lkk688/ofdm-preamble-intelligence

OFDM Preamble Intelligence Dataset

Paper: "Multi-Task Preamble Intelligence: Joint Packet Detection, Modulation Classification, and SNR Estimation with Phase-Rotation Augmentation for OTA Generalisation" — IEEE Globecom 2026 (submitted)

Code: github.com/lkk688/AIsensing

---

Dataset Overview

This dataset contains raw IQ preamble captures from a PlutoSDR OFDM testbed operating at 2400 MHz, collected across two experimental domains:

Domain	Description	NPZ files
CoaxSweep	Cable (coaxial) link, gain sweep 25–72 dB	8,605
AirLink	Over-the-air (OTA) at ~40–50 cm, 2.4 GHz WiFi environment	8,864

Total: ~17,469 captures, ~31 GB raw IQ data

---

Experimental Setup

• TX: NVIDIA Jetson Orin running rf_stream_tx_step6phy.py (PlutoSDR USB)
• RX: x86 host running rf_stream_rx_step10phy.py (PlutoSDR USB)
• Carrier frequency: 2400 MHz
• Sample rate: 3 MHz
• FFT size: 64 subcarriers, CP length 16
• Preamble: 800 samples (STF + LTF), Schmidl–Cox CFO estimation
• Modulations: QPSK, QAM16 (alternating per packet)
• TX gain: swept 25–72 dB (CoaxSweep), fixed 72 dB (AirLink)
• RX equalization: MMSE with pilot_weight=0.5, auto threshold for QAM16

---

Sub-datasets

CoaxSweep (coaxial cable baseline)

Collected under controlled coaxial cable conditions with known SNR ground truth. CFO: stable +330 Hz (Pluto TCXO offset). BER=0 on all decoded packets.

Sub-split	Directory	Runs	NPZ	Notes
CoaxSweep-I	coax/ber_sweep_v2/	6	1,403	Step8 PHY, QPSK only
CoaxSweep-II	coax/ber_sweep_v3/	6	3,282	Step9 PHY, QPSK + QAM16
CoaxSweep-III	coax/ber_sweep_v4/	3	3,078	Step10 MMSE, pilot_weight sweep
CoaxSweep-IV	coax/ber_sweep_v6/	4	842	QAM16 mixed, hard-negative gate examples

AirLink (over-the-air)

Collected at 2400 MHz in a 2.4 GHz WiFi-congested indoor environment. CFO: −3000 Hz (sign reversal vs. CoaxSweep — the domain shift studied in the paper). WiFi-triggered fail captures serve as hard-negative examples for the gate head.

Sub-split	Directory	Runs	NPZ	Notes
AirLink-QPSK-I	airlink/airlink_qpsk/	1	768	First OTA collection
AirLink-QPSK-II	airlink/airlink2_qpsk/	2	753	Panel antenna, repeat
AirLink-QAM16	airlink/airlink_qam16/	1	807	OTA QAM16, auto_z_th
AirLink-QPSK-III	airlink/airlink3_qpsk/	7	6,536	50 cm, WiFi env, large collection

---

NPZ File Format

Each .npz file is one packet capture attempt. Keys:

Key	Shape	Dtype	Description
rxw	(262144,)	complex64	Full receive window (IQ samples at 3 MHz)
corr	(N,)	float32	STF correlation magnitude
corr_norm	(N,)	float32	Normalized correlation
topk_idx	(K,)	int32	Top-K preamble candidate positions
topk_corr_norm	(K,)	float32	Correlation at top-K positions
topk_ncc	(K,)	float32	NCC scores at top-K positions
energy_ds	(N,)	float32	Downsampled energy envelope
meta_json	scalar	str	JSON string with metadata

meta_json fields

{
  "stf_idx": 12345,       // preamble start sample index
  "snr_db": 22.4,         // estimated SNR (dB), from pilot comparison
  "mod": "QPSK",          // "QPSK" or "QAM16"
  "gate": 1,              // 1 = CRC pass, 0 = fail
  "gain_tx": 72,          // TX gain (dB)
  "gain_rx": 40,          // RX gain (dB)
  "cfo_hz": -3012.5,      // estimated CFO (Hz)
  "timestamp": "2026-05-03T12:18:05"
}

Filename convention: cap_NNNNNN_ok.npz (CRC pass) or cap_NNNNNN_bg.npz (fail / background).

---

Preamble Extraction

The training pipeline extracts 800 complex samples starting at stf_idx and applies Schmidl–Cox CFO correction before converting to real/imag interleaved float32 (1600 values):

import numpy as np

def extract_preamble(npz_path, pream_len=800):
    with np.load(npz_path, allow_pickle=True) as d:
        meta = json.loads(d["meta_json"].item())
        rxw  = d["rxw"].astype(np.complex64)
    seg = rxw[meta["stf_idx"] : meta["stf_idx"] + pream_len]
    # CFO correction (Schmidl-Cox)
    L = 32  # STF half-period
    r = np.dot(seg[:L*(len(seg)//L-1)].conj(), seg[L:])
    cfo_hz = np.angle(r) / (2*np.pi * L / 3e6)
    t   = np.arange(len(seg), dtype=np.float32)
    seg = (seg * np.exp(-1j * 2*np.pi * cfo_hz * t / 3e6)).astype(np.complex64)
    out = np.empty(pream_len * 2, dtype=np.float32)
    out[0::2] = seg.real
    out[1::2] = seg.imag
    return out  # shape (1600,)

---

Models

Pre-trained model checkpoints are available on the companion model repository: lkk688/mt-preamcnn

Model	Params	CoaxSweep Mod	AirLink Mod	AirLink Gate AUC
MT-PreamCNN (CNN, mixed+aug)	362K	100%	89.6%	0.9847
MT-PreamCNN-Attn (mixed+aug)	428K	100%	92.3%	0.9850

---

Citation

---

License

Dataset: CC BY 4.0 Code: MIT License