PD-pattern recognition of on-site-measurement data

Digital systems for partial discharge (pd)-measurement allow the detection of pd characteristics according to IEC 270 as well as the classification of pd-sources. For pd-fault recognition a neural network based on the back propagation algorithm was implemented in C++ on a personal computer with Windows 95. The program uses input data files from PRPDA (phase resolved pd analysis) system ICM from Power Diagnostix. Training of the neural network is accomplished by a powerful preselected database with about 500 input files. All data were recorded in on-site-measurement, mainly GIS up to V/sub n/=170 kV (partly already in operation for 30 years) PE-X cables and its accessories, up to V/sub n/=400 kV and medium voltage air insulated switchgears with resin type insulators and instrumental transformers. The neural network has been optimized with respect to all its parameters and first practical applications showed very good recognition results.