Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
net/publication/3340257
CITATIONS READS
24 42
2 authors, including:
Tomasz Piotrowski
Lodz University of Technology
12 PUBLICATIONS 34 CITATIONS
SEE PROFILE
All content following this page was uploaded by Tomasz Piotrowski on 20 May 2015.
ABSTRACT
This paper presents a statistical method for generation of the signals warning of a
possibility of failure in oil filled1 power transformers. The method uses the dis-
solved gas analysis (DGA) data Itn produce quantitative values warning of the be-
ginning and progress of failure. S'tatisticaltests are employed tu test the bimodality
of four standard distributions: N,ormal, Log-normal, Weibull and Gumbel. Numer-
ical examples are also presented.
Index Terms - Dissolved gas analysis, statistical method, failure, transfurm-
ers.
List of Symbols
Symbol Variable, Unit
2 THEORETICAL BACKGROUND
P Median (for Gauss distribution), ppm The following assumptions are used in the DGA pro-
D Standard deviation. oom
~~~~~~
. I .
Number of distribution function parameters 4. Reliable tools for statistical analysis of random vari-
Where ppm is pans per million. ables are available.
of hundreds of transformers of different types. Each time, Table 2. Measurement data and results of statistical descriptors
-
no matter what designs or manufacturers of the trans-
formers were considered, the three-parameter Weibull
method for methane (CH,).
Gas Global “ b a d
concentration p + u p + 2 u p + 3 0 probability
distribution described the analyzed populations in the best NO. PPm PPm PPm PPm %
way. The parameters of Weibull distribution were pre- 2 5 5.0 6.5 8.0 1.6
sented in [SI for transformers belonging to sound and 3 6 6.0 7.7 9.4 1.9
faulty populations. Therefore, assumptions (2) and (3) 4 29 213 32.1 42.8 6.3
5 17 21.8 31.8 41.7 4.2
could he treated as proven. 6 15 21.5 30.7 39.9 3.8
7 16 21.5 30.1 38.7 4.0
There are many statistical computer tools available on
8 24 23.1 -32.0 40.8 5.4
the market. However, the authors developed numerical’ 9 27 24.9 34.2 43.5 5.9
programs that allow fast analysis of the samples of data [61 10 26 26.0 35.3 44.7 5.8
and which perform necessary analysis [4,12-14]. There- 11 35 28.7 39.0 49.3 7.2
. 12 50 34.2 47.4 60.6 9.4
fore, we may say that assumption (4) is also met. 13 77 44.9 64.5 84.1 13.0
14 6 43.4 63.0 82.5 1.9
On the basis of the above assumptions, it is possible 15 127 61.7 94.6 126.4 18.6
using statistical analysis to develop the quantitative signals 16 138 77.8 118.2 158.5 19.7
for graduate fault warnings. 17 732 246.4 414.4 582.4 55.5
3 STATISTICAL DESCRIPTORS
METHOD At the same time, the very low global probability that the
At the beginning we analyzed the changes of statistical transformer belongs to the faulty population, as shown in
descriptors such as the mean, median, mode, variance, the last column, allows us to disregard the warning. From
standard deviation, kurtosis, skewness, shape coefficient the 71h measurement signals disappeared and appeared
and running averages, for several classes of transformers. again starting from the 141h measurement, but still at a
Then, on the basis of this analysis, partially described in low global probability. For the measurement number 17
[2,31, the mean and standard deviation were chosen as the we get a vety strong warning signal (at the level of p + 3 6 )
most suitable descriptors to be used in a more complex with a sudden increase of the probability that the trans-
method. former belongs to the faulty population. Later, the tested
In this method, the current gas concentration value is transformer failed.
compared with the limits made as a sum of average and Data in Table 2 concems concentration of methane
one, two or three standard deviations ( p + 6 , p + 2 & /I (CH,) and gives similar information to that obtained from
+ 3 6 ) , respectively. This method is similar to statistical Table 1. The difference is that we get a very strong warn-
process control methods. The above method is illustrated ing signal (at the level of @ + 3 6 ) at the lSth measure-
in Table 1. In Table 1 the weak warning sienals (on the ment and at lower global
. probability.
