Passive Sonar Signal Processing Applied Research
Abstract
The proposed efforts will explore some of the latest concepts in machine learning especially deep learning due to its potential for enhancing performance in detection, classification, and localization of acoustic sources in shallow water environments. The stu dy will develop new concepts and network structures suitable for extracting large scale temporal-spectral context for detecting and localizing low signature targets over a long range. A hybrid CNN and RNN models, such as Long Short Term Memory (LSTM) or Gated Recu rrent Unit (GRU) will serve as the baseline model while architectures based on attention mechanism (or self-attention based transfor mers) will be exploited for adaptation to underwater acoustic environments. The effort proposed here also includes development of ef fective denoising algorithms for adaptively enhancing target signals in unseen noise environments. A Denoising Autoencoder (DAE) we developed previously for Automated Speech Recognition (ASR) will be examined for its suitability in an underwater environment. The e fforts for reducing noise will also include development of a learning based dynamic filter generation network that can be trained to generate a mask adaptive to input noise. Additionally, we will develop network architectures based on the concept of slot attention modules for effectively separating signals from a mixture of multiple acoustic sources.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 07, 2021
- Source ID
- N000142112790
Entities
People
- David K. Han
Organizations
- Drexel University
- Office of Naval Research
- United States Navy