Seoul Researchers Refine SSD Failure Prediction Using AI

The methodology employs Multiple Instance Learning to categorize drive data into failure bags based on rack location and reporting dates. By utilizing a temporal convolutional network, the system analyzes S.M.A.R.T. logs to estimate risk levels for individual units despite the presence of noisy, group-level training data. This development, conducted in partnership with Samsung Electronics using Alibaba Cloud data, addresses a significant vulnerability in industrial diagnostics.

Performance testing revealed a marked improvement in reliability. While conventional models saw their F1 scores plummet from 0.731 to 0.261 when faced with a 40% false-failure rate, the new system maintained an F1 score of 0.717 under identical conditions. Assistant Professor Jaewoong Shim notes that the model successfully prioritizes genuine failures, ranking them significantly higher in risk assessment than incorrectly flagged drives. Beyond server infrastructure, the team believes this logic is applicable to industrial settings like battery packs and heavy machinery where component-level identification remains a persistent challenge.