Deep-learning based multi-node channel quality estimation for underwater acoustic communications

In multi-node underwater acoustic (UWA) communications using autonomous underwater vehicles, efficient resource allocation is required due to the limited available bandwidth. However, obtaining UWA channels for resource allocation is challenging due to the long propagation delay of acoustic waves, which makes it difficult to collect feedback from multiple nodes. To address this issue, this paper proposes a deep learning-based channel quality estimator that simultaneously estimates the UWA channel qualities of multiple nodes at the transmitter by utilizing environmental information without feedback. The proposed transformer-based multi-node channel quality estimation network (MCENet) consists of a transformer-based feature extraction layer (FEL) and a fully connected layer-based channel quality estimation layer. The channel quality estimation and BER performance of the proposed transformer-based MCENet were evaluated using real-world data collected from sea experiments off the South Sea of Korea, and compared with MCENet models applying LSTM, GRU, and attention layers to the FELs and the conventional channel response generative adversarial network (CRGAN). The evaluation results demonstrated that the proposed transformer-based MCENet achieved superior performance compared to the conventional CRGAN. Specifically, it reduced the channel quality estimation mean square error by 95.7% and the bit error rate by 93.8%, while requiring merely 8.5% of the trainable parameters compared to the conventional CRGAN.