Supervisor: Dennis K. Peters
Modern fiber manufacturing plants rely heavily on the use of automation. Automated facilities use sensors to measure fiber state and react to data patterns, which correspond to physical events. Many patterns can be predefined either by careful analysis or by domain experts. Instances of these patterns can then be discovered through techniques such as pattern recognition. However, pattern recognition will fail to detect events that have not been predefined, potentially causing expensive production errors. A solution to this dilemma, novelty detection, allows for the identification of interesting data patterns embedded in otherwise normal data. In this thesis we investigate some of the aspects of implementing novelty detection in a fiber manufacturing system. Specifically, we empirically evaluate the effectiveness of currently available feature extraction and novelty detection techniques on data from a real fiber manufacturing system.
Our results show that piecewise linear approximation (PLA) methods produce the highest quality features for fiber property datasets. Motivated by this fact, we introduced a new PLA algorithm called improved bottom up segmentation (IBUS). This new algorithm produced the highest quality features and considerably more data reduction than all currently available feature extraction techniques for our application.
Further empirical results from several leading time series novelty detection techniques revealed two conclusions. A simple Euclidean distance based technique is the best overall when no feature extraction is used. However, when feature extraction is used the Tarzan technique performs best.
Last modified: Mon 2000.01.17 at 21:52 NST by Dennis Peters