7 things that DGA interpretation methods could do better
Interpreting DGA results isn't easy. A number of methods have sprung up over the years, promising to help. Some of them (like Duval, Dörnenburg, Rogers) have become quite famous and have certainly proven their use over the years. Still, many decades after their inception, we can start to think about how we can do even better. Here is our top 7 list of possible improvements, in the hope that it will guide future developments of new approaches.
1 - They don't detect faults
With some rare exceptions (like the Rogers method), most DGA interpretation methods don't actually detect faults, but classify them.
This means that you have to be sure that any fault has actually occured before using the method to identify the fault.
2 - They can be quite sensitive
2 - They can be quite sensitive
If DGA-values change a small amount due to measurement uncertainties, we expect the result of a DGA interpretation method to change a small amount as well. This is illustrated in the following figure, using a Duval-like triangle as an example.The yellow hexagons represent the measurement uncertainty. The darkness of the grey background represents the model output. On the left, the background is smoothly graded. As a result, the shades of grey included in the hexagon are at least somewhat similar. On the right, there are sharp steps. If a hexagon intersects one or more of these steps, the model results for repeated measurements can fluctuate between categories.
3 - They don't consider
environment conditions
The same DGA values on two different transformers can mean very different things if the environment conditions are distinct. A new but highly loaded transformer used in the context of wind power has some excuses for increased hydrogen values, a slightly older and lightly loaded transmission transformer has not. Most DGA interpretation methods don't consider this.
4 - They have a long reaction time
4 - They have a long reaction time
A faulty transformer could have low DGA values today, slightly increased DGA values tomorrow and strongly increased DGA values the day after. The existence of a fault can be concluded early if these trends are considered. But handling trends is difficult, especially if a method has to work on varying time scales. Many methods don't consider trends at all. If they do, they often use heavy filtering to decrease noise. This greatly reduces the advantage in reaction time that motivated the usage of trends in the first place.
5 - They don't quantify uncertainty
There are many sources of uncertainty in DGA interpretation. Measurement uncertainty, measurement age, ignorance of environmental conditions to name a few.
In many situations, the DGA numbers are unambiguous, and the importance of these uncertainties is small. In other situations, the muncertainties can be significant. In an ideal world, the model confidence should be transparent to the operator.
6 - They don't tell you
what to do
6 - They don't tell you
what to do
Knowing the type of a transformer fault is great, but its not the final step. In the end, action (or non-action) has to be derived from this information, otherwise what would be the point?
For instance, a recognition of a low temperature thermal fault could motivate a load reduction until a scheduled maintenance is executed. A seasoned transformer expert can translate
the result of the methods into an asset management decision. But chances are that a seasoned expert doesn't strictly need any algorithmic help in the first place.
7 - They can't handle the superposition of multiple fault types
If a transformer has a thermal fault, it doesn't mean that the same transformer has no other faults. On the contrary: The existance of failure cascades increases the probability of additional faults conditional of the existance of a prior fault. Most DGA interpretation methods cannot indicate fault mixtures in the general case, with some exceptions like the DT-field in the Duval triangle.
