No Labeled Data or GPU Needed: DINOv2-Powered Unsupervised Morphological Analysis System for Gravitational Wave Signals

The dante-gravi-signal-ml project proposes a fully unsupervised morphological analysis framework that uses frozen DINOv2 visual features to perform unsupervised clustering on LIGO/Virgo gravitational wave data. It requires no labeled data or GPU acceleration, can efficiently process observation data using only a CPU, automatically identifies astrophysical chirp signals and instrumental noise glitches, and provides a new signal screening paradigm for gravitational wave astronomy.

Gravitational wave detection is one of the cutting-edge fields in modern astronomy, but instrumental noise glitches mixed in massive datasets have long plagued researchers. Traditional methods rely on manually labeled training data, which is time-consuming and labor-intensive, and it's difficult to cover all noise types. Thus, a more efficient solution is urgently needed.

Data Preprocessing

The system obtains raw data from GWOSC, splits it into 32-second windows, generates time-frequency spectrograms via Q-transform after whitening and band-pass filtering, and uses the perceptually uniform cividis colormap to reduce visual artifacts.

DINOv2 Feature Extraction

It uses Meta's open-source DINOv2 with Registers encoder, with frozen weights (no fine-tuning needed). Register tokens clean the embedding representations, and the CLS token output is L2-normalized to get a 384-dimensional vector.

Two-Stage UMAP Clustering

First stage: PCA → UMAP (10 dimensions) → DPMM clustering; Second stage: UMAP (2 dimensions) for visualization. DPMM combined with cosine metric avoids density bias and outperforms HDBSCAN.

The project establishes a multi-level validation system:

• Similarity checker: Morphological validation based on KNN cosine similarity
• Ablation study: Stability assessment under perturbations via ARI
• Hyperparameter robustness: ARI consistency test across hyperparameter ranges
• Time sliding test: Calculation of p-values for coincident events between H1 and L1 detectors These methods ensure that clustering results reflect real physical differences rather than artifacts or parameter sensitivity.

The project has real-time processing capabilities:

• Threshold calibrator: Class-by-class threshold calibration based on intra-class cosine similarity distribution
• Real-time scanning: Producer-consumer mode for classifying 4096-second data blocks The classifier labels signals as "KNOWN" (known type) or "NOVEL" (potential novel pattern). When the number of NOVEL samples exceeds the threshold, it recommends initiating in-depth analysis.

The project documentation points out current limitations:

• UMAP distance distortion may cause abnormal cluster separation to reflect preprocessing artifacts
• The transfer of DINOv2 from natural images to gravitational wave spectrograms relies on heuristic validation
• Fixed Q-transform windows may mask high-frequency transients or slow broadband structures
• The ARI with Gravity Spy manual labels is low, as morphological features differ from human classification conventions
• Processing the full O4a dataset with pure CPU takes hours to days This reflects a rigorous scientific attitude.

The project brings important value to gravitational wave astronomy:

• Zero labeling requirement lowers the threshold for discovering unknown noise
• Pure CPU operation is suitable for deployment in resource-constrained institutions
• Isolated session mechanism ensures reproducibility of runs
• Open-source Apache 2.0 license and detailed documentation promote community reproduction and expansion.

dante-gravi-signal-ml focuses on interpretable morphological analysis, does not rely on expensive computing resources, achieves efficient processing through an ingenious architecture, and honestly faces limitations to point the way for subsequent improvements. It has important reference value for researchers in time-series signal analysis, scientific data mining, or anomaly detection.