Time-Series Classification for Plasma Diagnostics
Methods for labeling multichannel, non-stationary plasma signals in real time and offline.
The data
A tokamak discharge produces dozens to hundreds of synchronized signals sampled at kilohertz to megahertz rates. Many tasks reduce to classifying windows of this multichannel time series: disruptive versus safe, confinement mode, instability type.
Feature-based versus end-to-end
Feature-based pipelines compute physics-motivated quantities (mode amplitudes, gradients, radiated-power fractions) and feed them to a classifier. End-to-end deep models learn features directly from raw signals using convolutions over time or recurrent state. Feature-based methods are more interpretable and data-efficient; end-to-end methods can capture patterns humans did not engineer.
- 1D convolutional networks over time windows
- Recurrent networks (LSTM, GRU) for long dependencies
- Temporal convolutional networks with dilations
- Transformers with attention over time steps
Non-stationarity
Plasma statistics change through a shot: ramp-up, flat-top, ramp-down each look different. A classifier must either be conditioned on the phase or robust to it. Normalization that uses only past data avoids leaking future information into a causal, real-time model.
Causality and latency
Real-time classifiers may only use samples up to the present, and must produce an output within a fixed latency budget set by the control loop. This rules out models that need the full shot, and favors architectures with bounded per-step compute.
Evaluation
Split by shot, not by time slice, so windows from one discharge do not appear in both training and test sets. Report per-class metrics because plasma-signal classes are usually imbalanced. Where labels come from human experts, measure label noise and its effect on the ceiling of achievable accuracy.