The acceleration parameters of the car during elevator operation can reflect the health state of the elevator, and they can be used for elevator fault diagnosis. However, there are some problems such as noise interference of acceleration parameters and serious unevenness between different fault samples, so the traditional methods cannot meet the application requirements. Therefore, an elevator fault diagnosis method based on particle swarm optimization and a one-dimensional convolutional neural network is proposed. Firstly, the acceleration signals are pre-processed to extract the characteristics of the starting and braking sections of the elevator. A 1DCNN one-dimensional convolutional neural network model with starting and braking signals as input and fault types as output is constructed to diagnose elevator faults. Then PSO is introduced to optimize the number of the convolution kernel and step size hyperparameters of the model. Experimental results show that the proposed method is superior to the traditional method in both fault detection and classification index.