Partial Discharge Identification in MV Switchgear Using Scalogram Representations and Convolutional AutoEncoder

This work proposes a methodology to automate the recognition of Partial Discharges (PD) sources in Electrical Distribution Networks using a Deep Neural Network (DNN) model called Convolutional Autoencoder (CAE), which is able to automatically extract features from data to classify different sources. The database used to train the model is constructed with real defects commonly found in MV switchgear in service, and it also includes noise and interference signals that are present in these installations. PD sources consist of defective mountings, such as the loss of sealing cap of cable terminations, or an earth cable in contact with cable termination insulation. Four sources were replicated in a Smart Grid Laboratory and on-line measurement techniques were used to obtain the PD signal data. The Continuous Wavelet Transform (CWT) was applied to post-process the PD signal into a time-frequency image representation. The trained model predicts with high accuracy new data, demonstrating the effectiveness of the methodology to automate the recognition of different partial discharges and to differentiate them from noise and other interference sources.