Aiming at the problems that scarcity of fault samples in bearing fault diagnosis and low diagnosis accuracy of deep neural network models under the condition of small samples,a framework for extending deep neural networks into Siamese network structure was proposed to improve the fault diagnosis performance with small samples.Siamese network extracted the features of the sample pairs through the weight sharing backbone network and the similarity was compared based on the L1 distance to achieve fault classification.Different from the traditional deep neural network,the Siamese network adopted the method of input sample pairs,which could improve the performance of bearing fault diagnosis in the case of insufficient fault data.The convolutional neural network (CNN) and long short-term memory (LSTM) network with different layers were respectively expanded into Siamese network structure.A small sample fault diagnosis experiment was conducted on the measured bearing data set.The experimental results show that the accuracy of fault diagnosis results can be improved by expanding into Siamese network structure.The accuracy of the Siamese CNN network is 1.08% higher than that of the corresponding CNN network,and the accuracy of the Siamese LSTM network is 4.78% higher than that of the corresponding LSTM network.