Coarse-to-Fine Doppler Compensation in 6G VLEO: Dual-Head Temporal Convolutional Network Acquisition with CP-Aided Tracking

초록

The evolution toward 6G wireless networks has positioned Non-Terrestrial Networks (NTN) as critical enablers for ubiquitous global connectivity. Among NTN architectures, Very Low Earth Orbit (VLEO) satellites operating in Ka-band frequencies offer superior link budgets and reduced latency. However, extreme orbital velocities reaching 7.7 km/s introduce severe carrier frequency offset (CFO) challenges spanning multiple subcarrier intervals far beyond conventional estimation capabilities. Unlike terrestrial systems where CFO remains within fractional subcarrier bounds, VLEO systems experience massive Doppler shifts that degrades OFDMA subcarrier orthogonality, transforming carefully designed waveforms into interference-dominated signals. Existing approaches exhibit significant performance degradation because they treat integer and fractional CFO estimation as independent sequential problems, ignoring their fundamental coupling and creating error propagation leading to sustained synchronization failures. This paper presents a dual-head Temporal Convolutional Network (TCN) architecture that addresses the coupled integer-fractional CFO estimation problem through unified machine learning. Our two-stage framework employs Stage 01 for wide-range coarse acquisition using dual-head TCN to jointly estimate integer and fractional components, while Stage 02 provides precision tracking through optimized phase-locked loop techniques for small residual offsets. Comprehensive simulations across extreme VLEO scenarios for both ground and aerial users demonstrate strong system-level performance: 97.7% integer classification accuracy, 0.1229 fractional estimation RMSE, and bit error rate within 0.1 dB of ideal compensation, showing a substantial performance improvement over existing methods that suffer system performance degradation under high-Doppler conditions. Furthermore, backbone comparisons against CNN, LSTM, and Transformer architectures confirm the TCN’s superior accuracy-complexity trade-off, and ablation experiments verify the necessity of each architectural component. © 1972-2012 IEEE.

키워드