Answers are grounded in this paper and cite their sources.
Critical review analyzing AI-based DGA diagnostic accuracy and robustness in power transformers, suggesting improvements.
Dissolved gas analysis (DGA) remains a cornerstone technique for the early detection of internal faults in power transformers. However, the interpretation of gas signatures is inherently complex due to coupled thermochemical processes, operating variability, and measurement uncertainties. Over the past three decades, artificial intelligence (AI) techniques have been widely applied to DGA-based diagnostics, evolving from conventional machine learning to advanced deep learning and hybrid models. Although many studies report diagnostic accuracies exceeding 95%, these results often rely on limited datasets and heterogeneous validation practices, raising concerns about their robustness in real-world operating conditions. This review provides a comprehensive and critical analysis of the evolution of AI-based approaches, with emphasis on methodological foundations rather than algorithmic performance alone. Key aspects examined include dataset construction, validation protocols, class imbalance handling, and experimental design. The analysis highlights a structural Accuracy-Robustness Gap, in which high reported accuracies often coexist with limited evidence of generalisation under realistic industrial variability. To address this issue, a conceptual framework is proposed linking the physical complexity of the oil–paper insulation system, dataset characteristics, and algorithmic modeling strategies. The review further identifies several key research gaps in benchmarking practices, dataset availability, robustness evaluation, and model interpretability. Based on these findings, a standardized evaluation framework is proposed to support more rigorous and comparable assessments of AI-based DGA diagnostic models. Future research directions are discussed, highlighting the need for physics-informed learning, distribution-aware validation, and shared datasets to enable reliable deployment of AI-based transformer diagnostics.
No takes yet. Share an insight, caveat, or question.
Bouzar et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: