Underwater Acoustic Target Classification with Joint Learning Framework and Data Augmentation

As speech recognition technologies have garnered increasing attention recently, the possibility of integrating related technologies into the field of underwater acoustic target classification (ATC) has piqued researchers’ curiosity. However, the classification of underwater acoustic targets confronts several obstacles due to the time-varying and complicated nature of the underwater environment,, such as the low signal-to-noise ratio and background interference. To address these concerns, this paper proposes a combined feature extraction framework and a deep convolutional neural network learning framework. This combination features include MFCC, CQT, Gammatone, Log-mel which are extracted from the underwater acoustic signal and fed into the joint learning framework. The classifier has a deep learning architecture, which is a deep neural network system that can learn differentiating features from physical feature representations to abstract concepts in a hierarchical pattern. Multiple datasets under diverse settings have been used to verify the performance of our proposed architecture. Using the acoustic data gathered in several underwater experiments conducted in a shallow sea setting, the proposed framework achieves a recognition rate of 89.9% and a competitive classification accuracy in the underwater scenario.