Dissolved gas analysis (DGA) is the industry’s most established transformer diagnostic, but in 2026 its center of gravity is shifting. The change is not in how gases are measured — chromatographic and optical analyzers remain the workhorses — but in how the resulting time series is interpreted. Across research publications, vendor platforms and utility pilots, DGA is converging with artificial intelligence (AI) and digital twin technology, moving from threshold alarms toward predictive maintenance decisions.
This round-up covers the 2026 developments that define the trend, and — just as importantly — separates what is peer-reviewed research from what is vendor-published capability.
The shift: threshold alarms to AI-assisted decisions
Conventional online DGA is threshold-based: a gas concentration crosses a limit, an alarm fires, and a human interprets the trend using IEC 60599, IEEE C57.104 and graphical methods such as the Duval triangle. That workflow remains the interpretive foundation and is not being replaced.
The 2026 direction is to add a machine-learning layer on top: models trained on DGA histories that flag anomalous trajectories, rank fault likelihood, and recommend inspection priority. A systematic review published in 2026, covering research from 2016 to 2026, describes the broader shift from corrective to AI-driven predictive maintenance, with model families including support vector machines, random forests, XGBoost, LSTMs and generative adversarial networks applied to DGA plus thermal and vibration data. Reported research accuracies are high — in many studies above 90% — but these are research comparisons on specific labeled datasets, not a guarantee of field performance.
Digital-twin-assisted fault diagnosis
A 2026 study in Frontiers in Smart Grids proposed a digital-twin-assisted transformer fault diagnosis framework that combines a 3D transformer model, online monitoring data and DGA-based diagnosis using a Snake Optimization–Random Forest (SO-RF) model. The approach is representative of a wider trend: instead of treating DGA readings as isolated numbers, a digital twin correlates them with load, oil temperature and hotspot estimates, then applies machine learning to distinguish fault types.
The practical value is context. A hydrogen rise during a heavy-load summer day carries a different meaning than the same rise on a lightly loaded unit. Digital twins provide that context, and DGA provides the chemical ground truth.
What vendors are shipping in 2026
Multiple vendors have moved AI-assisted DGA from roadmaps into shipping products. The following are vendor-published claims and should be read as such:
| Vendor | Platform / product | Reported capability |
|---|---|---|
| Camlin Energy | TOTUS suite + Asset Insights | Continuous DGA and moisture modules (G1/G5/G9); predictive health indices to shift from time-based to condition-based maintenance |
| Power Technologies | VizionEye AI platform | Fuses DGA, partial discharge, acoustic and thermal sensors; company reports >95% early fault detection and 30–40% downtime reduction in pilots |
| HERTZINNO | HZ-GRID platform + 3/5/9-gas DGA | Hourly photoacoustic gas measurement with early warnings on key gases such as acetylene |
These products differ in sensor fusion and analytics depth, but they share a common design: online DGA as the sensing backbone, AI as the interpretation layer, and the SCADA or asset-management system as the consumer of recommendations.
Utility pilots and case studies
Real deployments are beginning to appear in the literature and at industry forums. At an EPRI Transformer and Switchyard User Group meeting, Xcel Energy presented a case study using Vaisala’s OPT100 mobile online DGA monitor alongside conventional offline laboratory methods to analyze abnormal gassing on an in-service transformer, enabling more confident maintenance decisions. The pattern — online monitor for trend, offline laboratory for confirmation — is the same complementary workflow the industry recommends for gas measurement, now augmented by analytics.
What this means for asset managers
For operators deciding whether to engage with AI-assisted DGA, four points are worth keeping:
- Keep the standards as the base. IEC 60599 and IEEE C57.104 define how to judge gas data; ML should assist, not override, engineering judgment. Expert sign-off remains the norm.
- Validate on your own data. Research accuracies are dataset-specific; ask vendors for evidence on assets similar to yours, and run a pilot before scaling.
- Start with data quality. AI models are only as good as the DGA time series they learn from — calibration traceability and consistent sampling matter more than model choice.
- Pilot → evaluate → scale. A single critical transformer is a sensible first pilot; measure alert precision and missed events before rollout.
For a deeper look at the modeling techniques and their known limitations, see our guides on machine-learning DGA diagnosis and digital twins for remaining useful life.
Where PAS DGA fits
AI-assisted diagnostics need clean, continuous gas data, and that is the layer PAS DGA builds. The DGA-900 measures nine gases plus moisture with laser photoacoustic spectroscopy — no carrier gas, no consumables, an acetylene detection limit of ≤0.1 ppm (vendor data) — and streams a calibrated time series over MODBUS TCP or IEC 61850 to your analytics or asset-management platform. That data is the fuel for the digital twin layer described above.
Whether you are piloting AI now or later, the monitoring backbone is the same. Contact PAS DGA to discuss the right analyzer configuration for your pilot.