Small Sample Coherent DOA Estimation Method Based on S2S Neural Network Meta Reinforcement Learning.

Area of Science:

• Signal Processing
• Machine Learning
• Array Signal Processing

Background:

• Traditional neural network-based Direction of Arrival (DOA) methods require extensive training data for signal-scene adaptation and accurate estimation, especially in coherent signal environments.
• Existing methods struggle with limited sample data, hindering performance in complex scenarios.

Purpose of the Study:

• To develop a novel meta-reinforcement learning (MRL) method for accurate coherent signal DOA estimation using minimal training samples.
• To convert the coherent signal DOA problem into an angle interval estimation problem for improved performance with small sample sizes.

Main Methods:

• The proposed MRL method models coherent signal angle interval estimation as a Markov decision process.
• A sequence-to-sequence (S2S) neural network is employed within the inner loop to process angular interval feature sequences, leveraging contextual relevance.
• The approach learns optimal strategies for angle interval identification from limited spatial spectral sequence data.

Main Results:

• The MRL algorithm successfully estimates the angle interval of incident signals under small sample conditions.
• The S2S neural network facilitates strategy learning by utilizing the context relevance of spatial spectral sequences.
• The method achieves accurate DOA estimation through a subsequent one-dimensional spectral peak search within the determined angle interval.

Conclusions:

• The meta-reinforcement learning algorithm based on S2S neural networks demonstrates rapid convergence to optimal states.
• It requires only minor gradient updates of S2S neural network parameters with small sample sets when encountering new signal environments.
• This approach offers an efficient solution for accurate coherent signal DOA estimation in data-scarce scenarios.