Feature importance-guided multi-regression ensemble with application to remaining useful life prediction

Ensemble methods have been applied to many problems, with the three typical steps of the ensemble process, namely, ensemble generation, ensemble pruning and ensemble integration, being widely studied. This paper proposes the use of multi-model ensemble in generating the final predictions of the Remaining Useful Life (RUL), where the primary focus are in ensemble generation (to produce multiple base models) and ensemble integration (to design weighting methods for the base models). Upon generating the multiple base models, each base model is then weighted by incorporating the importance information of the features that are used. Six weighting methods were implemented and experiments were conducted on real data from eight induction motors. Results comparison with the commonly used equal weighting method showed the benefits of such ensemble methods incorporating feature importance information.