Learned FFT Windowing for Device Classification from Unintended Conducted Emissions Data

摘要

Characterization of Unintended Conducted Emissions (UCE) from electronic devices is important when diagnosing electromagnetic interference, performing nonintrusive load monitoring (NILM) of power systems, and monitoring electronic device health, among other applications. Prior work has demonstrated that UCE analysis can serve as a diagnostic tool for these goals. UCE collections from 18 commercial devices were augmented with high levels of additive white Gaussian noise and used for proof of concept and analytic experimentation with the proposed technique. This dissertation describes a novel means of using deep neural networks (DNN) for the classification of low power electronic devices from UCE data. The author has conceived of a novel means of automatically generating a fast Fourier transform (FFT) window function that is shown by this work to have the ability to explain aspects of what the DNN classifier (ResNet) sees concerning the features and noise in the data set (important in unintended emission types of applications). The method back-propagates the classification loss/error through the network and the FFT, which is embedded in the network as a "fixed" layer, to arrive at a window function that is appropriate for the data set and classifier. This method can be used partially for explainability of the classifier, as the window is a mathematical function that can be analyzed using signal processing theory as a guide to determine general characteristics of spectral features and noise. This method produced on average a 1.79% better performing FFT window across five different types of initial window functions.