The current fault diagnosis of the grinding system mainly relies on the experience of the workers, which adds a lot of uncertainty to the fault diagnosis. In addition, the grinding system data is too complex, which makes it hard for the workers to judge, and brings about poor performance for traditional machine learning algorithms due to the linear indivisiblity. In order to solve the problem of linear indivisibility, we use neural network for fault classification. In face of the high complexity of fault data, we use autoencoder to increase network depth in order to improve the expression ability of neural network. To reduce the over-fitting brought by deep network, we introduce DropOut and denoising autoencoder. Finally, the experimetal results show that the DropOut denoising autoencoder(DDA) can make the accruacy rate of the fault classification be 90.4% fault classification.