A general extractor for coherent patterns across time-frequency measurements

TokEye presents a signals-first self-supervised framework for the automated extraction of coherent and transient modes from high-noise time-frequency data across various sensors This framework enables real-time mode identification and large-scale automated database generation for advanced plasma control

Signal Taxonomy Development

The authors propose a taxonomy of generally observed signal types based on the physical behavior of nonstationary time series and diagnostics This classification categorizes signals into five distinct classes: coherent, quasi-coherent, transient, broad, and stochastic measurements Coherent and quasicoherent modes share a general category since they both involve analysis of a localized time-frequency signature just with added uncertainty for quasicoherent structures. Transient and broad events can be thought of as another more general category where measurements of interest span a wide range of frequencies. Stochastic measurements can be thought of as uncertain variance.

Data Preparation Pipeline

The process developed involves several steps to clean and enhance the data before mode extraction The pipeline includes:

1. Automated baseline identification for unified broadband turbulence and TF-impulse separation without explicit labels.

2. A non-linear multichannel cross-spectrum inspired self-supervised signal enhancement.

3. Robust thresholding for spectrograms in Section 2.5.

4. Label correction in Section 2.6.

Signal Separation Techniques

Separating broadband observations from coherent signals is achieved by estimating the baseline V(f, t) and subtracting it to isolate coherent structures from the background This baseline estimation is done by solving an optimization problem involving a smoothness parameter λ controlling the second-order derivative regularization term. Furthermore, separating coherent signals from stochastic noise is addressed by extending multichannel correlation using a non-linear deep learning estimator, specifically a U-Net. This network learns optimal nonlinear combinations of the input components to reconstruct the coherent target.

Model Training and Validation

A surrogate model is trained on the processed results, initially using a U-Net trained on coherent and transient portions of the spectra The training incorporates spectrogram-targeted augmentations like SpecAugment to improve robustness. To deal with parameter uncertainty, multi-period multi-scale surrogate training on bicubic resampled labels is employed. The inference latency of 0.5 seconds enables real-time mode identification.

Generalizability and Performance

The framework demonstrates strong performance across multiple diagnostic types, including ECE, CO2, MHR, and BES measurements on DIII-D Benchmarking on TJ-II stellarator data yielded a recall of 0.825 on expert labeled spectrograms. The model also shows generalizability to non-fusion datasets, achieving a recall of 0.7708 on the DCLDE dataset. Training a high quality surrogate semantic segmentation model takes around 12 hours.

Future Directions

Future improvements include combining the denoising run with a scheme closer to Self-inspired Noise-to-Noise for full coverage The authors plan to include information on phase and turbulence in future additions Extending the method to extract inherent correlation between diagnostics in one pass is a planned extension. The performance on CPUs for variable size input spectra can be investigated through quantization and pruning.

Acknowledgments

This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Fusion Energy Sciences The authors also gratefully acknowledge financial support from the Princeton Laboratory for Artificial Intelligence under Award 2025-97. Special thanks to Marie Roch for providing access to the DCLDE dataset and annotations Andres Bustos and Enrique Zapata-Cornejo for providing access to the TJ-II dataset. Thanks to Jalal Butt, SangKyeun Kim, Ricardo Shousha, and Frederik Simons.

References

[1] J. S. Kim, D. H. Edgell, J. M. Greene, E. J. Strait, and M. S. Chance

[58] Marie A Roch, T Scott Brandes, Bhavesh Patel, Yvonne Barkley, Simone BaumannPickering, and Melissa S Soldevilla

[41] Yeong Il Jang, Keuntek Lee, Gu Yong Park, Seyun Kim, and Nam Ik Cho

[59] Weisong Zhao

--- Page 1 ---

TokEye presents a signals-first self-supervised framework for the automated extraction of coherent and transient modes from high-noise time-frequency data across various sensors; This framework enables real-time mode identification and large-scale automated database generation for advanced plasma control

This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Fusion Energy Sciences The authors also gratefully acknowledge financial support from the Princeton Laboratory for Artificial Intelligence under Award 2025-97 Special thanks to Marie Roch for providing access to the DCLDE dataset and annotations Andres Bustos and Enrique Zapata-Cornejo for providing access to the TJ-II dataset Thanks to Jalal Butt, SangKyeun Kim, Ricardo Shousha, and Frederik Simons.

