Deep Learning Method for Prediction of Planar Radiating Sources from Near-Field Intensity Data

As the frequency of operation and packaging density rapidly increase, the contemporary issues of intricate electromagnetic interference (EMI) cannot be effectively addressed by conventional electromagnetic analysis techniques. Predicting the intensity and location of radiation sources on a printed circuit board (PCB) is of great interest to the EMI/EMC community. Deep learning models could potentially be leveraged to deal with the non-linearity and complexity of present EMI issues. In this paper, we propose a method based on deep learning using cascaded convolutional neural networks to predict the frequency-dependent intensity distribution of planar radiation sources from frequency-dependent near-field intensity data. In particular, two convolutional neural networks are utilized. The first estimates the amplitude of radiation fields on a plane in close range to the radiation sources, taking input as the available near-field amplitude data. Using the distribution obtained as input, the second estimates the intensity of the planar radiation sources. The proposed deep learning method can be applied at any frequency point within the frequency range used to train the networks. Experimental results demonstrate that our approach achieves high accuracy with mean squared errors of 0.0062 and 0.0023 for the first and second networks respectively, validating its effectiveness for EMI diagnostic applications.