- Y
level I*. + 6 ) for carbon monoxide (CO) can be observed
starting from the 31d measurement (it should be noted that 4 TEST OF SIGNIFICANCE METHOD
that method is applicable for 2 and more measurements). As we mentioned above, a gas concentration in the oil
is a random variable. Statistical distribution of a particular
gas in a single transformer is called local distribution to
Table 1 . Measurement data and results of statistical descriptors
method far carbon monoxide (COX
differentiate it from the global distribution, which is ap-
Gas Global “bad” plied for gas concentrations obtained from many different
concentration p + u p + 2 c p + 3 u probability transformers. Local and global distributions for a particu-
No. ppm ppm ppm ppm 70 lar gas could be found from two different populations. The
2 15 15.0 17.0 19.0 6.2 first is applied for gas concentration in sound trans-
3 34 30.0 40.1 50.1 12.2
4 59.8 85.5 111.3 22.8
formeds) and the other for gas concentration in faulty
5 2 7 85.8 123.0 160.2 29.4 one(s).
6 112 100.5 141.9 183.3 30.4
7 79 100.9 139.8 178.8 23.5
If the number of measurements is sufficient, at least 5,
8 103 106.0 144.8 183.7 28.6 because a three-parameter Weibull distribution should
9 93 107.5 145.0 182.6 26.6 have a nonzero number of degrees of freedom, then it is
10 93 108.6 144.8 181.1 26.6
11 104 110.9 146.7 182.4 28.8
possible to estimate the distribution for the local gas con-
12 83 110.1 144.4 178.7 24.4 centrations: Obviously, the older the transformer, the more
13 84 109.5 142.5 175.5 24.6 measurement data is available and the more reliable are
14 113 112.3 145.4 178.6 30.6
15 143 119.1 155.0 190.8 36.0
the results of the statistical analysis
16 158 127.1 166.2 205.3 38.6 On this basis, starting from the SIh measurement it is
17 533 225.5 336.9 448.3 75.0
possible to estimate parameters of the statistical distribu-
262 Mosinski and Piotrowski: New Statistical Methods for Eualuation of DGA Data
Table 3. Measurement data and rrsults of test significance method for acetylene (C,H,).
Gas Gumbel Global “bad“
concentration Gauss distribution distribution Weibull distribution probability
H MO H Xo Xm k H
NO. PPm P.”. %
~
1 0.1 0.1
2 0.1 0.1
3 20.7 11.1
4 0.1 0.1
5 20.4 5.0 23.0 0.198 0.046 13.4 0.197 11.0
6 0.1 0.2 25.7 0.179 0.1139 9.5 0.178 0.1
7 22.0 4.2 24.5 0.232 0.043 13.2 0.228 1.8 0.13 0.338 11.7
8 24.0 7.5 23.3 0.293 0.048 16.0 0.279 3.6 0.15 0.445 12.6
9 29.0 10.5 22.4 0.298 0.053 18.7 0.259 6.2 0.16 0.567 14.7
10 0.1 7.4 24.3 0.292 0.046 16.5 0.258 3.0 0.14 0.550 0.1
11 22.0 9.2 24.5 0.474 0.052 17.6 0.423 4.2 0.16 0.863 11.7
12 0.1 8.2 21.8 0.492 0.052 16.5 0.441 0 3.0 0.15 0.818 0.1
13 0.1 7.4 21.2 0.539 0.052 15.6 0.487 0 2.4 0.15 0.792 0.1
14 26.0 9.1 20.6 0.574 0.056 17.1 0.503 0 3.4 0.16 0.911 13.4
15 26.0 10.7 20.5 0.705 0.059 18.5 0.619 0 5.8 0.17 1.124 13.4
16 27.0 11.9 19.6 0.785 0.1163 19.5 0.674 0 5.8 0.17 1323 13.9
17 6.0 11.3 19.1 0.656 0.062 18.6 0.564 0 6.8 0.21 1.297 3.9
18 68.0 12.4 24.8 0.686 0.035 23.3 0.956 0 8.4 0.22 13115 28.5
19 57.0 14.2 27.6 0.509 0.032 25.4 0.665 0 10.1 0.23 1346 24 9
20 113.0 15.5 39.1 0.850 0.021 34.3 1.171 0 12.6 0.26 1.294 40.7
21 93.0 18.1 45.2 0.621 0.019 36.7 0.827 0 15.1 0.28 1.292 35.6
22 87.0 21.1 18.3 0.531 0.l119 40.8 0.658 0 17.7 0.29 13111 34.0