[1] J. S. Kim, D. H. Edgell, J. M. Greene, E. J. Strait, and M S Chance

[58] Marie A Roch, T Scott Brandes, Bhavesh Patel, Yvonne Barkley, Simone BaumannPickering, and Melissa S Soldevilla

[41] Yeong Il Jang, Keuntek Lee, Gu Yong Park, Seyun Kim and Nam Ik Cho

[59] Weisong Zhao

[1] J S Kim D H Edgell J M Greene E J Strait and M S Chance

[58] Marie A Roch T Scott Brandes Bhavesh Patel Yvonne Barkley Simone BaumannPickering and Melissa S Soldevilla

[41] Yeong Il Jang Keuntek Lee Gu Yong Park Seyun Kim and Nam Ik Cho

[59] Weisong Zhao

The authors propose a taxonomy of generally observed signal types based on the physical behavior of nonstationary time series and diagnostics This classification categorizes signals into five distinct classes: coherent, quasi-coherent, transient, broad, and stochastic measurements Coherent and quasicoherent modes share a general category since they both involve analysis of a localized time-frequency signature just with added uncertainty for quasicoherent structures. Transient and broad events can be thought of as another more general category where measurements of interest span a wide range of frequencies. Stochastic measurements can be thought of as uncertain variance

Separating broadband observations from coherent signals is achieved by estimating the baseline V(f, t) and subtracting it to isolate coherent structures from the background This baseline estimation is done by solving an optimization problem involving a smoothness parameter λ controlling the second-order derivative regularization term. Furthermore, separating coherent signals from stochastic noise is addressed by extending multichannel correlation using a non-linear deep learning estimator, specifically a U-Net This network learns optimal nonlinear combinations of the input components to reconstruct the coherent target.

1. Correcting data labeling bottlenecks by developing a signals-first self-supervised framework that automatically creates labeled datasets across a wide range of different diagnostics, including incorporating an automated baseline identification for unified broadband turbulence and TF-impulse separation without explicit labels. This allows AI models to be trained on richer, more generalized data rather than being biased towards well-modeled phenomena.

2. Enabling real-time mode identification and large scale automated database generation with an inference latency of 0.5 seconds, which allows for advanced plasma control applications in fusion facilities like ITER by enabling real-time mode identification.

3. Improving generalizability across different fusion devices by deploying the model on diverse data, achieving a recall of 0.825 on expert labeled TJ-II spectrograms without retraining and a recall of 0.7708 on the DCLDE dataset for marine mammal detection, demonstrating performance beyond fusion domains.

4. Enhancing signal separation in complex diagnostics by implementing baseline removal using an optimization problem defined in equation 5 to isolate coherent structures from broadband background, effectively flattening the signal such that broadband structures are an independent set of measurements while coherent modes and stochastic noise are preserved.

5. Suppressing stochastic noise through a multichannel cross-spectrum inspired self-supervised signal enhancement, which uses a U-Net to map input components to target signals by minimizing mean absolute error, allowing the network to learn optimal nonlinear combinations of the input components analogous to RxRy + IxIy.

6. Automating candidate event segmentation by employing an anomaly detection problem using the knee point of the cumulative distribution function of power spectrograms, which is parameter-free and relies on the shape of the intensity distribution rather than absolute values.

7. Improving segmentation accuracy by training a U-Net on 5-fold randomized single channel measurements with data augmentations including flips and elastic deformation, ensuring that the final database of segmented plasma events has higher fidelity by capturing edge effects.