This paper investigates the impact of variations in excitation frequency variations on the dielectric properties of coal gangue to enhance identification accuracy in challenging underground conditions. A dielectric identification methodology based on frequency shift characteristics is proposed, focusing on coal gangue samples collected from Zhujidong mine in Huainan mining area, for which a dielectric identification system was designed and constructed. This system facilitates the collection of dielectric response signals from coal gangue at frequencies of 0.4 MHz and 0.5 MHz. To analyze these response signals, a Variation Mode Decomposition (VMD) combined with a median noise reduction technique was implemented. The results indicate that this approach yields optimal noise reduction was achieved, improving the signal-to-noise ratio and root mean square error metrics. The effective voltage values and frequency components of the coal gangue at the specified frequencies were extracted, leading to the development of eight identification models. Notably, the support vector machine (SVM) identification model achieved an accuracy of 98.3%, significantly outperforming K-Nearest Neighbor (KNN), Probabilistic Neural Network (PNN), and Back Propagation (BP) models. This research presents a novel approach to addressing the identification challenges associated with the separation of coal gangue in underground environments.