Detection of partial discharges in high voltage DC cable systems using current pulse waveform analyses and generative adversarial networks

Inhalt:
Since common phase related partial discharge (PRPD) detection techniques used in high voltage (HV) AC systems, can only be used to a limited extend in high voltage DC systems, other techniques are required. One common approach is the current pulse waveform analyses (CPWA). The CPWA is used to extract parameters from the pulse waveform of the partial discharge (PD). The target of these method is to detect, classify, and interpret physical properties of different partial discharges due to waveform parameters like rise time, fall time, pulse width, local charge, energy, etc. Instead of using classical algorithms to extract features from the current pulse waveform and analyses these features in an analytical manner, the signal is directly given to an artificial intelligence (AI) based system. First the measured signal is windowed to fixed size vectors, which supply the input of the AI Network. The Network is able to extract features from the given signal by itself. AI systems need much of measured signals for training. A Generative Adversarial Networks (GAN) trained with a reduced dataset is used to produce additional signals for training of other Neural Networks (NN). The generator of the GAN can later be used as simulation tool, while the discriminator model can be used to decide if a signal is an PD or not. First results applying the GAN principle to PD signals are presented.