Convolutional neural network based filter bank multicarrier system for underwater acoustic communications

This paper presents a convolutional neural network (CNN) based filter bank multicarrier (FBMC) system for underwater acoustic (UWA) communications. Different from traditional FBMC receivers that only detect the transmitted symbols after expliciting channel estimation and equalization, the proposed system takes a pre-trained CNN model as a receiver to recover the transmitted symbols directly and avoid the inherent imaginary interference. At the offline training stage, the CNN model takes the known transmitted data as the labels and the received data as input for iterative learning. At the testing stage, unlabeled received data are directly output to CNN to estimate the online transmitted symbols. The CNN based UWA FBMC system is analysed under various system parameters such as input permutations, convolution kernels and prototype filters, and sufficient communication data are generated according to the measured UWA channel impulse responses. Simulation results show that the proposed system achieves admirable performance for signal detection compared to the previous UWA FBMC system.