Embedded one-class classification on RF generator using Mixture of Gaussians

In this paper we apply a specific machine learning technique for classification of normal and not-normal operation of RF (Radio Frequency) power generators. Pre-processing techniques using FFT and bandpower convert time-series system signatures into single feature vectors. These feature vectors are modeled using k-component Mixture of Gaussians (MoG) where components and corresponding parameters are learned using the Expectation Maximization (EM) algorithm. Data is obtained from three different generator models operating under normal and multiple different not-normal conditions. Exploration into algorithmic parameter effects is conducted and empirical evidence used to select sub-optimum parameters. Robust testing is reported to achieve a 3s classification accuracy of 95.91% for the targeted RF generator. Additionally, a custom C++ library is implemented to utilize the learned model for accurate classification of time-series data within an embedded environment such as a RF generator. The embedded implementation is reported to have a small storage footprint, reasonable memory consumption and overall fast execution time.