5.1.1. Sample Data

Gas concentration of DGA method is analyzed in one part per million (PPM). To facilitate the presentation and analysis of the characteristics in the figure, logarithmic processing is performed for each monitoring value in Figures 2 and 3. The box diagram of the monitored data value in the sample data is shown in Figure 2. Sample type distribution and monitoring value distribution of each basic feature are presented in Figure 3.

**Figure 2.** Box diagram of the monitored data value in the sample data.

**Figure 3.** Sample type distribution and monitoring value distribution of each basic feature.

5.1.2. Basic Features by DGA Condition Monitoring

Typical gases measured in the DGA method of the main transformers include *H*2, *CH*4, *C*2*H*2, *C*2*H*4, *C*2*H*6, *CO*, *CO*<sup>2</sup> as illustrated in Figures 2 and 3.

$$TH = CH\_4 + C\_2H\_2 + C\_2H\_4 + C\_2H\_6$$

Other features commonly used in transformer condition diagnosis methods are shown in Table 2.


**Table 2.** Common features of transformer DGA condition diagnosis methods.

The symbols in the Duval triangle method shown in Table 1 are denoted as %*C*2*H*<sup>2</sup> = 100*x*/(*x* + *y* + *z*); %*C*2*H*<sup>4</sup> = 100*y*/(*x* + *y* + *z*); %*CH*<sup>4</sup> = 100*z*/(*x* + *y* + *z*); with *x* = (*C*2*H*2); *y* = (*C*2*H*4); *z* = (*CH*4) in PPM.

*ϕ*(*H*2), *ϕ*(*CH*4), *ϕ*(*C*2*H*6), *ϕ*(*C*2*H*4) and *ϕ*(*C*2*H*2) in Table 1 represent the contents of five characteristic gases, respectively, and Total Combustion Gases (TCG) as in: TCG = *H*<sup>2</sup> + *CH*<sup>4</sup> + +*C*2*H*<sup>4</sup> + *C*2*H*<sup>6</sup> + *C*2*H*2; *ϕ*(*H*2) = *H*2/*TCG*; *ϕ*(*CH*4) = *CH*4/*TCG*; *ϕ*(*C*2*H*6) = *C*2*H*6/*TCG*; *ϕ*(*C*2*H*4) = *C*2*H*4/*TCG*; *ϕ*(*C*2*H*2) = *C*2*H*2/*TCG*;

And %*H*<sup>2</sup> = 100 ∗ *H*2/(*H*<sup>2</sup> + *C*2*H*<sup>6</sup> + *CO* + *CO*2).

As shown in Figures 2 and 3, the class of the sample is not balanced and several basic features are similar in distribution. Unsupervised feature extraction is adopted to obtain the feature set with maximum correlation and minimum redundancy.
