A Synchronous Processing Method for Underwater Sound Source Recognition and Range Estimation Based on Multi-Task Learning and Transfer Learning

This study focuses on underwater sound sources and investigates the application of multi-task learning (MTL) in the synchronous processing of target recognition and range estimation to achieve systematic detection. To enhance the applicability of the MTL model with limited underwater acoustic data,...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Circuits, systems, and signal processing Jg. 44; H. 9; S. 6845 - 6872
Hauptverfasser: Wang, Yong, Yao, Qihai, Yang, Yixin
Format: Journal Article
Sprache:Englisch
Veröffentlicht: New York Springer US 01.09.2025
Springer Nature B.V
Schlagworte:
Acoustics
Broadband
Circuits and Systems
Datasets
Deep learning
Effectiveness
Electrical Engineering
Electronics and Microelectronics
Engineering
Instrumentation
Learning
Localization
Machine learning
Marine environment
Neural networks
Regularization
Signal,Image and Speech Processing
Sound sources
Target recognition
Underwater acoustics
Underwater construction
Wavelet transforms
Target recognition
Transfer learning
Multi-task learning
Range estimation
ISSN:0278-081X, 1531-5878
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:This study focuses on underwater sound sources and investigates the application of multi-task learning (MTL) in the synchronous processing of target recognition and range estimation to achieve systematic detection. To enhance the applicability of the MTL model with limited underwater acoustic data, this study integrates transfer learning (TL) with MTL, forming the MTL-TL model. The proposed approach extracts input features of underwater acoustic targets, constructs an MTL model, pre-trains it using a large dataset from a pre-selected marine area, retrains it with a small dataset from the detected sea area based on TL, and performs target recognition and range estimation on test sets. The effectiveness of the proposed MTL-TL method is validated using data from the S5 voyage of SWellEX-96 experiment and is compared with three models: single-task learning (STL), STL-TL, and MTL. The results indicate that the combination of the regularization effect between subtasks of MTL and the empirical information advantage of TL enables the MTL-TL model to achieve accurate target recognition and range estimation, outperforming other models. Furthermore, the model remains effective in broadband target scenarios and cases involving two sound sources of the same frequency. Even at low signal-to-noise ratios, the model remains applicable, achieving the recognition accuracy above 79% and mean absolute percentage error below 8% at 0 dB.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0278-081X
1531-5878
DOI:10.1007/s00034-025-03127-4