DOA Estimation via Meta-Learning under Array Sensor Failures

In the applications of direction-of-arrival (DOA) estimation, random array sensor failures can cause a serious performance degradation. Although deep learning-based techniques have offered an efficient solution, a critical limitation is that the network training has to be typically carried out from scratch for a new sensor failure scenario, resulting in a large amount of training time and resources. To this end, we propose a meta-learning-based DOA estimation method under array sensor failures. In particular, we formulate DOA estimation problem for different sensor failure scenarios as distinct tasks. The proposed method can make full use of the inherent relationship among individual tasks and train the deep neural network (DNN) to find a common global parameter vector. This vector enables the DNN to achieve fast convergence to the optimal state for DOA estimation when a new sensor failure scenario emerges. Simulation results demonstrate that the proposed method can speed up the adaptation process and achieve enhanced DOA estimation accuracy compared to the conventional methods.