A Heuristic-IRM Method on Hard Disk Failure Prediction in Out-of-distribution Environments

Abstract

The hard disk drives (HDD) are essential devices lying in primary layers of diverse information infrastructure. Long-term disk failure predictions are crucial to the stability and robustness of storage systems for data centers. In this paper, a domain adaption method is developed to improve prediction performance in out-of-distribution disk datasets. We propose heuristic invariant risk minimization (HIRM) with a new loss function to deal with imbalanced data. The HIRM combined with machine learning models are verified to promote the accuracy and stability in out-of-distribution (OoD) data. When hard disks with new SMART feature distribution are introduced into the data center, the proposed HIRM algorithm achieves better results than vanilla neural networks. A numerical example using the data from the BackBlaze data center is shown to illustrate the application of our HIRM model. The aims of each person are different.