Systemic health evaluation of RF generators using Gaussian mixture models

We propose an application of specific machine learning techniques capable of evaluating systemic health of a Radio Frequency (RF) power generator. System signatures or fingerprints are collected from multivariate time-series data samples of sensor values under typical operational loads. These fingerprints are transformed into feature vectors using standard scaling/translation methods and the Fast Fourier Transform (FFT). The number of features per fingerprint are reduced by banding neighboring features and Principal Component Analysis (PCA). The reduced feature vectors are used with the Expectation Maximization (EM) algorithm to learn parameters for a Gaussian Mixture Model (GMM) to represent normal operation. One-class classification of normal fingerprints is achieved by thresholding the likelihood of a fingerprint feature vectors. Fingerprints were collected from normal operational conditions and seeded non-normal conditions. Preprocessing methods and algorithmic parameters have been selected using an iterative grid search. Average robust true positive rate achieved was 94.76% and best specificity reported is 86.56%.