A characterization methodology for distribution system abnormalities using wavelet packets and self-organizing map neural networks

Automatic data processing and information retrieval is a desirable approach in dealing with the curse of the large volume of data recorded for the condition assessment of equipment in power systems. This approach involves developing a methodology that utilizes the inherent relationships of the data to automate the characterization of system behavior. This paper discusses a practical approach to process incipient abnormality data in an underground distribution system to detect failing equipment. The core components of the proposed approach consist of feature extraction, data mapping, clustering, and rule extraction. Wavelet packet analysis was used to extract the informative features from high frequency current signals and the self-organizing map was utilized to model the data and produce prototype vectors. Descriptive rules for the identified patterns were extracted from the k-means clustering results. The results can be employed in rule-based classifiers to automatically characterize the data in an unsupervised fashion