Enterprise AI Analysis
TokEye: Fast Signal Extraction for Fluctuating Time Series via Offline Self-Supervised Learning From Fusion Diagnostics to Bioacoustics
TokEye introduces a signals-first self-supervised learning framework for automated extraction of coherent and transient modes from high-noise time-frequency data. It leverages non-linear optimal techniques with a fast neural network surrogate, demonstrating strong performance across diverse sensors in fusion (DIII-D, TJ-II) and bioacoustics. With a low inference latency of 0.5 seconds on GPU, TokEye enables real-time mode identification and large-scale automated database generation for advanced plasma control, addressing the 'data deluge' challenge at next-generation fusion facilities like ITER.
Key Metrics & Impact
TokEye delivers significant operational advantages through enhanced speed and accuracy in scientific data analysis.
0s
GPU Inference Latency
0 hours
Training Time
0
Data Fragments Generated
0
TJ-II Recall Score
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Automated Data Extraction Pipeline
Acquire Signal
→
Resample (or Subsample)
→
T-F Transform (STFT, Scalogram,...)
→
Extract Vertical Observations
→
Extract Coherent Observations
TokEye's self-supervised pipeline systematically processes raw signals to extract coherent observations, as illustrated in Figure 6 of the paper. This robust framework ensures consistent and accurate feature extraction across diverse data sources, from initial signal acquisition to the final coherent event identification.
Baseline Removal Process
Raw STFT Power Spectra
→
Estimate Broadband Baseline V(t,f)
→
Subtract Baseline (Whitening)
→
Multichannel Denoising (Suppresses Stochastic Noise)
This process, detailed in Section 2.3 and Figure 2, effectively 'whitens' the signal by removing broadband background noise, revealing faint, coherent structures that would otherwise be obscured. It's critical for achieving high fidelity in mode extraction, especially in high-noise environments.
| Category | Characteristics | Examples |
|---|---|---|
| Coherent |
|
|
| Quasi-Coherent |
|
|
| Broadband (Transient & Nonstationary) |
|
|
| Stochastic |
|
|
This formalized taxonomy (Section 2.1, Figure 1) provides a systematic way to classify and process complex fusion plasma signals, enabling a more targeted and effective approach to feature extraction than traditional methods.
0.5s
GPU Inference Latency Per Full Shot
TokEye achieves a remarkable inference latency of 0.5 seconds per full shot on GPU. This near real-time performance is critical for active plasma monitoring and advanced control systems in next-generation fusion facilities where rapid decision-making is essential (Section 3.4).
12 hours
Total Training Time on GPU
A high-quality surrogate semantic segmentation model with full 5-fold validation trains efficiently in approximately 12 hours. This optimized training duration allows for rapid model updates and iteration, supporting agile development and deployment cycles (Section 3.4).
DIII-D Magnetic Spectrogram Analysis
Context: Applied to Magnetic High Resolution (MHR) diagnostics from DIII-D shot 170008, showcasing TokEye's capability in identifying plasma instabilities.
Key Findings:
- Clean mode identifications enable direct projection of mode numbers.
- Segmentation allows for direct amplitude measurements of interest.
- Transient events can be effectively filtered out, enhancing clarity for coherent mode extraction.
This demonstrates TokEye's ability to provide high-fidelity insights into complex fusion plasma behavior from fast magnetics data (Figure 8).
DIII-D CO2 Interferometer Analysis
Context: Applied to CO2 interferometer density spectrograms from DIII-D shot 185781, demonstrating its ability to handle different diagnostic types.
Key Findings:
- Clean extraction of both high and low frequency structures, including Alfvén-like and low frequency modes.
- Captures very fine chirping structures, which are often obscured in raw data.
- Robust performance despite strong low-frequency amplitudes in the raw signal.
TokEye successfully isolates critical density fluctuation phenomena, providing clearer data for analysis and control (Figure 9).
DIII-D ECE Tearing Mode Control Experiments
Context: Analysis of ECE spectrograms from DIII-D shots 199597 (tearing instability) and 199607 (ECCD suppressed), focusing on mode behavior during control experiments.
