Resumen
In shallow water, passive sonar usually has great difficulty in discriminating a surface acoustic source from an underwater one. To solve this problem, a supervised machine learning method using only one hydrophone is implemented in this paper. Firstly, simulated training data are generated by a normal mode model KRAKEN with the same environment setup as that in SACLANT 1993 experiment. Secondly, the k-nearest neighbor (kNN) classifiers are trained and evaluated using the scores of precision, recall, F1 and accuracy. Thirdly, the random subspace kNN classifiers are finely trained on three hyperparameters (the number of nearest neighbors, the number of predictors selected at random and the number of learners in the ensemble) to obtain the best model. Fourthly, a deep learning method called ResNet-18 is also applied, and it reaches the best balance between precision and recall, while the accuracies of both simulation and experimental data are all 1.0. Further, data from all 48 hydrophones of the vertical linear array (VLA) are analyzed using the three kinds of machine learning methods (kNN, random subspace kNN and ResNet-18) separately, and the results are compared. It is concluded that the performance of random subspace kNN is the best. Both the simulation and experimental results suggest the feasibility of machine learning as a surface and underwater acoustic source discrimination method even with only a single hydrophone.