Aiming at the problems of poor noise reduction effect, failure to distinguish the importance of multi-scale features, and insufficient extraction of temporal features in the current fault diagnosis of chemical processes, this paper proposes a chemical fault diagnosis method based on multi-scale fusion model. In this method, the attention mechanism is combined with soft threshold method and multi-scale learning respectively, and a multi-scale deep residual shrinkage network (MDRSN) is constructed. Moreover, the extracted multi-scale spatial features are sent to the bidirectional gated cyclic unit (BIGRU) to further extract temporal features. Compared with the single-channel network, BIGRU can not only complete the learning of past information, but also complete the learning of future moment information, so more temporal correlation information can be obtained. Finally, the modified Tennessee-Eastman process data were used to verify the classification accuracy of 95.08% and the recall rate of 94.76%, which was obviously better than the comparison method, and the effectiveness of the method was proved.