Abstract: Hard disk is the primary storage device in cloud data centers,and hard disk failure prediction is crucial for ensuring data security.However,there is a significant imbalance between failure and healthy SMART samples,which can lead to model bias.Moreover,hard disk models exhibit varying data distributions.Prediction models trained on specific hard disk data may not be suitable for other hard disks.To address these issues,this paper proposes a method for out-of-distribution hard disk failure prediction by combining the AP clustering algorithm and the broad learning system.To tackle the sample imbalance problem,this paper uses the AP clustering algorithm to cluster samples close to failure and treats all samples in the cluster containing determined failure instances as additional failure samples.To address the distribution differences of hard disk models,this paper combines the manifold regularization framework and the broad learning system to learn the low-dimensional structure of hard disk data,thereby improving the model’s generalization ability to unknown data.Experimental results show that,on the dataset resampled by the AP clustering algorithm,the F1_Score of multiple methods increases by an average of 0.2 compared to the datasets resampled by comparative methods.Additionally,in the task of predicting out-of-distribution hard disk failures,the F1_Score of the proposed model increases by 0.1~0.2 compared to other methods.

Key words: Hard disk failure prediction, Class imbalance, Out-of-distribution generalization, Affinity propagation clustering, Broad learning system, Manifold learning