A Knowledge-Graph-Guided Residual Graph Neural Network for Recommending Additional Diagnostic Tests for Power Transformers
Year: 2026
Type: article
Abstract: Power transformer condition assessment involves successive inspections to clarify abnormalities and obtain information for engineering decisions. Because inspection methods examine different components and fault mechanisms, selecting the next test requires interpreting available observations alongside inspection history and unresolved diagnostic questions. A recommendation framework should therefore connect the available evidence with the diagnostic purposes of candidate inspections. This study presents a knowledge-graph-guided residual graph neural network (KG-GNN) for ranking nine follow-up inspections. A frozen multilayer perceptron (MLP) provides observation-based scores, while the GNN learns corrections through inspection–hypothesis relations. Qualified completed-case evidence is incorporated through conditional relation updating, with applicability checks and training-side validation governing adoption. Evaluation used 62 retrospective decision episodes from 50 nominal assets and 21 source groups, with source-cross-fitted residual training and nested residual-scale selection. The KG-GNN increased mean Recall@3 from 0.5048 to 0.5927, an improvement of 8.79 percentage points over the MLP. Conditional updating activated the updated graph for one held-out episode, maintained cohort-level Recall@3 and preserved original outputs outside the activated scope. The framework provides a structured basis for reviewing follow-up inspection candidates by combining observation-based ranking, explicit engineering relations and controlled incorporation of case evidence.
Source: Electronics
Institution: Mokpo National University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
Awards: G9001358788, G9001398721
OpenAlex ID: W7236855612