__ ~ ~ ~ ~
IEEE Transactions on Dielechics and Electrical Insulation Vol. 10, No. 2; April 2003 263
0.5 '
10
I
15
I
20
No of samples
I
25
I
30
and show that the transformer in this case should he
treated as a new unit because the value of the gas concen-
trations are the same at 0.1 ppm,
Figure 3. Heterogeneity coefficient as a result of significance test The example in Figures 3 and 4 concerns a transformer
far the Log,normal distribution for the ethane (C,H,) versus num-
ber of samples; the 30th measurement was performed after trans- for which every measurement gave vety low concentration
former failure. values of the selected gases. But the new analytical method
Table 4. Results of the quotients (only quotients with statistical irreaularities) and H (for Weibull distribution) calculations far the same case
C=H2/CH, F=C,H,/C,H,
. Quotient Quotient Quotient Quotient Quotient Quotient
NO. c H F. H L4
_______
H I H G H E H
1 n ns nom 1.000 os00 20 2.000
2 0.02 0.033 0.600 0.200 50 3.400
3 0.017 0.050 0.333 0.167 60 4.941
4 0.0034. - 0.013 0.276 - 0.241 ~
2911 - 13.158 -
5 0.294 0.1076 0.020 - 0.294 0.2070 0.353 0.0079 3.4 0.0851 12.411 0.1040
6 0.2 0.0713 0.025 0.0082 0.267 0.0206 0.267 0.0784 5 0.0982 9.143 0.0814
7 0.125 0.0482 0.025 0.0655 0.250 0.0689 n.zo 0.0737 8 0.0289 12.405 0.1803
8 0.083 0.0286 0.250 0.6909 0.167 0.2322 0.292 0.1224 12 0.0403 14.699 0.2286
9 0.074 0.0401 0.014 0.5318 0.259 0.3380 0.296 0.1239 13.5 0.3091 18.118 0.2097
in 0.19 0.0686 0.017 0.5084 0.231 0.3782 0.269 0.2011 5.2 0.3458 16.387 0.2546
11 0.0857 0.0946 0.250 0.3718 0.114 0.3850 0.286 0.2823 11.7 0.4894 16.183 0.3321
12 0.23 0.1200 0.056 0.3677 0.419 0.3321 0.233 0.2982 4.3 0.3349 23.185 0.2838
13 0.15 0.1443 0.056 0.1286 0.450 0.2862 0.275 0.4061 6.7 0.5968 21.605 0.2943
14 0.08 0.1630 0.048 0.4142 0.420 0.2771 0.240 0.4359 12.5 0.7318 19.771 0.3243
15 0.05 0.1106 0.0006. 0.5446 0.288 0.3610 0.153 0.4114 19.7 0.2886 19.611 0.3854
16 0.068 0.1359 0.037 0.5894 0.365 0.4361 0.189 0.3235 14.8 0.3783 21.017 0.4408
17 0.1 0.1515 0.024 0.6494 0.472 0.3572 0.202 0.2952 9.9 0.4672 20.8116 0.5166
18 0.077 0.2025 0.0002 0.7937 0.571 0.3236 0.253 0.3928 13 0.7641 21.976 0.5817
19 11.065 0.2267 0.025 0.9446 0.519 0.2921 0.208 0.3386 15.4 0.906 18.881 0.6705
20 0.089 0.2829 0.034 1.0114 0.644 0.2061 0.278 0.4007 11.25 1.5935 22.979 0.7205
21 0.039 0.2497 0.0002 1.0257 0.373 0.2444 0.157 0.3787 25.5 0.8740 24.250 0.7460
22 I 0.7380 0.017 1.0926 9.667 1.7283 4.333 2.0988 1 1.1097 20.973 0.8557
23 0.056 0.7679 0.0002 1.0716 0.472 2.2767 0.200 1.9426 17.8 1.0958 22.812 0.9386
24 0.07 0.8688 0.m 1.1494 0.441 2.0687 0.205 1.9075 14.1 1.2699 16.826 0.9707
25 0.07 1.0281 0.018 1.2349 0.438 2.5568 0.203 2.4863 12.8 1.4692 23.820 1.0159
26 0.03 0.9433 0.0016 1.0342 0.496 2.2677 0.228 2.0132 31.75 1.4228 16.874 1.0522
27 0.057 1.2798 0.0002 1.0371 0.398 2.4369 0.187 2.5807 17.57 1.4980 17.903 1.1156
28 0.058 1.4521 0.0015 0.89~1 0.486 2.5779 0.232 2.2990 17.25 1.5628 16.405 1.1509
29 0.0797 1.4465 0.016 0.9704 0.457 >5 0.217 2.7845 12.5 1.7751 24.553 1.1680
30 1.175 2.1104 0.792 0.9030 0.918 ,z 0.138 2.7601 0.851 2.2572 4.844 1.1260
~
~
264 Mosinski and Piotrowski: New Statistical Methods for Evaluation of DGA Data
6 SUMMARY
1. New statistical methods useful in DGA process and
I I I
I5 20 25 30
examples of their utilization were presented.
