Abstract: The rapid development of the new energy vehicle industry has increased the requirements for bearing reliability, making defect detection crucial. In actual testing, the data on bearing defects in new energy vehicles is often unbalanced. To address this, a scheme for bearing defect detection based on unbalanced data is proposed. First, the Empirical Mode Decomposition （EMD） method is used to extract features from vibration signals; then, a dynamic firefly feature selection algorithm based on neighborhood granularity conditional entropy is employed for screening; subsequently, a natural neighborhood hyperspherical over-sampling method tailored for unbalanced datasets is applied to enhance the signal; finally, reliable data results are obtained through an integrated learning method based on cloud models. This process constructs a system-effective data imbalance problem-solving framework. Future research can explore advanced data enhancement and model improvement strategies to adapt to complex application scenarios.