To address the significant performance degradation of anomaly detection models under the coupled scenarios of class imbalance and class overlap, this paper proposes an adaptive oversampling algorithm named Multi-Class Overlap-Guided Combined Cleaning and Resampling (MCOG-CCR), which integrates dynamic metric fusion and sphere expansion. The method achieves data distribution balancing and mitigates class overlap through a three-phase collaborative optimization mechanism. Firstly, a multi-dimensional overlap metric pool is constructed from feature-based, structural, and instance-based perspectives, where genetic algorithms dynamically optimize the combination of metrics to resolve the limitations of traditional single-dimensional metrics in capturing complex data distributions. Subsequently, an adaptive weight function is designed to dynamically adjust the contributions of metrics based on imbalance ratio (IR), dataset size (N), and comprehensive overlap degree (Overlap). Finally, tailored to the characteristics of each dataset, an improved energy-driven sphere expansion strategy is proposed: it selects an appropriate proportion of overlap regions to form a majority-class subset while generating minority-class samples, simultaneously pushing majority-class samples out of overlap zones to reduce overlap severity. Experiments on 4 challenging datasets from the ICS repository demonstrate that MCOG-CCR outperforms seven state-of-the-art methods. The average improvements in MAUC and AUCm reach 2.76% and 3.52%, respectively.