No of ramplcr 2. Proposed statistical methods utilized all gathered
Figure 4. Heterogeneity coefficients as a result of significance test
far the Log-normal (solid), Gumbel (dotted) and Weibull (dashed)
measurements of gases dissolved in oil and not only the
distributions for total combustible gases (TCG) VCISUS number of last one or two as conventionally used in DGA methods.
samples; the 30Ih measurement was performed after tran8;former The reliability of the results of the statistical analysis in-
failure. creases with the age of the transformer (more data are
available). No such regularity can be observed for other
described here signals irregularities. Figure 3 shows data methods.
for ethane (C,H,) and Figure 4 shows data for total'com- 3. Statistical descriptors method requires very simple
bustible gases (TCG). Statistical analysis perform<:d for calculations and a minimal knowledge of statistics. Test of
TCG indicates first problems 5 years before and for ethane significance method requires specialized numerical pro-
7 months before the transformer failure occurred. If the gram and much deeper knowledge of statistics..
owner of the transformer utilized this information, he
4. The proposed method of analyzing the DGA signals,
would he able to shorten periods between measurements
which is based on statistical test of significance, may result
and could potentially avoid transformer failure. All other
in a warning at an early stage of a problem.
DGA methods failed because of a very low gas concentra-
tion.
REFERENCES
5 POPULATION OF THE RATIOS O'F
DISSOLVED GASES [ l ] F. Mosidski, M. Khalaf, and J. Galoch: "rozklad praw-
dopodobiebshva zawartoSci gazdw rozpuszczonych w deju trans-
Table 4 shows the characteristic ratios used in various formatorowy"'(Cumulative distribution function for gases dis-
solved in transformer oil), IX Konferencja Remontowa Ener-
diagnostic methods. The numbers in hold represent the getyki, Szczyrk, 14-15 wrzesnia, 1995.
case when x 2 statistics for the particular ratios p ~ p u l a - [2] F. Mosihski and T. Piatrawski, "statystyczna analiza zawartoici
tion exceeding the critical values of xi-,-
,,a (H> ).I with gaz6w rozpuszczonych w aleju transfotmatora energetycznego"
the significance level a = 0.05. We can observe the signs (Statistical Analysis of Gases Dissolved in Power Transformer
of irregularities in the investigated populations from the Oil), I11 Konferencja Naukawo-Technicma TRANSFORMA-
TORY ENERGETYCZNE I SPECJALNE, Diagnostyka, Ek-
sample number 20 for the ratios F and G , from the sam- sploatacja, Modemizacja, Kazimierz Dolny 11-13 paidziernika,
ple number 22'for ratios IA a n d ' J , and from the sample str. 193-199, 2000.
number 25 for the ratios C and E. These values increase 131 F. Mosidski and T. Piotrowski,"statystyczna analiza iloSci gaz6w
with subsequent sample numbers, indicating increasing rozpuszczonych w oleju transformatoro~m"(Statistica1Analysis
of Gas Dissolved in Power Transformcr Oil), IW-2000, Pornad-
statistical irregularity, until the transformer failure occurs Kiekrz, 22-24 maja, str. 275, 2000.
