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

The paper introduces TokEye, a fast, self-supervised deep learning framework that enables real-time, automated extraction of coherent and transient modes from noisy multi-sensor time-series data in fusion facilities and other domains, addressing the challenge of massive data volumes in next-generation experiments like ITER.

The Gist

Imagine you are trying to listen to a single violinist playing a beautiful, complex melody in the middle of a massive, chaotic rock concert. The crowd is screaming, the drums are booming, and the bass is rattling your teeth. To the untrained ear, it's just a wall of noise. But you need to find that specific violin melody, understand how it changes, and know exactly when it starts and stops.

This is the daily challenge for scientists studying fusion energy (the same power that fuels the sun). They use giant machines called tokamaks to create super-hot plasma. These machines are covered in sensors that record millions of data points every second. The problem? The data is a "data deluge." It's a mix of the "violin" (important plasma signals) and the "rock concert" (noise, turbulence, and random glitches).

Currently, scientists have to manually sift through this noise, which is like trying to find a needle in a haystack while wearing blindfolds. It takes forever, and they often miss the important stuff.

Enter TokEye. Think of TokEye as a super-smart, AI-powered "noise-canceling headphone" and "music translator" rolled into one. Here is how it works, broken down into simple steps:

1. The "Signal-First" Approach: Listening to the Music, Not the Crowd

Instead of trying to guess what the plasma is doing based on old theories, TokEye starts by looking at the raw sound waves (signals) themselves. It treats the data like a musical score.

• The Analogy: Imagine a messy room. Some things are furniture (the important signals), some are dust bunnies (random noise), and some are giant, messy piles of laundry (broad turbulence). TokEye doesn't just sweep the whole room; it first separates the laundry piles from the furniture, then cleans the dust off the furniture.

2. Step One: Removing the "Background Hum" (Baseline Removal)

Fusion data often has a low, rumbling hum (turbulence) that drowns out the quiet, important signals.

• The Analogy: Imagine you are trying to hear a whisper in a room where the air conditioner is roaring. TokEye first figures out the exact pattern of the air conditioner's roar and subtracts it from the recording. Suddenly, the room is quiet, and the whisper is clear. This is called baseline removal.

3. Step Two: The "Group Chat" Trick (Self-Supervised Learning)

This is the coolest part. TokEye doesn't need a teacher to tell it what is noise and what is signal. It learns by itself.

• The Analogy: Imagine you have 10 friends recording the same concert from different spots in the crowd.
  • Friend A hears the drums loud but the violin quiet.
  • Friend B hears the violin loud but the drums quiet.
  • Friend C hears mostly the crowd noise.
  • If you ask Friend A to guess what Friend B heard, Friend A can use the information they share to "fill in the blanks."
• TokEye does this with sensors. It looks at 10 different sensors, asks them to predict what the others are hearing, and uses that to figure out what is real (the signal) and what is just random static (noise). It's like a group of detectives cross-referencing their notes to solve a mystery without ever having seen the crime scene before.

4. Step Three: The "Smart Highlighter" (Thresholding)

Once the noise is gone, TokEye needs to decide: "Is this a real event or just a tiny glitch?"

• The Analogy: Imagine you have a photo of a starry night. Most of the photo is black (darkness), and there are a few bright stars. TokEye looks at the photo and finds the "knee" of the curve—the point where the darkness suddenly turns into starlight. It highlights everything above that line. It doesn't need to be told what a star looks like; it just knows that stars are significantly brighter than the background.

5. The Result: A Super-Fast "Surrogate"

After learning from thousands of hours of data, TokEye builds a tiny, fast version of itself (a surrogate model).

• The Analogy: Think of this like a student who studied for a final exam for 12 hours, then took a 30-second nap and woke up ready to take the test instantly.
• Speed: It can analyze a full experiment (a "shot") in 0.5 seconds. That's faster than it takes to blink. This means it can be used in real-time to help control the fusion reactor, preventing it from crashing before the scientists even have time to react.

Why Does This Matter?

• For Fusion: It helps scientists understand the "moods" of the plasma. Is it about to explode? Is it stable? TokEye spots these changes instantly, helping us get closer to clean, limitless energy.
• For Everything Else: The authors tested TokEye on whale calls and dolphin clicks (bioacoustics) and it worked great there too! It's a universal tool for finding patterns in noisy data, whether that's in a nuclear reactor, the ocean, or even a stock market chart.

In short: TokEye is a fast, self-teaching AI that turns a chaotic wall of noise into a clear, readable story about what's happening inside a fusion reactor, allowing scientists to finally hear the "violin" over the "rock concert."

Technical Summary

1. Problem Statement

Next-generation fusion facilities, such as ITER, face a "data deluge," generating petabytes of multi-diagnostic signals daily. Analyzing these signals manually is impossible due to:

• High Noise Levels: Signals are inundated with stochastic noise, integrated noise, and overlapping transient burst events that obscure faint but physically significant coherent modes (e.g., Alfvén eigenmodes, tearing modes).
• Complexity of Signal Types: Fusion diagnostics contain superpositions of coherent modes, quasi-coherent modes, transient events, broadband turbulence, and stochastic noise. Distinguishing between these is difficult, especially when boundaries between signal and noise are blurred.
• Lack of Unified Annotation: Existing methods rely on manual post-processing, specific filterbanks for each sensor, or simulation-based datasets that introduce bias and fail to generalize to novel plasma scenarios.
• Real-Time Constraints: Current AI solutions are often computationally expensive, preventing real-time mode identification required for advanced plasma control.

