Abstract
Dell provides a monitoring HDD support to predict the failure and anticipate the change of the component. Given a 15 million operational database, which include 63 features and more than 30 thousand drives, we had to build an end-to-end predictive model to forecast the failure of a drive within a time window.
Challenges
- Data processing – large database with lot of missing days and values.
- Time series modelling – data are sequential by drive with very different lengths and characteristics.
- Train test split – be sure to not use future information to train the model and try to replicate the production test given a specific date.
Achievements (according to KPIs)
- Feature selection and engineering.
- Built a LSTM regression model to fill the missing values and both a HMM and a LSTM sequential models to predict the failure given multiple time windows.
Further development
- Train and test the models on the entire data.
- Deepen both data exploration and data modelling.