Effective Convolutional Neural Network for Wideband Frequency Invariant Beam Pattern Synthesis

Abstract

In this letter, a convolutional neural network (CNN) based framework is proposed to achieve efficient wideband frequency invariant (FI) beampattern synthesis. The CNN is employed to acquire a set of optimized finite-impulse-response (FIR) filter coefficients corresponding to the desired wideband FI pattern. In the proposed CNN framework, a set of initialized filter coefficients obtained by the upper bound of the desired pattern is taken as the unlabeled input sample, with two input channels representing the real and the imaginary parts of the filter coefficients, respectively. By carefully designing loss function, low sidelobe level (SLL) and frequency invariant beam pattern are well achieved. Then, by minimizing the loss function during the training process, the filter coefficients corresponding to the desired pattern are acquired. Numerical examples involving wideband FI pencil-beam pattern, shaped beam pattern, and scannable beam pattern are provided to validate the flexibility and effectiveness of the proposed method.