In order to solve the problem of data redundancy and low diagnosis rate in transformer fault diagnosis, this paper combines the improved genetic algorithm rough set attribute reduction algorithm with probabilistic neural network (PNN), and establishes a PNN neural network transformer fault diagnosis model based on the improved genetic algorithm rough set attribute reduction. In this paper, the traditional genetic algorithm rough set attribute reduction algorithm is improved, the attribute kernel is added to the genetic algorithm population initialization code, and the attribute dependency is added to the genetic fitness function to increase the speed and accuracy of attribute reduction. Through this algorithm, the reduced fault data is smaller and more reliable, and the PNN neural network training simulation can effectively reduce the complexity of the network, reduce the network training time, and improve the practicability and diagnosis rate of transformer fault diagnosis.