Remaining Useful Life Prediction Based on Modified Relevance Vector Regression Algorithm

Abstract

Remaining useful life (RUL) is a significant and challenging task in prognostics and health management (PHM) of engineered systems. For data-driven prognostics, machine learning algorithms are nowadays attracting the attentions of researchers. This paper introduces relevance vector regression (RVR) algorithm into RUL prediction, as it models the nonlinearity and uncertainty of the degradation process very well. However, the conventional RVR model cannot recognize the overall degradation pattern. When applying it for long-term prediction to estimate RUL, the result might deviate from the real situation greatly. This paper proposes a modified RVR model with a new design matrix (RVR-NDM) with an additional column vector which represents the overall degradation pattern. For an RVR-NDM model, both of the kernel width and normalization of input vector have impacts on the learning results. We propose a strategy for model optimization. For demonstrating the proposed method, a case study for turbofan engine RUL estimation is given. The results show that the RVR-NDM is effective for RUL prediction and better than the basic RVR and generalized linear regression methods.