Underwater ship engine sound classification using Tensor Factorized Neural Network

Existing studies primarily utilize spectrograms generated through the short-time Fourier transform (STFT) as features for underwater ship engine sound classification. However, these features are constrained by the parameters defined during the STFT process, potentially restricting the information required for classification. To address this limitation, this paper proposes constructing multi-resolution spectrograms generated using various window sizes into a three-dimensional tensor, which is then classified using a Tensor Factorized Neural Network (TFNN). This three-dimensional tensor combines features across multiple time and frequency resolutions, enabling more efficient learning. Experimental results using a publicly available ship engine sound dataset demonstrate that employing tensor features structured in three dimensions achieves superior classification performance compared to previously used features.