Machine Learning based Glitch Veto for inspiral binary merger signals using Linear Chirp Transform

Technical Summary: Machine Learning based Glitch Veto for Inspiral Binary Merger Signals using Linear Chirp Transform

Problem Statement
Transient non-Gaussian noise artifacts, known as "glitches," pose a significant challenge in gravitational wave (GW) detection. These artifacts can mimic genuine compact binary coalescence (CBC) signals, leading to increased false-alarm rates in detection pipelines. While matched filtering is the standard technique for identifying inspiral-merger signals, certain glitches possess characteristics close enough to merger waveforms to bypass initial veto mechanisms. Existing veto methods, such as the traditional Chi-square test or those based on unphysical waveform templates, face limitations regarding the diversity of glitch classes (approximately 22 distinct types) and reliance on predefined parameter spaces that may not capture the full range of observed anomalies. There is a critical need for robust discrimination techniques that can accurately distinguish astrophysical signals from instrumental noise to ensure the reliability of GW observations.

Methodology
The authors propose a hybrid deep learning framework that combines advanced signal processing with neural network classification. The methodology proceeds through the following stages:

1. Feature Extraction via Linear Chirp Transform (LCT):
   Unlike conventional Fourier transforms which analyze time and frequency, the authors utilize the Joint Chirp Rate Time Frequency Transform (JCTFT), a generalization incorporating a linear chirp rate parameter ($\gamma$). This transforms the 1D strain time series into a three-dimensional "chirp-volume" spectrogram spanning time, frequency, and chirp-rate. This approach is motivated by the nature of inspiral signals, where frequency evolves rapidly (chirps) as objects merge. The LCT captures this frequency evolution more effectively than standard time-frequency representations, particularly for signals with low signal-to-noise ratios (SNR).

2. Data Preparation:
   The dataset comprises confirmed CBC events and glitch samples from the LIGO O1–O4 observing runs (Hanford and Livingston detectors).

   • Glitches: 708 instances sourced from GWOSC, clipped to 6-second windows centered on the Time of Arrival (TOA).
   • Merger Signals: 215 confirmed events (430 instances including both detectors), also 6 seconds in duration.
   • Preprocessing: The raw LCT output generates large 3D arrays (approx. 1.2 GB per signal). To make this tractable for deep learning, the authors downsample the 61 chirp-rate slices (each originally $471 \times 6000$) to $64 \times 128$ using linear interpolation. This reduces the storage requirement to ~1.9 MB per signal while preserving the dominant spectro-temporal morphology. The data is normalized to the range $[0, 1]$.

3. Classification Architecture:
   A hybrid Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) architecture with an attention mechanism is employed:

   • CNN Layers: Three blocks of convolutional layers (with 32, 64, and 128 filters respectively) extract local spectro-temporal features from each 2D chirp-rate slice.
   • GRU Layers: A two-layer GRU processes the sequence of 61 feature vectors (one per chirp-rate slice), modeling correlations and the evolution of signal morphology across the chirp domain.
   • Attention Mechanism: An attention pooling layer assigns learnable weights to each chirp-rate slice, allowing the network to focus on the most informative regions of the chirp volume.
   • Output: A fully connected head with a softmax activation outputs the probability of the signal being a merger or a glitch. The model is trained using Focal Loss to handle class imbalance and difficult examples.

Key Contributions

• Application of LCT to Glitch Veto: The paper introduces the use of the Linear Chirp Transform as a feature extraction tool specifically for glitch classification, moving beyond standard time-frequency representations to include the chirp-rate dimension.
• Hybrid Deep Learning Framework: The development of a CNN-GRU-Attention architecture that simultaneously exploits local spectrogram structures and long-range correlations across the chirp domain.
• Efficient Data Representation: A preprocessing pipeline that reduces massive 3D chirp-volume datasets (from ~1.2 GB to ~1.9 MB) without losing the discriminative features necessary for classification.
• Empirical Validation: The study demonstrates that chirp-domain representations provide highly discriminative information, outperforming traditional methods in separating merger signals from glitches.

Results
The proposed model was evaluated on an independent validation set comprising 308 signals (20% of the total dataset).

• Performance: The model achieved a best validation accuracy of 95.45%. After threshold tuning (optimal threshold 0.44), the overall classification accuracy reached approximately 96%.
• Metrics: Both merger and glitch classes achieved comparable precision and recall values of roughly 95–96%. The macro-averaged F1-score was approximately 0.96.
• Error Analysis: Out of 308 validation samples, only 13 were misclassified (6 glitches and 7 mergers), indicating a balanced performance with no significant bias toward either class.
• Efficiency: The trained model classifies a signal in approximately 30 ms on standard hardware, while the chirp transform and preprocessing take about 32.7 seconds.

Significance and Claims
The authors claim that this work demonstrates the potential of combining chirp-based signal processing with deep learning to improve glitch mitigation in current and future GW observatories. The study suggests that the chirp-volume representation contains highly discriminative information that allows for the separation of merger signals from glitches, even when glitches mimic merger waveforms.

The paper positions this method as a complementary veto technique to be applied after the matched filtering stage to validate candidate events. While the model was trained on LIGO data, the authors suggest the underlying logic of signal separation in the chirp domain could be applicable to future detectors (e.g., Einstein Telescope, Cosmic Explorer, LISA) and other signal processing domains with varying frequency profiles (e.g., RADAR). The authors remain modest regarding future extensions, noting that deriving exact analytical solutions for higher-order or fractional chirp terms is a daunting task but could further enhance signal processing capabilities. The work is presented as a step toward eradicating false alarms and improving the accuracy of GW detection pipelines.