The invention discloses a multi-level optimization PCC-SDG-based chemical process fault diagnosis method. According to the invention, firstly, a full-process network topology structure is adopted as a reference point, and selected variables are preliminarily optimized based on the statistical indexes of Pearson correlation coefficient (PCC). Secondly, based on the PCA weight idea, special variables, large in weight, are selected out of multi-level correlation coefficients. In combination with a signed directed graph (SDG), an optimal PCC-SDG network is established. Finally, for the optimal PCC-SDG network, a rule of an aggregated weight coefficient Q is established. The invention provides a novel fault diagnosis method, and perfects the SDG modeling method. Meanwhile, the efficiency of the working staff to detect the states of multiple variables is improved. The influence of time lag and other non-information synchronization factors is avoided. The rate of false alarms is more effectively reduced and the type of a fault can be accurately identified. The occurrence of production safety accidents is greatly reduced.