Abstract: The molecular pump provides a clean vacuum environment for the EAST device, and its running state affects the smooth development of the EAST experiment. During the operation of the EAST experiment, the molecular pump equipment may suffer from foreign matter falling in or vacuum leakage, causing secondary hazards to the device. Aiming at the problem of low fault diagnosis accuracy and model over-fitting caused by the imbalance of molecular pump fault data set, an algorithm based on the combination of time and frequency domain preprocessing and improved BP was proposed to realize molecular pump fault diagnosis. On the basis of BP neural network, particle swarm optimization (PSO) is introduced and combined with five-fold cross-verification to optimize the model. Firstly, vibration signals of normal, vacuum leakage and foreign body falling faults are collected on a destructive test platform simulating molecular pump faults, and then the data are extracted and fused in the time domain and frequency domain. The obtained feature vector set is used as the input of the optimization algorithm to train the model and realize molecular pump fault diagnosis. The experimental results show that the diagnostic accuracy of the proposed improved BP algorithm can reach 96.84%, which is superior to support vector machine (SVM), K-nearest neighbor (KNN) and BP algorithm.