eess.SPFeb 23, 2026

TokEye: Fast Signal Extraction for Fluctuating Time Series via Offline Self-Supervised Learning From Fusion Diagnostics to Bioacoustics

Authors: Nathaniel Chen, Kouroche Bouchiat, Peter Steiner, Andrew Rothstein, David Smith, Max Austin, Mike van Zeeland, Azarakhsh Jalalvand, 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Next-generation fusion facilities like ITER face a "data deluge," generating petabytes of multi-diagnostic signals daily that challenge manual analysis. We present a "signals-first" self-supervised framework for the automated extraction of coherent and transient modes from high-noise time-frequency data across a variety of sensors. We also develop a general-purpose method and tool for extracting coherent, quasi-coherent, and transient modes for fluctuation measurements in tokamaks by employing non-linear optimal techniques in multichannel signal processing with a fast neural network surrogate on fast magnetics, electron cyclotron emission, CO2 interferometers, and beam emission spectroscopy measurements from DIII-D. Results are tested on data from DIII-D, TJ-II, and non-fusion spectrograms. With an inference latency of 0.5 seconds, this framework enables real-time mode identification and large-scale automated database generation for advanced plasma control. Repository is in https://github.com/PlasmaControl/TokEye.

Figures & tables

Figure 1: Signal taxonomy with example modes and spectra.
Figure 2: Signal processing pipeline for ECE shot 178631 demonstrating progressive separation of coherent modes from broadband background and stochastic noise. (left) Raw STFT power spectra. (center) Welch periodogram, which is the time averaged spectrogram. (right) Thresholding on power spectra.
Figure 3: (left) Averaging signals can introduce a bias that removes information about each individual signal. (right) Reconstructing each channel with the information from all other channels leads to a better estimate of the true signal.
Figure 4: Blind-spot denoising example with AT-BSN, a more efficient form of AP-BSN. (top) Original spectra with small kernel. (bottom) denoised spectra with large kernel. Although noise is reduced in both cases, the small kernel method still retains largely observable measurements of noise, while the large kernel method removes fluctuations especially at the top.
Figure 5: (Top) Original denoised spectra. (Middle) CDF with threshold set at corner. (Bottom) Thresholded spectra.
Figure 6: The automated data extraction pipeline and surrogate model training process. For real-time usage, the surrogate model can automatically extract vertical and coherent observations from signals.
Figure 7: U-Net architecture used for all three models in this paper. For self-supervised training, the input and output are multichannel depending on the number of samples used. For robust label expansion, the input and output is one single channel. For surrogate model training, the input is single channel and the output is two channel for binary segmentation.
Figure 8: (top) example magnetic spectrogram for shot 170008 with only coherent mode extraction. (middle) individually extracted events using a model threshold. (bottom) gated spectrogram, which captures amplitudes of interest. transient events here can be filtered out, allowing direct coherent mode extraction.
Lblfmin⁡f_{\min}fmax⁡f_{\max}tmin⁡t_{\min}tmax⁡t_{\max}Amp.
 (kHz)(kHz)(ms)(ms)(dB)
176818340965335
328963762358546
21991721295179642
43295231829881026
462643011137131632
Table 1: Database of detected signal regions.
Figure 9: (top) example CO 2 density spectrogram for shot 185781. (middle) the individually extracted events using a model threshold. (bottom) gated spectrogram, which captures amplitudes of interest. Different events including alfven-like and low frequency modes can be captured, along with very fine chirping structures.
Figure 10: (top) ECE spectrogram segmentation in a tearing mode shot 199597. (middle) individually extracted events using a model threshold. (bottom) gated spectrogram, which captures amplitudes of interest. high frequency modes appear starting at 2 seconds, rapidly ending near 3.4 seconds when tearing mode occurs.
Figure 11: (top) ECE spectrogram segmentation in a tearing-suppressed shot 199607. (middle) individually extracted events using a model threshold. (bottom) gated spectrogram, which captures amplitudes of interest. high frequency modes appear starting at 2 seconds, and do not diminish as tearing mode is suppressed.
Figure 12: Example of TJ-II ECE spectrogram with expert annotations and model output. The threshold used here is the recommended 50% threshold.
Figure 13: Example of DCLDE 2011 dolphin call spectrogram with expert annotations (top) and model output (bottom) at the recommended 50% threshold.
Figure 14: Example of DCLDE 2011 with missed overlap. While the general segmentation region is in the correct vicinity, there are only three pixels of overlap between annotation and model output.
More in this topic

Time-Frequency Analysis

See all 17 papers

Explore similar work

CardsList

  1. TokaMind: A Multi-Modal Transformer Foundation Model for Tokamak Plasma Dynamics

    Feb 16, 2026Tobia Boschi, Andrea Loreti, Nicola C. Amorisco +13Multimodal Foundation ModelsMultimodal Pretraining

    IBM Research Dublin, Ireland · UK Atomic Energy Authority Abingdon, UK · STFC Hartree Centre Daresbury, UK

  2. Self-Supervised Noise2Noise-Enhanced Denoising for Continuous-Scan Air-Plasma THz Spectroscopy

    Aug 17, 2026Adam Umra, Oways Alsoloh, Oliver Nagy +2Self-Supervised Image Denoising

    Institute of Digital Communication Systems, Ruhr University Bochum, Germany · Photonics and Ultrafast Laser Science, Ruhr University Bochum, Germany

  3. Benchmarking Sensor Robustness in Plasma Diagnostic Models: A Systematic Evaluation on TokaMark

    Jul 5, 2026Neerav GuptaIncomplete Data Imputation

    Independent Researcher