Enhancing Traditional Underwater DoA Estimation Techniques Using Convolutional Autoencoder-Based Covariance Matrix Reconstruction

Accurate Direction-of-Arrival (DoA) estimation is highly dependent on the correctness of the covariance matrix, which reflects the spatial characteristics of the signal. In underwater environments, noise and multipath interference can significantly impair the covariance matrix's quality, thus affecting DoA accuracy. To address these challenges, this study introduces a convolutional autoencoder (CAE) for reconstructing the covariance matrix. The reconstructed covariance matrix (RCM) is then employed for DoA estimating with traditional covariance-based DoA estimation algorithms, which is crucial for validating the effectiveness of the covariance reconstruction. By reducing noise and multipath interference impacts, our approach provides a cleaner input for DoA estimation algorithms, enhancing accuracy and robustness while preserving interpretability and flexibility for postprocessing tasks. We assess the performance of DoA estimation using the RCM compared to the original, unprocessed covariance matrix with traditional methods. The effectiveness of the CAE in reconstructing the covariance matrix is evaluated through root-mean-square error and resolution probability, demonstrating its ability to mitigate the effects of noise and multipaths.