A One-Class Anomaly Detection Method for Drives based on Adversarial Auto-Encoder

Abstract

In recent years, hard disk drives (HDDs) and solid-state drives (SSDs) are both widely deployed in data centers. As a proactive warning technology, drive anomaly detection can detect anomalies in advance so that it is of good practicability. The existing anomaly detection methods are generally based on supervised learning, which requires a large amount of abnormal data as the basis. However, in the early stage of establishment of a data center or the deployment of new devices, anomalies or faults rarely occur, resulting in serious data imbalance problem, which causes many difficulties in the use of classification algorithms. To tackle the above problems, we propose a one-class drive anomaly detection method-AAD, which combines adversarial auto-encoder (AAE) and long short-term memory (LSTM). AAD model is trained on normal data (healthy drives) only in an unsupervised way. Since the pattern of normal data and abnormal data is obviously different, AAD learns the distribution of normal data and calculates the anomaly score of input data by integrating the reconstruction error and discrimination error to determine whether the anomalies exist. The experimental results show that AAD outperforms multiple traditional one-class classifiers, and compared with the two baseline methods, the fault detection rate (FDR) and false alarm rate (FAR) of our method on HDD dataset are 32.44%/6.885% higher and 2.3%/0.06% higher, respectively. The results on SSD dataset are 22.54%/0.01% higher and 5%/-0.004% higher, respectively. AAD outperforms the baseline methods in most cases.