Time series clustering for fault detection and isolation
Proceedings of the 12th IASTED International Conference on Control and Applications, CA 2010
Clustering, Fault diagnosis, Time-series segmentation
Fault detection and isolation (FDI) algorithms have been widely studied in recent years. Most of the existing algorithms are supervised. The modified Gath-Geva (MGG) algorithm has been recently introduced as an unsupervised method for time series segmentation and condition monitoring. This algorithm is applied here for FDI purpose on DAMADICS benchmark. However, it fails to classify the faults properly because of the high dimensionality of the data. To tackle this problem, dynamic PCA (DPCA) is then utilized as a preprocessing step to enhance the informative richness of the data set and to reduce the dimension of the data. The derived DPCA-MGG clustering approach is used to detect and isolate the faults by organizing the DPCA-transformed data in different clusters. This method results in better isolation of the faults than the original MGG algorithm. Another methodology to overcome the high dimensionality problem is feature weighting which is also incorporated here to enhance the monitoring task. For this purpose, a new clustering method is introduced here which provides weights for different features in different clusters through an optimization procedure. This method can properly detect and isolate the faults of the DAMADICS benchmark while outperforming the other discussed methods.
Bahrampour, Soheil and Saif, Mehrdad. (2010). Time series clustering for fault detection and isolation. Proceedings of the 12th IASTED International Conference on Control and Applications, CA 2010, 490-496.