Novelty Detection in Underwater Acoustic Environments for Maritime Surveillance Using an Out-of-Distribution Detector...

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Underwater acoustic sensing systems require robust detection of unknown signals for maritime security and autonomous navigation.
• Conventional deep learning models struggle with overconfidence and lack of uncertainty quantification for unknown signals.
• Existing methods fail to provide reliable novelty detection in dynamic underwater environments.

Purpose of the Study:

• To develop a novelty detection framework that enhances the reliability and stability of underwater acoustic sensing systems.
• To address the limitations of conventional deep learning models in handling unknown signals.
• To improve the quantification of predictive uncertainty in signal detection.

Main Methods:

• Integration of an out-of-distribution detector for neural networks (ODIN) with Monte Carlo (MC) dropout.
• ODIN calibrates softmax probabilities to reduce overconfidence and improve signal separability.
• MC dropout introduces stochasticity for estimating predictive uncertainty, with outputs modeled using Gaussian mixture models and Kullback-Leibler divergence for deviation quantification.

Main Results:

• The proposed method demonstrated an average increase of 9.5% in the area under the receiver operating characteristic curve compared to MC dropout baseline.
• A 5.39% increase in the area under the receiver operating characteristic curve was observed compared to the ODIN baseline.
• Significant reductions in false positive rates were achieved, with 7.82% and 2.63% improvements over MC dropout and ODIN baselines, respectively.

Conclusions:

• The integration of stochastic inference with ODIN significantly enhances the stability and reliability of novelty detection in underwater acoustic environments.
• The proposed framework effectively mitigates model overconfidence and quantifies predictive uncertainty, crucial for real-world applications.
• This approach offers a promising solution for robust signal detection in maritime security and autonomous navigation.