MDCEN: Feedback-Free Multidistance Channel Estimation for Underwater Sensor Networks

초록

The advancement of underwater acoustic (UWA) communication holds great promise for the Internet of Underwater Things (IoUT). Accurate estimation of the channel impulse response (CIR) between underwater sensor nodes is crucial for optimizing the link performance of UWA communications. Due to the long propagation delay of acoustic waves, it is challenging to obtain UWA CIRs from multiple sensor nodes located at different distances through feedback. This study proposes deep-learning-based multidistance UWA CIR estimation methods at the transmitter without feedback. The proposed multidistance CIR estimation network (MDCEN) is designed by integrating a feature sharing module (FSM) based on parallel gated recurrent unit (GRU) and a CIR estimation layer (CEL) based on fully connected layers (FCLs) and applies a weighted fusion structure that learns the physical characteristics of sound speed profile (SSP) and seabed topography, and the positions and depths of the sensor nodes separately, and integrates them into a unified representation for CIR estimation. Also, the proposed MDCEN reflects the physical continuity and spatial overlap properties of the underwater environment. A distance-ordered, progressive accumulation approach is adopted to reduce complexity while improving the accuracy of multidistance CIR estimation. The convergence speed, CIR estimation performance, and computational complexity of the proposed MDCEN were evaluated using real-world data collected from the southern coast of Korea. Comparative evaluations were conducted against MDCEN variants that apply an FCL, a 1D-Conv layer, long short-term memory (LSTM)-based FSMs, and the conventional separate bidirectional feature pyramid network CIR estimation model. As a result of the evaluation, the proposed MDCEN shows a fast convergence speed, reducing the normalized mean square error (NMSE) of CIR estimation by up to 94.8%, increasing delay match accuracy (DMA) by an average of 24.7 percentage points compared to the conventional methods. In addition, the number of parameters was reduced to 1.6%, demonstrating the proposed method's superior performance.

키워드