Comparison of a Few Fault Diagnosis Methods on Sparse Variable Length Time Series Sequences
Accurate gas turbine engine Fault Detection and Diagnosis (FDD) is essential to improving aircraft safety as well as in reducing airline costs associated with delays and cancellations. This paper compares broadly three methods of fault detection and diagnosis (FDD) dealing with variable length time sequences. Chosen methods are based on Dynamic Time Warping (DTW), k-Nearest Neighbor method, Hidden Markov Model (HMM) and a Support Vector Machine (SVM) which makes use of DTW ingeniously as its kernel. The time sequences are obtained from Turbo Propulsion Engines in their nominal conditions and two faulty conditions. Typically there is paucity of faulty exemplars and the challenge is to come up with algorithms which work reasonably well under such circumstances. Also, normalization of data plays a significant role in determining the performance of the classifiers used for FDD in terms of their detection rate and false positives. In particular spherical normalization has been explored considering the advantage of its superior normalization properties. Given sparse training data how well each of these algorithms performs is shown by means of tests performed on time series data collected at normal and faulty modes from a turbofan gas turbine propulsion engine and the results are presented.