A. Data collection
The PD data was collected from the power utility industry.
The data are captured into pulses from HFCT sensors, each
pulse is captured within a window that has a fixed length
of 1000 readings. Each pulse is only recorded if its the
largest reading (trigger) is above a predefined threshold. For
consistency, the trigger is positioned at the 200the reading
of each pulse. A sample waveform from one pulse can be
seen in Figure 2 A. Then, we aggregate 300 data pulses into
a data instance. In power industry domain, experts usually
analyse and label a PD instance based on its phase-resolved
representation, where the maximum value and phase angle of
each PD data pulse are extracted for visualisation. In total, we
have 476 data instances (which corresponds to 142.28k data
pulses), including 256 noise instances and 220 PD instances.