Feature Analysis and Optimization of Underwater Target Radiated Noise Based on t-SNE

The analysis and optimization for underwater target radiated noise recognition are addressed. The t-SNE (t-distributed stochastic neighbor embedding) algorithm is taken for dimension reduction of the underwater target radiated noise spectrum bands segmented by frequency and the visual results can be obtained. The average distance between classes and coincidence rate are constructed as indicators of the visualization results. By analyzing the separability of each frequency band, the optimal features can be obtained. Then the optimal features are recognized by three classification algorithms which are Random Forest, SVM (Support Vector Machine) and AdaBoost. The processing results of experimental signal spectrum with two types of target are as follows. The separability of the target signal spectrum decreases with the increase of frequency. The spectrum band of 10-150Hz has the best separability. SVM is sensitive to the increase of the data dimension with non optimal separability. Random Forests and AdaBoost can apply separability components in wider frequency bands and have greater tolerance. The overall recognition accuracy of AdaBoost is the highest, but the computation efficiency is the lowest. These analysis results shows that the t-SNE algorithm can be used to optimize the underwater target radiated noise spectrum features for the purpose of improving the accuracy and efficiency of the classification algorithm.