Abstract：In order to improve the accuracy of non-Gaussian variables monitoring in multi-modal chemical process, aiming at the problem that the normal transient fluctuation of the process variable was misjudged as the process error caused by the static control limit in the traditional multivariate statistical monitoring method, a dynamic multi-point fault monitoring method was proposed. Firstly, the chemical process was divided into stationary mode and transition mode. Based on the independent component analysis algorithm of the autoregressive model and particle swarm optimization, the single-point monitoring statistic and the multi-point anomaly statistic of the stationary mode were constructed, and then the stationary mode non-Gaussian monitoring model was founded. Based on the particle swarm optimization.Independent component analysis algorithm, the non-Gaussian monitoring model of the transition mode was structured. Dynamic monitoring strategy was used in both the stationary mode monitoring model and the transition mode monitoring model to realize on-line fault monitoring. The multi-modal dynamic multi-point monitoring method was applied to the propylene metering tank device. Test results showed that false negative rate was less than 0.8%; false alarm rate was reduced by 6.33% compared with the 3σ threshold value monitoring method, and well controlled within 1.2%.