Development and Evaluation of a CNN-LSTM Architecture based Neural Network for Time Optimization during EMI Measurements

In this paper, an approach is proposed to find the worst-case positions during the final measurement phase on critical frequencies in electromagnetic interference (EMI) measurements according to 47 CFR § 15.209 by using a developed measurement software and deep neural networks (DNN). Firstly, because of its advantage of incomplete connection, relatively simple model structure and strong data features extraction, a dimensional convolution neural network (1D CNN) was present to predict the positions that meet the maximum radiation emission level. Secondly, a hybrid deep learning neural network framework, that combines CNN with long short term memory(LSTM) was adopted to forecast the worst-case of the high variance emission levels. The DNNs were trained using real EMI measurements for different equipment under test (EUT) in a Semi Anechoic Chamber (SAC) by Cetecom GmbH in Essen, Germany. By predicting the position azimuth of the turntable and the height of the antenna, the required time to carry out the final measurement phase is effectively reduced.