A Machine Learning Label-Free Method for Underwater Acoustic OFDM Channel Estimations

In this paper, a machine learning label-free scheme is proposed for underwater acoustic (UWA) orthogonal frequency division multiplexing (OFDM) channel estimation, which avoids the necessity of the real UWA channel label as in the traditional training process. To this end, a label-free loss function is developed, based on which the training process requires only the received pilot symbols without true channel information. The experiments indicate that, with sufficient training, the proposed label-free network can perform a near-optimal channel estimation.