Localization and detection of underwater acoustic communication signals using convolutional recursive neural networks

Abstract This paper employs a multi-task neural network (UWAELD) to simultaneously perform the Underwater Acoustic Communication Event Detection (UACED) and direction-of-arrival (DOA) mission. Typically, the UACED and DOA tasks are executed separately, as the UACED task is identified through time-frequency analysis, while the DOA task is identified through the amplitude and phase differences of the signal, making it very difficult to jointly perform both tasks. To overcome this, we utilize a logarithmic spectrogram for time-frequency analysis and stack the normalized eigenvectors obtained by processing the signal covariance matrix. We ultimately evaluate the network’s feasibility using simulation data and lake test data, utilizing acoustical signals measured in actual experiments. Finally, for the UADCED task, the error rate is only 6%, with an F1 score of 0.83; for the DOA task, the localization error is only 1.72°, with a localization recall score of 0.87. Additionally, for the DOA task, we employ the classical MUSIC algorithm to demonstrate the network’s superiority.