Partial-discharge diagnosis with artificial neural networks

Abstract

The new diagnosis method employs a classical PD measurement system consisting of a coupling capacitor, measuring impedance, and a "wideband" integrator, cascaded by an artificial network evaluation. Upon passing a polarity detection unit, the output signal of the "wideband" integrator is recorded via a digital storage oscilloscope which simultaneously serves as an interface to the subsequent computer-aided evaluation. The personal computer stores the PD-values in a phase resolving PD-matrix. After sufficient learning with training matrices the system recognizes different fault types with high probability. The recognition likelihood of trained patterns is almost 100 percent and of a new pattern approximately 90 percent, depending on both the number of training matrices and the repetition rate. The implemented artificial neural network is composed of a three layer backpropagation algorithm with threshold units and a recognition volume of up to 16 fault types. To guarantee the highest individual detection rate, each fault type must be trained with the same number of matrices. Thereafter, the network is able to recognize previously learned fault types without any other data pre- or post-processing, i.e. the diagnosis system relies exclusively on pattern recognition.<>