DCGAN with Quadratic Potential for Hard Disk Failure Prediction in Operational Data Center

Hard disk failure is one of the most common storage system failures, and hard disk drive (HDD) manufacturers are actively reducing the failure rate to improve data center security by reducing the incidence of HDD failures in all types of operational data centers. At present, the research on intelligent prediction of hard disk failure is mainly based on traditional machine learning methods as well as neural networks, but it still faces the problems of extreme data imbalance, limited label availability and low fault detection accuracy. In this paper, an unsupervised adversarial learning method-Deep Convolution Generative Adversarial Networks with Quadratic Potential (DCGAN-QP) is proposed to improve the accuracy of HDD failure prediction, with labels that are incomplete, inaccurate, and learns anomaly feature representation to enhance the effect of weakly supervised anomaly detection. The failure prediction experiment using the Self-Monitoring and Reporting Technology (SMART) of the hard disk verifies the effectiveness and robustness of the proposed method, which achieves higher recall, precision and accuracy than the traditional supervised and semi-supervised methods.