An improved similarity-based residual life prediction method based on the dynamic variable combination

Remaining useful life (RUL) prediction is essential for preventive maintenance and industrial safety operation. As an emerging data-driven method, the similarity-based residual life prediction (SbRLP) method plays a vital role in the RUL prediction. However, its existing research usually adopts fixed variables without considering the sensitivity difference of the same variable in different degradation states, thus making locally insensitive variables unusable and resulting in a waste of valuable monitoring information. Therefore, to fully mine useful monitoring information to improve the RUL prediction accuracy, this paper proposed an improved SbRLP method based on the dynamic variable combination. Firstly, the K-means algorithm was employed to divide state intervals, and combined with state intervals, the dynamic variable combination was determined through the local sensitivity analysis. Then, the state interval was recognized by the multi-SVDD algorithm with a novel discrimination criterion, and united with the dynamic variable combination, the RUL was predicted through the SbRLP method. Eventually, a case study is provided to demonstrate the effectiveness and superiority of the proposed SbRLP method. The results show that the proposed SbRLP method has better prediction performance, especially during the equipment's early and mid-term performance degradation. Moreover, implementing the new discriminate criterion can help improve its prediction accuracy.