Researchers at Seoul National University of Science & Technology, in collaboration with Samsung Electronics, have developed an AI method that can predict SSD failures even when maintenance data contains inaccurate failure reports. The approach uses Multiple Instance Learning (MIL) to learn from groups of SSDs reported as failed together, while still estimating the individual failure risk of each drive. The researchers tested the method using real-world SSD data from an Alibaba Cloud data center. When 40% of the training data contained false failure labels, a conventional model’s F1 score dropped from 0.731 to 0.261. In contrast, the proposed mean-pooling approach achieved an F1 score of 0.717, showing much greater resilience to inaccurate labels. The method could help data center operators identify SSDs that are more likely to fail and prioritize inspections, backups, monitoring, and replacements. The researchers also suggest that the approach could be applied beyond SSDs to other industrial systems where a problem can be detected within a group, but identifying the exact faulty component is difficult.
