Transfer Learning for Bayesian Networks with Application on Hard Disk Drives Failure Prediction
Francisco Lucas F. Pereira, Fernando Dione S. Lima, Lucas Goncalves de Moura Leite, João P. P. Gomes, Javam C. Machado · 2017
Predicting failure in Hard Disk Drives (HDD) is of fundamental importance for data loss avoidance as well as to lower the downtime costs of a system. As a consequence, an increasing effort may be observed from both universities and industry to find suitable failure prediction methods. Despite the encouraging performances achieved by various methods one important aspect that shall be noticed is the lack of data available to build reliable models. Considering the HDD failure prediction task, this problem arises with the numerous HDD models. To overcome such problem, transfer learning strategies offer a valid alternative since it can be used to transfer learning from HDD model with enough data to build failure prediction methods for HDD models with lack of data. In this work we evaluate several transfer learning strategies in the task of HDD failure prediction. Additionally we propose a new strategy to build information sources based on the clustering of similar HDD models. This approach may be a valid alternative when no HDD model has enough data to generate a reliable model. Results showed that all transfer learning scenarios can improve the performance of HDDs failure prediction methods, mainly for HDDs with very limited data. Moreover, the clustering-based information source also results in performance gains in all transfer methods and HDD models tested.