Ship-radiated noise feature extraction using multiple kernel graph embedding and auditory model

The analysis of underwater acoustic signals, especially ship-radiated noise received by passive sonar, is of great importance in the fields of defense, military, and scientific research. In this paper, we investigate multiple kernel learning graph embedding using auditory model features in the application of ship-radiated noise feature extraction. We use an auditory model to get auditory model features for each signal sample. In order to have more effective features, iterative multiple kernel learning methods are adopted to conduct dimensionality reduction. Validated by experiments, the proposed method outperforms ordinary kernel-based graph embedding methods. The experiments show that the multiple kernel learning method can automatically choose relatively appropriate kernel combinations in dimensionality reduction for ship-radiated noise using auditory model features. In addition, some worthwhile conclusions can be drawn from our experiments and analysis.