Unsupervised Machine Learning for Blind Separation of Multiple PD Sources

Abstract

We tested the ability of unsupervised machine learning approaches to separate an unknown number of artificial PD sources simultaneously acting, recorded by a multi-antenna system in the Ultra-High Frequeny range. Each pulse has been described with a minimum set of features and successively undergone to an unsupervised clustering procedure via Hierarchical Agglomerative Clustering (HAC) technique. The results have been compared, in a fully blind performance test, with those obtained from a cross-correlation analysis of the fully sampled pulses that clearly separates three groups of pulses radiated from the three PD sources. A limited set of mono-dimensional features (maximum amplitude, standard deviation, phase, power) shows a cluster detection accuracy of 58%. A different set of features (including maximum amplitude, standard deviation, phase, power, skewness value, kurtosis value), shows an accuracy of 64% while performance parameters as Precision and F1-Score show better results. Despite the sub-optimal results, the false-negative detection rate is low and the computational cost of this simplified approach is dramatically lower than the cross-correlation approach, thus allowing possible development in real-time supervising devices.