Deep Learning based Framework for Underwater Acoustic Signal Recognition and Classification

The recognition and classification are important to the research contents like underwater acoustic signal processing. Generally, the state-of-the-art signal recognition systems depend on feature extraction that is based on the knowledge or experience of experts, so that, it can efficiently represent target signatures. In contrast, this paper introduces a new framework based on deep learning methods to preprocess signals and extract features for the recognition and classification of underwater acoustic signals, which present the targets of ships and torpedoes. In our framework, signals are firstly transformed to spectral images of LOFAR, so that, those images can be extracted and classified by convolutional neural networks (CNN). The experimental results show that our deep learning based framework can obtain high classification accuracy in underwater acoustic signals case with the transformation to LOFAR spectrum. The accuracy of our best version reaches 97.22%, higher than those that use other networks, and achieved the expected objectives for real applications.