Temporal Dynamic Frequency Autoencoder: Enhancing the Performance of Unsupervised Dimensionality Reduction for Gravitational Wave Glitch Clustering

With the ongoing advancements in gravitational wave detection technology, the results are becoming increasingly precise. However, it is precisely this high precision that has led to the detection of noise signals that should not have been recorded, which are then mixed with real gravitational wave signals. As a result, efficient and accurate algorithms are needed to distinguish between these signals and noise. Past research has mostly relied on fully supervised or semi-supervised learning methods to classify and cluster these glitches, thereby improving the accuracy and reliability of glitch detection. However, neural networks typically require a large amount of training data to perform well, and in future work, it is impractical to manually label a vast number of experimental samples. Additionally, as time progresses, new unlabeled glitches may emerge. To address this issue, we adopted a Temporal Dynamic Frequency Autoencoder (TDFAE) based on a Kolmogorov-Arnold Network (KAN) that integrates Vision Transformers (ViT) and a four-branch encoder. To enhance feature extraction accuracy, we utilized a multi-branch fusion approach with the CLS (Class) token. This innovative unsupervised learning technique focuses on dimensionality reduction and clustering of gravitational wave glitches. Our approach, tested on the GravitySpy O3 dataset, outperforms current semi-supervised learning methods in clustering tasks. Additionally, our work on unsupervised clustering methods for traditional LIGO datasets offers significant potential, representing a key advancement in gravitational wave research.