician fading evolves continuously within each scene (no state reset between timesteps)

Source

Generated using the ChangShuoRadioData (CSRD) MATLAB simulation framework with the LeWM dataset pipeline.

Citation

@software{csrd_rf_spectral_trajectories_2026,
  title = {RF Spectral Trajectories},
  author = {Ozlabs},
  year = {2026},
  url = {https://huggingface.co/datasets/ozlabs/rf-spectral-trajectories}
}

Related

• ChangShuoRadioData — MATLAB RF simulation framework
• Joint Signal Detection and AMC via Deep Learning — IEEE TWC paper using CSRD data