LWCM: a Lookahead-Window Constrained Model for Disk Failure Prediction in Large Data Centers

Disk failure seriously affects the reliability and availability of the storage system, leading to additional maintenance costs. However, current studies on disk failure prediction have yet to address challenges such as inaccurate sample labeling, unstable performance within the lookahead window, and improper sample segmentation. Consequently, the need for more robustness and applicability of existing prediction models persists. In this paper, we propose LWCM (a Lookahead- Window-Constrained Model), which optimizes the multilayer perceptron algorithm through dynamic backpropagation with batch training, dynamic sample relabeling and reweighting techniques to improve the model performance further. Experimental results show that LWCM has better failure prediction performance. Its true positive and false positive rate surpass the existing schemes by 38.7% and 92.4%, respectively. Furthermore, LWCM demonstrates its applicability across various datasets while maintaining stability within the lookahead constraint window.