Detecting Gravitational Waves using Constant-Q Transform and Convolutional Neural Networks

Dingyun Zhang

Abstract: The discovery of gravitational waves from the mergers of binary black holes has opened doors to an unprecedented revolution in the fields of physics and astronomy. However, the signals of gravitational waves with tiny magnitudes are inevitably buried in detector noise, leading to a great demand for the accurate analysis of gravitational wave data. In this paper, based on a set of time-series data containing simulated gravitational waves in a Kaggle competition, we propose a deep learning method by combining constant-Q transform and convolutional neural networks (CNNs), to achieve a promising performance for the detection of gravitational waves. In our method, the gravitational wave signal is firstly transformed into a spectrogram by the constant-Q transform, and is subsequently classified by the CNN network. In particular, EfficientNet-B3 and EfficientNetV2-L are both utilized as the CNN backbones to extract features from spectrograms. After applying an ensemble average of two backbones and the K-Fold cross validation technique, our model reaches an AUC score 0.8786 on the private test set. This result ranks top 5% (63/1219) in the Kaggle leaderboard, and can get a bronze medal in the G2Net Gravitational Wave Detection competition. This work would help increase the sensitivity of interferometers to gravitational wave signals, and potentially accelerate the development of next-generation detectors to explore the Universe.