FinalKAN: A Dual‐Stream KAN Network Integrating Multisource Features for Radio Telescope Pointing Correction

This paper investigates the degradation of pointing accuracy in the Kunming 40‐m radio telescope due to long‐term equipment aging and environmental disturbances. Conventional linear pointing models are constrained by their linear modeling framework, making it difficult to accurately represent the nonlinear errors induced by temperature, wind speed, and gear backlash, thereby limiting high‐frequency, high‐resolution observations. We propose a dual‐stream Kolmogorov–Arnold network (FinalKAN) that integrates multisource environmental dynamics. Guided by the Kolmogorov–Arnold representation theorem, our heterogeneous dual‐stream architecture employs stream‐specific feature selection for angular and environmental inputs, a cross‐stream attention mechanism, and a multistrategy fusion module to jointly model and compensate for errors from multiple disturbance sources. Experiments on a historical dataset from the Kunming 40‐m telescope (6,734 samples, 74 radio sources, and 8 feature variables) demonstrate that FinalKAN achieves a root mean square error (RMSE) of 28. 26 ″ in azimuth and 14. 89 ″ in elevation under five‐fold cross‐validation, improving over least‐squares modeling by 21.6% and 20.9%, respectively, and outperforming baselines including extreme gradient boosting (XGBoost) and multilayer perceptron (MLP). Ablation studies confirm the contributions of the dual‐stream design, cross‐stream attention, and directional encoding. The method provides a data‐driven, interpretable solution for high‐precision pointing control in radio telescopes, with practical implications for enhancing observational performance.