Robust Demodulation of eLoran Tri-State PPM via Multi-Scale Spatiotemporal Modeling

Enhanced Loran (eLoran) provides a terrestrial complement to Global Navigation Satellite Systems (GNSSs), but tri-state pulse position modulation (PPM) demodulation remains difficult under composite interference. This study proposes a label-independent Nominal-Center (NC) sampling strategy and a Multi-Scale Spatiotemporal Network (MSST-Net). NC extracts each pulse around its predicted nominal timing anchor, thereby avoiding target-dependent actual-center alignment while preserving the displacement that encodes the PPM state. MSST-Net combines multi-scale one-dimensional convolutions, Squeeze-and-Excitation channel recalibration, bidirectional long short-term memory, temporal attention, and a distance-aware mixed ordinal loss. Under a predefined controlled simulation protocol with a group- and frame-disjoint partition, the seed-42 model was evaluated on 61,008 held-out channel-realized windows generated from 372 test pulse identities belonging to 62 pulse groups across 164 nominal SNR–Stage configurations. It achieved 93.71% accuracy, 93.85% balanced accuracy, and 93.81% macro-F1. Under a separate parameter-matched cross-entropy protocol, MSST-Net (CE) achieved 92.56% accuracy, versus 91.23% for the Transformer, 89.34% for CNN-BiLSTM, and 88.29% for ResNet1D. A label-free diagnostic on 1440 receiver-recorded windows indicated a remaining simulation-to-real gap; without independently verified tri-state labels, no real-waveform accuracy was claimed. Thus, the supervised results apply to the controlled simulated benchmark and do not establish operational receiver accuracy.