Assistant Professor Jaewoong Shim and a research team at Seoul National University of Science and Technology (SEOULTECH) have developed a machine learning method that predicts solid-state drive failures more reliably when training data contains errors. The approach, published September 1 in Computers & Industrial Engineering, addresses a common problem in large-scale data centers: failure reports that incorrectly label healthy drives as faulty, which degrades the accuracy of conventional AI models.
The work was conducted with researchers from Samsung Electronics and used real-world SSD data from an Alibaba Cloud data center. When the team simulated a 40% false-failure rate in training data, a standard model's F1 score dropped from 0.731 to 0.261. The new method held at 0.717 under the same conditions.
Why data center failure labels are unreliable
SSDs continuously generate S.M.A.R.T. (Self-Monitoring, Analysis and Reporting Technology) logs that record errors, wear, and operating conditions. Machine learning models typically use these logs alongside failure reports to estimate which drives are at risk. The problem is that failure reports are often imprecise. When an abnormal event occurs, data center operators may not be able to identify the exact faulty SSD quickly. Instead, several drives in the same rack get reported as failed, and healthy drives end up with failure labels in the training data.
"Industrial AI has to work with the data that are actually available in the real world, and those data are not always perfectly labeled," said 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."
How Multiple Instance Learning handles the noise
The researchers applied Multiple Instance Learning (MIL), a technique that groups related examples into "bags." SSDs from the same rack with failure reports on the same date were grouped together. A temporal convolutional network (TCN) then analyzed each drive's S.M.A.R.T. data over time to produce individual failure risk scores. During training, the model combined predictions at the bag level, but during inference it still produced a risk score for each SSD.
This structure meant the model learned from group-level patterns without needing to know which specific drive in the bag had actually failed. The mean-pooling variant of the approach proved most resilient. Under the 40% false-failure condition, it ranked genuine failures an average of 1.6 compared with 3.5 for drives incorrectly reported as failed, giving operators a clear way to prioritize inspections.
Beyond SSDs: group-level fault detection
The method is not limited to storage hardware. Dr. Shim pointed to other industrial settings where a fault can be traced to a group but not to the exact component responsible. "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," he said. Research teams working with noisy equipment logs or sensor arrays may find the approach adaptable to their own domains, including predictive maintenance for manufacturing lines or energy storage systems. For those building expertise in this area, the AI Learning Path for Research Scientists covers techniques relevant to working with imperfect real-world data.
Why this matters for science and research professionals
The study addresses a gap between clean academic benchmarks and the messy data found in operational environments. Many research scientists working on industrial AI projects encounter labeled datasets where a significant fraction of labels are wrong, yet most published methods assume near-perfect annotations. This work provides a concrete, tested architecture for learning under label noise without discarding data or requiring expensive manual re-labeling. The full paper is available in Computers & Industrial Engineering via DOI 10.1016/j.cie.2026.112229. Researchers exploring similar applications can find additional resources under AI for Science & Research.
Your membership also unlocks:
