TY - JOUR
T1 - High-accuracy DOA Estimation Method Using Acoustic Vector Sensor Array under Non-uniform Noise
AU - Wang, Weidong
AU - Wang, Tianyou
AU - Li, Hui
AU - Zhang, Kai
AU - Shi, Wentao
AU - Ali, Wasiq
N1 - Publisher Copyright:
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2026.
PY - 2026
Y1 - 2026
N2 - To address the challenge of direction-of-arrival (DOA) estimation with acoustic vector sensor array (AVSA) in non-uniform noise, a graph neural network-based differential self-attention (GNNDSA) architecture is proposed in this paper. By transforming conventional covariance matrices into a graph-structured representation consisting of nodes and their associated edges, a new data preprocessing method is introduced to explicitly model the inherent correlations among the output channels of the acoustic vector sensor (AVS). Furthermore, a physics-aligned sliding-window node reduction strategy is incorporated to effectively mitigate the exponential growth of graph nodes, enabling the pre-trained model to be seamlessly generalized to arrays with varying numbers of sensors without retraining. To accommodate the assumption of unknown non-uniform noise, a differential self-attention (DSA) mechanism is designed to dynamically calculate the aggregated weights of adjacent nodes by introducing node difference information, thereby suppressing the adverse effects of overly similar node representations. Through progressive multi-layer information aggregation to extract features most relevant for classification, the GNNDSA architecture is ultimately proposed to enhance DOA estimation performance. Simulation results validate that the presented GNNDSA network provides markedly higher DOA estimation accuracy than both conventional subspace-based approaches and existing deep learning methods in non-uniform noise environments, especially under conditions of low signal-to-noise ratio and a restricted number of snapshots.
AB - To address the challenge of direction-of-arrival (DOA) estimation with acoustic vector sensor array (AVSA) in non-uniform noise, a graph neural network-based differential self-attention (GNNDSA) architecture is proposed in this paper. By transforming conventional covariance matrices into a graph-structured representation consisting of nodes and their associated edges, a new data preprocessing method is introduced to explicitly model the inherent correlations among the output channels of the acoustic vector sensor (AVS). Furthermore, a physics-aligned sliding-window node reduction strategy is incorporated to effectively mitigate the exponential growth of graph nodes, enabling the pre-trained model to be seamlessly generalized to arrays with varying numbers of sensors without retraining. To accommodate the assumption of unknown non-uniform noise, a differential self-attention (DSA) mechanism is designed to dynamically calculate the aggregated weights of adjacent nodes by introducing node difference information, thereby suppressing the adverse effects of overly similar node representations. Through progressive multi-layer information aggregation to extract features most relevant for classification, the GNNDSA architecture is ultimately proposed to enhance DOA estimation performance. Simulation results validate that the presented GNNDSA network provides markedly higher DOA estimation accuracy than both conventional subspace-based approaches and existing deep learning methods in non-uniform noise environments, especially under conditions of low signal-to-noise ratio and a restricted number of snapshots.
KW - Acoustic vector sensor array (AVSA)
KW - Differential self-attention (DSA) mechanism
KW - Direction of arrival (DOA) estimation
KW - Graph neural network (GNN)
KW - Non-uniform noise
UR - https://www.scopus.com/pages/publications/105052503224
U2 - 10.1007/s00034-026-03774-1
DO - 10.1007/s00034-026-03774-1
M3 - 文章
AN - SCOPUS:105052503224
SN - 0278-081X
JO - Circuits, Systems, and Signal Processing
JF - Circuits, Systems, and Signal Processing
ER -