at sample number 30. We can also observe that the value [4] F. Mosihski, "zastasowanie metod statystyczuych dla iniyniedw
of the variable H increases and exceeds the value of 5 for elektlyk6w" (Statistical method for electrical engineen), Wyd.
the ratio IA, a value of 'almost 3 for the ratio J and a PL,' ?Adz; 2000.
value somewhat greater than 2 for the ratios C and G. [SI K. Y. S. AI-Mualla, "The Analysis of Diagnosis Method of Power
Transformer Insulation on the Basis of Chromatographic Mea-
The last two ratios are the inverses of each other. surements", Ph.D. dissertation, Palitechnika Iddzka, 1997.
The analysis of various ratios allows drawing some con- 161 F. Mo&ski and T. Piotrowski, "numelyczna kanoteka transfor-
clusions regarding .the slow degradation process i n , the matora energetycznego" (Data Base for Power Transformers),
IV Konferencja Naukowo-Techniczna ELEKTROWNIE
transformer insulation. The first signal was obtainmed for CIEPLNE, Eksploatacja-Moderniracje-Remonty,Slok 23-25
the sample'number 20 from ratios G and F. According to wfzebnia, s. 141-14, 1999.
the Japanese diagnostic method [7, 101, this may indicate [7] CEIflEC 605991999 "Guide 10 the Interpretation of Dissolved
partial discharges with very small energy. and Free Gases Analysis".
IEEE Transactions on Dielectrics and Electrical Insulation Val. 10, No. 2; April 2003 265
IEEE Std. C57.104-1991, “IEEE Guide for the Interpretation of Franciszek Mosinski was bom in Widen.
Gases Generated in Oil-Immersed Transformers”. Poland, in 1946. He studied Electrical Engi-
V. G. Arakielian and E. D. Sienkiewicz, “rannaja diagnostika neering at the Technical University of LodG
maslonapelnionnogo vysokovoltnogo oborudovanija”, Elek- where he received the MSc., Ph.D. and
triczeskije Stancji, No. 6, p. 50, 1985. DrSc. degrees in 1972, 1976 and 1984, re-
1101 T. Kawamura, M. Yamaoka, H. Kawada, K. Audo, T. Maeda, spectively. In 1984 and 2Wl he became As-
and T. Takatsu, “Analyzing Gases Dissolved in Oil and its Ap- sistant Professor and Professor, respectively.
plication to Maintenance of Transformers”, C E R E Sess. rep. His main fields of interest are electrical
12.05, 1989. withstand voltage and diagnostics of electri-
M. Duval, “Dissolved Gas Analysis. It Can Save Your Trans- cal insulation of power transformen, appli-
formers”, IEEE Electrical Insulation Magazine, Vol 5, No. 6, cation of mathemati cal statistics and the influence of electrical
”” 77-27
rr. -- -’.19Y9 facilities on environinent. Prof. Marinski is author of about 120
[12] G. B. Brown, ”Method of Mmimum LikelihoodApplied to the publications, including papers. books and monographs. He is also
halysis of ~~~~h~~~~ D ~ ~ ~I E. ,~, Errans, vel, pAs 88, pp, author and co-author of about 90 unpublished expertise, technical
1823-1830, 1979. elaborates and numerical programs.
[13] G. Carrara and W. Hauschild, “Statistical Evaluation of Dielec-
tric Test Results”., Studv, Committee
~~
- ~~~ ~~.
33. Electra No.
~. 131.
~~~, 00.
r.
109-131, 1990: Tamaaz F’iotmwski was born in Lo& Pol-
[14] W. Hauschild and W. Mosch, SIalisrical Techniques for Higk- and, in 1965. He studied Electrical Engi-
Voltage Engineering, Peter Peregrinus Ltd., Bristol, 1992. neering at the Technical University of Lodz,
where he received the M.Sc. and Ph.D. de-
, grees in’ 1990 and 2001, respectively. His
main field of interest is diagnostics of elec-
trical insulation of power transformers and
utilization of the IT in electrical engineer-
ing.