Remaining Useful Life Prediction with Missing Data Using Relevance Vector Regression

This paper presents a lifetime prediction method for electromechanical equipment, addressing the common problem of missing data during long-term operation. The method, based on the Relevance Vector Regression (RVR), involves three key processes: data normalization, RVR model development, and lifetime estimation by determining the first occurrence of a specified threshold. Missing data, often caused by sensor malfunctions or transmission issues, disrupts the reliability of operational analysis. To overcome it, the proposed RVR model employs a Gaussian kernel function to capture complex nonlinear relationships in the data and applies Bayesian inference to highlight critical data samples that influence predictions. This process effectively identifies hidden data correlations, enabling accurate lifetime estimation even under incomplete data conditions. In case study, trend data from a clinker cooler fan in cement production is analyzed. An incremental accuracy evaluation validates the model's prediction performance, achieving a 94% accuracy rate compared to actual lifetime data, despite data loss.