Automatic classification pipeline for glitches in the Virgo detector

Tiago Fernandes; Francesco Di Renzo; Antonio Onofre; Alejandro Torres-Forné; José A. Font

gr-qcApr 15, 2026

Automatic classification pipeline for glitches in the Virgo detector

Authors: Tiago Fernandes, Francesco Di Renzo, Antonio Onofre, Alejandro Torres-Forné, José A. Font

Organizations: Centro de Física das Universidades do Minho e do Porto (CF-UM-UP), Universidade do Minho, 4710-057 Braga, Portugal · Università di Firenze, I-50019 Sesto Fiorentino, Firenze, Italy · INFN, Sezione di Firenze, I-50019 Sesto Fiorentino, Firenze, Italy · Observatorio Astronómico, Universitat de València, C/ Catedrático José Beltrán 2, 46980, Paterna (València), Spain

Abstract

Glitches frequently contaminate data in gravitational-wave detectors, complicating the observation and analysis of astrophysical signals. This work introduces VIGILant, an automatic pipeline for classification and visualization of glitches in the Virgo detector. Using a curated dataset of Virgo O3b glitches, two machine learning approaches are evaluated: tree-based models (Decision Tree, Random Forest and XGBoost) using structured Omicron parameters, and Convolutional Neural Networks (ResNet) trained on spectrogram images. While tree-based models offer higher interpretability and fast training, the ResNet34 model achieved superior performance, reaching a F1 score of 0.9772 and accuracy of 0.9833 in the testing set, with inference times of tens of milliseconds per glitch. The pipeline has been deployed for daily operation at the Virgo site since observing run O4c, providing the Virgo collaboration with an interactive dashboard to monitor glitch populations and detector behavior. This allows to identify low-confidence predictions, highlighting glitches requiring further attention.

Figures & tables

Figure 1: Distribution of the Gravity Spy labels for Virgo O3b glitches. There are 22 classes, with a highly imbalanced representation.

Figure 2: UMAP projection of the curated glitch dataset, before and after the re-labeling of some glitches. The re-labeling grouped two sets of closely related glitch families into two single families.

Figure 3: Distribution of the labels for our Virgo O3b glitch dataset. There are 7 classes, with less imbalance than in the original data.

Class	Training	Validation	Testing
Scattering	2800 (35.7%)	600	600
Resonance_Line	1400 (17.9%)	300	300
Tomte	1261 (16.1%)	270	270
Blip	810 (10.3%)	173	173
No_Glitch	700 (8.9%)	150	150
Extremely_Loud	656 (8.4%)	141	141

Table 1: Number of samples of each class in the three subsets of the Virgo O3b glitch dataset. Percentages (shown for the training set) indicate the fraction of each subset represented by the given class. Stratified sampling ensures that the class distribution is the same across all subsets.

Figure 4: Examples of the different glitch classes in our Virgo O3b dataset.

Figure 5: Scheme of the construction of the encoded134 view. The three grayscale spectrograms with different time windows are stacked along the colour dimension.

Figure 6: Confusion matrices for the validation dataset obtained with the best-performing configuration of each model (Decision Tree, Random Forest, XGBoost, and ResNet). The DT baseline shows several misclassifications, particularly for the minority Koi_Fish class and between some class pairs. RF and XGBoost improve the classification performance, although some confusion between morphologically similar classes remains. The ResNet model provides the best overall results, with fewer misclassifications across all classes and a clear improvement for Koi_Fish and previously problematic class pairs.

Figure 7: Feature importance of the Omicron parameters for each of the best tree-based models. While peak frequency is the primary predictor for all models, the ensemble methods (RF and XGBoost) exhibit a more distributed feature importance, reducing the dominance of peak frequency in favor of a more balanced contribution from SNR, amplitude, and duration.

Figure 8: Glitches from the validation dataset where the ResNet predictions differed most from the ground truth. In most cases, the ground-truth labels appear incorrect, and the model predicts either the correct label or a more appropriate label.

model	F1 score	accuracy	$t_{\textrm{train}}$ (s)	$t_{\textrm{inf}}$ (s)
Decision Tree	0.856	0.903	0.07	$<0.01$
Random Forest	0.893	0.924	1.86	$<0.01$
XGBoost	0.901	0.930	4.38	0.01
ResNet	0.971	0.975	940	0.04

Table 5: Performance of the models in the validation dataset. The table reports the F1 score and accuracy, together with the training time (

t_{\textrm{train}}

) and the CPU inference time per glitch (

t_{\textrm{inf}}

). The ResNet achieves the highest performance by a large margin, at the cost of significantly longer training time, while maintaining a comparable inference time.

Figure 9: Confusion matrix obtained for the best ResNet model, for the held-out test dataset. The model shows precision and recall above 0.95 for every class.

Figure 10: Distribution of the maximum softmax probabilities (confidence) of the ResNet model for the test dataset. The threshold used to identify low-confidence glitches is also shown. The model presents a clear skew to very high confidence levels.

Figure 11: Confusion matrix obtained for the best ResNet model, for the test dataset, after applying the threshold to separate low-confidence glitches into a dedicated Low_Confidence class. This increases the precision of the other classes while slightly reducing recall, and highlights the small fraction of glitches (2.7%) where the model is less certain.

Figure 12: Examples of glitches from two out-of-domain classes used to test the generalization of the ResNet model. The first class, Air_Compressor , shows short thick lines around 50 Hz, while the second class, 32_Hz_Blobs , contains mostly round shapes centered at 32 Hz.

