Can DGA Predict Transformer Failure? What the Evidence Shows

The premise of transformer DGA monitoring is that it provides early warning of developing faults before those faults cause failure. This premise is substantially correct, but the detail matters. What DGA can detect, with what lead time, under what monitoring conditions, and with what analytical method all affect the practical value of the warning it provides.

This article examines the question directly: what does the published evidence show about DGA's predictive capability, and what does it take to extract the maximum available warning from a DGA programme?

What DGA Detects: The Physical Basis

DGA is effective for fault detection because the primary fault mechanisms that lead to transformer failure produce characteristic dissolved gases in the insulating oil before they cause catastrophic failure.

Thermal faults arise from localised overheating of conductors, core steel, or insulation, typically caused by poor connections, circulating currents in core or tank, or insulation degradation increasing contact resistance. Thermal decomposition of oil at temperatures above approximately 150°C produces hydrogen (H₂), methane (CH₄), and ethane (C₂H₆); at temperatures above approximately 300°C, ethylene (C₂H₄) becomes the dominant product. The gas profile shifts toward higher ethylene/methane and ethylene/ethane ratios as the fault temperature increases [1].

Partial discharge, low-energy electrical discharge in gas-filled voids or near high-stress points, generates primarily hydrogen and methane, typically with low hydrocarbon concentrations and minimal acetylene [1].

Electrical arcing, sustained high-energy discharge, generates acetylene (C₂H₂) as a primary product, alongside hydrogen and other hydrocarbons. Acetylene production requires temperatures above approximately 700°C and is highly specific to high-energy electrical events. Even small concentrations (above 1–2 ppm) in a sample that previously showed zero are diagnostically significant [2].

Cellulose insulation degradation produces carbon monoxide (CO) and carbon dioxide (CO₂) as the cellulose polymer chains break down under thermal or oxidative stress. CO and CO₂ trends indicate insulation paper condition independently of the oil decomposition gases.

Duval [1] demonstrated that these gas patterns are consistent enough across transformer types and sizes that fault type classification from gas ratios achieves reasonable accuracy for a large proportion of documented fault cases.

Detection Lead Time by Fault Type

Detection lead time, the interval between when DGA first shows a detectable abnormality and when the fault would cause a failure requiring intervention, varies substantially by fault mechanism.

Slow thermal faults (overheated joints, core circulating currents at moderate temperature) typically develop over months to years. Gas generation is gradual; the transformer's DGA record shows slowly rising ethylene and methane over consecutive quarterly samples. Quarterly sampling provides reasonable lead time for planned maintenance response. These faults are reliably detectable by DGA well in advance of failure.

Fast-developing thermal faults (severe insulation-to-conductor contact, rapid carbonisation of paper at very high hotspot temperatures) can develop in weeks. The gas generation rate is high enough that a fault initiating shortly after a quarterly sample may be substantially advanced by the next one. CIGRE TB 630 [3] cites this class of fault as a primary justification for continuous online monitoring on high-consequence assets.

Partial discharge typically develops over extended periods, months to years, with low-level gas generation. Standard quarterly monitoring provides adequate lead time if the trend is recognised. Because hydrogen and methane from partial discharge can be confused with modest thermal activity, accurate fault type classification matters for appropriate follow-up.

Arcing faults are the most dangerous and the fastest-developing. A transformer experiencing sustained internal arcing can progress from first acetylene appearance to catastrophic failure in days. CIGRE TB 812 [4] identifies catastrophic failure events from arcing as among the most consequential transformer failures in terms of grid impact. For a transformer sampled quarterly, the probability of catching a fast arcing fault with sufficient lead time for planned intervention is limited. Continuous online monitoring that alerts on acetylene detection within hours is the appropriate risk management tool for high-consequence assets.

The Rate-of-Change Problem With Threshold Methods

The most important limitation of threshold-based DGA interpretation for failure prediction is not that it fails to detect faults. It detects them, eventually. The limitation is that it detects them late, at the point where a threshold is crossed, rather than early, when the rate of change first departs from baseline.

Consider a transformer whose hydrogen has been rising at 5 ppm per month for two years (now at 120 ppm) and suddenly begins rising at 30 ppm per month. Against IEEE C57.104-2019 [2] thresholds, the transformer is still below the 150 ppm Condition 2 boundary. The threshold method generates no alert. But the sixfold increase in generation rate represents a significant change in fault behaviour, one that demands investigation regardless of absolute concentration level.

Dukarm et al. [5] demonstrated that this trajectory-insensitive characteristic of threshold methods produces systematically delayed detection. The Hazard Factor (HF) metric in R-DGA methodology is specifically designed to capture trajectory changes: it is sensitive to the current rate of fault severity accumulation relative to the historical baseline, and flags transformers where this rate has increased significantly, independently of whether any absolute threshold has been crossed. Dukarm [6] further established that measurement uncertainty in laboratory DGA is substantial enough to make single-point threshold comparisons unreliable for precisely the borderline cases where early detection matters most.

What DGA Cannot Predict

Certain transformer failure modes are not detectable by DGA or only partially so:

Mechanical failures from through-faults. A sudden through-fault event, such as a fault current surge from a downstream line fault, can cause winding deformation or conductor displacement that does not generate gases but compromises the transformer's ability to withstand subsequent fault events. Frequency response analysis (FRAX/FRA testing) is the appropriate method for detecting mechanical deformation from through-fault events.

Bushing failures. Power factor testing and partial discharge measurement on bushings provide the primary early warning for bushing insulation deterioration. DGA may show anomalies if bushing oil is in circuit with the main tank, but this is not always the case for oil-impregnated paper bushings.

Tap changer faults in separate compartments. Some load tap changers are housed in separate oil compartments not connected to the main tank. DGA on the main tank oil will not detect fault gases generated in the tap changer compartment. Separate tap changer DGA sampling is required.

Moisture-driven dielectric failure. Progressive moisture ingress can reduce dielectric withstand to failure-threatening levels without generating significant fault gases prior to failure. Moisture in oil analysis and dielectric breakdown voltage testing provide the relevant early warning.

This does not diminish DGA's centrality to transformer condition monitoring. It simply defines the boundaries of the method. A complete programme combines DGA with the complementary tests appropriate for each failure mode.

Maximising Predictive Value

The practical upshot from the evidence:

DGA works best as a continuous programme with consistent sampling intervals, long-term records, and trajectory-sensitive analysis, rather than a point-in-time check. Trends over time are more informative than any single result [5].

For the highest-consequence transformers, those where fast-developing arcing faults or rapid thermal deterioration would be most catastrophic, continuous online monitoring through Monitor Watch provides the real-time detection capability that quarterly sampling cannot. CIGRE TB 630 [3] provides the framework for selecting which assets justify this level of monitoring investment.

And the analytical method matters. Threshold comparison answers "has this concentration crossed a limit?" Population-based severity assessment with trajectory sensitivity [5] answers "is this transformer accumulating fault severity at a rate that the population data associates with elevated failure probability?" The second question produces earlier warnings and fewer false alarms.

For technical background on R-DGA methodology and the evidence base for its performance advantages, visit the Science page. For product details, visit the TOA page and Monitor Watch page.