A Subsystem-Aware Graph Autoencoder with uncertainty guided learning framework for explainable power transformer health classification and health-index estimation
Year: 2026
Type: article
Abstract: This paper presents a Subsystem-Aware Graph Autoencoder uncertainty guided learning (SA-GAE (ug)) framework for explainable power transformer health assessment. The framework is developed to address the limitation of conventional autoencoder-based health-classification models, which compress condition-monitoring variables as ordinary tabular features without explicitly representing the physical relationships among transformer subsystems. In the proposed approach, assessment data from 405 power transformers described by 31 condition-based monitoring (CBM) parameters are first grouped into diagnostic subsystems, including dissolved gas analysis, oil quality, internal components, insulation/impurity, bushing, on-load tap changer and operating factors. These variables are then formulated as nodes in a subsystem-aware diagnostic graph, while subsystem membership, statistical correlation and physical degradation relationships define the graph edges. A graph autoencoder is used to extract compact latent health embeddings, which are then passed into uncertainty-guided multi-task learning heads for simultaneous estimation of transformer health class, continuous health index, confidence level and subsystem-level risk drivers. The proposed framework is compared with direct machine-learning classifiers, conventional MLPA-autoencoder baselines, graph convolutional networks and graph attention networks. The proposed model achieved a mean accuracy of 95.2% (95% CI: 94.8–95.6%), macro-F1 of 0.960 (95% CI: 0.956–0.964) and poor-class recall of 0.940 (95% CI: 0.930–0.950). This indicated improved class discrimination and reduced missed detection of poor-condition transformers. The framework therefore demonstrates strong internal performance and potential as an interpretable, maintenance-oriented decision-support approach for transformer asset management.
Source: Energy Informatics
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,490
OpenAlex ID: W7236329735
