Abstract: The rapid development of the new energy vehicle industry has raised the requirements for bearing reliability, making defect detection increasingly critical. However, in practice, bearing defect data in new energy vehicles often suffer from imbalance. To address this issue, this paper proposes a bearing defect detection scheme based on imbalanced data. First, features are extracted from vibration signals using Empirical Mode Decomposition (EMD). Next, a dynamic firefly feature selection algorithm based on neighborhood granularity conditional entropy is employed to screen the features. Then, a natural neighborhood hypersphere oversampling method tailored for imbalanced datasets is used to augment the data. Finally, reliable results are obtained through an ensemble learning approach based on cloud models. This workflow establishes an effective system for addressing data imbalance issues. Future work may explore advanced data augmentation and model improvement strategies to accommodate more complex and variable application scenarios.