<p>Granular computing(GrC), as an information processing method with both interpretability and robustness, has been widely used in classification scenarios. However, most of the existing granule-based methods are offline, and it is difficult to handle data streams with high real-time requirements. In this paper, an online self-organizing learning framework based on hypersphere granules is proposed, including four parts: granule generation, parameter update, detection and elimination of conflicting granules, and identification and correction of noisy granules. A vital feature of the proposed method is that the system can generate or update the granules online with very limited known information by judging the neighborhood granules of the newly arrived data. In order to achieve better performance in complex environments, a novel granules quality measure based on conflict and noise detection is introduced to help SOHGBS find and eliminate semantically contradictory granules, as well as identify abnormal granules caused by noisy data, and make reasonable correction. This helps maintain a healthier granular knowledge base, ensuring that models remain highly interpretable and robust. Very importantly, the fuzzy rules generated by the existing granules for classification decisions not only take the distance factor into account, but also consider the geometric characteristics of the granules, enabling the system to perform accurate classification. Compared with the offline granular computation methods, classical machine learning classification algorithms, as well as the existing state-of-the-art online data stream processing methods, the proposed approach achieves good performance in terms of classification accuracy, stability, robustness, and time efficiency.</p>