Due to the long-term operation of centrifugal pump in harsh environment, affected by field working conditions, medium corrosion and other factors, many fault signals represent obvious nonlinearity and time-varying nonstationarity, large amount of data, and it is difficult to predict the operation state in real time and accurately. In this paper, a centrifugal pump state prediction method based on PSO（Particle Swarm Optimization） optimized LS-SVM（Least Squares Support Vector Machines） was proposed. Firstly, the kernel parameter g and penalty factor C of least squares support vector machine are quickly and automatically optimized by using the global search characteristics of particle swarm optimization algorithm. Secondly, the average absolute error, average relative error and root mean square error are determined as the prediction accuracy evaluation indexes. Finally, the prediction method in this paper was verified by the real-time collected data. The results show that compared with LS-SVM prediction model, PSO optimized LS-SVM model reduces the computational complexity, has the advantages of strong generalization ability and high prediction accuracy, and the average absolute error, average relative error and root mean square error are reduced by 52%, 56% and 44% respectively. This method can provide a theoretical basis for predictive maintenance and has a good application prospect in engineering practice.