DiskTransformer: A Transformer Network for Hard Disk Failure Prediction

The hard disk drive is one of the most commonly damaged components in data centers, which can cause a lot of losses such as unexpected shutdowns and data loss. Therefore, the prediction of hard disk drive failures has received widespread attention in data center management. Existing work has made remarkable progress in the accuracy of failure prediction. However, the prediction performance for long-term failure and small-sample disks is not satisfactory. To address this issue, we propose a new method for hard disk failure prediction. The framework consists of a time-series feature extraction network and a prediction network. The time-series feature extraction network is composed of a Temporal Convolutional Network used to extract different dimensional relationships and a set of Long Short Term Memory networks used to extract independent dimensional time-series features. The prediction network uses a Transformer model, which can fully utilize the high-quality fused features extracted by the time-series feature extraction network for disk drive failure prediction. Experimental results on public datasets have demonstrated that our proposed method can not only predict long-term failure but also has reliable prediction performance when facing small-sample disk data.