Abstract: To address the issue of sample imbalance in the operational status assessment of ultra-high voltage converter valve cooling systems, this paper proposes a DB-MAHAKIL data balancing method that integrates density-based clustering and genetic sampling. Based on this method, an XGBoost status assessment model incorporating the gray wolf optimization algorithm (GWO) is constructed. First, multi-dimensional feature quantities reflecting the operational status of the valve cooling system are extracted. Feature weights are calculated using association rules, and system operational status levels are defined according to status evaluation guidelines. Second, outliers are removed using the density-based noise space clustering (DBSCAN) algorithm, and new samples within the boundary are generated using the MAHAKIL method to reduce sample distribution bias. Finally, a GWO-optimized XGBoost model is constructed to achieve accurate identification of various operational states. Through validation with actual measurement data, the proposed method significantly outperforms traditional methods in identifying severe state samples, with an overall model accuracy rate of 98.4%.