Generative adversarial networks(GAN) are widely used in various fields. However, when generating text data with contextual correlation characteristics such as fault data, GAN has many limitations. On the one hand, discrete data output makes it difficult to pass gradient updates from the discriminator to the generator; on the other hand, it is difficult for the discriminator to process incompletely generated sequences. In this paper, we propose a fault data generation algorithm based on GAN and policy gradient mechanism. Using the reinforcement learning method aims to solve the gradient update transfer problem, and the policy gradient algorithm is used to directly update the parameters of the generator; at the same time, by using the Upper Confidence Bound Apply to Tree(UCT) algorithm to simulate the incomplete sequence into a complete sequence so that the discriminator can evaluate its reward value. The simulation results show that our fault data generation algorithm based on GAN and policy gradient mechanism performs better in the fault data generation task.