An improved nuclear Fisher（ KFDA） diagnosis method is proposed to solve the problems of non-linear data

complex fault types and local extremum in the neural network in transformer faults. On the basis of nuclear Fisher

the distance between classes is weighted by Euclidean distance

which reduces the problem of overlapping data projection to some extent and improves the classification performance. In addition

for the shortage of single kernel function

compound kernel function is used to make it has better nonlinear data processing ability. It is proved by experiments that the KFDA classifier does not has local maximum value and has the advantage of high recognition accuracy

which is an effective fault diagnosis method.