A data mining algorithm for automated characterisation of fluctuations in multichannel timeseries

We present a data mining technique for the analysis of multichannel oscillatory timeseries data and show an application using poloidal arrays of magnetic sensors installed in the H-1 heliac. The procedure is highly automated, and scales well to large datasets. The timeseries data is split into short time segments to provide time resolution, and each segment is represented by a singular value decomposition (SVD). By comparing power spectra of the temporal singular vectors, singular values are grouped into subsets which define fluctuation structures. Thresholds for the normalised energy of the fluctuation structure and the normalised entropy of the SVD are used to filter the dataset. We assume that distinct classes of fluctuations are localised in the space of phase differences between each pair of nearest neighbour channels. An expectation maximisation clustering algorithm is used to locate the distinct classes of fluctuations, and a cluster tree mapping is used to visualise the results.

💡 Research Summary

The paper introduces a fully automated data‑mining workflow designed to identify and classify fluctuating structures in multichannel time‑series data, with a concrete demonstration on magnetic probe arrays installed in the H‑1 heliac stellarator. The authors begin by segmenting the continuous recordings into short, overlapping windows (typically on the order of a few milliseconds) to retain temporal resolution while keeping each segment short enough for stationary analysis. For every window they compute a singular‑value decomposition (SVD) of the N‑channel data matrix, yielding spatial singular vectors (U), temporal singular vectors (V) and singular values (Σ). The temporal vectors are transformed to the frequency domain, and their power spectra are compared using a similarity metric (e.g., cosine similarity). Singular values whose associated spectra are sufficiently alike are grouped together, forming what the authors call a “fluctuation structure”.

Two quantitative filters are then applied to each structure: (1) a normalized energy, defined as the sum of the squared singular values in the structure divided by the total energy of the window, and (2) a normalized entropy, derived from the distribution of singular‑value contributions. Structures with low energy or high entropy are discarded as either noise‑dominated or overly mixed modes. This dual‑threshold approach automatically removes spurious components without manual inspection.

The remaining structures are mapped into a low‑dimensional feature space built from phase differences between nearest‑neighbour sensor pairs. For each structure the phase of the corresponding temporal singular vector is evaluated at each channel; the differences Δφij = arg(v(i)) – arg(v(j)) for all adjacent pairs are concatenated into a feature vector. The authors argue that distinct physical modes (e.g., Alfvén waves, drift‑waves) occupy localized regions in this phase‑difference space because the spatial phase pattern encodes wavelength and propagation direction.

Clustering in the phase‑difference space is performed with an Expectation‑Maximisation (EM) algorithm that fits a Gaussian mixture model (GMM). The number of clusters is selected by minimizing the Bayesian Information Criterion (BIC) while also checking the Akaike Information Criterion (AIC) for robustness. Each fluctuation structure is assigned to the cluster with the highest posterior probability. To visualise the hierarchical relationships among clusters, the authors construct a “cluster tree map”: nodes represent clusters, edges encode similarity (e.g., Kullback‑Leibler divergence) or temporal adjacency, and the tree layout reveals how modes split, merge, or evolve over the course of the discharge.

Applying the method to 10 seconds of data from 64 magnetic probes, the authors identify five to seven dominant clusters. Each cluster exhibits a characteristic frequency band (e.g., 5–10 kHz, 15–20 kHz) and a consistent phase‑difference signature that matches previously reported plasma eigenmodes. The cluster tree shows that certain clusters bifurcate abruptly at times that coincide with known plasma transitions (such as L‑H mode changes), suggesting that the algorithm can capture dynamic mode evolution.

Key strengths of the approach include: (i) scalability – the SVD and EM steps have computational costs that grow modestly with the number of channels, making the pipeline suitable for very large sensor arrays; (ii) automation – the energy/entropy thresholds and the EM clustering require minimal human tuning; (iii) physical interpretability – phase‑difference features directly relate to wave‑vector information, facilitating a clear link between statistical clusters and underlying plasma physics.

Limitations are also acknowledged. The choice of window length influences frequency resolution and the ability to resolve rapid transients; too short a window smears spectral peaks, while too long a window averages over non‑stationary events. The EM algorithm assumes Gaussian clusters, which may be violated if the phase‑difference distribution is multimodal or highly skewed, potentially leading to over‑ or under‑segmentation. Moreover, SVD is inherently linear, so strongly nonlinear interactions (e.g., mode coupling, turbulence cascades) may not be fully captured. The authors suggest future extensions such as wavelet‑based time‑frequency analysis, nonlinear dimensionality reduction (t‑SNE, UMAP), or deep‑learning autoencoders to address these shortcomings.

In summary, the paper presents a comprehensive, end‑to‑end framework that combines SVD‑based dimensionality reduction, spectral similarity grouping, energy‑entropy filtering, phase‑difference feature extraction, and EM‑driven Gaussian mixture clustering, capped by a hierarchical visualisation. The methodology demonstrates high reproducibility, efficient handling of large multichannel datasets, and clear physical insight, making it a valuable tool not only for fusion plasma diagnostics but also for any field that deals with high‑dimensional, oscillatory time‑series data such as geophysics, neuroscience, or structural health monitoring.