Abstract

Acoustic emission (AE) detection is a nonintrusive technique for monitoring transformer conditions by capturing emissions from partial discharges (PDs), hotspots, and noise. Machine learning (ML) has been widely used in PD diagnostics in power transformers. While effective in managing complex data characteristics, traditional ML algorithms cannot quantify uncertainty in classification. This study uses Bayesian networks integrated with a classification algorithm to quantify uncertainty in acoustic signal classification. Bayesian deep learning (BDL) and ensemble models are the two commonly used techniques for uncertainty quantification (UQ). BDL models with multiple Bayesian layers are more prone to convergence issues with difficulty interpreting the sources of uncertainty. The performance of the ensemble method is based on model diversity with a higher risk of overfitting and does not offer insights into aleatoric uncertainty. The present work proposes architectures with a single Bayesian layer to quantify both aleatoric and epistemic uncertainties in acoustic signal classification. A Bayesian convolutional neural network (CNN) layer combined with an ensemble architecture demonstrates comparatively higher performance among the considered Bayesian architectures. Raw acoustic signals are transformed into spectrogram images to enhance feature representation, capturing time-frequency characteristics. The proposed method is assessed using laboratory and field-measured data, demonstrating significant improvements in estimating uncertainty in classification.