A Deep Learning Receiver for Underwater Acoustic OTFS Communications with Doppler Squint Effect

A deep learning (DL)--based detector is proposed for underwater acoustic (UWA) communication systems using orthogonal chirp division multiplexing with index modulation (OCDM-IM). The proposed high-performance and lightweight network integrates the detection of the index bits and the carrier bits as a whole, employing a squeeze-and-excitation (SE) mechanism enhanced residual neural network (ResNet) cascaded with a bidirectional gated recurrent unit (BiGRU) to detect OCDM-IM signals. For the accuracy of data recovery, the SE mechanism in the proposed network can focus more on the useful information of OCDM-IM signals to improve signal detection performance. For the complexity of the network, the lightweight ResNet with depthwise separable convolution and BiGRU with a simple structure can significantly reduce network parameters. Simulation results demonstrate that the SE-ResNet-BiGRU has a lower bit error rate (BER) and fewer parameters than the previously proposed DL-based detectors. Moreover, the proposed network exhibits strong robustness in the presence of carrier frequency offset.