The Practice of Few-Shot Learning on Electromagnetic Susceptibility Evaluation

Abstract

Few-shot learning (FSL) inspired by deep neural network and its practice on electromagnetic susceptibility evaluation are reported. By the assistance of model, metric or optimization, FSL can not only make use of the prior information, but also gain the ability of learning the pattern for distinguishing and classifying input vectors from only a few samples. For transient electromagnetic disturbances, the recurrent neural network (RNN) is involved to help with the memory of the input signal with the style of 1-D time sequence. In order to learn the proper feature of transient disturbances, a Siamese neural network containing two RNNs is built and trained, and the distance of two outputs can be measured by a preselected distance function. This makes the proposed model capable to learn the difference between normal cases and failure cases. In order to demonstrate this method, a pulse current injection testing on a certain electronic system is conducted and evaluated with the metric-based FSL. The result confirms the practicability of the proposed model, for the correct classification rate can be 83.3% with only 36 samples.