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

Orthogonal time frequency space (OTFS) modulation offers reliable communication in time- and frequency-selective channels and is a promising technique in underwater acoustic (UWA) communication. This letter proposes a deep learning receiver for UWA OTFS communication that addresses the Doppler squint effect (DSE). Conventional UWA OTFS systems tend to ignore DSE, which leads to significant performance degradation. The proposed receiver stacked convolutional neural network (CNN) and ResNet as an receiver, which is abbreviated as S-CNN-ResNet. To provide additional effective features to the network, first, the pilot data is processed using a CNN to obtain effective channel features, and then channel features are stacked together with the received data. Finally, the stacked data is input into the improved ResNet to recover the transmitted symbols. Specifically, to deal with the complex DSE UWA channel, the scheme employs the idea of data augmentation technique, utilizing CNN to capture the effective channel features from pilot data and then stack these with the received data. This approach expands the training features of the network input data, enhancing the network’s learning capability. Subsequently, the ResNet focuses on the effective information of the input data through the residual structure, thereby facilitating accurate symbol recovery. Simulation results demonstrate that the proposed method achieves a lower bit error rate (BER) compared to single network receivers, cascaded network receivers, and classic algorithm-based receivers. Additionally, it strikes an effective balance between complexity and performance.