A comparative study of Data-driven Prognostic Approaches: Stochastic and Statistical Models

Prognostics and health management (PHM) has become a crucial component of many machines and engineering systems, enabling fault diagnostics, prognostics yielding remaining useful lifetime (RUL), and health management [1]. Research on how to predict RUL accurately has achieved popularity due to the accelerated development of prognostic techniques and algorithms [2]. In recent years, a considerable amount of research has been undertaken in data-driven prognostic approaches for RUL prediction, and various novel prediction techniques have been proposed in order to improve estimation accuracy. However, a systematic discussion and comparison of data-driven approaches and their contributing elements are lacking. While several reviews provide excellent overviews of RUL prediction methods, these are typically covered from a holistic perspective[3][4]. The contributing process elements in prognostic methods, such as health indication construction, are typically not compared in detail. To fill this gap, this paper proposes a generic prognostic method breakdown with five contributing technical steps. In addition, case studies are implemented to compare the technical steps of health indicator (HI) construction, degradation model, and RUL prediction. This paper comprehensively compares the approaches, in a systematic manner, covering the proposed five technical steps, and evaluates the results using a consistent set of criteria. This research contributes towards formulating a decision framework to guide a PHM system designer towards selection of an appropriate prognostic approach.