Figure 13: Distribution of the maximum softmax probabilities (confidence) for the ResNet model, for the two out-of-domain classes. Glitches with confidence below the threshold are flagged as Low_Confidence , corresponding to 38% of Air_Compressor glitches and 29% of 32_Hz_Blobs .

Figure 14: Confusion matrix for the out-of-domain glitches, before and after applying the threshold on the model confidence. After thresholding, Air_Compressor glitches are mostly classified as Scattering or No_Glitch , while 32_Hz_Blobs are labeled as Tomte or Scattering .

Figure 15: Example plots from the glitch dashboard for March 22, 2020. The top left panel shows the glitchgram, with the glitch time in the $x$ -axis and the peak frequency in the $y$ -axis; marker size encodes SNR, and marker color indicates the glitch family. The top right panel shows the daily glitch count bar plot, reporting the number of glitches per family on that day. Finally, the bottom panel shows the monthly glitch heatmap for March 2020, summarizing daily glitch counts for each family. This highlights unusual behavior, such as March 16, which had relatively few glitches overall but a disproportionately high number of Resonance_Line glitches. An interactive version of the dashboard is accessible at https://tttiago.github.io/glitchgram/dashboard_demo.html .

Gravitational-Wave Astronomy

Gravitational-wave astronomy uses ripples in spacetime to detect cosmic events, measure compact objects, and test gravity.

See all 20 papers

MADGRAV: a multilevel anomaly-detection pipeline for gravitational-wave searches applied to LIGO data
Sep 30, 2026Gianluca Inguglia, Huw Haigh, Ulyana Dupletsa +1Gravitational-Wave Astronomy Anomaly Detection
Marietta Blau Institute for Particle Physics, Austrian Academy of Sciences, Vienna, Austria · Universit`a degli Studi di Urbino ‘Carlo Bo’, I-61029 Urbino, Italy · INFN, Sezione di Firenze, I-50019 Sesto Fiorentino, Firenze, Italy
Testing General Relativity Through Gravitational Wave Classification: A Convolutional Neural Network Framework
May 4, 2026Lavinia Heisenberg, Shayan Hemmatyar, Hector Villarrubia-RojoGravitational-Wave Astronomy Binary Classification
Institute for Theoretical Physics, Heidelberg University, Philosophenweg 16, 69120 Heidelberg, Germany · Departamento de F´ısica Te´orica and IPARCOS, Facultad de Ciencias F´ısicas, Universidad Complutense de Madrid, Plaza de las Ciencias 1, 28040 Madrid, Spain
A Detection-Gated Pipeline for Robust Glottal Area Waveform Extraction and Clinical Pathology Assessment
Mar 2, 2026Harikrishnan Unnikrishnan, Rita PatelMedical Image Segmentation Endoscopy
¹Orchard Robotics, San Francisco, California, USA. · ²Department of Speech, Language, and Hearing Sciences, Indiana University, Bloomington, Indiana, USA. · ³Department of Otolaryngology–Head and Neck Surgery, Indiana University School of Medicine, Indianapolis, Indiana, USA.

Sep 30, 2026J/K move · Enter open · S save

MADGRAV: a multilevel anomaly-detection pipeline for gravitational-wave searches applied to LIGO data

Gianluca Inguglia, Huw Haigh, Ulyana Dupletsa, Alessandro Longo

Marietta Blau Institute for Particle Physics, Austrian Academy of Sciences, Vienna, Austria · Universit`a degli Studi di Urbino ‘Carlo Bo’, I-61029 Urbino, Italy · INFN, Sezione di Firenze, I-50019 Sesto Fiorentino, Firenze, Italy

Gravitational-Wave Astronomy Anomaly Detection

Open paper AskPDF arXiv

We present the results of MADGRAV, a deep-learning-based search for high-mass compact binary coalescences, applied to the data collected by the LIGO interferometers during the third observing run and during the first and second part of the fourth observing run. The MADGRAV pipeline consists of a series of sequential convolutional neural networks that perform anomaly detection, glitch classification, coherence testing, and signal ranking. Data from the Hanford and Livingston LIGO detectors are studied (both individually and in coherence) by way of 1 second Q-transform windows. Of the candidates that survive every stage of the pipeline, 48 reach the significance threshold, and we report 47 gravitational wave detections characterised by a false alarm rate below

1\,{\rm yr}^{-1}

with a probability of astrophysical origin

p_{\rm astro}>0.9

. Of the 47 detections, 44 are shared with the minimally modelled coherent WaveBurst search. The observed total source-frame masses, extracted from official gravitational wave transient catalogues, are in the

14-236 M_{\odot}

range with a median of

69 M_{\odot}

, and a median SNR of 16. We note that the recovered fraction of confident detections rises with mass: for LIGO detectors network SNR

>10

the pipeline recovers

8.1\%

of confident catalog events below

30 M_{\odot}

39.8\%

between

30

and

100 M_{\odot}

, and

53.3\%

above

100 M_{\odot}

, corresponding to

33.3\%

45.5\%

and

53.3\%

of the events detected by coherent WaveBurst in the same bins. These results suggest that anomaly detection pipelines can serve as an independent detection channel complementary to matched filtering in the high-mass high-SNR regime.

Abstract

Gravitational-Wave Astronomy

Explore similar work

MADGRAV: a multilevel anomaly-detection pipeline for gravitational-wave searches applied to LIGO data

Testing General Relativity Through Gravitational Wave Classification: A Convolutional Neural Network Framework

A Detection-Gated Pipeline for Robust Glottal Area Waveform Extraction and Clinical Pathology Assessment