Solid-state drives (SSDs), which store data in computers and servers, are critical to data center reliability. When they fail, they can contribute to service disruptions, data loss and high maintenance costs, making early failure prediction important. SSDs continuously generate monitoring information through S.M.A.R.T. (Self-Monitoring, Analysis and Reporting Technology) logs, including indicators related to errors, wear and operating conditions, which machine-learning models can analyze to estimate failure risk.
However, failure reports used to train these models are not always perfectly accurate. When an abnormal event occurs, it can be difficult and time-consuming to pinpoint the exact faulty SSD, so several drives in the same rack may instead be reported as failed. As a result, healthy drives can receive failure labels in the training data, potentially misleading conventional AI models.
Now, a research team led by Assistant Professor Jaewoong Shim from the Department of Data Science, Seoul National University of Science & Technology (SEOULTECH), addressed this challenge using Multiple Instance Learning (MIL). The developed approach groups SSDs reported from the same rack on the same date, allowing the model to learn from group-level information while estimating the failure risk of individual drives. The study was conducted in collaboration with researchers from Samsung Electronics, South Korea, using real-world SSD data from an Alibaba Cloud data center. The research was made available online on July 06, 2026 and was published in Volume 219 of Computers & Industrial Engineering in September 01, 2026.
"Industrial AI has to work with the data that are actually available in the real world, and those data are not always perfectly labeled," explains Dr. Shim. "Our goal was to develop a way for AI to learn from these imperfect failure reports without assuming that every reported SSD failure is correct."
Researchers grouped related SSD data into "failure bags" using MIL. SSD sequences from drives in the same rack with failure reports on the same date were grouped together. A temporal convolutional network (TCN) then analyzed each SSD's S.M.A.R.T. data over time to estimate individual failure risk. During training, the model combined these predictions at the bag level while still producing individual SSD predictions during inference.
The researchers evaluated performance using an F1 score, which balances precision and recall. Scores range from 0 to 1, with higher scores indicating better performance. Under a 0% false-failure condition, the conventional model achieved an F1 score of 0.731. When the researchers simulated a 40% false-failure rate in the training data, its score fell to 0.261. In contrast, the mean-pooling variant of the proposed approach achieved an F1 score of 0.717 under the same 40% condition, demonstrating greater resilience to inaccurate failure labels.
The model also distinguished genuine failures from healthy SSDs incorrectly reported as failed, ranking true failures an average of 1.6 compared with 3.5 for incorrectly reported failures. Overall, the approach could help data-center operators prioritize inspections, backups, monitoring and SSD replacements, enabling them to focus resources on drives most likely to require attention.
"The idea is not limited to SSDs," said Dr. Shim. "It could also be useful for other industrial settings such as battery packs, industrial machinery and other systems where a problem can be identified within a group, but the exact component responsible is difficult to determine."
Reference
Title of original paper: Multiple Instance Learning for SSD Failure Prediction Under Customer Failure-Biased Labels
Journal: Computers & Industrial Engineering
DOI: 10.1016/j.cie.2026.112229
About the institute Seoul National University of Science and Technology (SEOULTECH)
Seoul National University of Science and Technology, commonly known as 'SEOULTECH,' is a national university located in Nowon-gu, Seoul, South Korea. Founded in April 1910, around the time of the establishment of the Republic of Korea, SEOULTECH has grown into a large and comprehensive university with a campus size of 504,922 m2.
It comprises 10 undergraduate schools, 35 departments, 6 graduate schools, and has an enrollment of approximately 14,595 students.
Website: https://en.seoultech.ac.kr/
About Assistant Professor Jaewoong Shim
Dr. Jaewoong Shim is an Assistant Professor of Industrial Engineering and Data Science at Seoul National University of Science & Technology (SeoulTech), South Korea. His research aims to build AI systems that can understand complex industrial environments, adapt to changing conditions, and support reliable decision-making. Before joining SeoulTech, he worked as an engineer in the Quality Assurance Team of the Memory Business at Samsung Electronics and later served as an AI Lead at RTM. He received his Ph.D. in Industrial Engineering from Seoul National University in 2021.
