TL;DR
This paper introduces a novel deep learning approach for pinpointing acoustic sources in shallow ocean environments. By developing enhanced input features (MTPEF) and a specialized neural network architecture (IRSNet), the method significantly improves localization accuracy and robustness against environmental variations, outperforming traditional techniques and existing deep learning models.
What Problem Does This Paper Address?
Accurately locating underwater acoustic sources, particularly in shallow water, is a critical but challenging task for applications like sonar and marine surveillance. Traditional methods, such as Matched-Field Processing (MFP), are highly sensitive to environmental mismatches (e.g., variations in water depth, sediment properties, sound speed) and can be computationally expensive. While deep learning offers promise, existing approaches often apply generic architectures and features, failing to fully leverage the unique characteristics of underwater acoustics, leading to suboptimal accuracy and robustness. This paper aims to bridge this gap by proposing a tailored deep learning solution that addresses these limitations.
Methods
The core of this work lies in two main innovations:
- Improved Input Features (MTPEF): The authors introduce a "Multi-Time Pressure and Eigenvector Feature" (MTPEF) that combines complex pressure amplitudes and phases with eigenvector features derived from covariance matrices. This feature set is designed to capture richer information from the acoustic signals. Crucially, they also incorporate a "feature augmentation" step. This involves strategically perturbing environmental and target motion parameters during training to expose the network to a wider range of scenarios, thereby enhancing its robustness.
- Specialized Network Architecture (IRSNet): A novel deep neural network, termed "Inception and Residual in Series Network" (IRSNet), is designed. It combines the strengths of Inception modules (for capturing multi-scale features) and Residual modules (for mitigating overfitting and enabling deeper networks). Unlike some prior work that nests these modules, IRSNet connects them in series, which the authors found to be more suitable for the available underwater acoustic data and less prone to complexity issues. The network is designed to be lightweight, balancing localization accuracy with training efficiency.
Additionally, a network visualization method is proposed. This technique uses features from the last hidden layer to generate a 2D ambiguity plane, similar to MFP outputs. This visualization aids in understanding the network's learned relationships and allows for the definition of a Localization Confidence Interval (LCI) by setting a correlation threshold, providing a range of high-confidence source positions.
Datasets
The research utilizes simulated datasets built to mimic the SWellEx-96 experiment environment, featuring a multi-layered shallow waveguide. Two primary training datasets, Train-A and Train-B, were created. Train-A uses fixed target motion parameters, while Train-B incorporates variations in both environmental and target motion parameters to improve robustness. Validation and test datasets were also generated with fixed depth/variable CPA range and fixed CPA range/variable depth targets. The SWellEx-96 real-world experiment data was used for final validation.
Key Insights and Ideas
- Feature Engineering is Crucial: Simply applying generic features to underwater acoustics is insufficient. The MTPEF, by combining pressure and eigenvector information and incorporating multi-time aspects, significantly outperforms simpler feature sets.
- Augmentation for Robustness: Exposing the network to variations in environmental and motion parameters during training is vital for achieving robust localization, especially in dynamic ocean environments.
- Tailored Architectures Matter: A specialized network like IRSNet, which intelligently combines Inception and Residual modules in series, outperforms generic deep learning architectures and even complex nested structures for this specific problem.
- Visualization Aids Trust: The LCI, derived from network visualization, provides valuable confidence intervals for localization results, offering a more interpretable output than a single point estimate.
Key Results
- Improved Accuracy: The proposed MTPEF features achieved the lowest RMSE for both depth (7.722 m) and CPA range (0.921 km) compared to other tested features (PEF, MTPF, MTEF) when used with the OriginNet baseline.
- Enhanced Robustness: Training with the augmented dataset (Train-B) significantly improved robustness against variations in source velocity and time difference, as demonstrated by consistently lower RMSE compared to training with Train-A.
- Superior Network Performance: The IRSNet achieved the best localization accuracy among various tested network structures, with an RMSE of 3.657 m for depth and 0.523 km for CPA range, while also maintaining a competitive training time.
- Outperforming MFP: In simulation experiments with environmental mismatches, the proposed method showed significantly stronger robustness than improved MFP methods (MFP-M and MFP-PDS), maintaining low RMSE even with parameters not seen during training.
- Real-World Validation: The SWellEx-96 experiment results confirmed the effectiveness of the proposed method, showing comparable or better accuracy and faster computation than MFP methods.
Limitations and Future Work
- Deep Water Extension: The current research focuses on shallow water. The authors suggest that extending the method to deep water environments, which have different modeling requirements, is a future research direction.
- Parameter Quantity Optimization: The network's parameter quantity was determined empirically. Further study is needed to optimize this based on changes in input feature types and dimensions.
- Multi-Target Localization: The current work addresses single-target localization. Future research could explore extending the method to multitarget scenarios.
Summary updated