Remaining Useful Life Estimation by Empirical Mode Decomposition and Support Vector Machine

Knowledge on reliability and maintenance performance of physical assets can bring competitive advantages for companies (e.g., shorter downtimes and higher operation continuity). Prognostic and Health Management (PHM) can be used to improve not only the availability, but also the quality of the system. Furthermore, the assessment of important measuresin PHM, such as the Remaining Useful Life (RUL), facilitates the development of action plans such as maintenance programs or inventory acquisition. A well-known method that has been successfully used to estimate RUL is Support Vector Machines (SVM), which present the advantage that knowledge about the function's behavior and the relationship between input and output are not required. In addition, SVM solves a quadratic and convex optimization problem on the training set for which the Karush-Kuhn-Tucker (KKT) first order conditions assure a global maximum. However, SVM does not provide satisfactory results when dealing with non-stationary series with monotonic trend. In this context, Empirical Mode Decomposition (EMD) raises as an alternative to tackle this issue. In this paper, a EMD+SVM based approach is used for predicting the RUL based on non-stationary series, and is applied in two cases with different behaviors. The results indicate that EMD+SVM presents an improved performance in comparison to stand-alone SVM for both case studies.