Hierarchical feature extraction of intrinsic modes for ship-radiated noise classification

The complexity of marine acoustic environment, including background noise interference and variability in underwater transmission paths, presents significant challenges for accurately identifying ship-radiated noise (S-RN). To enhance the adaptability and precision of S-RN recognition, we propose a novel S-RN identification system that utilizes an improved empirical mode decomposition (EMD) algorithm combined with a hierarchical intrinsic mode function (IMF) selection and feature fusion approach. This system leverages the adaptive decomposition capabilities of the improved EMD algorithm to decompose original signals into a set of IMFs. It then selects those IMFs with high discriminative power based on the characteristics of the sample data, constructing a hierarchical feature extraction framework. Entropy and energy features are extracted separately from different IMFs to capture the diversity of underwater signals. The entropy features reveal the complexity and dynamic characteristics of high-frequency IMFs, while the energy intensity reflects the amplitude information in lower-frequency modes. By comparing various feature fusion strategies, the complementary features are optimally combined to enhance the discriminatory power for S-RN classification. The generated feature set is fed into different classifiers, and the classification accuracy and computational efficiency are evaluated based on the DeepShip and ShipsEar datasets. The experimental results demonstrate that the random forest (RF) model achieves a superior balance between the performance and efficiency, displaying high classification accuracy and rapid computation, and validating the potential of this method for practical signals recognition applications.