In this paper, the use of a neuro-fuzzy approach to PD pattern recognition in cross-linked polyethylene (XLPE) insulated medium voltage (MV) cables is proposed. The proposed neuro-fuzzy classifiers are based on the Min-Max model and are trained by both the adaptive resolution mechanism (ARC) and a novel modified version called nonexclusive ARC (NARC). The resulting neuro-fuzzy classifiers are compared with traditional neural architectures, as the Multilayer Perceptron and the Cascade Correlation, which up to now performed better than early statistical operators. Data sets were built through an experimental procedure, which simulated four typical defects on the joints and terminations of a common XLPE insulated MV cable. The resulting PD pulses are suitably measured and processed, so that each pattern is identified by a set of numerical features relevant to well-known physical attributes of the PD pulses. As evidenced in the experimental results, the proposed neuro-fuzzy networks significantly improve the performances of other neural networks in terms of classification accuracy and of robustness of the learning process.
A Neuro-Fuzzy Approach to Partial Discharge Pattern Recognition of XLPE Insulated MV Cables / Panella, Massimo; Rizzi, Antonello; FRATTALE MASCIOLI, Fabio Massimo; Martinelli, Giuseppe. - STAMPA. - (2003), pp. 184-191. (Intervento presentato al convegno International Conference on Engineering Applications of Neural Networks tenutosi a Malaga, Spagna nel 8-10 settembre 2003).
A Neuro-Fuzzy Approach to Partial Discharge Pattern Recognition of XLPE Insulated MV Cables
PANELLA, Massimo;RIZZI, Antonello;FRATTALE MASCIOLI, Fabio Massimo;MARTINELLI, Giuseppe
2003
Abstract
In this paper, the use of a neuro-fuzzy approach to PD pattern recognition in cross-linked polyethylene (XLPE) insulated medium voltage (MV) cables is proposed. The proposed neuro-fuzzy classifiers are based on the Min-Max model and are trained by both the adaptive resolution mechanism (ARC) and a novel modified version called nonexclusive ARC (NARC). The resulting neuro-fuzzy classifiers are compared with traditional neural architectures, as the Multilayer Perceptron and the Cascade Correlation, which up to now performed better than early statistical operators. Data sets were built through an experimental procedure, which simulated four typical defects on the joints and terminations of a common XLPE insulated MV cable. The resulting PD pulses are suitably measured and processed, so that each pattern is identified by a set of numerical features relevant to well-known physical attributes of the PD pulses. As evidenced in the experimental results, the proposed neuro-fuzzy networks significantly improve the performances of other neural networks in terms of classification accuracy and of robustness of the learning process.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.