Regression Kernel for Prognostics with Support Vector Machines

Estimating the remaining useful life (RUL) of systems and/or equpipments has been an important goal for reliable, safe, and profitable operation of industrial plants. However, traditional mathematical and statistical modeling based approaches are difficult to design and they adapt poorly to the ever changing operating and environmental conditions in real-world industries. With recent developments in computational technologies, data storage, and industrial automation recording and storage of large amounts of historical plant data from embedded sensors and maintenance records have become easy. Availability of large data sets together with advancements in data driven machine learning algorithms has been the key driver for prognostic and diagnostic research in the industry as well as by academia. Nevertheless, developing generalized machine learning algorithms for the prognostic domain has been challenging due to the very nature of the problem. This paper describes some of these challenges and proposes a modified regression kernel that can be used by support vector regression (SVR) for prognostic problems. The method is tested on a simplified simulated time-series data set that is modeled to represent the challenges presented.