Deep Convolutional Network-Assisted Multiple Direction-of-Arrival Estimation

Multiple direction-of-arrival estimation is one of the core functions in array signal processing and has many engineering applications. It is proposed that it can be realized using a two-stage strategy which consists of multiclass classification for region segmentation and successive cancellation-based fine estimation in subregions. In the first stage, the deep convolutional network (DCN) is introduced to classify the 2D angles into discrete subregions of the arrival plane. To enhance the DCN performance, the K-means clustering algorithm is applied to label 2D angles. The DCN is trained using the dataset with only one signal which can considerably reduce training complexity and dataset size. The trained DCN can be generalized for multisignal scenarios. In the second stage, the orthogonal matching pursuit algorithm is utilized to estimate the 2D angles in each subregion. To estimate multiple signals, orthogonal projection transformation is employed to successively cancel the estimated signals. Numerical results demonstrate that the proposed DCN-assisted multiple direction-of-arrival estimation method has lower complexity with better performance than the reduced-dimension MUSIC algorithm.