Optimization of Ultra-Wideband MIMO Antenna Design Using a Novel Convolutional Neural Network Approach.

Aiming at the low efficiency of antenna modeling and solving the problems of slow training speed and high resource consumption when multiple antenna structural parameters are optimized simultaneously, a new deep multi-layer convolutional neural network (DMCNN) is proposed. The DMCNN model uses four layers of ladder-type fully connected layers combined with convolutional layers to improve the learning and classification capabilities of data features. It adds a LeakyReLU activation function to solve the problem of neurons disappearing in negative value areas and alternately uses maximum pooling and average pooling operations to update training parameters to reduce the amount of calculation and maintain feature invariance. Apply Adam optimizer combined with dropout technology to adjust the learning rate and reduce weight fluctuations. The DMCNN model extracts samples from the geometric construction parameters of the ultra-wideband multiple-input multiple-output (UWB MIMO) antenna as feature input to predict the return loss S11 and insertion loss S21. Simulation results prove that the DMCNN model’s prediction fit for S11 and S21 reached 98.48% and 95.02%, respectively. Compared with the multi-layer perceptron (MLP) and Fully Connected Neural Network (FCNN) models, the training effect was improved by at least 11.66%. It solves the shortcomings of traditional methods, and has excellent UWB MIMO antenna modeling capabilities. [ABSTRACT FROM AUTHOR]