F. Esposito et al. (Eds.): ISMIS 2006, LNAI 4203, pp. 84–90, 2006. © Springer-Verlag Berlin Heidelberg 2006
Particle Swarm Optimization-Based SVM for Incipient
Fault Classification of Power Transformers
Tsair-Fwu Lee1,2, Ming-Yuan Cho1, Chin-Shiuh Shieh1, Hong-Jen Lee1, and Fu-Min Fang2,*
1
National Kaohsiung University of Applied Sciences, Kaohsiung, Taiwan 807, ROC 2
Chang Gung Memorial Hospital-Kaohsiung Medical Center, Chang Gung University College of Medicine, Kaohsiung, Taiwan, ROC
Abstract. A successful adoption and adaptation of the particle swarm
optimiza-tion (PSO) algorithm is presented in this paper. It improves the performance of Support Vector Machine (SVM) in the classification of incipient faults of power transformers. A PSO-based encoding technique is developed to improve the ac-curacy of classification. The proposed scheme is capable of removing mislead-ing input features and, optimizmislead-ing the kernel parameters at the same time. Ex-periments on real operational data had demonstrated the effectiveness and effi-ciency of the proposed approach. The power system industry can benefit from our system in both the accelerated operational speed and the improved accuracy in the classification of incipient faults.
Keywords: particle swarm optimization, incipient fault, classification.
1 Introduction
According to the IEC 599/IEEE C57 [1], degradations of the transformer oil and insu-lating materials produce gases, such as hydrogen (H2), oxygen (O2), nitrogen (N2), methane (CH4), ethylene (C2H4), acetylene (C2H2), ethane (C2H6), carbon dioxide (CO2), and carbon monoxide (CO). They all can be detected from a DGA examina-tion. In the presence of these gases, meaning corona or partial discharge, thermal heating and arcing may occur. The ratios between different gas concentrations, then, can be used to classify faults. However, any DGA approach is inherently imprecise and demands other supporting techniques to obtain reliable results. Some unidentified cases occurred in IEC 599, for example, when ratio codes calculated from actual DGA results did not correspond to any of the codes associated with a characteristic fault. One must then refer to Publication 60599, which contains an in-depth descrip-tion of the five main types of faults usually found in electrical equipment in service. Publication 60599 classifies faults according to the main types that can be reliably identified by visual inspection of the equipment, after the fault has occurred in service [2]: (1) Partial discharges (PD) of the cold plasma (corona) type, (2) Low-energy
*
PSO-Based SVM for Incipient Fault Classification of Power Transformers 85
discharges (D1), (3) High energy discharges (D2) with power follow-through, (4) Thermal faults below 300℃ if the paper has turned brownish (T1), and above 300℃ if the paper has carbonized (T2); (5)Thermal faults above 700℃ (T3).
The three basic gas ratios described in IEC 599 (C2H2/C2H4, CH4/H2, and C2H4/C2H6) are also used in IEC 60599 to identify the characteristic faults. The ratio limits have also been made more precise, in order to reduce the number of unidenti-fied cases from around 30% in IEC 599 to nearly 0% in IEC 60599.
This paper proposes a method to improve the performance of Support Vector Ma-chines (SVM) for fault diagnosis of power transformers by a particle swarm optimiza-tion. It improved the accuracy of classification by removing redundant input features which might cause confusion.
2 Particle Swarm Optimization
PSO algorithm optimizes an object function by conducting population-based search. The population consists of potential solutions, called particles, which are metaphor of birds in bird flocking. These particles are randomly initialized and then freely fly across the multi-dimensional search space. During the flying, every particle updates its velocity and position based on the best experience of its own and the entire popula-tion. The updating policy will drive the particle swarm to move toward region with higher object value, and eventually all particles will gather around the point with highest object value. The detail operation of PSO is as follows [3]:
Step 1. Initialization
The velocity and position of all particles are randomly set to within pre-specified or legal range.
