摘要
This study presents an interpretable machine learning (ML) framework for diagnosing and mitigating input/output operations per second (IOPS) degradation in hard disk drives (HDDs) through slot-aware placement optimization. The framework integrates decision trees (DTs), which provide model transparency, with neural networks capable of capturing nonlinear patterns, thereby enabling the identification of complex interactions between HDD performance and structural slot characteristics. A series of 36 slotreassignment experiments (12 slots & times; 3 vendors & times; 1 trial each), conducted across 12 physical positions and involving three HDD vendors (Seagate, HGST, and WD), demonstrate consistent IOPS recovery, achieving an average maximum improvement of 23.3% without hardware modification. The proposed approach facilitates proactive, data-driven drive placement planning and presents a scalable solution for improving system reliability and sustainability in large-scale data center environments.