Speaker
Description
Gravitational-wave detectors are affected by short-duration non-Gaussian noise transients, known as glitches, which can resemble astrophysical gravitational-wave signals and reduce the quality of detector data. Reliable classification of these glitches is therefore important for gravitational-wave data analysis. In this work, we investigate a computationally efficient approach to glitch classification by combining deep learning-based feature extraction with conventional machine learning algorithms. Pre-trained ResNet-50 and Inception-v3 networks are used through transfer learning, and their extracted visual features are classified using logistic regression, extreme gradient boosting, and support vector machines. Two transfer-learning strategies are compared: fine-tuning the pre-trained deep networks on the glitch dataset and using the networks directly as feature extractors for subsequent machine learning classification. The results show that using pre-trained deep models as feature extractors can substantially reduce training time while maintaining good classification performance. The fastest configuration achieved a classification accuracy of 93.98% with a training time of 37.6 seconds. Compared with fine-tuning, the feature-extraction approach reduced training time by factors of approximately 24 and 31 for ResNet-50 and Inception-v3, respectively.These results show that combining transfer learning with lightweight machine learning classifiers can provide an efficient alternative for classifying transient noise in gravitational-wave detector data, particularly when computational cost and training time are important considerations.