Direction finding method via acoustic vector sensor array under time-varying non-uniform noise

To address the performance degradation of direction-of-arrival (DOA) estimation for an acoustic vector sensor array (AVSA) under time-varying non-uniform noise, a cross-channel adaptive fusion long short-term memory with information aggregation (CAFLIA) framework is proposed in this paper. First, to address the spatial non-uniformity of noise power across the channels within a single acoustic vector sensor (AVS), a residual-compensation-based cross-channel adaptive fusion mechanism (CAFM) is developed, which exploits the inherent coupling between the acoustic pressure and acoustic particle velocity channels to mitigate the impact of spatially non-uniform noise. Second, to suppress time-varying noise disturbances, a segmented temporal refinement strategy is introduced, which divides the observation sequence under short-term stationarity assumptions and employs a manifold-prior-guided long short-term memory network (MG-LSTM) to enhance stable directional feature extraction while suppressing temporal noise fluctuations. Finally, a reliability-aware information aggregation mechanism (IAM) is designed to adaptively select informative features using cosine similarity for neighbor selection and employing multi-head self-attention for aggregation, mitigating noise across temporal segments. Extensive simulation results and anechoic water tank experiments demonstrate that the proposed CAFLIA method achieves superior DOA estimation accuracy compared with representative subspace-based and deep learning-based approaches, particularly under low signal-to-noise ratio and limited snapshot conditions in time-varying non-uniform noise environments.