FPH2D: a failure prediction model for heterogeneous hard drives based on feature similarity

With the increasing scale and heterogeneity of data centers, accurate failure prediction in mixed hard drive models poses critical challenges. Existing studies based on single-model SMART datasets perform well in homogeneous settings but suffer degradation in mixed-model scenarios due to inconsistent attribute distributions and failure patterns across models. To address this, we propose FPH2D, a fusion prediction model for multi-model hard drives. Dynamic Time Warping (DTW) measures inter-model feature similarity, and a feature alignment method maps heterogeneous data into a unified space to reduce bias. Raw SMART features are preserved to avoid normalization-induced distortion. The model is trained using an improved LightGBM with Enhanced Gradient-based One-Side Sampling (E-GOSS) for enhanced accuracy and stability. Experimental results show optimal performance with a 16–20-day time window, a relatively high failure rate, a count ratio of 0.7–1.0, and DTW similarity of 0.28–0.36, balancing feature sufficiency and sample representativeness. Experiments on ST12000NM0007 and ST14000NM0138 drives demonstrate strong generalization, achieving average AUC of 0.9984, Precision 0.9926, Recall 0.9054, and F1-score 0.9479, enabling reliable early failure prediction and supporting maintenance optimization for heterogeneous storage systems.