A Unified Two-Dimensional Direction Finding Approach for Sensor Arrays Based on Deep Neural Networks With Inhomogeneous Angle and Frequency Partition

This paper develops a deep neural networks (DNNs) based direction finding approach for 2-D direction finding in multiple frequencies. The approach is frequency-unified and is capable of direction-of-arrival (DoA) prediction for multiple frequencies. A model output is proposed and proved for 2-D direction finding. In addition, different from the common mean squared error (MSE) loss, a label-dependent loss is proposed motivated by the different Cramer-Rao Lower Bound (CRLB) of direction finding accuracy for different DoAs, and an inhomogeneous angle and frequency partition strategy is suggested for efficient implementation. Following the model output and the partition strategy, the general DNN architecture consists of a series of classification networks for angle zoom partition and regression networks for DoA prediction. Several useful tips for model inputs and zoom partition are also suggested. Numerical investigations are included to evaluate the performance of the proposed approach and show desirable performance for both ideal and nonideal scenarios. The added value of treating samples of different frequencies in a unified manner is also shown.

Automated summary

This paper proposes a deep neural networks (DNNs) based approach for 2D direction finding in multiple frequencies. The approach is frequency-unified and capable of predicting direction of arrival (DoA) for multiple frequencies. The paper introduces a model output for 2D direction finding and suggests a label-dependent loss motivated by different Cramer-Rao Lower Bound (CRLB) of direction finding accuracy. An inhomogeneous angle and frequency partition strategy is also proposed for efficient implementation. Numerical investigations show that the proposed approach performs well in both ideal and nonideal scenarios, and demonstrates the added value of treating samples of different frequencies in a unified manner.