Leveraging overcomplete autoencoder for improved direction-of-arrival estimation using sensor array

The Direction of Arrival (DoA) estimation using linear sensor arrays is a critical task in array signal processing, with applications spanning SONAR, RADAR, and more. Traditional methods often struggle with the non-linear and high-dimensional nature of the data involved. This paper proposes a novel approach that leverages the strengths of both deep learning and machine learning classifiers. By combining an overcomplete Auto Encoder (AE) with a Support Vector Machine (SVM), the proposed method effectively extracts intricate features from the high-dimensional input signals, enabling accurate DoA estimation. The model was trained and evaluated using field data collected from the Indian Ocean. Sensor delays were introduced experimentally to the collected single-channel data by positioning acoustic sources at various angles within 0 to 180-degree relative to a linear microphone array with 16 channels. In DOA estimation, where experimental data are often noisy or missing features, the AE learns a detailed latent space representation that helps to smooth the data before it is classified. While AE + SVM is not new, its application to DoA estimation with noisy, incomplete, or sparse experimental data is novel. The overcomplete AE addresses the unique challenges of DoA estimation by improving feature quality and enabling more accurate, robust predictions. Experimental results demonstrate the efficacy of this unified model, achieving a remarkable 94.7% accuracy in DoA estimation tasks.