Enhancing Information Extraction in EMC Measurements through Artificial Intelligence

In this publication, a machine learning model of the relation between design parameter and EMC properties is derived from near field scans of printed circuit boards. The obtained near field data are transformed to the far field, where relevant information such as peak field strength or the radiation pattern are derived. These information are used as annotations to the corresponding design parameters (the data) during learning. Techniques for data augmentation are considered as well as adding simulation data to speed up data generation. The trained model can be used as surrogate model in optimization or further machine learning procedures. In addition, the corresponding near field data are analyzed in order to find relevant patterns that can be related to possible EMC problems via the known far field. Auto-encoders are employed to reduce the high dimension of the near field radiation data, which are thus mapped to a low dimensional latent space with minimum information loss and with respect to EMC properties. This allows for dimensional reduction of the design space and for an identification of sensitive parts of the near field.