Abstract: At present, the fault diagnosis of grinding systems is evaluated by humans, which easily causes low efficiency, low accuracy, high cost, and casualties. The classification ability of the traditional method is considered unsatisfactory for sample datasets with high dimension and high temporal correlation. Considering the above problems, a deep learning method based on recurrent neural network and long short term memory (RNN-LSTM) and that realizes the intelligent fault diagnosis of grinding systems is proposed. The dataset is batche for inputs of LSTM networks, and the temporal correlation of the dataset is extracted, which is then used to complete the fault classification by comparing input feature vectors before and after the moment. We compare RNN-LSTM-based networks and autoencoder-based networks via a simulation. The RNN-LSTM-based method significantly outperformed the autoencoder-based method for a dataset with high dimension and high temporal correlation, and the result of the error rate is as low as 3%.