Abstract:To effectively extract the fault characteristics hidden in the vibration signal and accurately identify the type of machinery fault, a fault diagnostic model based on convolutional residual sharing weight long short term memory neural networks (Conv-Res-SWLSTM) was proposed. A convolutional network was used to capture the local spatial features of vibration singal; a weight-sharing long and short term memory neural network (SWLSTM) was built by fusing the gate structures, which could reduce the parameters and training time required for optimization, and the hidden time features were discovered more efficiently in the output signals of the upper layer network. Simultaneously, the scaled exponential liner unit function was introduced to improve the self-normalization property of network, while the residual module was implanted to improve the fault features perception and feature extraction ability of network. Finally, the comparative experiment was conducted based on the measured data set of mechanical faults。 The experimental results show that the diagnostic accuracy of the proposed model reaches 99.30% at four speeds, which has better diagnostic accuracy and stability compared to other models.