Spectral-Based SPD Matrix Representation for Signal Detection Using a Deep Neutral Network.

Area of Science:

• Signal processing
• Machine learning
• Matrix analysis

Background:

• Symmetric positive definite (SPD) matrices are crucial in classification due to their Riemannian manifold structure.
• Non-linear geometric metrics offer superior discrimination and reduced information loss over Euclidean metrics for SPD matrices.
• Existing spectral methods for signal detection have limitations in performance.

Purpose of the Study:

• To propose a novel spectral-based SPD matrix signal detection method utilizing deep learning.
• To transform signal detection into a binary classification problem on a manifold.
• To evaluate the performance of the proposed method against state-of-the-art techniques.

Main Methods:

• Constructing SPD matrices from time-frequency spectra using spectral covariance and spectral transformation models.
• Employing a deep SPD matrix learning network for binary classification on the manifold.
• Utilizing simulated and semi-physical signal datasets for validation.

Main Results:

• The proposed spectral-based SPD matrix signal detection method with deep learning achieves a 1.7-3.3 dB gain under specific conditions.
• The deep learning approach demonstrates superior detection performance compared to state-of-the-art spectral methods using convolutional neural networks.
• The method effectively transforms signal detection into a manifold-based binary classification task.

Conclusions:

• The developed deep learning framework offers a powerful new approach for SPD matrix-based signal detection.
• This method provides significant performance improvements over existing spectral detection techniques.
• The use of SPD matrices and deep learning on Riemannian manifolds shows great promise for advanced signal processing applications.