Multi-Task Learning Based Underwater Acoustic OFDM Communications

Deep learning (DL) based receivers are trained under simulation dataset and tested in the real world. Unfortunately, there is no accurate channel model in underwater acoustic (UWA) communications. The mismatch between the training data set and the test environment degrades the receiver performance for UWA communications. Considering that the classical receiver contains channel estimation, equalization and demodulation tasks that are mutually dependent, in this paper, we propose a novel receiver, called TaskNet, based on multi-task learning (MTL) to improve the DL-receiver's generalization performance. To demonstrate the advantages of TaskNet, we establish a baseband UWA channel impulse response (CIR) dataset, which includes both simulated and measured CIRs, and make comparison with the classical single-task based DL-receivers inlcuding FC-DNN and ComNet. Through analysis and simulation results, the proposed MTL based DL-receiver is proved to have a better generalization performance than the single-task based ones, whether the signal suffers nonlinear damage, or the training set does not match the real channel.