Bearing fault diagnosis methods based on data driven have been considered as a research focus in the field of bearing fault diagnosis. However, it is low in the accuracy of bearing fault diagnosis based on data drive because the bearing fault of hydraulic dynamometer is rare. It leads to low accuracy of bearing fault diagnosis based on data-driven. To solve this problem, an on-line fault diagnosis method of hydraulic dynamometer bearing based on improved generative adversarial neural network (GAN) was proposed. Firstly, GAN training method was improved. The distribution properties of the raw data were learned by the discriminator and the generator, and alternately trained with an improved GAN. The data enhancement model of bearing fault of hydraulic dynamometer was established to obtain synthetic data. Then the bearing fault diagnosis model based on SVM was obtained by combining the original data and synthetic data training. Finally, the bearing fault diagnosis model was adopted to realize the bearing fault diagnosis of hydraulic dynamometer on line. Through the simulation results, the real-time accuracy of bearing fault diagnosis is greatly improved by the proposed online fault diagnosis method through improved GAN enhanced training. And it has the characteristics of strong anti-noise interference.