MT-PreamCNN: Multi-Task Preamble Intelligence

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

Dataset: lkk688/ofdm-preamble-intelligence Code: github.com/lkk688/AIsensing

---

Models

Checkpoint	Arch	Training	CoaxSweep Mod	AirLink Mod	AirLink Gate AUC
mt_preamcnn_attn_mixed_phaseaug.pt	Attn	mixed+aug	100%	92.3%	0.9850
mt_preamcnn_cnn_mixed_phaseaug.pt	CNN	mixed+aug	100%	89.6%	0.9847
mt_preamcnn_attn_mixed.pt	Attn	mixed	100%	74.6%	0.9820
mt_preamcnn_cnn_mixed.pt	CNN	mixed	100%	72.0%	0.9826
mt_preamcnn_attn_zeroshot_phaseaug.pt	Attn	zeroshot+aug	100%	61.0%	0.8725
mt_preamcnn_cnn_zeroshot_phaseaug.pt	CNN	zeroshot+aug	100%	62.9%	0.8754
mt_preamcnn_attn_zeroshot.pt	Attn	zeroshot	100%	48.9%	0.8851
mt_preamcnn_cnn_zeroshot.pt	CNN	zeroshot	100%	32.7%	0.8397

ONNX exports (opset 17, dynamic batch) are in onnx/.

Inference Latency

Platform / Backend	CNN	Attn
x86 PyTorch-CPU	5.01 ms	2.43 ms
x86 ORT-CPU (ONNX Runtime)	0.61 ms	0.28 ms
Jetson Orin ORT-CPU	0.62 ms	1.45 ms
Jetson Orin TRT-FP16	0.15 ms	0.25 ms

Usage

import torch, numpy as np
from huggingface_hub import hf_hub_download

# Download best checkpoint
path = hf_hub_download("lkk688/mt-preamcnn",
                        "mt_preamcnn_attn_mixed_phaseaug.pt")

# Rebuild model (copy class definition from train_multitask_v2.py)
ckpt  = torch.load(path, map_location="cpu")
# model = MultiTaskPreamCNNAttn(embed_dim=256, tasks=("gate","mod","snr"))
# model.load_state_dict(ckpt["state_dict"])
# model.eval()

# Input: 1600-float32 vector (800 complex IQ samples, interleaved real/imag)
# x = torch.from_numpy(preamble_float32).unsqueeze(0)   # (1, 1600)
# with torch.no_grad():
#     out = model(x)   # {"gate": ..., "mod": ..., "snr": ...}

Citation

@inproceedings{liu2026multitask,
  title     = {Multi-Task Preamble Intelligence: Joint Packet Detection,
               Modulation Classification, and {SNR} Estimation with
               Phase-Rotation Augmentation for {OTA} Generalisation},
  author    = {Liu, Kaikai},
  booktitle = {Proc. IEEE Global Communications Conference (Globecom)},
  year      = {2026}
}