A new feature extraction method of ship radiated noise based on variational mode decomposition, weighted fluctuation-based dispersion entropy and relevance vector machine. (15th December 2022)

• Record Type: Journal Article
• Authors: Liu, Feng; Li, Guohui; Yang, Hong
• Is Part Of: Ocean engineering. Volume 266(2022)Part 5
• Publication Date: 2022-12-15
• Subjects: Variational mode decomposition -- Weighted fluctuation-based dispersion entropy -- Relevance vector machine -- Ship radiated noise -- Feature extraction; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2022.113143
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: Due to the changeable circumstance and the continuous improvement of ship manufacturing technique, ship radiated noise (SRN) has become more and more complex, which makes it more essential to extract its features. Moreover, the target ships can be quickly classified by feature extraction, thus occupying the advantage of maritime confrontation. For the sake of improved feature extraction technology, a new feature extraction method of SRN based on optimized variational mode decomposition by tuna swarm optimization (TSO-VMD), weighted fluctuation-based dispersion entropy (WFDE) and optimized relevance vector machine by sparrow search algorithm (SSA-RVM) is proposed, moreover, the features of seven types of SRN are extracted respectively. To solve the weakness that decomposition level K and penalty factor α must to be preset artificially in VMD, TSO-VMD is proposed. To select optimal width factor and super parameter of RVM, SSA-RVM is proposed. In the first place, decompose SRN into a range of intrinsic modal functions (IMFs) by TSO-VMD, calculate the K-nearest neighbor mutual information (KNN-MI) value between each IMF and the original signal, and normalize it to obtain normalized KNN-MI (nor-KNNMI). Next, select the IMF with the largest nor-KNNMI value as the feature vector, and 60 samples are randomly selected from the feature vector and their fluctuation-based dispersion entropy (FDE) are calculated. Afterwards, select the nor-KNNMI value corresponding to the feature vector to weight the FDE value to obtain the WFDE value, and use the WFDE value to classify SRN. In the end, input the WFDE value of each sample into SSA-RVM for identification. The experimental results show that the identification rate of the proposed method is over 90%. Therefore, the proposed method can accurately and efficiently extract the features of SRN. Graphical abstract: Image 1 Highlights: A new feature extraction model is proposed. TSO-VMD is proposed, which solves the weakness that decomposition level K and penalty factor α must to be preset. WFDE is proposed, which solves the weakness that FDE can't connect each IMF and can't accurately extract the features. SSA-RVM is proposed, and the optimal width factor and super parameter are selected.