An interpretable CNN model for classification of partial discharge waveforms in 3D-printed dielectric samples with different void sizes

Regular maintenance of power equipment in high voltage power systems is essential for avoiding outages. An effective way to maintain such systems is the measurement of partial discharges in the insulation material. Voids in solid dielectrics may result from many causes including defects taking place during the manufacturing of the dielectric. These voids induce PDs. Classifying different void sizes is challenging since traditional classification tools used for partial discharge (PD) classification do not work properly. For instance, phase resolved partial discharge (PRPD) patterns resulting from different void sizes will be roughly the same since the source of the partial discharge is the same. Using existing clustering techniques such as Time–Frequency (T–F) map or analysis of statistical features extracted from the PRPD patterns presents their own limitations. T–F map restricts the use of Fast Fourier Transform, while working with PRPDs is only applicable for AC measurements. In this paper, a convolutional neural network (CNN) attention-based model has shown superior capability over traditional classification technique (T–F map) to classify partial discharge (PD) waveforms resulting from different voids in PLA 3D-printed samples. 1D-CNN has classification accuracy of 98.7% with an increase of 21.42% compared to the T–F map technique. Extensive investigation of the learned model has been conducted in order to interpret the decisions made by the proposed neural network. In particular, adding an interpretable attention model such as GRAD-CAM to our CNN shows that while making the decision the neural network learns to focus more on the regions of the waveform corresponding to the rise of the pulse.