BTM Interference Suppression Based on Multi-Scale Linear Attention

• 1、School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：High-speed railway BTM balise message signals are susceptible to non-stationary aliasing interference, such as traction transient pulses, rail radiation, and in-band harmonics, which can lead to message loss. To address the insufficient utilization of balise message signal features in existing methods, this paper proposes a Multi-Stage Hybrid Downsampling and Multi-Scale Linear Attention Network (MSHD-MLANet), which adopts an encoder-bottleneck-decoder architecture. The encoder compresses 2FSK symbol redundancy through hybrid downsampling while preserving waveform details. The bottleneck layer employs parallel multi-scale dilated convolutions and linear attention to extract long-term temporal dependencies with linear complexity. The decoder combines sub-pixel convolution with skip connections to achieve high-precision waveform reconstruction. Experimental results show that, under the experimental conditions in this paper, the proposed model improves the precision of BTM message signals under strong interference from 87.615% to 99.717% and increases the signal-to-noise ratio by 18.11 dB. Ablation experiments verify the effectiveness of each module. Comparison experiments further demonstrate that the proposed method can effectively suppress complex aliasing interference in BTM message signals while balancing computational cost and inference time.

Keywords： Signal and Information Processing Balise Deep Learning Interference Suppression