Abstract
Dissolved gas analysis (DGA) is central to the condition assessment of oil-immersed power transformers, yet conventional ratio and rule-based methods can be inconclusive near decision boundaries and supervised models depend on scarce, reliable fault labels. This paper presents an unsupervised autoencoder framework for screening abnormal DGA behaviour as a precursor to engineering diagnosis. Eight gas measurements are chronologically ordered, cleaned, standardised using training-set statistics, reconstructed by a dense autoencoder, and scored through sample-wise mean-squared reconstruction error. A data-adaptive threshold is defined at the 95th percentile of the reference error distribution. Evaluation on an open UK power-station transformer DGA dataset produced mean training and validation reconstruction errors of 0.0926 and 0.1802, respectively. With a threshold of 0.2296, 105 observations were flagged, while the maximum error reached 1.0369, approximately 4.52 times the threshold. The analysis shows that reconstruction error can provide a transparent ranking of unusual operating records and a practical inspection workload. Importantly, because the dataset has no confirmed fault labels and the threshold fixes the upper-tail rate, the flagged observations are candidate anomalies rather than proven thermal faults. The resulting workflow is therefore positioned as a reproducible decision-support layer that complements , rather than replaces, IEC/IEEE DGA interpretation and asset-engineering review.