| Abstract: | The detection of gravitational waves has revolutionized astrophysics, providing a novel way
to observe cosmic events such as binary black hole mergers and neutron star collisions. This
project aims to develop a robust machine-learning model for classifying gravitational wave
signals using the Gravity Spy dataset from Kaggle. The dataset contains 31,868 images
divided into 22 categories, including both true gravitational wave events and various types
of noise interference.
Our objective is to accurately distinguish between genuine gravitational wave signals and
noise, while minimizing false positives and missed detections. To address this, we designed a
comprehensive preprocessing pipeline to manage class imbalance, ensuring that the model
is trained on a representative sample of signals. We also applied feature extraction
techniques to emphasize key waveform characteristics. Through extensive experimentation,
deep neural networks (DNNs) were identified as the most effective approach for
differentiating between noise and true gravitational wave signals.
The model was evaluated using metrics like accuracy, precision, and F1 score, focusing on
optimizing performance through careful tuning of parameters. In addition, the project
explored real-time detection enhancements, essential for astronomers to observe cosmic
events more quickly and with greater precision.
Ultimately, transfer learning models were incorporated to further boost performance,
leveraging pre-trained networks to refine the classification of gravitational waves. This work
not only advances automated gravitational wave classification but also lays the groundwork
for improving real-time detection systems, which is critical for the growth of multi-messenger
astrophysics. |