Step 2. Velocity Updating
The velocities of all particles are updated according to the following rule:
1 1 ( , ) 2 2 ( )
i i i best i best i
v ← ⋅ + ⋅ ⋅w v c R p −p + ⋅ ⋅c R g −p (1)
where pi and v are position and velocity of particle i, respectively; i pi best, and
best
g are the position with best object value found so far by particle i and the entire population, respectively; w is a parameter controlling the dynamics of flying; R1 and R1 are random variables from the range [0,1]; c and 1 c are 2 factors used to control the related weighting of corresponding terms.
After the updating, v should be checked and clamped to pre-specified range i
to avoid violent random walking. Step 3. Position Updating
Assuming unit time interval between successive iterations, the positions of all particles are updated according to the following rule:
i i i
p ←p +v (2)
After the updating, the p would also be checked and clamped to legal range i
PSO-Based SVM for Incipient Fault Classification of Power Transformers 89
5.2 Performance for the Taipower Company Dataset
Table 3 shows the classification results for ANN’s and SVM’s with the Taipower Company data. The test success rate of stand alone ANN’s were 82.9% and 89.7% for SVM’s. With our proposed method, for ANN’s, the number of neurons in the hidden layer was automatically selected from 10–36; and for SVM’s, the RBF kernel width was automatically selected between 0.1–2.0, the test success rate improved substan-tially, 99.7% for ANN’s and 100% for SVM’s which were shown in Table 4. This is because of our method is capable of removing redundant input features that may be confusing the classifier, and hence to improve the generalization performance. The remaining features are 4 for both ANN and SVM classifiers respectively.
Table 3. Performance without PSO optimization for the Taipower Company dataset
Stand alone Classifier No. of inputs Training success (%) Test success (%) Training time (s) ANN 36 77.6 82.9 20.221 SVM 36 79.0 89.7 4.104
Table 4. Performance with PSO optimization for the Taipower Company dataset
With PSO Classifier No. of inputs Training success (%) Test success (%) Training time (s) ANN 4 100 99.7 6.445 SVM 4 100 100 2.676
6 Conclusions
In this paper, the selection of input features and the appropriate classifier parameters were optimized by using a PSO-based approach. Experiment results on real data dem-onstrate the effectiveness and high efficiency of the proposed approach. Without the incorporation of PSO, the classification accuracy of SVMs was better than of ANNs. With PSO-based selection, the performances of both classifiers were comparable at nearly 100%. We conclude that our method is not only able to find out optimal pa-rameters but also minimize the number of features that SVM classifier should process and consequently maximize the classification rates of DGA.
References
1. IEEE Std.C57.104, IEEE Guide for The Interpretation of Gases Generated In Oil-Immersed
Transformers, 1991.
2. M. Dong, “Fault diagnosis model power transformer based on statistical learning theory and DGA,” in Proceedings of the IEEE 2004 International Symposium on Electrical Insulation, p85-88, 2004.
90 T.-F. Lee et al.
3. S.-C. Chu, C.-S. Shieh, and J. F. Roddick, “A tutorial on meta-heuristics for optimization,” in J.-S. Pan, H.-C. Huang, and L.C. Jain (Eds.), Intelligent Watermarking Techniques, World Scientific Publishing Company, Singapore, Chapter 4, pp. 97-132, 2004.
4. V. Vapnik, S. Golowich, A. Smola, “Support vector method for function approximation, regression estimation, and signal processing” in M. Mozer, M. Jordan, T. Petsche (Eds),
Neural Information Processing Systems, vol. 9, MIT Press, 1997.
5. L.B. Jack and A.K. Nandi, “Fault detection using SVM and artificial neural networks: aug-mented by genetic algorithms,” Mechanical Systems and Signal Processing, pp. 373-390, 2002.
6. P.F. Pai and W.C. Hong, “Forecasting regional electricity load based on recurrent support vector machines with genetic algorithms”, Electric Power System Research, vol. 74, pp. 417-425, 2005.
7. M. Duval, “Interpretation of Gas-In-Oil analysis using new IEC Publication 60599 and IEC TC 10 Databases,” IEEE Electrical Insulation Magazine, vol. 17, no. 2, 2001.
