ResNetKAN1D: Attention-Enhanced Hybrid Architecture for Seismic Signal Classification in STEAD

The accurate classification of seismic signal and noise is a crucial part of earthquake early warning systems. However, the complex and changeable noise poses severe challenges to signal classification. In this study, an innovative hybrid architecture named ResNetKAN1D is proposed, which combines the Residual Network (ResNet), dual attention mechanism, and Kolmogorov-Arnold Network (KAN). In this architecture, the dual attention mechanism, namely channel attention and spatial attention, is embedded in the residual blocks, which significantly enhances the ability to focus on and select key features. At the same time, the KAN layer is used to replace the traditional Multilayer Perceptron (MLP), and for the first time, KAN is applied to seismic signal classification, greatly improving the nonlinear modeling ability of the model. The dual attention mechanism dynamically enhances the key spatiotemporal features. Combined with the nonlinear modeling ability of KAN, it significantly improves the feature selection efficiency of ResNet. Experimental results on STEAD show that the ResNetKAN1D model significantly improves the effect of seismic signal classification.