VIGILant: an automatic classification pipeline for glitches in the Virgo detector

This paper introduces VIGILant, an automatic pipeline deployed at the Virgo detector that utilizes a high-performing ResNet34 model to classify and visualize gravitational-wave glitches, enabling real-time monitoring and improved data quality analysis.

The Gist

VIGILant: The "Bouncer" for the Virgo Detector

Imagine the Virgo detector as a super-sensitive microphone in a quiet room, trying to hear the faintest whisper of two black holes colliding billions of light-years away. This is the job of gravitational wave astronomy.

But here's the problem: The room isn't quiet. It's full of noise. A truck drives by, a door slams, a fan vibrates, or the detector itself has a hiccup. In the world of physics, these annoying, sudden bursts of noise are called "glitches."

These glitches are like static on a radio or a fly buzzing in a recording studio. They are so loud and frequent that they often drown out the real cosmic signals or trick scientists into thinking they heard a black hole collision when it was just a truck passing by.

This paper introduces VIGILant (Virgo Glitch Identification and Learning), a new, automatic system designed to act as a smart bouncer for the Virgo detector. Its job is to look at the noise, figure out what kind of noise it is, and tell the scientists, "Ignore this, it's just a glitch," or "Pay attention to this, it might be real."

---

The Problem: Too Much Noise, Too Many Types

Before VIGILant, scientists had a tool called "Gravity Spy" that tried to sort these glitches. But Gravity Spy was like a bouncer who only knows the rules for a different club. It was trained mostly on data from the American LIGO detectors. Since Virgo (in Italy) is built differently, the "glitches" there look different.

Imagine trying to sort a pile of mixed-up toys using a guidebook written for a different set of toys. The bouncer might confidently say, "That's a teddy bear!" when it's actually a robot. This overconfidence was dangerous because it could lead scientists to waste time studying fake signals.

The Solution: VIGILant's Two Brains

To fix this, the researchers built VIGILant with two different "brains" (machine learning models) to see which one was better at sorting the noise:

1. The "Stats Brain" (Tree Models):

   • This brain looks at a list of numbers describing the glitch (how loud it is, how long it lasted, what frequency it was).
   • It's like a detective looking at a police report. It asks simple questions: "Is it loud? Yes. Is it short? Yes. Therefore, it's a 'Blip'."
   • Pros: It's fast and easy to understand.
   • Cons: It misses the big picture. It's like trying to identify a song just by reading the sheet music without hearing the melody.

2. The "Visual Brain" (ResNet/CNN):

   • This brain looks at pictures of the glitches. Scientists turn the sound data into "spectrograms," which look like colorful heatmaps or sonograms (like the ones you see in movies for bat echolocation).
   • This is like a human art expert looking at a painting. They don't just look at the colors; they see the shapes, the patterns, and the flow.
   • Pros: It's incredibly good at spotting subtle patterns that the "Stats Brain" misses.
   • Cons: It takes longer to "learn" (train), but once it's ready, it works instantly.

The Results: The Visual Brain Wins

The researchers tested both brains on a massive dataset of 11,000 glitches from the Virgo detector.

• The Stats Brain was decent, getting about 90% of them right.
• The Visual Brain (ResNet) was a superstar, getting 98% right.

The Visual Brain was so good that it could even tell the difference between glitches that looked almost identical to the human eye. It learned that a "Whistle" glitch (caused by a vibrating mirror) looks slightly different from a "Scattered Light" glitch (caused by a laser bouncing off dust), even if their numbers were similar.

How It Works in Real Life

VIGILant isn't just a one-time experiment; it's now working every single day at the Virgo site. Here is its daily routine:

1. The Morning Scan: Every night, it grabs all the noise data from the previous day.
2. The Sorting: It runs the data through its Visual Brain.
3. The Dashboard: It updates a colorful, interactive map (a dashboard) that scientists can look at.
   • The Map: Shows every glitch as a dot on a graph. The color tells you what kind of glitch it is (e.g., Red for "Scattering," Blue for "Tomte"). The size tells you how loud it was.
   • The "Low Confidence" Flag: Sometimes, the AI sees a glitch and thinks, "I've never seen this before, or I'm not sure." Instead of guessing, it flags it as "Low Confidence." This is like the bouncer saying, "I don't know who this person is, let's check their ID manually." This helps scientists find new types of glitches they didn't know existed.

Why This Matters

Think of the Virgo detector as a gold miner.

• The gold is the real signal from black holes.
• The dirt and rocks are the glitches.

Before VIGILant, the miners were sifting through the dirt by hand, often picking up rocks and thinking they were gold. Now, VIGILant is a high-tech sieve that automatically separates 98% of the rocks from the gold.

This allows the scientists to:

• Trust the data more: They know exactly what kind of noise they are dealing with.
• Fix the detector: If they see a sudden spike in "Scattering" glitches, they know to check the lasers.
• Find new gold: By flagging the weird, unknown glitches, they can discover new problems with the machine or even new types of cosmic events.

In short, VIGILant is the tireless, super-smart assistant that keeps the Virgo detector's ears clean, ensuring that when we finally hear the universe whisper, we know it's the real thing and not just a fly buzzing in the room.

