关键词: 分布式网络, 多终端, 零陷优化, 集成学习

Abstract:

In recent years, distributed cooperative communication systems have shown great potential due to their small size, lightweight, low power consumption, and cost efficiency, thus expanding their range of applications for various tasks. In multi-terminal scenarios, signal interference from multiple terminals degrades the demodulation performance of distributed cooperative communication. To overcome these limitations, an ensemble learning combining convolutional neural network-bidirectional long short-term memory （CNN-BiLSTM） and Transformer architectures is proposed to predict the weights for zero-steering optimization, with the aim of eliminating interference caused by the overlap of signals with multiple terminals and achieving the final demodulation. The CNNBiLSTM network focuses on local linear problems and deals with linear phase variations caused by residual frequency deviations, while the Transformer network focuses on global nonlinear problems and deals with nonlinear phase variations caused by mobility. Both networks have memory capacities that improve the accuracy of continuous prediction of zero-steering weights. Compared to convex optimization, which only provides local optima, the neural network can achieve a near-global optimal solution. The simulation results show that the proposed zero-steering algorithm based on optimized ensemble learning can accurately predict zero-steering weights and effectively suppress mutual interference, with the signal gain difference between different terminals reaching 45 dB. Compared with the zero-steering method based on the linearly constrained minimum variance criterion, it significantly improves the demodulation performance of overlapping signals from multiple targets.

Key words: distributed network, multi-terminal, zero-steering optimization, ensemble learning