Incorporation of RGB Data into Gravitational Wave Data
The pursuit of detecting gravitational waves, as predicted by Einstein's General Relativity in 1915, has been significantly advanced by the Laser Interferometer Gravitational Wave Observatory (LIGO) at Caltech. To distinguish these elusive waves from Earth's noisy background, neural networks are employed, trained on time-frequency and Signal-to-Noise Ratio (SNR) data. However, the heavy memory load posed by the traditional .mat file format necessitates multiple, time-consuming training sessions. This project proposes a novel approach—converting the data into three-channel .png files, significantly reducing file size while maintaining data integrity. Machine learning models, including U-Net, Residual Unet (ResUnet), and ResUnet++, were trained on these png files, demonstrating comparable performance to models trained on .mat data. Despite a slight numerical discrepancy in quantitative evaluations, the gains in memory efficiency and training speed outweigh this difference. The png files, at 0.023% of the size of .mat files, enable training on nearly 1500 images per trial. The achieved balance between model accuracy and reduced training costs suggests that the adoption of png data for preprocessing in gravitational wave detection is a promising and efficient alternative.