Synthesis of Linear Sparse Array Using DNN-Based Machine-Learning Method
Xin Yang, Deqiang Yang, Yanwen Zhao, Jin Pan, Yongpin Chen
Abstract
In this article, a deep neural network (DNN)-based machine learning (ML) approach is presented to synthesize linear sparse arrays (LSAs). Due to the powerful fitting capability of neural networks, a DNN framework is established to map the relationship between the desired array pattern and LSA geometry. In this manner, the array synthesis is accomplished by training a proper DNN that can generate the desired LSA geometry. To obtain better training, the density tapering (DT) technique is introduced to the training process as the prior domain knowledge to specify the varying boundaries of each element in the LSA. Consequently, an efficient DT-assisted DNN (D-DNN) synthesis method is developed. The proposed D-DNN method has the ability to simultaneously optimize the element positions and excitations, and directly constrain the inter-element spacings. Furthermore, by introducing the active element pattern (AEP) technique, the presented method can synthesize both ideal arrays and arrays with mutual coupling. Numerical results of a wide variety of synthesis scenarios are presented to demonstrate the validity, efficiency, and flexibility of the proposed method.