Schematic of the proposed dual-scale neural network architecture for general trigger classification. The model processes input light curves through two parallel pathways: a long-scale encoder (blue) and a short-scale encoder (orange), each composed of four consecutive 2D convolutional units (Conv Units 1–4 and 5–8, respectively) for hierarchical feature extraction. The extracted feature vectors, f_L and f_S, are fed into the central CSFF. The zoomed-in view details the CSFF’s operation: a 1D convolutional layer models cross-scale interactions, generating attention maps which adaptively weight and combine f_L and f_S to produce a unified representation f_F. This fused feature is aggregated via global average pooling (GAP) and finally classified by a fully connected (FC) layer with dropout. This end-to-end architecture enables the joint analysis of complementary temporal information.