An Interpretable Predictive Model for Early Detection of Hardware Failure
Artsiom Balakir, Alan S. Yang, Elyse Rosenbaum · 2020
This paper develops an accurate yet interpretable machine learning framework for predicting field failures from time-series diagnostic data with application to datacenter hard disk drive failure prediction. Interpretable models are accountable: model reasoning can be verified by a domain expert for critical reliability tasks. We develop an attention-augmented recurrent neural network that visualizes the temporal information used to generate predictions; visualizations correlate with physical expectations. Finally, we propose a clustering-based method for discovering failure modes.