A Ship Radiated Noise Recognition Method Applicable to Incomplete Training Data Sets

Abstract

To effectively classify ship radiated noise signals under incomplete training data sets, this article proposes a ship radiated noise recognition method. This method consists of a feature selection method based on the Bhattacharyya distance measurement (FS-BD) and the fuzzy support vector machine (SVM) based on fuzzy support vector center (FSVM-fsv). In the feature extraction stage, FS-BD removes redundant dimensions based on the similarity of feature probability distribution, aiming to improve the separability of features. However, the complex underwater acoustic environment leads to heterogeneity between similar samples, which further causes the incompleteness of the learning feature set. In the classifier design, under the framework of SVM, the FSVM-fsv method uses the distance between samples and their respective fuzzy support vector centers to optimize the decision hyperplane's spatial division capability, thereby reducing the impact of outliers and noise on classifier performance. The experimental results show that the FS-BD method can significantly improve the difference and robustness of the target features under the condition of about 900 samples of single-class targets and containing multiple environments. At the same time, the performance of FSVM-fsv under incomplete training sets is better than other machine learning methods, such as SVM, FSVM based on class center, FSVM based on estimated hyperplane, FSVM based on actual hyperplane, and FSVM based on support vector center (FSVM-sv), and the average recognition accuracy reaches 84.65%. This recognition method provides a reliable solution for target recognition in unknown environments.