Context‐Aware Reconstruction for DVL Outlier Detection in Robust AUV Navigation

In complex shallow‐water environments, environmental and acoustic interferences may lead to anomalous Doppler Velocity Log (DVL) measurements, significantly degrading Autonomous Underwater Vehicle (AUV) navigation performance. To address this issue, this article proposes a robust AUV navigation strategy based on context‐aware reconstruction learning. First, a Context‐Aware Reconstruction Network (CAR‐Net) based on deep Autoencoder (AE) architecture is proposed for DVL outlier detection. By constructing a deep network composed of a heterogeneous AE, a multi‐head self‐attention mechanism, and multi‐scale dilated residual convolutions, the model is capable of achieving global perception over the entire temporal domain while progressively and accurately reconstructing fine‐grained local temporal features. Subsequently, a context‐aware reconstruction mechanism is introduced, which incorporates historical data as contextual constraints, thereby enabling focused modeling of recent time‐series data. Meanwhile, to enhance the ability to fit critical recent moments, a time‐weighted loss is proposed, increasing sensitivity to predictions at the end of the sequence and thereby further improving the reconstruction accuracy of complex temporal patterns. Real‐world AUV sea trial data thoroughly validate that the Extended Kalman Filter (EKF) framework integrated with CAR‐Net demonstrates outstanding performance in extreme outlier detection and robust navigation, providing a novel deep‐learning paradigm for AUV operations in complex underwater environments.