Adapting nearest neighbors-based monitoring methods to irregularly sampled measurements

Abstract


 
 
 
 Irregularly spaced measurements are a common quality problem in real data and preclude the use of several feature ex- traction methods, which were developed for measurements with constant sampling intervals. Feature extraction methods based on nearest neighbors of embedded vectors are an example of such methods. This paper proposes the use of a time- based construction of embedded vectors and a weighted similarity metric within nearest neighbor-based methods in order to extend their applicability to irregularly sampled measurements. The proposed idea is demonstrated within a method of univariate detection of transient or spiky disturbances. The result obtained with an irregularly sampled measurement is benchmarked by the original regularly sampled measurement. Although the method was originally implemented for off-line analysis, the paper also discusses modifications to enable its online implementation.