Introduction
Direction of Arrival (DoA) estimation is a fundamental task in signal processing, essential for applications ranging from autonomous vehicle radar to seismic monitoring. Traditional subspace methods, such as MUSIC and ESPRIT, are computationally efficient but often fail in environments characterized by low signal-to-noise ratios (SNR) or coherent (multipath) signals. Conversely, modern deep learning (DL) approaches offer high accuracy but are computationally intensive and lack the transparency required for safety-critical systems.
HYPERDOA addresses these challenges by leveraging Hyperdimensional Computing (HDC), a brain-inspired paradigm that represents data as high-dimensional vectors. By reframing signal localization as a pattern recognition problem in hyperspace, HYPERDOA achieves state-of-the-art accuracy while significantly reducing energy consumption.
Methodology
The HYPERDOA framework replaces traditional matrix decompositions and deep neural networks with an HDC-based architecture. The process is divided into four primary stages:
- Feature Extraction: The system utilizes two specialized feature extraction methods: Mean Spatial-Lag Autocorrelation for general efficiency and Spatial Smoothing to handle coherent source interference.
- HDC Encoding: Numeric features are mapped into a high-dimensional space (typically 10,000+ dimensions) using a Fractional Power Encoder based on Fourier Holographic Reduced Representations.
- Associative Memory: During training, the model constructs "centroids" representing prototype hypervectors for each possible angle. A novel "positive-only" update rule is employed to prevent destructive interference when learning multiple signal sources simultaneously.
- Multi-Source Decoding: Inference involves comparing input hypervectors against stored centroids using a greedy peak-selection algorithm with non-maximum suppression to identify multiple signal sources.
Experimental Results
The performance of HYPERDOA was evaluated using a synthetic narrowband Uniform Linear Array (ULA) with 8 antennas. The model was tested against state-of-the-art (SOTA) methods, including SubspaceNet and DeepMUSIC, across various SNR levels and source counts ($M=2, 3, 4$).
Accuracy and Robustness
HYPERDOA demonstrated superior performance in challenging scenarios. In cases involving coherent sources and low SNR (1–5 dB), HYPERDOA achieved up to a 53.13% improvement in accuracy over SOTA methods. Across all tested low-SNR ranges, the framework maintained an average accuracy improvement of 35.39%.
Computational Efficiency
A key advantage of HYPERDOA is its extreme energy efficiency. When deployed on an NVIDIA Jetson Xavier NX embedded board, the "lag" version of HYPERDOA consumed only 135 mJ per inference. This represents a 92.93% reduction in energy consumption compared to the SubspaceNet+MUSIC baseline (4645 mJ), making it approximately 32.5 times more efficient.
Conclusion
HYPERDOA represents a significant advancement in DoA estimation, providing a robust, transparent, and energy-efficient alternative to both classical and deep learning methods. By utilizing the holographic properties of HDC, the system achieves high resilience to noise and interference. Future research will focus on validating the framework with real-world antenna array data and exploring its performance under physical array imperfections.
Summary updated