EPLNet: An Efficient Convolutional Neural Network Framework Utilizing Large Kernels for Continuous Gravitational Wave Detection

Gravitational wave detection remains challenging due to low signal amplitudes often obscured by instrumental noise, particularly in low signal-to-noise ratio scenarios where conventional Convolutional Neural Networks (CNNs) exhibit limited effectiveness due to their restricted receptive fields. This paper presents an enhanced EfficientNet-based architecture for continuous gravitational wave classification that addresses these limitations through several key innovations. We transform input data into spectrograms using constant Q transformation and augment the baseline EfficientNet architecture with an additional convolutional layer featuring enlarged kernel sizes to expand the receptive field. This modification enables improved detection of weak gravitational wave signals embedded in noisy datasets while enhancing texture feature learning capabilities. Through K-fold cross-validation on a private dataset, our proposed model achieves Area Under the Curve (AUC) scores ranging from 0.878 to 0.978, demonstrating significantly enhanced sensitivity to gravitational wave signals. These results indicate substantial potential for advancing next-generation detection methodologies in astrophysical applications and improving the reliability of gravitational wave detection systems.