基于旋转位置编码与多尺度自适应融合的水声信道均衡方法OA

To address the severe inter-symbol interference induced by multipath effects in underwater acoustic(UWA)channels,a rotary multi-scale adaptive fusion network(RoMAFNet)for underwater acoustic channel equalization was proposed.Inter-symbol relative time delays were multiplicatively embedded by rotary position embedding(RoPE)into the attention computation of the Transformer encoder,whereby the time-varying adaptability was improved.In a multi-scale adaptive multipath equalization module,multipath components at different delay scales were characterized by paral-lel branches with increasing dilation rates and dynamically weighted by a global context-driven gating mechanism.Fur-thermore,a closed-form error bound of the fusion mechanism was derived under Lipschitz smoothness and related as-sumptions,based on which the online fine-tuning stability was analyzed.Under an"offline pre-training and online fine-tuning"paradigm,per-frame updates with about 50 training symbols sufficed to adapt to new simulated channels,and on measured at-sea channels,training symbols were incrementally utilized significantly more efficiently than in the compari-son methods.