FPTSF: A Failure Prediction of Hard Disks Based on Time Series Features Towards Low Quality Dataset

Hard disk failures cause data loss, reducing storage system reliability. While many machine learning models predict hard disk failures, they often neglect temporal characteristics and rely on high-quality datasets like Backblaze. In real-world industrial applications, data may be missing or of low quality. Current models perform poorly with such datasets. To address this, we propose a prediction technique tailored for low-quality datasets. We create a low-quality dataset, Backblaze-, by deleting 10% to 80% of data from the original Backblaze dataset. We introduce time series features like the Absolute Sum of First Difference (ASFD) to highlight fluctuations in hard disk SMART data, enhancing the distinction between positive and negative samples. Our approach achieves near-original dataset performance, with a TPR of 86.7%, an AUC score of 0.93, and an f1-score of 0.89, predicting failures 9.75 days in advance and maintaining an FPR below 0.01% even with 80% data missing.