Reliable start-of-packet (SOP) detection is essential for packet-based wireless receivers, as boundary errors can affect subsequent synchronization and symbol detection. The problem is particularly challenging in multiple-input singleoutput (MISO) systems, where independently faded signal replicas, phase offsets, Doppler shifts, propagation effects, and noise distort the received preamble. Conventional correlation-based detectors are sensitive to these impairments, while existing deep learning (DL) methods primarily exploit local signal features. To address this issue, this paper proposes SSPDet, a lightweight LSTM-based detector that formulates SOP detection as sequential binary classification. SSPDet processes oversampled in-phase and quadrature samples using overlapping windows to learn the transition from the noise-only region to the packet preamble, followed by a probability-based SOP decision. Simulation results show that SSPDet outperforms the conventional and considered DL-based baselines in Pd, Pfa, DER, and MAE, while requiring ≈ 0.96M FLOPs and 5.631μs detection time. Over-the-air MISO experiments further validate its practical effectiveness through improved BER performance.

wireless communications, packet detection, deep learning, MISO, SDR