2. Methodology

The authors propose TokEye, a "signals-first," self-supervised framework designed to automatically extract coherent and transient modes from high-noise time-frequency data across various sensor types without explicit manual labeling. The pipeline consists of the following stages:

A. Signal Taxonomy & Preprocessing

• Taxonomy: Signals are classified into five categories: Coherent (narrowband, e.g., MHD instabilities), Quasi-coherent, Transient (instantaneous broadband spikes, e.g., ELMs), Broad (non-stationary drifts/turbulence), and Stochastic (noise).
• Transformation: Raw time-series data is resampled to 500 kHz and converted to spectrograms using the Short-Time Fourier Transform (STFT) with a Hann window.
• Decomposition: The STFT representation $Z(t, f)$ is decomposed into Coherent ($M$), Broadband ($V$), and Stochastic ($\eta$) components.

B. Baseline Removal (Separating Broadband from Coherent)

• Challenge: Strong broadband turbulence can obscure faint coherent modes.
• Solution: An Asymmetric Least Squares (AsLS) algorithm is used to estimate and subtract the broadband baseline $V(t, f)$.
  • This treats the baseline as a "color" of the signal. By removing it, the signal is "whitened," isolating coherent structures and transient spikes.
  • A pre-emphasis filter is applied to mitigate edge effects in low-frequency bins.

C. Multichannel Self-Supervised Denoising

• Challenge: Removing stochastic white noise without averaging out faint, transient, or non-stationary signals (a failure point of linear filters like Wiener or BM3D).
• Solution: A U-Net architecture is trained using a self-supervised approach based on non-linear multichannel cross-spectrum estimation.
  • Mechanism: Instead of averaging channels (which suppresses transients), the network predicts one channel ($X_N$) using the information from all other channels ($X_1...X_k$).
  • Theory: Since noise is independent across channels while the physical signal is correlated, the network learns to reconstruct the coherent signal by minimizing the Mean Absolute Error (MAE) between the prediction and the target.
  • Stopping Criterion: Total Variation (TV) is used as a stopping criterion to prevent the network from learning residual noise during training.

D. Adaptive Thresholding

• Challenge: Standard image thresholding (e.g., Otsu) fails on spectrograms due to sparse, high-intensity signals against a dense background.
• Solution: An anomaly detection approach using the Knee Point of the Cumulative Distribution Function (CDF).
  • The algorithm identifies the "knee" where the distribution shifts from background noise to signal, providing a parameter-free global threshold.

E. Surrogate Model Training

• A final U-Net surrogate model is trained on the processed, labeled data (generated automatically by the pipeline) to perform direct, interpretable event analysis.
• Augmentation: Techniques like SpecAug and elastic deformation are used to improve robustness.
• Architecture: A modified U-Net using upsample double convolutions to prevent checkerboard artifacts.

3. Key Contributions

1. Unified Signal Taxonomy: Formalized a classification system for fusion signals (Coherent, Quasi-coherent, Transient, Broad, Stochastic) to guide processing.
2. Self-Supervised Framework: Developed a pipeline that generates its own labeled dataset from raw data, eliminating the need for manual annotation or simulation bias.
3. Non-Linear Multichannel Denoising: Introduced a neural network approach that outperforms classical linear cross-spectrum methods by preserving transient events and avoiding the suppression of faint modes.
4. Generalizability: Demonstrated that the method works across diverse sensor types (Magnetics, ECE, CO2 Interferometers, BES) and even generalizes to non-fusion domains (Bioacoustics).
5. Real-Time Capability: Achieved inference latencies suitable for active plasma monitoring.

4. Results

The framework was tested on data from DIII-D, TJ-II, and the DCLDE 2011 bioacoustic dataset.

• DIII-D Performance:
  • Successfully extracted coherent modes and transient events from Electron Cyclotron Emission (ECE), CO2 interferometers, and Magnetic High Resolution (MHR) data.
  • Case Study: In tearing mode control experiments, the model correctly identified the suppression of tearing modes and the simultaneous rise of high-frequency Alfvén-like modes, correlating with Shapley value analysis of core vs. edge measurements.
• Generalization (TJ-II):
  • Deployed on TJ-II stellarator data (different noise characteristics) without retraining.
  • Achieved a recall of 0.825 against expert annotations, demonstrating strong cross-device generalizability.
• Generalization (Bioacoustics):
  • Tested on the DCLDE 2011 dataset (marine mammal calls).
  • Achieved recalls of 0.77 (Delphinus capensis) and 0.79 (Delphinus delphis) in a zero-shot setting, proving the method's applicability beyond fusion physics.
• Performance Metrics:
  • Database Generation: Creating a training set of 5,000 shots takes ~5 hours (A100 GPU).
  • Inference Latency: 0.5 seconds per shot on GPU; 5–10 seconds on CPU. This enables real-time mode identification for advanced plasma control.

5. Significance

• Scalability: TokEye addresses the critical bottleneck of data analysis in future fusion reactors (like ITER) by automating the extraction of physics-relevant features from petabytes of data.
• AI Readiness: By generating large-scale, high-quality, automatically labeled databases, the framework provides the necessary training data for future, more complex AI models in fusion research.
• Real-Time Control: The sub-second inference time makes it feasible to integrate mode identification directly into feedback control loops for plasma stability.
• Cross-Domain Utility: The "signals-first" approach demonstrates that advanced time-frequency signal processing techniques developed for fusion can be effectively applied to other fields dealing with fluctuating time series, such as bioacoustics and seismology.