Improved Saliency Maps for Interpretability of Acoustic Signal Classification in Condition Monitoring of Power Transformers

Partial discharge (PD)-based condition monitoring is a critical aspect of high-voltage equipment diagnostics, where timely detection and interpretation of acoustic emission (AE) signals are pivotal in preventing system failures. This article presents a novel framework for obtaining saliency maps, a visually explainable AI (XAI) technique, to enhance PD AE data’s interpretability and diagnostic accuracy. A novel architecture based on multilevel subspace approximation with adjusted bias (Saab) transform is proposed for feature extraction and then trained on a convolutional neural network (CNN) based regressor to predict saliency maps. The AE signals acquired in a laboratory setup are used to train the model for predicting the saliency maps of noise, hot wave, and PD signals. The trained model is also tested on the onsite field-acquired AE signals. Experimental results reveal that the predicted saliency maps effectively highlight diagnostic features of PD events compared to traditional techniques. When evaluating the best saliency maps based on the metric values, the proposed method achieves a 21.5% higher metric score than the following best method on the test dataset. An ablation study is carried out to effectively analyze each feature extraction module’s contribution and the effect of ablation of ground truth XAI methods. Furthermore, the wavelet decomposition of energy levels is used to validate the alignment of predicted maps with the physical characteristics of the fault signal. This work allows for creating more effective saliency maps, leading to better characterization of PD events.