Cost-sensitive learning considering label and feature distribution consistency: A novel perspective for health prognosis of rotating machinery with imbalanced data

Intelligent operation and maintenance methods based on data-driven concepts provide a new development direction for the field of mechanical prognostics and health management. Unfortunately, most current models are designed based on the assumption of data balance, while data collected from industrial sites usually show an unbalanced state. In addition, the current research based on the imbalance problems only stays in fault classification, and the regression prediction of remaining useful life (RUL) under imbalance data has not been fully discussed. In view of the above, this paper takes imbalanced regression as the research proposition for the first time, aiming to develop a framework for health prognosis of mechanical equipment under imbalanced data. First, we generalize the deep imbalanced classification (DIC) problems to the regression problems, formally define the deep imbalanced regression problems (DIR), and propose two conjectures about DIR. Second, based on two conjectures, label distribution normalization and feature distribution normalization are proposed to locally calibrate the implicit distribution of label space and deep feature representation space. Then ranking similarity optimization is designed to globally match the label space and the deep feature representation space. Finally, a cost-sensitive learning framework considering label and feature distribution consistency is introduced for end-to-end RUL prediction under imbalanced data. Experiments verify the effectiveness of the proposed prediction framework, which also provides a new perspective for realizing regression prediction under imbalanced data.