Characterization and detection of incipient underground cable failures

Abstract

For distribution systems, failure of an underground cable results in an unscheduled outage. An unscheduled outage costs a utility manpower and materials, and affects their reliability index. Thus, the need for an on-line, non-destructive, incipient fault detection system, which can also predict the remaining life of an underground cable, is prevalent. Such an incipient fault detection system could enable utilities to reconfigure the distribution system or replace the cable before a catastrophic failure. The long term-objective of the research project being conducted by the Power System Automation Laboratory (PSAL) at Texas A&M University is to develop such a non-destructive, on-line monitoring system that can not only detect incipient faults but also predict the remaining life of the cable. Earlier research work conducted by the Power System Automation Lab included the development of two prototype monitoring systems. The first monitoring system was used to monitor field aged and intentionally damaged cable sections in a controlled environment. These controlled monitoring (CE) experiments were conducted at the Downed Conductor Testing Facility at Texas A&M University Riverside Campus. The second monitoring system was used to monitor for short periods of time various cable laterals feeding residential and commercial areas of the Dallas, Texas area. As part of the on-going research, this thesis presents: a) Development of an on-line, non-destructive experimental setup to monitor an distribution underground cable lateral and record real time data over a long period of time. b) Formation of a data library including data from controlled experiments (CE), short-term monitoring (ST) experiments, and the long-term monitoring experiments. c) Analysis of the recorded data implementing advanced signal-processing techniques (wavelet packet analysis) to characterize incipient behavior. d) Implementation of artificial neural network (self-organizing map) to identify groups of similar abnormalities and conduct comparison study among them. Wavelet packet analysis yielded dominant frequency ranges from the incipient fault cases in the data library. Features were extracted from the time-frequency information of wavelet packet analysis results for training the SOM artificial neural networks. Categories of similar incipient behavior were identified by the SOMs and a comparison study was conducted among the various categories. Similarities were found between the data recorded at various sites.