Intent classification plays a crucial role in the development of dialogue systems. Most existing research assumes that training data is correctly labeled; however, real-world annotated data often contains noise due to the time-consuming and labor-intensive labeling process. Noisy training data degrades model representations, making noise-robust intent classification essential. Existing approaches for handling noisy data primarily rely on coarse-grained semantic representations, failing to capture the true semantic distribution. Moreover, they do not account for the uncertainty in noise recognition and label correction, instead directly classifying a sample as either clean or noisy. To address these challenges, we propose a Robust Granular Ball Three-Way Decision (RGB3WD) method. Specifically, we leverage granular ball clustering to represent samples with multiple granular balls, capturing fine-grained semantic structures within clusters. For noise recognition, we introduce a label-consistency-based three-way decision rule, categorizing samples into positive, boundary, and negative regions. Finally, we design region-specific strategies and develop a loss function to learn noise-robust representations. Extensive experiments on multiple datasets demonstrate the effectiveness of the proposed method.