Advancing robust underwater acoustic target recognition through multitask learning and multi-gate mixture of experts

Area of Science:

• Underwater acoustics
• Signal processing
• Machine learning

Background:

• Limited authentic underwater acoustic recordings impede data-driven target recognition models.
• This data scarcity compromises the stability and generalization of current acoustic recognition systems.

Purpose of the Study:

• To introduce a novel recognition framework, M3 (multitask, multi-gate, multi-expert), designed to enhance pattern recognition capabilities.
• To improve the robustness and generalization of underwater acoustic target recognition models.

Main Methods:

• Implemented a multitask learning approach by incorporating an auxiliary task for estimating target properties (e.g., size).
• Utilized multi-expert and multi-gate mechanisms to allocate distinct parameter spaces for diverse underwater signals.
• Developed a framework that enables fine-grained and differentiated processing of intricate signal patterns.

Main Results:

• The M3 framework demonstrated superior performance compared to advanced single-task recognition models.
• Achieved state-of-the-art results on the ShipsEar underwater ship-radiated noise dataset.
• Validated the model's ability to capture robust target patterns through enhanced generalization.

Conclusions:

• The M3 framework effectively addresses the limitations of data scarcity in underwater acoustic target recognition.
• The proposed approach significantly enhances model stability and performance in practical applications.
• M3 represents a significant advancement in achieving state-of-the-art underwater acoustic target recognition.