Integrated optimization of underwater acoustic ship-radiated noise recognition based on two-dimensional feature fusion

Abstract(#br)Feature fusion methods are introduced to ship-radiated noise recognition in this paper. Wavelet packet (WP) decomposition is used to decompose the ship-radiated noise into multiple different subbands. By considering the features extracted from the different subbands reflecting different characteristics of the ship-radiated noise, a two-dimensional feature fusion (2DFF) scheme is proposed to fuse the features extracted from the different subbands. Principal component analysis (PCA) and canonical correlation analysis (CCA) are used in the 2DFF scheme. Then, a so-called discriminative ability improving (DAI) strategy is proposed to improve the discriminative ability of the extracted features. Starting at the 2DFF, a processing chain of feature fusion and ship-radiated noise recognition is designed and jointly optimized to the task. The 2DFF scheme and DAI strategy are tested on real ship-radiated noise data recorded. Experimental results indicate that compared with the baseline, the 2DFF scheme can improve 7.25% of recognition accuracy. Experimental results also show that the DAI strategy can further improve the recognition accuracy of 13.10%