Weighted-feature and cost-sensitive regression model for component continuous degradation assessment
Résumé
Conventional data-driven models for component degradation assessment try to minimize the average estimation accuracy on the entire available dataset. However, an imbalance may exist among different degradation states, because of the specific data size and/or the interest of the practitioners on the different degradation states. Specifically, reliable equipment may experience long periods in low-level degradation states and small times in high-level ones. Then, the conventional trained models may result in overfitting the low-level degradation states, as their data sizes overwhelm the high-level degradation states. In practice, it is usually more interesting to have accurate results on the high-level degradation states, as they are closer to the equipment failure. Thus, during the training of a data-driven model, larger error costs should be assigned to data points with high-level degradation states when the training objective minimizes the total costs on the training dataset. In this paper, an efficient method is proposed for calculating the costs for continuous degradation data. Considering the different influence of the features on the output, a weighted-feature strategy is integrated for the development of the data-driven model. Real data of leakage of a reactor coolant pump is used to illustrate the application and effectiveness of the proposed approach.
Origine : Accord explicite pour ce dépôt