Technical Summary

1. Problem Statement

Gravitational wave (GW) detectors, such as Virgo, are plagued by transient non-Gaussian noises known as "glitches." These artifacts, originating from environmental, anthropogenic, or instrumental sources, frequently contaminate data, mimicking astrophysical signals or overlapping with genuine events. This complicates detection and parameter estimation.

While the LIGO-Virgo-KAGRA (LVK) collaboration utilizes tools like Gravity Spy for glitch classification, these tools face significant limitations when applied to the Virgo detector:

• Domain Mismatch: Gravity Spy's training corpus consists almost entirely of LIGO glitches. Due to substantial hardware and sensitivity differences between LIGO and Virgo, Virgo-specific glitch populations often differ, leading to out-of-distribution data.
• Overconfidence: The existing ML classifiers often produce high-confidence predictions even when incorrect, which can mislead detector characterization studies.
• Lack of Automation: Previous efforts to characterize Virgo glitches (e.g., GWitchHunters) were not integrated into a continuously deployed, automated pipeline.

2. Methodology

The authors developed VIGILant (Virgo Glitch Identification and Learning), an automated pipeline that evaluates two distinct machine learning approaches for classifying Virgo O3b glitches:

A. Dataset Construction

• Source: Virgo O3b observing run data (Nov 2019 – Mar 2020).
• Preprocessing: The authors re-evaluated the original Gravity Spy labels. Through visual inspection and UMAP dimensionality reduction, they identified overlapping classes (e.g., Scattered Light, Low Frequency Burst, Paired Doves) and merged them into broader categories.
• Final Classes: The dataset was reduced from 22 to 7 classes:
  1. Scattering (merged from 5 original classes)
  2. Resonance Line (merged Whistle and Violin Mode)
  3. Tomte
  4. Blip
  5. No Glitch
  6. Extremely Loud
  7. Koi Fish
• Size: ~11,200 glitches, split into 70% training, 15% validation, and 15% testing (stratified sampling).

B. Feature Representation

Two input modalities were tested:

1. Structured Parameters: 6 Omicron parameters (peak frequency, SNR, amplitude, central frequency, duration, bandwidth).
2. Spectrogram Images: Q-transform representations of the glitches. To capture multi-scale morphologies, three spectrograms (0.5s, 2.0s, and 4.0s windows) were stacked into a single "encoded" image (170x140 pixels, 3 channels).

C. Machine Learning Models

1. Tree-Based Models (Structured Data):
   • Algorithms: Decision Tree (DT), Random Forest (RF), and XGBoost.
   • Goal: Evaluate interpretability and performance on tabular Omicron data.
2. Convolutional Neural Networks (Image Data):
   • Architecture: ResNet family (ResNet18, ResNet34, ResNet50).
   • Training: Used AdamW optimizer with a 1-cycle learning rate policy. The best performing architecture was selected via hyperparameter tuning.

3. Key Contributions

• VIGILant Pipeline: The first fully automated, daily-operational pipeline for Virgo glitch classification, deployed since the O4c observing run.
• Curated Virgo Dataset: A re-labeled, high-quality dataset specifically for Virgo, addressing the domain shift issues of previous LIGO-centric models.
• Interactive Dashboard: A visualization tool providing real-time monitoring of glitch populations, detector behavior, and low-confidence predictions via an interactive "glitchgram," daily counts, and monthly heatmaps.
• Confidence Thresholding: Implementation of a confidence threshold (set at the 2.5th percentile of validation scores) to flag "Low Confidence" glitches. This allows operators to identify ambiguous or novel glitches that require manual review, mitigating the risk of overconfident misclassifications.

4. Results

The models were evaluated using the Macro-averaged F1 score and Accuracy.

Model	F1 Score (Val)	Accuracy (Val)	Training Time	Inference Time (CPU)
Decision Tree	0.856	0.903	0.07 s	< 0.01 s
Random Forest	0.893	0.924	1.86 s	< 0.01 s
XGBoost	0.901	0.930	4.38 s	0.01 s
ResNet34	0.971	0.975	940 s	0.04 s

• Performance: The ResNet34 model significantly outperformed tree-based models, achieving a test set F1 score of 0.9772 and accuracy of 0.9833.
• Generalization: The model struggled with "out-of-domain" glitches (e.g., Air Compressor and 32 Hz Blobs), correctly flagging ~57% and ~40% of these as "Low Confidence," respectively. This demonstrates the utility of the confidence threshold for identifying novel glitch types.
• Error Analysis: Most misclassifications were attributed to label noise in the ground truth rather than model failure. The model often predicted a more appropriate label than the original ground truth.
• Efficiency: Despite longer training times, the inference time is extremely fast (~40ms per glitch), making it suitable for low-latency operations.

5. Significance

• Operational Impact: VIGILant is currently running daily at the Virgo site, providing the collaboration with critical insights into detector health and noise sources.
• Improved Data Quality: By identifying low-confidence predictions, the pipeline helps operators focus on glitches that may represent new noise sources or require further investigation, improving the overall data quality for GW searches.
• Future Roadmap: The authors propose extending VIGILant to include unsupervised clustering for the discovery of new glitch classes and advanced reconstruction methods (e.g., Sparse Dictionary Learning) for glitch subtraction, moving from mere classification to active noise mitigation.

In summary, VIGILant represents a significant step forward in detector characterization, successfully adapting deep learning techniques to the specific needs of the Virgo detector and providing a robust, automated framework for managing the complex noise environment of gravitational wave astronomy.