Detection of Partial Discharges under Repetitive Pulsed Voltage Excitation using Deep Learning

Premature failure of motor winding insulation is a critical issue for variable speed drives (VSDs), with partial discharges (PD) being a primary cause. The high-frequency switching of the pulse width modulation (PWM) voltage driving the motor introduces strong interference, complicating the measurement to detect PD for winding insulation assessment. This study applies deep learning to identify the presence of PD within the switching remnants of the repetitive pulsed excitation voltage. Raw data, collected from actual experimental circuits, are classified using a recurrent neural network (RNN) architecture. The different classes comprise the measured signals with only the interference from the switching remnants of the pulsed excitation voltage with varying duty cycles $(10 \%$, $30 \%, 50 \%, 70 \%$) and with the additional PDs submerged in the interference. Results demonstrate that deep learning can effectively classify switching sequences with and without PD. Specifically, the long-short-term memory (LSTM) model achieves a recognition accuracy of up to ${9 9. 4 \%}$. Additionally, an autoencoder is employed to filter out the switching remnants from the PD signals, further enhancing the capability of the proposed method to isolate and identify PD occurrences. This demonstrates the potential of deep learning methods not only in classification but also in denoising applications, providing a robust approach for PD detection in the presence of highfrequency switching noise.