Effective underwater acoustic target passive localization of using a multi-task learning model with attention mechanism: Analysis and comparison under real sea trial datasets
Abstract
Sound source ranging (SSR) and depth estimation are fundamental tasks in passive localization, aiming to determine target locations using passive sonar. In the past, most neural network models only focus on the SSR. The importance of depth estimation is ignored. Recently, multi -task learning (MTL) models have been used for SSR and depth estimation. These tasks share layers to utilize task knowledge, however, communication between tasks and adjacent parameters is still lacking. Moreover, MTL overlooks cross -hierarchical relationships between range, depth, and the environment. The data -driven learning process lacks interpretability, limiting performance due to the predictions not adhering to physics information. Thus, we present MLF-TransCNN, a hybrid datamodel driven multi -task learning model combining Transformer and convolutional neural network (CNN) with a multi -level dilated convolutional fusion (MLF) module, for passive localization. Compared with the experiments performed on real -world and simulated datasets by the state-of-the-art methods, the MLF-TransCNN achieving the best range and depth mean absolute errors of 0.0834 and 0.2718, respectively. Ablation studies confirmed that the performance gains result from our designed modules.