OCEAN-Lab/SS-Sine-Data

Photonic Time-Stretch ADC Dataset for Equalization and Denoising

Dataset Summary

This dataset provides 126,000 preprocessed captures (~1.4 TB) from a photonic time-stretch (PTS) analog-to-digital converter (ADC) testbed, collected across a systematic sweep of 56 RF frequencies and 15 power levels, with 150 repetitions per sweep point. Each data point contains a paired distorted measurement and undistorted clean target, along with an unmodulated baseline (calibration) window, enabling supervised training and benchmarking of machine learning models for PTS-ADC signal equalization, denoising, and spectrum sensing.

PTS-ADC front ends extend the usable digitization bandwidth of electronic ADCs well beyond their native limits by stretching RF signals in time using a chirped mode-locked laser pulse and a dispersive optical medium, before photodetection and electronic digitization. This architecture introduces its own nonlinear distortions (modulator nonlinearity, Gaussian pulse envelope effects, photodetector saturation), which this dataset is designed to help correct via machine learning.

Supported Tasks

• Signal equalization: train models to recover the clean target waveform from the distorted measurement.
• Denoising: use the unmodulated calibration windows as paired baseline-noise segments, independent of equalization.
• Spectrum sensing: PTS-ADC front ends are central to wideband spectrum sensing in radar and wireless systems; this dataset supports training and evaluating ML-based spectrum sensing methods directly, across a swept frequency-power grid.

Dataset Structure

Sweep Grid

Parameter	Value
RF frequency	0.5 – 6 GHz (56 values)
Power level (at MZM)	-14.7737 to 13.2262 dBm (15 values)
Repetitions per sweep point	150
Total data points	126,000 (56 × 15 × 150)

The power level reported is the RF power delivered to the modulator (MZM), computed by applying the RF chain gain (RF amplifier, cable losses, and a fixed attenuator) to the power programmed at the AWG. Two attenuator configurations were used: a 30 dB attenuator (net chain gain 15.2263 dB) and a 20 dB attenuator (net chain gain 25.2263 dB). Files collected with the 20 dB attenuator are marked high_power in the filename.

Data Fields

Each repetition is stored as an HDF5 group containing:

Field	Description
noisy_mod	Distorted RF capture after the PTS path (the signal to be corrected)
clean_mod	Reconstructed undistorted ground-truth target, sample-aligned to noisy_mod
calib_unmod	Unmodulated (gap) calibration window, shared per repetition
rf_phase	Reference phase of the window, extracted via the Hilbert transform of the clean target

File Organization

Data is sharded into one HDF5 file per (frequency, power) sweep point, with repetitions stored as internal groups within each file. This keeps individual files at a manageable size and allows any single (frequency, power, repetition) data point to be accessed directly without downloading the full dataset.

shards_dir/
  {stem}_point{NNNN}_f{frequency:.4g}GHz_p{power:+.1f}dBm[_high_power].h5
    attrs: power_dbm, frequency_ghz, source_point_path,
           pulses_per_window, pulse_samples
    /rep_000/
      ├── calib_unmod   1D float32 (calib_len,)
      ├── clean_mod     2D float32 (n_windows, win_len)
      ├── noisy_mod     2D float32 (n_windows, win_len)
      └── rf_phase      1D float32 (n_windows,)
    /rep_001/...
    ...

power_dbm stored in the file attrs is the AWG-programmed power; apply the relevant net chain gain (see Sweep Grid above, or use --attenuator_db/--gain_db in the loading utility) to recover the actual power delivered to the MZM.

Data Collection

Raw waveforms were captured on four synchronized oscilloscope channels sampled at 16 GSa/s:

Channel	Role
Data (ch1)	RF signal after the PTS path — the distorted capture
Reference (ch2)	Same RF content without PTS-path distortion — source of the clean target
Marker (ch3)	AWG trigger output, delimiting modulated vs. unmodulated intervals
MLL (ch4)	Mode-locked-laser sync-pulse train, marking each stretch window

Raw captures were aligned, windowed, and processed into the released (noisy_mod, clean_mod, calib_unmod, rf_phase) format as described in the accompanying paper (citation below). Full preprocessing methodology is detailed there.

Loading and Visualizing the Data

A lightweight, open-source Python utility (visualize_SS_Sine_data_shards.py, MIT license) for loading and visualizing any data point by (frequency, power, repetition) is included in this repository.

Requirements

pip install h5py numpy matplotlib

Usage

List all available (frequency, power) sweep points in a shard directory:

python visualize_SS_Sine_data_shards.py --shards_dir /path/to/shards --list

Load and plot a specific data point:

python visualize_SS_Sine_data_shards.py --shards_dir /path/to/shards \
    --frequency 3.5 --power -10.0 --repetition 0 --window 0

To display the actual RF power delivered to the modulator (rather than the AWG-programmed value), pass the attenuator configuration used for that capture:

python visualize_SS_Sine_data_shards.py --shards_dir /path/to/shards \
    --frequency 3.5 --power -10.0 --attenuator_db 30

This plots three stacked subplots: the distorted capture (noisy_mod), the clean reconstructed target (clean_mod), and the shared per-repetition calibration block (calib_unmod).

The utility can also be used as a library:

from visualize_SS_Sine_data_shards import load_point, plot_point
point = load_point("/path/to/shards", frequency=3.5, power=-10.0, repetition=0)
plot_point(point, window=0)

Licensing

This dataset is released under the Creative Commons Attribution 4.0 (CC-BY-4.0) license. You are free to use, share, and adapt this data for any purpose, including commercially, provided appropriate credit is given.

Dataset Card Contact

For questions about this dataset, please open a discussion on this Hugging Face dataset repository rather than contacting the authors directly.