Hard Disk Drives Failure Prediction Using the Deep Learning Method Based on Attention Mechanism

Hard disk Drives failure is the main cause of data center downtime. Utilizing predictive techniques to extend drives’ remaining useful life (RUL) can significantly reduce data loss. Machine learning and deep learning methods are exploited to address these issues, and Long Short-Term Memory (LSTM) network has achieved good results. However, traditional LSTM networks are classified only by the features learned at the last step, resulting in poor performance. In addition, these techniques are susceptible to the high imbalance in existing datasets. This paper proposes an RUL prediction framework for hard disks based on the attention mechanism. By using the LSTM network, it can learn sequence features from the raw data. Meanwhile, the proposed attention mechanism can automatically learn the importance of time steps and estimate the health of hard disks based on their failure time. Experimental results on real datasets demonstrate that our proposed method can achieve better prediction performance than the existing methods.