Interpreting decisions of lightweight CNN for glitch classification in time frequency images: Enhancing black-hole merger detection in O1 and O2 runs

Abstract

When two or more black holes orbit each other and eventually merge, they emit significant gravitational waves (GWs). These waves, first detected by the Laser Interferometer Gravitational-Wave Observatory (LIGO), are recorded as time-frequency patterns in images. The classification and segmentation of these images are critical for conducting precise and thorough characterization analysis. A robust signal-versus-glitch classifier designed for automation must be capable of handling diverse background noise, new glitch sources, overlapping glitches, and gravitational wave signals. This challenge becomes increasingly complex, particularly in the context of lightweight devices. This study introduce a custom CNN architecture with 1,899,318 trainable parameters, effectively classifying 22 categories in the Gravity Spy dataset (time-frequency domain visualization of glitch behavior), even with the challenging nature of glitches, and showing strong resistance to adversarial attacks. Under 10-fold cross-validation, the proposed model achieves training, validation, and test precisions of 96.02 ± 0.2790%, 96.54 ± 0.2840%, and 97.12 ± 0.5240% respectively. Cohen’s Kappa score is 0.978 ± 0.0038, while the Mann–Whitney and Kruskal–Wallis tests yield 3214.59 ± 1.44 and 0.155 ± 0.003, respectively. The minimum AUC-ROC is 0.98 for only two categories—‘No Glitch’ and ‘None of the Above’ — whereas all other categories have higher values. To enhance interpretability, local interpretable model-agnostic explanations are generated as helping information for the decision-making process, and the model is deployed in a web environment using Flask to evaluate the performance of the CNN model when implemented on lightweight devices.