Predicting Failures in HDDs with Deep NN and Irregularly-Sampled Data

As information systems became basic requirements for essential human services, safeguarding stored data became a requirement to maintaining these services. Predicting Hard Disk Drive (HDD) failure can bring efficiency gains in HDD maintenance and reduce the risk for data loss. Recurrent Neural Networks (RNN) are powerful tools for predicting HDD failure but require complete data entries sampled at regular intervals for efficient model training, testing, and deployment. Data imputation is a baseline method to preprocess data for RNN models. However, typical data imputation methods introduce noise into datasets and erase missing data patterns that would otherwise improve model predictions. This article surveys existing RNN models robust to the presence of substantial amounts of missing data and benchmarks the predictive capabilities of these methods on HDD failure prediction using Self-Monitoring, Analysis, and Reporting Technology (SMART) data. To evaluate different missing data conditions, we simulate binomial and exponential sampling schema with varying levels of missing data. The successful implementation and comparison of these methods demonstrated that the GRU-D, phased-LSTM, and CT-LSTM methods are well-rounded methods for multiple missing data conditions, having achieved better performance than basic LSTM networks.