Key Findings:
- Shot 199597: High-frequency modes appear starting at 2 seconds and rapidly end near 3.4 seconds when tearing occurs.
- Shot 199607: High-frequency modes appear starting at 2 seconds but do not diminish as tearing mode is suppressed by ECCD.
- Reveals the upward movement of low-frequency modes during tearing mode suppression.
This highlights TokEye's capability to differentiate and track complex mode evolution under various plasma control scenarios, essential for understanding and optimizing fusion performance (Figures 10 & 11).
82.5%
Recall on TJ-II Spectrograms (Zero-Shot)
TokEye demonstrates strong generalizability to new fusion devices, achieving an impressive 82.5% recall on expert-labeled TJ-II stellarator ECE spectrograms without retraining. This indicates robust performance across different noise characteristics and event structures, making it highly adaptable (Section 3.2).
Benchmarking on TJ-II Stellarator Data
Context: Applied to ECE spectrograms from the TJ-II stellarator in Spain, a different fusion device from DIII-D, featuring distinct noise and event characteristics.
Key Findings:
- Achieved 82.5% recall on expert-labeled spectrograms without retraining, demonstrating strong generalizability.
- Successfully identifies structures not annotated by original labels and areas with no mode activity.
- Model can adapt to different cutoff frequencies through post-processing, showing flexibility in handling diverse data characteristics.
This validation confirms TokEye's broad applicability to different fusion research facilities and data types (Figure 12).
77.1%
Recall on Delphinus capensis (Zero-Shot)
79.5%
Recall on Delphinus delphis (Zero-Shot)
Beyond fusion, TokEye shows remarkable zero-shot transfer capabilities, achieving 77.08% recall on Delphinus capensis and 79.53% on Delphinus delphis in the DCLDE 2011 bioacoustic dataset (Section 3.3). This underscores its foundational strength in spectrogram analysis.
Bioacoustic Data from DCLDE 2011
Context: Evaluated on the Detection, Classification, Density Estimation, and Localization of Marine Mammals Using Passive Acoustics (DCLDE) 2011 Ocodonte dataset, which features localized time-frequency annotations of snapping shrimp and dolphin calls.
Key Findings:
- Achieved 77.08% recall on Delphinus capensis and 79.53% on Delphinus delphis, demonstrating strong zero-shot performance.
- Performs well even when annotations are highly localized, indicating robustness to varied labeling granularity.
- Highlights the model's applicability to other scientific domains with similar time-frequency analysis challenges.
This successful cross-domain application validates TokEye's underlying signal extraction principles as universally effective (Figures 13 & 14).
Calculate Your Potential ROI
Estimate the efficiency gains and cost savings your organization could achieve with TokEye.
Annual Savings
$0
Hours Reclaimed Annually
0
Our Implementation Roadmap
A clear path to integrating TokEye into your enterprise, ensuring a smooth transition and rapid value realization.
Phase 1: Discovery & Strategy
Detailed assessment of your existing data infrastructure, current signal analysis workflows, and specific operational challenges. We define key performance indicators (KPIs) and tailor a TokEye implementation strategy to your unique needs.
Phase 2: Integration & Customization
Seamless integration of TokEye with your diagnostic systems and data pipelines. This includes customizing the self-supervised learning framework to optimize for your specific sensor types and signal characteristics, leveraging our signals-first approach.
Phase 3: Validation & Deployment
Rigorous testing and validation of the deployed TokEye solution against your historical and real-time data. We ensure accurate mode identification, optimal latency, and provide training for your team, culminating in full operational deployment.
Phase 4: Ongoing Optimization & Support
Continuous monitoring, performance tuning, and updates to ensure TokEye evolves with your operational demands and data volume. Our dedicated support team is available to ensure peak performance and address any emerging needs.
Ready to Transform Your Data Analysis?
Schedule a free consultation to see how TokEye can revolutionize signal extraction for your complex time-series data.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat