Clustering of periodic multichannel timeseries data with application to plasma fluctuations

Abstract A periodic datamining algorithm has been developed and used to extract distinct plasma fluctuations in multichannel oscillatory timeseries data. The technique uses the Expectation Maximisation algorithm to solve for the maximum likelihood estimates and cluster assignments of a mixture of multivariate independent von Mises distributions (EM-VMM). The performance of the algorithm shows significant benefits when compared to a periodic k-means algorithm and clustering using non-periodic techniques on several artificial datasets and real experimental data. Additionally, a new technique for identifying interesting features in multichannel oscillatory timeseries data is described (STFT-clustering). STFT-clustering identifies the coincidence of spectral features over most channels of a multi-channel array using the averaged short time Fourier transform of the signals. These features are filtered using clustering to remove noise. This method is particularly good at identifying weaker features and complements existing methods of feature extraction. Results from applying the STFT-clustering and EM-VMM algorithm to the extraction and clustering of plasma wave modes in the time series data from a helical magnetic probe array on the H-1NF heliac are presented.