Dealing with missing data for prognostic purposes

Abstract

Abstract-Centrifugal compressors are considered one of the most critical components in oil industry, making the minimisation of their downtime and the maximisation of their availability a major target. Maintenance is thought to be a key aspect towards achieving this goal, leading to various maintenance schemes being proposed over the years. Condition based maintenance and prognostics and health management (CBM/PHM), which is relying on the concepts of diagnostics and prognostics, has been gaining ground over the last years due to its ability of being able to plan the maintenance schedule in advance. The successful application of this policy is heavily dependent on the quality of data used and a major issue affecting it, is that of missing data. Missing data's presence may compromise the information contained within a set, thus having a significant effect on the conclusions that can be drawn from the data, as there might be bias or misleading results. Consequently, it is important to address this matter. A number of methodologies to recover the data, called imputation techniques, have been proposed. This paper reviews the most widely used techniques and presents a case study with the use of actual industrial centrifugal compressor data, in order to identify the most suitable ones.

An improved k-nearest neighbours method for traffic time series imputation

“…A typical interpolation based method is regression spline, which is a piecewise polynomial function that tries to approximate the unknown function [16], [17]. In the statistical/machine learning category, k-Nearest neighbours (kNN) [9], [18] and principal component analysis (PCA) based methods [8], [19] have shown great performance [8], [14] and efficiency [11]. It has been shown that different PCA based methods perform similar with regards to accuracy [14].…”

“…Another work introduces an ad hoc category which includes mean, median and last observation carried forward (LOF) [13]. Among those methods, kNN is accurate and efficient [8], [11]. Fig.…”

“…There has been much work on imputation for time series data and kNN has produced imputations accurately and efficiently [8], [11]. However, two key topics are not found in the literature.…”