Identifying Parameters for Nonlinear Degradation-Based Prognostics: A Case Study With Exponential Degradation on Induction Motors

Identifying Parameters for Nonlinear Degradation-Based Prognostics: A Case Study With Exponential Degradation on Induction Motors

Feng Yang (Institute of High Performance Computing (A*STAR), Singapore) and Mohamed Salahuddin (Institute of High Performance Computing (A*STAR), Singapore)
DOI: 10.4018/978-1-5225-6989-3.ch012


Prognostics and health management (PHM) methodologies are increasingly playing active roles in improving the availability, reliability, efficiency, productivity, and safety of systems in many industries. In predicting the remaining useful life (RUL), this chapter introduces a prognostics framework with health index (HI) formulation, with specific emphasis on incorporating and validating nonlinear HI degradations. The key issue to the success of this framework is how to identify appropriate parameters in describing the behavior of the nonlinear HI degradations. Using exponential HI degradation as an example in predicting the RULs of induction motors, this chapter discusses three different explorations in verifying the existence of good parameter values as well as identifying the appropriate parameters automatically. Comprehensive experiments were carried out with degradation process (DP) data from eight induction motors, and it was discovered that good parameters can be automatically determined with the proposed parameter identification method.
Chapter Preview


Condition-based maintenance (or predictive maintenance) has received increased attention in many industries, due to its promise of cost savings over preventive maintenance (i.e. maintenance conducted routinely in a time-specific manner) and corrective maintenance (i.e. maintenance only conducted upon failure). By predicting the future trends of health using process information from past, current and even future (such as the planned operating condition), predictive maintenance allows for convenient and flexible scheduling of maintenance works. Undoubtedly, this will reduce costs, with other potential advantages including increased equipment lifetime and availability, increased plant safety and productivity, reduced disruption and downtime to equipment, reduced human work-load and optimized inventory handling. In achieving predictive maintenance, prognostics modelling is a requisite, which generally aims to predict the Remaining Useful Life (RUL) of equipment.

Among existing prognostics techniques, the use of data-driven prognostics is gaining much attention in both academia and industry. This is primarily due to its non-requirement of any prior knowledge related to the equipment’s degradation as well as the increasing availability of sensory data in practice. In essence, data-driven prognostics approaches can be divided into either direct or Health Index (HI)-based, depending on whether a HI is formulated and used as an intermediate expression of the health condition for predicting the RULs (Yang et al., 2016). Direct prognostics approaches attempt to predict the RULs directly from input sensory data, using modelling methods such as Support Vector Regression (SVR) (Khelif et al., 2017; Loutas, Roulias, & Georgoulas, 2013), Neural Network (NN) (Liu et al., 2010), Gaussian process regression (Goebel, Saha, & Saxena, 2008), and ensemble of multiple models (Hu et al., 2012; Baraldi, Mangili, & Zio, 2012; Patil et al., 2015). Although these approaches have been developed and applied to several practical problems, showing their simplicity and efficiency in implementation, the prediction accuracies generated from these direct approaches are generally not as effective as those obtained from HI-based approaches (see for comparison examples in Yang et al. (2016) and Khelif et al. (2017)).

Alternatively, HI-based prognostics approaches predict RUL in a two-stage process, by firstly predicting a HI from the input sensory data for representing the intrinsic health status and then followed by mapping the HI to the RUL. Compared to direct prognostics approaches, the merit of HI-based prognostics lies in its ability to ‘manipulate’ and improve the HI, thus, resulting in improved accuracy in predicting the final RUL. The prominence and effectiveness of HI-based prognostics for RUL prediction has been argued and shown in several existing work (Bechhoefer, 2012; He et al., 2012; Liu, Wang, & Tian, 2015; Tian et al., 2016; Liu et al., 2017). One specific case of ‘manipulating’ HI is in Yang et al. (2016), where the HI was assumed to be degrading linearly and was dynamically smoothed, using current and past information as well as domain knowledge of the HI. The success was illustrated in predicting RULs of eight induction motors, with substantial improvement in the prediction accuracy.

Based on the works of Yang et al. (2016), this chapter introduces a generic formulation of HI-based prognostics, with the ability to consider both linear and nonlinear HI degradations. In this formulation, modelling the degradation behaviour of HI is a pre-requisite, implying that the HI degradation curve needs to be determined a prior. In consideration of simplicity and ease of implementation, linear degradation of health has been studied and gained success in many literatures (Wang et al., 2011; Si et al., 2013; Yang et al., 2016). However, in most real and practical systems, nonlinearity is ubiquitous as most processes degrade in a nonlinear manner, which has driven the efforts in utilizing nonlinear degradation for prognostics (Si et al., 2012; Feng et al., 2013; Si, 2015; Yin & Zhu, 2015).

Complete Chapter List

Search this Book: