Based on the study of discharge properties

an automatic pattern recognition system for transformer ultra-high-frequency （UHF） PD is designed and developed in this paper.An adaptive genetic algorithm （AGA） is used to train the neural network （NN） in system. Using BP-NN and AGA-NN

we distinguish the basic types of defects appearing in transformers

such as corona

void

bubble

creeping discharge and floating discharge. Tests in laboratory show that the results are satisfactory. Compared with BP-NN

AGA-NN can overcome slow convergence and possibility of being trapped at local minimum value. Thus

the convergence

discrimination and generalization ability of AGA-NN are more powerful.