Clipping the target value or filtering data in predictive maintenance: example of C-MAPSS

The article explores methods to improve predictive maintenance by enhancing data preparation for Remaining Useful Life (RUL) predictions. The study compares the traditional "capping" approach, which limits RUL values, with a new "filtering" method that focuses exclusively on data from the degradation phase, excluding normal operation data. While capping reduces errors in the normal operation phase, it distorts predictions during the degradation phase. In contrast, the filtering approach improves accuracy by training the model solely on degradation data, resulting in more precise RUL estimates. The findings highlight the importance of selecting relevant data to improve the performance and reliability of predictive maintenance models.

The article explores the influence of data preparation on the performance of Remaining Useful Life (RUL) prediction models in predictive maintenance. Using the C-MAPSS dataset, which simulates the degradation of turbofan engines, the study compares different data processing approaches, including truncation and filtering. Truncation, while reducing errors during the normal phase, distorts predictions during the degradation phase, whereas filtering, which involves using only degradation data, results in more accurate predictions without disrupting the model. The proposed filtering approach thus improves model performance by preventing it from being influenced by normal operation data and focusing solely on degradation phases.