RCMGGDFuRDE features and BMBT LSTSVM classifier and their application in underwater radiated noise recognition

Feature extraction and recognition are key steps in underwater radiation noise (URN) recognition, which are of great significance for various underwater monitoring tasks. However, existing MFuDE suffers from coarse-grained detail loss and entropy instability issues, and the current LSTSVM series often have poor accuracy in real high-dimensional, multiclass, and noisy recognition tasks. Therefore, we propose a novel framework via refined composite multiscale generalized gaussian distribution fuzzy Rényi dispersion entropy (RCMGGDFuRDE) and Balanced Minimum-span-tree Binary Tree LSTSVM (BMBT LSTSVM). Firstly, the new single-scale entropy GGDFuRDE is constructed, which utilizes GGD mapping and Rényi entropy to overcome data sensitivity and enhance the flexibility of feature expression. Then, GGDFuRDE is extended to refined composite multiscale GGDFuRDE (RCMGGDFuRDE), alleviating the problem of coarse-grained information loss. Secondly, an automatic sorting mechanism of "class-center minimum-spanning-tree + depth-first search" and a "node-level negative class undersampling" strategy are introduced into the binary tree decision path, and BMBT LSTSVM is proposed to alleviate the shortcomings of class imbalance, square complexity and error accumulation in the current LSTSVM series. Finally, the collaborative RCMGGDFuRDE and BMBT LSTSVM are applied to two URN datasets. For each type of signal, the experimental dataset is sampled at equal intervals from a large number of raw signals, with 240000 sampling points. After feature extraction, each type of signal is segmented into 60 non overlapping samples with a length of 2048, and 60%/40% of the samples are used for training/testing sets. Our method achieved recognition accuracies of 99.51% and 99.86% on two datasets, respectively.