Feature Extraction of Underwater Acoustic Signal Target Using Machine Learning Technique

Abstract

Finding underwater sound sources is quite difficult because of how complex and interconnected the undersea world is. When trying to use ship-radiated noise for Underwater Acoustic Target Recognition (UATR), the complicated presence of aquatic creatures makes the process even more tough. An innovative DNN called the Audio Perspective Region-based Convolutional Neural Network (APRCNN) is presented in this study. In order to train APRCNN, we used Depth Wise Separable (DWS) Convolutional Deep Neural Network architecture using initial underwater acoustic wave data. After that, APRCNN is used to categories and forecast underwater sound waves. Novel aspects of the suggested APRCNN model include optimization, adaptive learning, processing in parallel with residual connections, and underwater environment adaptation. With the help of integration layers that were influenced by the perceptual processing of the SPDNN system. Machine translation approaches also improve the model's performance through the use of time-dilated convolution. In this research, we offer a technique for underwater target detection with a full-featured DNN architecture. To achieve optimal classification accuracy with minimal compute load in different underwater degradation conditions, the suggested architecture maximizes the reallocation of prior feature maps. In addition, the suggested method eliminates the requirement for time-frequency spectrum analyzer pictures by allowing the network to be fed real data from audio signals instead. With a remarkable 98.4 percent accuracy level at 0-dB Signal to Noise Ratio (SNR), our classification methodology outperforms both state-of-the-art DNN systems and classic ML approaches, as demonstrated by thorough examination of a real-world dataset of passive sonar waves. When compared to other DNN systems for processing underwater acoustic signals, PRCNN stands out due to its adaptive learning mechanisms, focus on perspective regions, integration of domain knowledge, learning of hierarchical representations, resilience to environmental variability, and region-specific feature extraction. Thanks to these upgrades, APRCNN is now better able to identify, classify, and localize sounds in underwater acoustic settings, as well as to handle the unique problems of underwater signal processing. [ABSTRACT FROM AUTHOR]